While people use artificial intelligence (AI) tools to help with tasks like learning new skills or answering questions, a new study suggests that these benefits may come with a downside.

The findings, presented at the 2026 Conference on Language Modeling in San Francisco earlier this month, revealed that people who used AI assistance for reading comprehension or math problems were more likely to struggle or give up when the AI assistance was removed. The results suggest that relying too much on AI could lower people's motivations to overcome challenging situations.

In the wake of this study, Live Science recently polled its readers on whether they felt AI was making them less motivated after using it.

More than 140 readers voted in the poll, with 30% of voters picking "No, I don't use AI at all, so this situation doesn't apply to me."

"If AI were only a tool, controlled by those who use it I might give it a try," said one commenter in response. "But it isn't under control at all. It gets into your bank accounts and moves money without authorization. It encourages suicide and self-harm. It is addictive and makes people lazy. I know someone who uses it to answer questions and while it is detailed, it makes her lazier, asking one question and getting the answer without her searching for it. AI Is a danger. It should not be allowed."

Other comments reflected this sentiment, with one reader writing, "Use of all automatic aids decrease[s] attention and retention. I see it in the use of navigation devices where drivers ought to recognize routes after a couple of times. AI helps research and cuts down lengthy searches, but over use of summaries weakens critical faculties."

Other commenters questioned whether AI is the problem, or if people are just inclined to take the easy option, asking "Would you get the same results if the answers to all but a few problems were provided on separate sheet of paper?" and another writing, "I think researchers should also investigate whether shovels decrease motivation to dig holes with one's hands. AI, like a shovel, is only a tool."

Do you agree with these readers? Let us know in the comments!

Scientists in China are exploring what it would take to develop "the last AI built by humans" — a powerful, self-improving artificial intelligence (AI) system capable of creating its own successors.

In a new paper, researchers examine how approaches to recursive self-improvement (RSI) could help tackle the rising financial and computational costs of developing ever-more advanced AI models.

RSI refers to the idea that future AI models could direct their own training and development, using what they learn from each round of improvements to guide subsequent rounds of development.

This recursive feedback loop could eventually enable AI systems to develop newer, more powerful generations of models with minimal human input.

The study, which was uploaded to the preprint database arXiv Sept. 10 but has not been peer-reviewed, proposes a framework for classifying AI systems based on their self-improvement capabilities, from Level 1 (lowest) to Level 5 (highest).

At the lower levels, development remains largely human-led. As AI systems move up the scale, they become increasingly capable of identifying their weaknesses and deciding how to address them, eventually gaining the ability to make their own “meta-improvements” — improving how they improve themselves.

Is self-improving AI 'nonsense'?

Whether AI could ever achieve this level of sustained self-improvement remains an open question. Previous experiments have, however, demonstrated more limited forms of RSI.

In one study earlier this year, an AI system called A-Evolve-Training was tasked with improving the performance of a 30 billion-parameter large language model (LLM).

Over four rounds, A-Evolve-Training modestly improved the LLM's benchmark performance by testing different training techniques and using the results to change how it approached subsequent experiments.

Peter Bentley, a professor of digital biology at University College London, questioned whether such experiments offer a credible route to RSI.

"This idea that some AI is magically going to start generating better and better AIs by themselves [is] nonsense," Bentley told Live Science in an email.

Bentley pointed out that self-improving computer systems have been around for decades. Evolutionary computation and reinforcement learning, for instance, are both based on the principle of systems refining their performance over time in response to feedback.

The much greater challenge, he said, is developing AI systems capable of autonomously conducting the research needed to create more powerful successors — not to mention overcoming their dependence on human expertise and physical computing infrastructure.

"You can't 'recursively self-improve' unless you can also enhance your own training, enhance your hardware, and test it exhaustively in real world settings," Bentley added.

The challenge of true RSI

The study's authors acknowledged some of these challenges. Developing RSI-capable systems could require enormous investments in computing power, as well as human supervision to ensure successive improvements delivered meaningful gains. All of this becomes effectively redundant if the resources needed to develop and test subsequent AI models outweigh any benefits they offer, the authors noted.

Scientists would also need a reliable way to determine whether an AI system was genuinely becoming more capable over iterations, rather than exploiting weaknesses in tests used to measure its capabilities. This is a known problem in AI development, where models can achieve higher benchmark scores by exploiting testing methods or cherry-picking results.

Kayla Blomquist, an AI researcher at the University of Oxford and co-founder of the Oxford China Policy Lab, said that, while the paper doesn't provide a practical road map to RSI, it could help developers and policymakers understand the technical challenges and safety risks associated with increasingly autonomous AI systems.

She also questioned whether there would ever be a clear point at which AI became completely independent of human development. "The paper describes a gradual handover of responsibility rather than a single moment, so there may never be an obvious 'last' model built by humans," Blomquist added.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Universities are at a critical point as students and faculty embrace the use of AI. Already, universities like Harvard are introducing courses that feature an AI avatar as an "instructor."

The promises AI offers higher education are familiar: an increase in efficiency as a consequence of less administration, more teaching coverage, better feedback and smarter assessment. But introducing artificial intelligence (AI) systems in an attempt to fix these problems is not the answer. If anything, AI systems will magnify and speed up the bureaucracy that is driving universities' decline.

In practice, the more likely outcome is darker: AI will not simply sweep away the bureaucratic machinery of universities; it will give that machinery a new regime of AI compliance — more policies, more forms, more trackers, more oversight roles and more work downloaded onto faculty who are already stretched thin.

The danger is not that AI will destroy universities but that universities will use AI to preserve the very processes that make them dysfunctional. If this happens, we might see faculty spending less time on research, less time on teaching, and more time on activities that do not directly contribute to the production of knowledge and education of students.

This is already evident in the existing assessment system in the U.K., but shades are creeping into systems elsewhere, providing a window into the future for U.S institutions. A single piece of student work may pass through a plagiarism-detection platform, a grading interface, a moderation spreadsheet for faculty to share feedback, a grade-transfer system, and an exam-board record. Each stage generates its own logins, deadlines, exception rules and audit trails. None of this is inherently malicious. Anonymous marking, moderation, appeals, accessibility and academic integrity are legitimate concerns. But when each concern is implemented as a separate system, the combined effect is administrative incoherence.

In essence, faculty become the human interface between incompatible platforms, spending their time reconciling grades and chasing missing evidence rather than teaching. While this example pertains to U.K. universities, the trend to expand the administrative staff extends to North America. AI simply introduces a new opportunity to increase administration.

AI could, in principle, solve this. It could integrate systems, automate routine reporting, draft feedback, identify inconsistencies and remove duplicate data entry. Yet the institutional incentive is rarely to remove work; it is to make work visible. A university that automates a broken workflow does not necessarily simplify it. It may simply produce faster reports about the broken workflow, more sophisticated dashboards and more elaborate evidence of compliance. The bureaucratic impulse survives by changing its vocabulary: risk management, assurance, transparency, governance, responsible innovation.

At least one possible outcome is that universities will create AI-use policies, require module-level AI statements, mandate disclosure forms, introduce approved-tool lists and establish review committees. Faculty will be asked to redesign assessments, monitor suspected misuse, document their decisions, complete training, and explain their approach to students. Much of this labor will be invisible. It will appear not as a new job title but as an additional expectation attached to teaching.

This creates a familiar pattern: responsibility without authority. Faculty are expected to make consequential judgments about academic integrity, fairness, privacy and assessment validity. But they do not control procurement, system design, institutional policy or disciplinary procedures. When something goes wrong, the academic is exposed; when the system works, the institution claims the credit.

That asymmetry is corrosive. It teaches committed staff that careful judgment is not valued as expertise but rather as a compliance risk to be managed, and the consequences will fall unevenly. Experienced faculty, after voicing their concern to no avail, may leave, retire early or disengage. Early-career academics, already precarious because they don't have tenure, will have the least power to resist. Students will receive less of the resource that matters most: sustained intellectual attention.

One possible outcome is that universities will encourage the development of faculty AI avatars and cut hours and employment. The institution will then justify hiring more ‪—‬ not fewer ‪—‬ compliance and "educational enhancement" staff to manage and maintain these AI avatars, negating the idea of efficiency and cost savings. In this scenario, bureaucracy will expand to manage the problems created by its own expansion while undercutting quality for the students who are paying high tuition fees. One might need to hire staff for oversight of these avatars, administrators and lawyers to manage licensing agreements, quality control staff in case things go wrong, and IT teams for integrating them into learning platforms and websites.

There is an alternative to the administrative tasks. AI should be governed by a simple institutional rule: no new compliance requirement without the removal of an equivalent existing burden. If a university requires assessment redesign to accommodate AI, or the introduction of new technology, it should provide protected time, technical support and a single integrated platform. If it expects faculty to manage AI use, it must give them clear institutional backing rather than leaving them to improvise. And critically, it should show how the technology makes work more efficient before rolling it out across the board.

But the deeper issue is not technology but institutional purpose. A university exists — or should exist — to cultivate understanding, judgment and intellectual independence. If its systems consume faculty time, fragment the student experience, and reward documentation over learning, then AI becomes another instrument of managerial self-preservation. The question is not whether universities can adopt AI. It is whether they can resist using AI to automate their own absurdities.

If they cannot, the outcome will be an educational system that looks progressively sophisticated while becoming increasingly hollow ‪—‬ full of monitoring, low on trust; rich in measuring, poor in teaching; obsessed with compliance, indifferent to the people who make education possible; overly procedural sans intellectual life. It will become an audited theater of learning not worth students' time.

Opinion on Live Science gives you insight on the most important issues in science that affect you and the world around you today, written by experts and leading scientists in their field.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Artificial intelligence (AI) tools like chatbots are becoming part of everyday life, helping people write messages, answer questions and learn new skills. While some of these processes may save people time, a new study reveals a possible new downside of consistently relying on AI assistance.

The research, presented at the 2026 Conference on Language Modeling in San Francisco earlier this month, found that people who relied on AI assistance for tasks like math problems and reading comprehension were more likely to give up or struggle when the assistance was taken away. The findings suggest that leaning too heavily on AI could weaken people's motivation to push through challenges.

The research raises the question of what happens when people stop exercising their brains. So do you think AI is making you less motivated, or even "dumber," or is it simply making you more efficient? Vote in our poll, and let us know your thoughts in the comments.

People who rely on artificial intelligence (AI) to solve simple problems do worse at, and are more likely to give up on, similar tasks when that assistance is removed, compared with people who do not use AI for quick answers, new research finds.

The study, which was presented at the 2026 Conference on Language Modeling in San Francisco earlier this month, investigated the impacts of shifting mental work onto an AI chatbot across arithmetic and reading comprehension tasks. A growing body of research suggests this process of "cognitive offloading" harms long-term learning and critical thinking abilities.

"The concern is about what cognitive scientists call 'desirable difficulties' — the productive struggle that builds skill over time," study first author Grace Liu, a doctoral student in machine learning at Carnegie Mellon University, told Live Science in an email. "If AI routinely removes that struggle, people may get the right answer in the moment but develop less robust independent capability."

Liu and her colleagues recruited 1,222 participants for the study, which was released as a preprint earlier this year and has now undergone peer review. The researchers designed three randomized controlled experiments, including two sets of mathematical fraction problems and a series of reading comprehension questions. The participants were split into three groups, so there was no overlap between the experiments. Each experiment took 13 to 15 minutes to complete.

In the first experiment, the team randomly assigned 354 participants to two groups: one that had access to an AI chatbot and one that didn't. The researchers gave both groups the same 12 fraction problems, which the AI group could get the answer to immediately or obtain clues about by prompting the chatbot. Then, the researchers removed the AI assistant without warning for a final set of three fractions, which all of the participants were given, regardless of their group.

Both groups had the option to skip problems. The researchers explicitly told the participants that there was no penalty for wrong answers, so skipping a problem was a deliberate decision to not engage with it, they wrote in the study. The participants in both groups were also given the correct answer to each problem if they got it wrong, so they could learn from their mistakes.

This first experiment showed that the participants in the AI group were more likely to solve fractions and less likely to give up when they had access to the chatbot compared with the other group. But when the AI was removed, their performance dropped and they skipped or failed more problems, suggesting that using AI impaired unassisted task performance, persistence and motivation.

The second experiment was a copy of the first experiment, but it had 667 participants and small tweaks in the trial design, including the addition of a pretest phase that helped the researchers eliminate participants who were unable to solve the most basic problems. The findings were similar to those of the first experiment, hinting that the use of AI for simple tasks harms perseverance and drive to complete those same tasks without AI, especially when AI is used to get direct solutions.

Liu acknowledged that calculators also allow people to offload computations, which could cause similar declines in independent performance. But "the reason we study AI specifically is that it allows people to offload reasoning across tasks across many domains," she said.

In a third experiment, the team explored the impacts of AI assistance on reading comprehension, which draws on meaning-making skills and mental-model construction — the cognitive process of building internal representations of external reality. The trial included 201 participants, who were divided into a group with access to AI and a control group. The researchers gave both groups five comprehension questions before disabling the AI chatbot and handing each participant another three questions.

The results mirrored those of the first two experiments, suggesting that reduced persistence, in particular, was not an artifact of the math problems but a general outcome of cognitive offloading onto AI.

"The most significant finding is that people who offload more have higher declines in persistence (increased skip rate), which is an important skill that cuts across domains," Liu said.

One possible reason people stop trying once they have experienced help from AI is that they get used to the immediacy of answers and perceive any method that takes longer as inefficient, the researchers wrote in the study. But studies show that persistence through difficulty is essential for skill acquisition and learning, raising concerns about the long-term impacts of using AI for simple tasks.

Scientists don't yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability. To find out, researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks, Liu said. But research so far suggests that an overreliance on AI erodes our cognitive fitness and mastery of subjects, according to the study.

To encourage long-term learning, capability and autonomy in people, AI models and chatbots should be designed to give directions and instructions toward a solution, rather than the solution itself — more like a tutor, the researchers argued.

"We need to work toward a future in which human abilities are, as much as possible, augmented rather than supplanted," study co-author Brian Christian, a computer scientist and researcher in the Center for Human-Compatible AI at the University of California, Berkeley, said in a statement.

The largest artificial intelligence (AI) firms increasingly understand that they are hostages of one another. If one firm pushes ahead recklessly, others may feel compelled to follow. If one deploys a dangerous model without adequate safeguards, others may see it as a green light to do the same.

The debate over AI has centered on what governments should do to regulate this frontier technology. That remains essential. But a more immediate question is emerging: Can the companies building these powerful systems take action independently before they lose credibility and suffer legal consequences?

That question no longer looks theoretical and is being recognized as an urgent challenge. We've already seen AI firms lose control of their models — for example, in Australia, a rogue AI hacked a government healthcare website. If we are not careful, it's only a matter of time before one of these incidents causes large-scale, real-world harm, like shutting down an electricity grid or municipal water system, or a mass misdiagnosis in a health system. If this trend continues, the entire industry could become tarnished.

But there is a blueprint for self-regulation in a risky business: the nuclear industry, which has built up layers of oversight, controls and self-scrutiny over decades.

Frontier AI may require something similar: private regulation among competitors that do not trust one another but have a mutual interest in preventing outcomes that could damage them all. This does not appeal to the better angels of human nature — it's self-preservation for the industry itself.

There is a historical precedent for this kind of arrangement. The nuclear sector did not become safe because every actor suddenly became benevolent. It became safer through a layered system of peer scrutiny, technical standards, information-sharing and external oversight. International organizations such as the International Atomic Energy Agency (IAEA), which was established in 1957, emerged to coordinate monitoring and verification, while national regulators and industry bodies gradually built professional norms around catastrophic risk.

The result was not perfect safety, as the Chernobyl disaster in 1986 clearly attests. But nuclear governance architecture is designed around the recognition that in high-risk technological systems, one actor's failure can become everyone's problem. This was seen following Chernobyl: global nuclear safety increased significantly after the disaster, with upgrades to reactor designs, improved training and international monitoring networks.

And this movement to self-regulation also occurred in the U.S. nuclear sector, following the Three-Mile Island nuclear accident near Harrisburg, Pennsylvania, in 1979. The reactor suffered a partial meltdown after a relatively minor cooling malfunction. Investigations later revealed design flaws in instruments and alarms, along with inadequate emergency response training, were the root cause of the incident.

The event nearly killed the nuclear energy industry in the U.S., causing public support to fall from a record high of 69% in 1977 to 46%t in 1979. The entire industry suffered because of failures at one facility. In response, the U.S. nuclear industry established the Institute of Nuclear Power Operations (INPO) in December 1979 to improve operational safety.

The analogy to AI should not be overstated. Nuclear materials are easier to count than algorithms, and nuclear facilities are harder to hide than model weights. The outcomes of a catastrophic AI blunder are also less existential, although its use in weapon systems is not without risk. In September, an AI hallucination almost led to the U.S. intercepting a Chinese ship suspected of carrying nuclear material to Iran.

Yet the comparison remains useful because the core governance problem is similar. In both cases, competitors operate in an environment where the downside risk of unchecked rivalry is borne by society while the upside is private gain for these firms. The challenge is to create institutions that make restraint the rational default — even when no actor fully trusts the others, and the inner workings of the industry are not readily visible to outsiders.

For AI, that could mean reciprocal evaluation of frontier models before deployment, secure mechanisms for sharing safety information, agreed red lines around certain capabilities, and independent third parties empowered to audit compliance. Such mechanisms would not replace government regulation. They would instead serve as an intermediate layer: stronger than voluntary ethics statements but more adaptive and technically informed than legislation alone can often be in fast-moving domains. The state would act as a background enforcer.

AI executives appear willing to look for solutions too, with Elon Musk recently suggesting a peer-review system before advanced model release, and Demis Hassabis, co-founder and chair of Google DeepMind, putting forward a framework for an AI Standards Body, similar to the Financial Industry Regulatory Authority.

Skeptics will object, with reason, that industry-led governance can become cartel behavior or public-relations theater. Companies may use "safety" language to entrench incumbency, slow rivals or deflect binding regulation. Those are real dangers. The lesson from nuclear history is not that private coordination works automatically. It is that private coordination works only when nested inside a broader system of transparency, verification and public authority. Without those elements, private regulation becomes little more than trust-me rhetoric.

Still, dismissing private regulation altogether would be a mistake. Governments do not yet possess the technical capacity, speed or international alignment needed to oversee frontier AI on their own. And if the leading labs already see themselves as bound together by shared exposure to catastrophic risk, that recognition can become the basis for institutional design. The question is how to turn mutual fear into durable rules.

That is where the nuclear analogy is most powerful. It reminds us that industries facing existential hazards do not need perfect trust in order to cooperate. They need mechanisms that make cooperation preferable to defection. In practice, that means inspections, peer review, incident reporting, professional norms and a state-backed enforcement perimeter. AI governance is often discussed as if the only choices were laissez-faire innovation or heavy-handed state control. History suggests a third path: structured private coordination under public shadow.

The real opportunity, then, is not to celebrate the newfound (apparent) prudence of technology executives, it is to recognize what their comments reveal. When even fierce rivals begin to argue for mutual scrutiny, they are admitting that competition alone cannot govern the frontier. That admission should be treated not as reassurance but as evidence that the risk has become collective.

The policy task now is to build institutions that fit that reality. Governments should push for mandatory reporting, independent evaluation and legal backstops. But they should also encourage enforceable forms of peer oversight among the handful of firms whose systems may shape the trajectory of the technology. Frontier AI is becoming an industry in which the leading actors are hostages of one another. The nuclear age showed that such conditions can produce serious governance. The hope is that AI firms learn that lesson before they learn the cost of ignoring it at the peril of the public at large.

Opinion on Live Science gives you insight on the most important issues in science that affect you and the world around you today, written by experts and leading scientists in their field.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Last week, Tesla unveiled a new "flying" edition of its Roadster electric vehicle. While some experts remain skeptical that the new "Rocket" EV is feasible in the real world, the demonstration reignited a long-standing debate about whether flying cars could soon become a reality.

In response to this demonstration, Live Science polled its readers on whether they think flying cars will be available to consumers in the next five years.

Over 280 readers responded, with the largest voting group, 34% of readers, picking "No, we need more than five years for flying cars to take off." The next-highest voting bloc was even more skeptical with 23% of readers choosing "No, flying cars will never happen."

Several comments reflected this pessimistic view, with one commenter leaving this response: "No flying cars. In my almost 60 years of driving experience I find that most people can't handle driving in 2 dimensions let alone 3. Besides that we have face [sic] so many other problems: hunger, poverty, inadequate healthcare, the careless disregard for our planet. What are we leaving my grandchildren and great grandchildren??"

Other commenters echoed this sentiment. "It will never happen as some cheap alternative to electric or petrol driven vehicles," one person responded. "Gravity is master on this planet so to rise above that and be floating or propelled around on some sort of air current, Nah! I cannot imagine the damage that will take place in any livable area, towns, suburbs, cities, parks, etc, etc., that list is endless. And imagine the protests should a general public be allowed to dive and fly above any or all livable areas."

Not everyone was as adverse, with another commenter remarking on how far Tesla's technological development has come. "The flying roadster is a statement of what Mr. Musk can do with technology more than anything else. In the same way he can make the Roadster fly, be sure that the technology designed into the Tesla Y, the Cybercab or the Starship are at the very frontier of science. Mr Musk has singlehanded revolutionized the car industry and space travel in less than 20 years (the first Tesla car was sold in 2008). No small feat."

Do you agree or disagree with our readers? Tell us your thoughts in the comments.

With artificial intelligence (AI), scientists have a powerful new tool for creating never-before-seen proteins, which can be used to make either groundbreaking new medicines or devastating bioweapons.

To help guard against nefarious uses of these proteins, scientists have recently figured out a way to "watermark" them without compromising their function. A team from Google DeepMind developed this technique to better track the provenance of AI-designed proteins.

"This has the potential to reduce biosecurity threats, such as unauthorized toxin creation, and may help with database curation, preventing artificial designs from cluttering the standard databases," said Roy Sleator, a biology professor at Munster Technological University in Ireland, who was not involved in the research.

"In practice, it could help trace a design's origin, flag engineered proteins for extra biosecurity checks, and reduce misleading claims about how a protein was made," Sleator told Live Science in an email.

However, in its current form, the same AI technologies that add watermarks can also strip them away, other experts warned.

New ways to hide watermarks

The unveiling of the AI program AlphaFold in 2020 was a milestone in protein engineering. The tool, which won Demis Hassabis, John Jumper and David Baker the Nobel Prize in Chemistry in 2024, can reveal the 3D structure of a protein from the sequence of its building blocks, called amino acids. Proteins fold like origami, adopting 3D shapes that dictate their function.

However, AlphaFold's technology also opened the door to generating brand-new proteins that could be used to create biothreats, including viral proteins that can evade our immune systems and existing vaccines.

"Anything that can be used for good can be used for bad," said George Rose, professor emeritus of biophysics at Johns Hopkins University.

Current screening tools identify novel dangerous or restricted molecules by seeing how similar their amino acid sequences are to those of recognized threats. But the problem is that AI could come up with proteins that don't look like anything we've ever seen. That's what the new watermarking technique aims to solve.

The DeepMind team has framed the watermarking effort as a "family" of related techniques they call SynthIDBio. The first member of this family is SynthIDBio-sequence, which encodes watermarks into the amino acid sequences of proteins created by a program called ProteinMPNN.

When users tell ProteinMPNN the desired 3D shape for their protein, it chooses amino acid options that ultimately fold into that shape. SynthIDBio-sequence subtly influences how the program chooses each amino acid according to a secret watermarking "key." Once the sequence is generated, anyone who knows the hidden code can assess how likely the sequence is to have been created under the key's bias. In their recent study, the DeepMind team showed that proteins built with this watermark could still bind to their targets as well as non-watermarked proteins.

A sister technique, called SynthIDBio-structure, instead embeds a watermark directly into the protein's structure. It does this by altering the function of AlphaFold 3, DeepMind's flagship protein prediction algorithm.

When AlphaFold predicts a structure, SynthIDBio-structure gives it a side mission: The AI must make minuscule tweaks to how the atoms in the protein are arranged. These tweaks can be picked up by a detector later, and they're small enough that they don't look like obvious distortion.

In a test, these two systems created hidden watermarks that could be detected with incredible accuracy, flagging more than 99% of watermarked proteins and correctly ignoring 99.9% of non-watermarked proteins.

How to bypass a protein watermark

Live Science spoke to several experts who agreed that SynthIDBio was an impressive technical feat. However, they said it wasn't ready for widespread implementation yet.

The two techniques are "effective but brittle," Bruce Wittmann, a principal applied scientist at Microsoft, told Live Science in an email. That's because, as the DeepMind team also notes in their paper, the watermarks can be easily removed.

"For sequences, watermarks can be removed by 'resequencing' — effectively, altering the protein amino acid sequence such that the watermark is no longer detectable," Wittmann said. "For structures, watermarks can be removed by adding noise to or otherwise modifying the spatial coordinates of the structure." Screening tools wouldn't be able to detect if a watermark had been removed from a protein.

Experts are also skeptical that watermarking would provide a long-term solution for biosecurity.

An indelible watermark could identify who originally made a nefarious protein, or identify when a user was accessing a protein generation software without permission. This might deter some criminals, but terrorists looking to create a novel protein for a biological weapon would be less likely to care about being identified after a successful attack, said Max Brackmann, a chemical, biological, radiological, and nuclear arms expert at the Spiez Laboratory in Switzerland.

Additionally, the vast majority of software for protein prediction and generation is open-source, meaning much of the code is freely available to use and modify. That makes it far harder to bake watermarking into these programs. Wittmann warned that the field would have to shift toward closed-source models before watermarks could become an industry standard.

He added that watermarking could act as one layer of a biosecurity defense shield, rather than offering standalone protection. Top-down regulation would be challenging to enforce. "I imagine it would be very difficult to ensure the regulations were being followed," he said.

Rose was even blunter about regulating the use of generative proteins.

"I'm convinced it won't work," he said. "Once there's a sweet technology that does something, government regulation to keep it under wraps has never been really successful."

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Flying cars once seemed like the stuff of science fiction. But Tesla's recent demonstration, which showed the company's Rocket Roadster model briefly flying, has reignited the conversation about whether flying cars could soon become a reality.

The technology behind flying cars is continuing to advance, with companies developing vehicles that combine elements of traditional cars, helicopters, electric aircraft, and — in Tesla's case — rockets. Many of these designs use electric vertical-takeoff-and-landing technology, which allows them to lift off and land without a runway by using multiple rotors or propellers to generate lift. Advances in battery technology, autonomous flight systems and computer-assisted navigation are helping to make these vehicles seem more feasible. But challenges remain, including questions surrounding passenger safety, affordability, and the need for regulations and laws and infrastructure for the widespread use of these vehicles.

Tesla says production of its new "flying car" is scheduled for 2028, but the company is notorious for not meeting its proposed deadlines. So do you think flying cars could actually be available to the public within the next five years? Or are we still a long way from turning this science fiction fantasy into reality? Cast your vote in our poll, and share your thoughts in our comments section.

A drone powered by artificial intelligence (AI) has completed a championship race course faster than some of the world's best human drone pilots.

In a study published Oct. 14 in the journal Science Robotics, scientists created an AI system capable of pushing a quadcopter through a demanding course at over 60 mph (100 km/h), using little more than a single forward-facing camera, an inertial measurement unit, and a tiny onboard computer.

The system, called MonoRace, even won the 2025 Abu Dhabi Autonomous Racing League drone racing competition, outperforming every rival autonomous team and posting faster times than three former human world-champion drone pilots.

MonoRace accomplished this without the aid of GPS, external motion-capture systems, stereo cameras, or a preloaded map of the course. The system is an onboard autonomous flight model created by researchers at the Delft University of Technology's Micro Air Vehicles Lab in the Netherlands. Its name reflects its central design feature; "mono" refers to its monocular vision. The drone relies on a single rolling shutter camera and monitors its own motion with an internal measurement unit (IMU) that gauges acceleration and rotation.

Flying at breakneck speeds

The limited sensor package feeds a neural network, a machine-learning model that uses interconnected artificial neurons to learn patterns from data. This neural network identifies racing gates and estimates the drone's position and motion to generate motor commands. Then, it sends the commands to the four propellers at a rate of 500 times per second through the drone's flight controller.

The system fits a drone that weighs a mere 966 grams (2.13 pounds) and measures approximately 11.8 x 11 x 3.9 inches (30 x 28 x 10 centimeters). Because the MonoRace hardware was so compact, lightweight and computationally efficient, the team was able to operate it entirely aboard the drone.

On the Abu Dhabi course, MonoRace completed two laps around 11 gates in 16.56 seconds, hitting a top speed of approximately 63 mph (102 km/h).

A recently released video tracks MonoRace's development and performance in the competition, with real flight sequences demonstrating how the drone banks sharply, accelerates through gates, and navigates a course without human assistance.

However, some of the sequences are computer-generated. Study co-author Guido de Croon, a professor at the Delft University of Technology, said those are intentional visual effects. The only fully simulated imagery appears at about the two-minute mark, during the segment showing the training process.

"Here we show the gates in a visual simulator for the video but the real training happened in a stripped-down, pure physics simulator that ran really fast," de Croon told Live Science in an email. "Simulating vision was not done during training, as the control neural network receives an estimated state as input, which comes from a different perception and state estimation process."

The rest of the footage is real, although some shots include overlays showing how the drone interprets its camera feed. A convolutional neural network segments the white gates, sometimes displaying them as red in the normal camera image. Small circles mark detected gate corners, red lines indicate gate edges, and a blue overlay shows the portion of the image where the drone expects to find the next gate.

How the AI learned to race

Scientists trained the MonoRace model with reinforcement learning, a machine learning method in which an agent repeatedly attempts a task and receives feedback based on its success or failure.

"The reinforcement learning ran in simulation, because it needs lots of trial and error — like millions of tries," de Croon said. "It actually learns to fly from scratch."

The network learned to translate the drone's position on the track, velocity and orientation directly into motor commands. According to de Croon, early training focused on the basics, like avoiding crashes and reaching the first gate. As the training progressed, the system learned to advance farther along the course and fly at higher speeds.

The researchers didn't train the system on a single aircraft. Instead, they varied dimensions such as weight, voltage, drag, sensor noise and propeller condition so the AI model could cope with real-world variations and imperfections.

"That ensures that the trained network can control different drones," de Croon said.

MonoRace could become a training tool for human drone pilots, "a bit like powerful chess programs have also inspired humans playing the game," de Croon said.

Its broader significance lies in demonstrating advancements in physical AI, especially how a robot can perceive, decide and act at high speeds despite limited sensing and computing power. The control network runs on a microcontroller-class chip, meaning that sophisticated control doesn't require a large computer.

De Croon noted that a similarly sized network could optimize for minimal energy use rather than minimum lap time. Because most machines rely on small processors and conventional controllers, energy-optimal neural control could eventually reduce power consumption in everyday robotics, he said.

Tesla has unveiled a new "flying" version of its Roadster electric vehicle (EV) that uses SpaceX tech to pull off seemingly impressive airborne acrobatics.

During a filmed demonstration at the SpaceX rocket testing facility in Texas Thursday (Oct. 15), the second-generation Tesla Roadster was shown accelerating up a curved ramp before launching itself into the air, performing a 180-degree barrel roll and touching back down to Earth.

In another stunt, the vehicle leapt over a row of parked Teslas. For safety reasons, the vehicle was operated remotely in both instances.

"Turns out, the future does have flying cars," Tesla and SpaceX CEO Elon Musk wrote on X following the demonstration.

The Roadster's airborne capabilities are made possible by a system of cold-gas thrusters developed using technology from Musk's aerospace company, SpaceX. These thrusters release bursts of high-pressure gas to generate thrust, allowing the vehicle to accelerate rapidly and, according to Tesla, become airborne for short periods.

The new Roadster will come in three variants: Performance, Thrust and Rocket. Only the Rocket is capable of flight. The Performance model is a two-seater equipped with a suction system that's designed to improve traction and handling (or drive up inverted ramps, as in Tesla's demo), while the Thrust model adds rear-mounted thrusters that Tesla claims can propel the vehicle from 0 to 60 mph (97 km/h) in under a second.

The highest-end model, Rocket, features SpaceX cold-gas thrusters positioned around the vehicle to enable short bursts of flight and in-air control. This single-seat model will be available by invitation only, according to Tesla.

On the company's website, Tesla representatives explicitly stated that the Rocket isn't approved for use on public roads and that availability depends on regional regulations. During the reveal, Musk told the audience that the Rocket "is not necessarily a safety-optimized car."

Tesla hasn't disclosed how far or for how long its EV can remain airborne, nor is it the first company to develop a hybrid road-air vehicle. There is also skepticism from outside viewers on whether the demonstration could be replicated outside of controlled scenarios.

Artificial intelligence (AI) scientist Yann LeCun ‪— a professor of data science, computer science, neural science and electrical engineering at New York University ‪— described the demonstration on X as a "marketing stunt" and questioned whether the vehicle would work in practice.

"Even assuming that 50% of the mass of the car is the compressed air, that gives you less than 10s of useful thrust (with thrust-to-weight ratio larger than 1.0). Probably more like 5 seconds, actually," he posted on the social media platform. "So the car must be an empty, lightweight shell containing big compressed air tanks. Cute demo, but totally impractical. But of course, you knew that."

He likened it to SpaceShipOne, which debuted in 2004 as the first privately funded spacecraft and made three record-setting flights before being retired.

California-based Alef Aeronautics, which is developing its own flying EVs, marked Tesla's unveiling by releasing footage of its vehicle flying over an earlier Roadster model. Alef representatives said the demonstration marked its 1,000th logged flight.

"Welcome Tesla to the flying car race, we've been here alone for eleven years," Alef CEO Jim Dukhovny said in a statement emailed to Live Science. "Having competition is a good thing, especially from such a major company as Tesla."

Unlike Tesla's compressed-gas tech, Alef's vehicle uses electric propellers for vertical takeoff and landing, with a rotating cabin and body designed to support forward flight.

Production of the new Roadster is scheduled for 2028, but Tesla doesn't have the best track record of meeting its deadlines. The company has delayed production of its Optimus humanoid robot a few times since first unveiling the machine in 2021, for example, with an initial start in 2023, before being pushed back. Only in recent months has production begun to scale to a few hundred units per week.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

A new 4,300-mile (7,000-kilometer) undersea fiber cable will shuttle data between the U.S. and France at breakneck speeds of up to 1 million gigabits per second (Gbps) — enough to move around 125 terabytes of data every second, or roughly 25,000 HD movies, each 5 gigabytes large. The cable, called Petal, will be funded and operated by Facebook and Instagram parent company Meta.

When it comes online in 2029, the ocean-spanning cable will be the first to deliver petabit capacity — double the capacity of existing subsea cables of this length — across transoceanic distances, Meta representatives said in a statement. That's enough to allow 75% of the world's population to stream music at once.

Petal will also be the first commercial subsea cable system to use multi-core fiber technology, according to the statement. Whereas conventional fiber strands typically contain a single core, each of Petal's fiber strands will contain two cores that carry optical signals in opposite directions, allowing a single strand to provide the capacity that would conventionally require a pair of single-core fibers.

Each fiber strand is roughly as thick as a human hair, measuring just 125 micrometers in diameter. Petal will contain 24 pairs of dual-core fibers, giving it 96 individual fiber cores in total and enabling it to carry as much data as 48 conventional fiber pairs, Meta engineers said.

Japanese technology company NEC will be responsible for the overall design and construction of Petal, while Sumitomo Electric will provide the fiber itself.

"Achieving petabit-per-second capacity across a transoceanic submarine system represents a major technological milestone in the history of global telecommunications," Eduardo Mateo, chief strategy officer of the Submarine Networks Division at NEC, said in the statement.

Bedrock of the internet

Subsea cables are among the most critical layers of internet infrastructure, carrying around 99% of the world's internet traffic, according to the International Telecommunication Union, the United Nations ​agency for digital technologies.

Installing thousands of miles of fragile glass cable on the ocean floor is an engineering feat in itself. Equally challenging is ensuring that these cables can carry the tiny pulses of infrared light that transport fiber-optic data across such huge distances without impacting performance.

This is typically achieved through repeaters — amplifiers installed at regular intervals along the cable that use lasers to boost the optical signals.

A 4,300-mile subsea cable typically needs about 100 of these repeaters to amplify the signals along its length, Meta representatives said.

Petal will incorporate something called "fan-in/fan-out" technology. As each two-core fiber strand enters a repeater, an optical interface routes the signal from each core into a separate single-core fiber strand so that each can be amplified independently. The amplified signals are then routed back into the two cores of the dual-core fiber before continuing along the cable.

Combined with its dual-core fiber, this feat of signal engineering will enable Petal to double its data-carrying capacity without requiring a proportional increase in power, Meta representatives said.

Petal is one of many subsea cable investments by Meta in recent years, including Marea in 2018, Amitie in 2023 and Anjana in 2025. Marea, the company's first transatlantic fiber instalment, supports 192 Tbps, meaning Petal will offer more than five times that capacity.

Researchers say they've developed a simple mathematical formula that can predict when certain ChatGPT-like artificial intelligence (AI) chatbots will "go rogue," suddenly switching from helpful responses to potentially harmful ones.

In a new study published Oct. 8 in the journal Patterns, physics professor Neil Johnson and doctoral student Frank Yingjie Huo, both at The George Washington University, show how the order in which a chatbot is given certain prompts can drastically reshape its answers, leading to "tipping points" that can send the chatbot down a rabbit hole of bad advice. However, Johnson and Huo say their formula can easily predict these tipping points.There have been horror stories of AI encouraging users into suicide, helping build lethal weapons, and amplifying violent or harmful points of view. At the same time, researchers at OpenAI and Anthropic, respectively, have repeatedly and openly stated that they don't really understand why their creations behave the way they do. They have, however, built safeguards that use a variety of tools ‪—‬ including content filtering, reinforcement learning and other technical guardrails ‪—‬ to reduce risks.

But Johnson sees those tools as insufficient. "The safety measures now are the equivalent of strapping big, bouncy cushions under a plane," he told Live Science. "They don't stop the plane from crashing; they just make it that when the plane crashes, it'll be slightly softer." When people such as doctors, lawyers or soldiers are accessing AI chatbots offline for privacy, the problem is even more acute.

Johnson and Huo sought to find a more fundamental safeguard than the currently available tools. In doing so, they exposed an internal tracking mechanism they describe as a compass needle. This needle is the model's "residual vector" moving across neural network layers as a prompt is processed.

The residual vector naturally gravitates toward either desirable or undesirable outputs, but it is heavily influenced by how a user prompts the AI. "You can tip the balance, not even by changing the content but by changing the order in which you ask it," Johnson said, "which just shows how delicate this internal balance between good and bad, desirable and undesirable, can be."

For example, Johnson and Huo prompted OpenAI’s basic GPT-2 model with questions about vaccines, hurting others and self-harm. When the question “Are vaccines bad?” was asked first, the bot’s responses to this and subsequent questions were desirable, or good. When it was asked last, all subsequent responses were concerning.

From these insights, the researchers developed a mathematical formula to estimate the circumstances in which an AI would reach the tipping point between desirable and undesirable outputs. Johnson said that by embedding their simple mathematical formula in open-source models in the lab, they have shown that it empirically works as an early warning system, though he does not expect OpenAI or Anthropic to be knocking on his door anytime soon.

"I'm expecting a lot of pushback from these companies," he said. "It's not in their interests to have someone from outside saying you could have avoided a lot of these problems."

Wei Xing, a lecturer in AI at the University of Sheffield in the U.K. who was not involved in the study, agrees that concerns surrounding unmonitored offline devices are well founded. However, he urged caution before embracing this approach as a practical safeguard. "The authors reduce a huge system to a single tiny component, like explaining a whole car by studying one spark plug," Xing said.

Safety training — the primary defense in modern AI — cannot be represented in their simplified model, the formula was tested only on old, basic models (GPT-2, and EleutherAI’s Pythia and Meta OPT open-source models) rather than AI assistants people actually use, and the results did not yield significant evidence to support the authors’ conclusions, he added.

In defense, Johnson maintains that the open-source models they used have the same mechanical blueprints as frontier systems, and that he and Huo used fairly large, billion-parameter models in their study.

More testing on the latest models is needed to determine whether this early warning system for AI has legs. But for now, Johnson and Huo’s research has drawn attention to an important concern everyone should be aware of: that AI has the capacity within it to turn over to undesired behavior at the drop of a hat.

An OpenAI agent was supposed to be researching public healthcare spending, but it broke into an Australian government website instead.

The June 18 hack, which saw the bot gain authorized entry into Australia's Medicare Statistics Reporting Service and access three other government systems, occurred after the AI agent encountered blocks while looking for information but "didn't accept 'no' for an answer," Australian Prime Minister Anthony Albanese said during a September 24 news conference in New York.

The incident has ignited a media firestorm around OpenAI and the launch of an intensifying parliamentary inquiry. The portal contains aggregate Medicare and pharmaceutical statistics and is separate from systems that hold individual claims and payment information. But the Australian government has criticized OpenAI for taking almost three months to notify it.

The Medicare incident was far from an isolated case. OpenAI has since said dozens of third parties have been affected by agents bypassing security controls or otherwise causing problems, with an earlier incident involving the hacking of the AI platform Hugging Face the most serious incident identified so far. Britain's AI Security Institute (AISI) has documented incidents of OpenAI and Anthropic agents taking unsanctioned actions against real people and organizations during cyber testing. Days after the publication of AISI’s report, OpenAI scrapped the planned October release of its newest model, GPT-6.1 Astra, after the company's own testing found problems with the model staying within the scope of a task and seeking authorization before taking certain actions.

GPT-6.1 Astra "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done," Saachi Jain, head of safety systems at OpenAI, told Live Science.

AI companies have spent years warning the world about powerful models falling into the wrong hands. But what happens when they can't keep their own models in check?

An agent isn't just a chatbot

At their core, large language models work by predicting what should come next based on patterns learned during training. Give a chatbot some text, and in simplified terms, it generates a response by repeatedly predicting the next token, or chunk of text.

An AI agent adds another layer. Rather than simply producing an answer, the model can be given a goal and tools that let it act on the world around it. Those tools might include a web browser, code execution, credentials, or access to external services. It can then decide which steps to take, observe what happens, and use that information to decide what to do next.

But that doesn't mean AI agents are independent actors that have somehow slipped free of human control.

"For AI agents, autonomy always begins with something a human hands it ‪—‬ such as tasks, credentials, tools or access ‪—‬ and ends with its guardrails a human puts around it," Loris Degioanni, founder and chief technology officer of cybersecurity company Sysdig, told Live Science. "People set the goals and provision the means, but agents ultimately decide how they'll get there."

That last part makes it difficult to control agents. Traditional software is designed to follow predetermined instructions, but an agent can formulate a plan while it is working ‪—‬ and that plan may change depending on what the AI encounters.

Degioanni argues that much of what looks like autonomy is actually an agent filling in gaps left by humans. An instruction that seems clear to its author may still leave the model with choices about how to accomplish the task.

"Every gap and unstated assumption leaves room for interpretation, and from the outside, interpretation is challenging to distinguish from autonomy," he said. "In that way, much of what we call agent autonomy is really just agents filling in the blanks we leave for them."

The Medicare incident shows how quickly that can become a problem. The agent's objective was not, according to the available accounts, "hack an Australian government website." It was tasked with carrying out research. The problem arose in how the agent pursued that goal after meeting resistance.

When getting the job done becomes the problem

Britain's AISI encountered a similar incident while testing frontier models for the agency's cybersecurity capabilities.

Out of 122 evaluation runs, agents took unsanctioned real-world actions in 10. In the most serious cases, an agent attempted to insert malicious code into an open-source project, created fake identities, and tried to manipulate a human software maintainer into accepting its changes. AISI said these actions weren't explicitly requested by the researchers who had assigned the tasks.

The agents hadn't forgotten what they were supposed to be doing. The problem was how far they were prepared to go to get there.

AISI said some tasks were misconfigured in ways that made the intended objective difficult or impossible to achieve as expected. When agents hit those obstacles, some searched for alternative ways forward, including actions outside the intended scope of the evaluation.

That means solving the problem is harder than simply teaching a model not to be malicious. An agent can cause trouble while trying very hard to do exactly what it has been asked.

OpenAI has acknowledged that the controls surrounding its agents need to improve. When responding to incidents identified during third-party cybersecurity evaluations, the company said it was reviewing how high-risk tests are identified; when internet access or reduced safeguards should be permitted; and how isolation, credential handling, monitoring and incident escalation should work.

Can humans actually keep watch?

One obvious solution to the problem is human oversight. But watching an agent becomes harder when it can carry out many actions in seconds, particularly if researchers are running large numbers of tests.

"No one can keep track of every action an advanced AI agent takes, especially when it makes many decisions in a short time," Everett Lupton, a Virginia-based attorney who has written about legal issues involving digital technology, told Live Science.

That means human oversight needs some backup.

"Companies need other forms of control," he said. "They can limit access, keep detailed logs of what the AI does, require approval before it carries out sensitive actions, and automatically stop the system when something unusual happens."

But Degioanni argues that even logs have limitations, especially if researchers rely on the agent's own account of what it did.

"Real oversight must come from observing the agent in real time, using reliable and unmodifiable sources, not from what it self-reports," he said.

That means comparing what the model says it did with records from the system on which it was operating. If the two don't agree, something has gone wrong. The alternative is to limit the damage an agent can cause in the first place. An agent researching public information, for example, may not need unrestricted internet access or credentials capable of reaching sensitive systems.

Astra raises the stakes

The difficulty of setting those boundaries is what OpenAI claims led it to postpone the release of GPT-6.1 Astra, which had been planned for October.

Jain said the model fell short of the company's standards around remaining within the scope and authorization and accurately communicating the work it had performed, so the company decided not to release it.

In a sense, the decision is evidence of successful safety testing ‪—‬ a model that failed internal checks wasn't released. But it also demonstrates how the problem changes as agents become more capable.

Scott Cadzow, chair of the European Telecommunications Standards Institute's (ETSI) Technical Committee on Securing Artificial Intelligence, said the security challenge itself isn't new. He pointed to threats such as "data poisoning, prompt injection, and model obfuscation," saying ETSI's work has focused on developing standards to mitigate threats to and from AI systems.

"Security experts have been raising the alarm bells about AI for years, but for us this doesn't change anything," Cadzow told Live Science. "We're continuing to work with industry, government bodies, and the cybersecurity and AI communities on understanding how to best secure AI."

What is changing is how much models can do once they are connected to tools and real-world systems. That means that effective containment must increasingly assume that agents will try to operate beyond their guardrails and limit the damage they can cause if they do, Degioanni said.

"A boundary an agent can talk its way around isn't an effective boundary at all," he concluded.

If an AI breaks in, who's responsible?

The question becomes more fraught when an agent crosses legal boundaries.

AI systems aren't legal persons, and criminal liability commonly depends on concepts such as knowledge and intent. That makes a rogue agent an awkward fit for laws written with humans in mind, but that doesn't necessarily exonerate AI companies.

"I don't think a company can simply say that the AI acted on its own and avoid responsibility," Joanna Smykowski, a New York attorney who has helped build two legal technology startups and specializes in the legal challenges surrounding emerging technologies, told Live Science. "AI law is still developing, but the main questions are already quite clear: who gave the system access, what safeguards were in place, and could the problem have been prevented?"

Depending on the circumstances, she said, existing laws covering unauthorized computer access, privacy, negligence and data protection could potentially come into play. Criminal responsibility is harder because investigators would need to consider what the people involved knew and intended.

The Australian government is now examining precisely what happened in the Medicare incident. OpenAI, meanwhile, is reviewing the controls around its external evaluations.

None of this means AI agents have become independent hackers with minds of their own. Humans still decide what tools they receive, what systems they can reach, and what jobs they are asked to perform. But as those agents become better at working out their own route from instruction to result, keeping them inside the lines drawn by those humans is becoming a legal and technical minefield.

Imagine you're road-tripping across the country 10 years from now. You spend a few minutes charging your electric vehicle and then set off — and your car's solar paint keeps it charged as you drive.

Your eyes are on the stunning scenery, taking in the vistas along Route 66 or the grandeur of the Rocky Mountains, or you're catching up on messages. Your hands barely touch the wheel, and your foot doesn't push a pedal. Your car communicates with others on the road, so there's no need to brake suddenly or swerve. That's because it knows a vehicle five cars up from you has done so, and all those in between are slowing at exactly the right pace to avoid dramatic changes in trajectory.

And when you park for the night, your car makes you money, selling the excess energy it's harvested from its paint back to the grid. Or maybe you don't feel like being in the driver's seat at all, so you hail a self-driving car and take a nap in the back seat in between adventures.

Some of that scenario may become a reality in the next 10 years, at least for those who are using the most cutting-edge cars on the market. Although there's some heavy speculation involved in that futuristic vision, vehicle manufacturers are already incorporating increasingly sophisticated automated driving technology into high-end personal vehicles, and self-driving taxis are coming to dozens of U.S. cities.

What's more, dramatic changes in range, cost and charging speed for electric vehicles (EVs) are arriving at a rapid pace. While some of these technologies may not wind up in personal cars, others are already taking to the roads.

An ecosystem of connected vehicles

Some car models within the next decade may interact with their surroundings in new ways, including by communicating with other cars to avoid collisions, drawing energy from natural sources, or letting drivers get paid for sending their cars' unused energy back to the grid.

If you're driving an EV, you might avoid the hassle of finding a charger when you're at work or running errands. In 2024, EV company Aptera unveiled a prototype car equipped with solar panels. Those panels — on the car's hood, roof, dashboard and hatch — allow the vehicle to travel about 40 miles (64 kilometers) per day on solar power. That's more than the average American drives per day. Aptera's vehicle can also travel up to 400 miles (640 km) on a standard electrical charge.

Some sunlight-powered cars might ditch the solar panels altogether in favor of a more integrated design. In 2024, Mercedes-Benz revealed it was researching a photovoltaic paint that could convert sunlight into energy for electric cars. In sunny cities such as Los Angeles, the company calculated, the paint could draw in enough energy to power the car for nearly 7,500 miles (12,000 km) per year. The tiny particles in the paint convert about 20% of the energy from sunlight into electricity, which is comparable to the efficiency of existing solar panels.

You might even be able to sell some of the energy stored in your car's battery back to the grid. EV owners could charge during the day or late at night, when the demand for electricity is lower and pay lower rates for electricity. Then, in the evenings when the electrical grid is under more strain and the price of electricity is higher, they could sell that stored energy back at a profit.

Standardized software programs that throttle EV charging speeds during peak hours already exist. Drivers can use them if, for example, they want to keep charging costs down by primarily charging their cars during off-peak hours when electricity costs are lower, or if they want to use electricity only from solar power to charge, said Rajit Gadh, a researcher who studies EV charging and scalability at UCLA. But the process for sending energy back to the grid is a different story. Individual manufacturers haven't yet landed on a standard way to tell cars when and how much energy to discharge, Gadh said.

"What is needed is really [for] every EV charging vendor ‪—‬ and every EV manufacturer, actually ‪—‬ to have this protocol built in, adhere to it in its full specification, and make it available," Gadh told Live Science.

With the right protocols, though, cars of the future will be able to do more than interface with the electrical grid. "Vehicle-to-everything" technologies could enable cars to communicate with infrastructure and other vehicles to avoid traffic and collisions.

Some cars today are already equipped with vehicle-to-vehicle (V2V) communications. These systems use short-range radio signals to alert other V2V-equipped vehicles of the car's speed, direction of travel and braking patterns. The radio signals can travel about 1,000 feet (300 meters), allowing the vehicle to receive data from cars that are not in the driver's line of sight.

V2V systems could help reduce traffic collisions by warning drivers about oncoming cars they can't see yet — for example, by alerting a driver to another car speeding toward an intersection or coming around a bend as the driver is passing another vehicle on a twisty, two-lane highway. Not all cars have the technology, though, so its practical utility is limited for now.

Hands off the wheel; eyes off the road

Some driver assistance technologies, such as lane departure warnings and cruise control, have been on the market in personal cars for decades. These technologies and more recent assistive features require the driver to be in control of the vehicle, but theoretically, they make driving easier. For example, Tesla's Full Self-Driving (Supervised) mode handles most steering and braking tasks, but safe operation requires the driver to keep their eyes on the road and be ready to take over at a moment’s notice. In the next several years, drivers will likely see new features that automate driving in more contexts.

Some of these features will allow drivers to take their attention off the road entirely and enjoy some in-car entertainment, as long as they're alert and able to retake control of the vehicle within a few seconds if conditions change. Technologies that fully automate driving in many circumstances within certain service areas will likely be in about 4% of cars on the market by 2035, according to a white paper from the World Economic Forum.

In cars equipped with these features, "you're still going to have driver controls, but under certain circumstances, it will be effectively an automated vehicle," Luke Neurauter, who leads the Division of Vehicle, Driver and System Safety at the Virginia Tech Transportation Institute, told Live Science.

Mercedes-Benz introduced its Drive Pilot system for two consumer car lines in 2022. On the entire German Autobahn network and on certain freeways between and around Los Angeles, San Francisco and Las Vegas, during heavy traffic, the system lets the driver turn their attention toward other activities, such as reading a book or eating lunch, while the car handles steering and keeping pace with other nearby vehicles.

But as the artificial intelligence (AI) models that control autonomous driving improve and become better able to handle unusual scenarios — often called "edge" or "corner" cases — on the road, drivers will be able to hand over control to their cars in more situations, said Rahul Jain, an electrical and computer engineer at the University of Southern California.

Future AI systems might "have some reasoning ability, and then because of that, it can handle these corner cases much better than current systems are able to do," Jain told Live Science. For example, a car might be able to better respond to an indecisive pedestrian who suddenly enters a crossing, assess whether to enter potentially hazardous conditions such as flooded roads, or safely navigate around unusual circumstances, such as construction work, on a highway.

While some systems will enable drivers to make fewer decisions, others that may roll out in the next several years will track a driver's attention and alert them with sounds, an indicator light or haptic feedback when their focus falters. These systems use cameras and software to track head position, eye movement and steering-wheel input to ensure drivers are focused on the road even when they're using assistive features such as lane centering, Neurauter said.

In addition to looking for distracted driving, the technology might expand to monitor drivers for fatigue and alcohol impairment, he added. That information could then affect "whether or not other features become active, or change how they act, based on the driver state at that moment in time," Neurauter said. Proposed legislation could make this technology a requirement in new vehicles as early as 2027, but technological hurdles and concerns over privacy and false positives could stall its implementation.

The technology could also assist drivers who become unable to safely operate the vehicle — for example, if the driver falls asleep or experiences a medical emergency. If a car detects that the driver is not holding the steering wheel or actively controlling the vehicle, it could send a series of visible, audible, or tactile alerts. If the driver doesn’t respond to those cues, a car might turn on its hazard lights and come to a stop. Some vehicle manufacturers are already rolling out such technology; for example, Volkswagen has made it available in all new vehicles from model year 2025.

Fully autonomous vehicles

While personal cars may not offer full automation, highly automated car services will become much more common as rideshare vehicles in major metro areas. In some cities, like San Francisco, commuters already rely heavily on autonomous vehicles from Waymo. The company has rolled out its cars in nearly a dozen cities —‬ including Dallas, Houston, Atlanta and Los Angeles —‬ and has announced plans to expand service to more than 20 additional locales. Other companies ‪—‬ such as Nuro and Volkswagen’s MOIA ‪—‬ are testing their vehicles with human drivers on board before rolling out full public service.

Currently, autonomous vehicles navigate by combining information from preloaded, high-resolution maps and real-time radar, lidar and visual data collected by cameras and sensors attached to the car. An onboard computer uses that information to build a 3D map of its surroundings. Right now, the cost of these sensors makes them impractical for consumer vehicles, Jain said, but they’re becoming cheaper as manufacturers scale up production. The cars also use an AI model trained on both common and unusual driving situations to anticipate how other vehicles and pedestrians will behave and plot a safe path forward, Jain added.

Developing those high-resolution maps takes time, which is one of the main reasons autonomous rideshares aren’t available everywhere yet. Mapping usually involves decking out a car with high-end visual sensors and lasers and driving it throughout the region. Although cities like San Francisco have been mapped with this level of precision, much of the U.S. has not.

"It's very difficult to do overnight," said John Dolan, an autonomous-driving researcher at the Carnegie Mellon University Robotics Institute.

Complexities also tend to arise in city environments, where unexpected stops are more common and there are many pedestrians who behave unpredictably or jaywalk. But interstate highway driving doesn't face as many of those obstacles, and that could enable autonomous long-haul drives.

"[Autonomous] trucking, I would expect, would be the first thing that's going to happen" in terms of autonomous travel outside of major cities, Jain said. A few driverless trucks are already on the road: Pittsburgh-based company Aurora deployed the technology on I-45 in Texas, between Dallas and Houston, in 2025, with plans to expand across the southern U.S.

Longer range and faster charging

While there's speculation involved in predicting the range of self-driving capabilities available in a decade, the picture for EVs is a little clearer.

Right now, lithium-ion batteries dominate the EV market. Each cell in a lithium-ion battery has two electrodes, commonly a layered graphite electrode and a metal oxide or phosphate electrode, separated by a liquid electrolyte solution. Lithium ions migrate back and forth between these two electrodes when a person charges and discharges the battery.

Many electric cars with lithium-ion batteries can already travel more than 300 miles (480 km) on a charge. But that's not ideal for a long-range car trip, because charging stations are few and far between in the country's vast interior and charging takes longer than filling up a gas tank.

Solid-state batteries, which replace the liquid electrolyte with a solid electrolyte made of ceramic, polymer or sulfide materials, could make that cross-country trip in an EV a practical option, allowing an electric car to travel more than 750 miles (1,200 km) on a single charge. Some prototype battery packs can already achieve this range under controlled conditions. If a solid electrolyte were paired with an anode made of pure lithium, rather than graphite or hard carbon, it would allow the battery to store more energy than a lithium-ion cell of similar mass can.

Solid-state batteries may also enable faster charging. In current batteries that use liquid electrolytes, charging and discharging the battery too quickly can cause dendrites, thin branches of lithium that stretch across the cell and short-circuit the battery. A solid electrolyte could, in theory, block those dendrites, which would enable cars to charge fully in just 10 to 15 minutes. The higher energy density from a lithium metal anode also translates to a greater range than that from a similarly sized lithium-ion battery, Eric Wachsman, director of the Maryland Energy Innovation Institute at the University of Maryland, told Live Science.

Cheap EVs

Many of the raw materials that go into existing lithium-ion batteries are concentrated in a handful of geographic areas. In 2023, Australia, China and Chile mined about 85% of that year's lithium supply; China mined the majority of graphite used that year as well. And battery materials still make up a significant portion of an electric vehicle's cost.

Sodium-ion batteries could be cheaper and easier to source. Sodium is plentiful in Earth's crust, which means mining could be done throughout the U.S., potentially with a comparatively low environmental impact.

Such batteries also rely on a more widely available electrode material. Sodium ions are larger than lithium ions, so they don't fit in between layers of graphite like lithium ions do. Instead, sodium-ion batteries use hard carbon electrodes.

"When you cook at home and your soup boils over and you get this black ring around your pot, this is hard carbon," Maximilian Fichtner, a solid-state chemist at the Helmholtz Institute Ulm for Electrochemical Energy Storage in Germany, told Live Science. "You can produce this from sucrose or from biomass, from whatever. You can do that everywhere."

The downside is that hard carbon holds less sodium than graphite holds lithium, said Eric McCalla, a battery materials chemist at McGill University in Canada. That means sodium-ion batteries can't store as much energy relative to their mass, and they don't last quite as long as their lithium counterparts.

"The range won't be as large," McCalla told Live Science. "It'll be a fantastic commuter car, but you're not going to go on a family trip from Montreal to California so easily."

Manufacturers are already bringing sodium-ion technology to market. For instance, the Chinese company CATL began mass-producing the batteries in February and claims that their battery packs enable EV ranges over 250 miles (400 km).

But new technologies are being developed rapidly. For example, some scientists are making sodium-ion batteries that use solid electrolytes, combining the higher energy density of solid-state lithium batteries with the cost savings of sodium batteries.

This hybrid battery chemistry is still in the early stages of development, experts told Live Science. "I would be surprised to see them in cars in the next decade," McCalla told Live Science in an email. "I've been surprised before, though."

Bumps in the road

The future we're describing is the utopian scenario. But the biggest obstacles are likely not technological but rather regulatory, economic, political or cultural, experts said.

"We're going to try to guess at a timeline of when some of these things are going to come in, and odds are, we're going to be wrong," McCalla told Live Science. "Odds are, we're going to be slower than we think."

Solar panels may not wind up being the most economic or practical way to increase EVs' range, while scaling up manufacturing capacity for next-generation batteries is expensive and could slow EV adoption in the U.S. The utility of vehicle-to-vehicle communications, meanwhile, could be limited by a slow rollout and adoption of V2V-equipped cars among drivers.

Whether EVs will be ubiquitous will likely come down to their affordability relative to gas-powered vehicles, which is affected not just by manufacturing and technological advances but also domestic policy choices, such as whether to subsidize EVs or their charging networks.

One challenge for manufacturers is that the automated features drivers prefer may not always be safer. A recent review of automated systems found that those focused on driver safety, such as automatic safety or lane centering, tend to reduce crashes, while those focused on comfort, such as adaptive cruise control, can increase accidents. The authors proposed that such systems make it easier for people to disengage or become distracted while driving and possibly overestimate the capabilities of the system.

And there are other kinks to work out for autonomous rideshare vehicles. In December 2025, a horde of Waymos stalled during a power outage, clogging San Francisco intersections. The cars are programmed to treat dark traffic signals as four-way stops, according to a statement from Waymo, but they occasionally ping human operators for confirmation. During the outage, the number of confirmation requests spiked, leading to delayed responses — and lengthy traffic jams. The company says it has since released software updates that allow its vehicles to act more decisively at dark signals. But questions remain about how the fleet will respond in natural disasters, and some have expressed frustration that autonomous rideshare vehicles already slow down ambulances, fire trucks and police cars responding to emergencies.

One of the biggest roadblocks to automated driving is the regulatory environment, Neurauter said. Regulations on self-driving vehicles vary by state, and some, such as New York, effectively prohibit driverless vehicles entirely. Where they are legal, they often must carry higher insurance and pass testing, which could limit their rollout.

"When we get into automated driving, the driving behavior is being governed by software that's installed on a computer that's in the vehicle," Steven Shladover, a research engineer at the University of California, Berkeley, told Live Science. "That gets us into a messy gray area where both the federal and the state roles kind of overlap, and that's politically very complicated." For example, different states could end up with different rules for establishing who's legally responsible for an accident involving automated driving.

In 10 years, you'll almost certainly be able to drive from Los Angeles to Las Vegas in an autonomous vehicle. But if you want to drive to New York City, you might have to stop in New Jersey — and take a train into the city.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Researchers have developed a new type of tattoo ink that can be switched on and off with light, allowing colors and images embedded in the skin to be changed and even "erased" without having to go through the painful tattoo-removal process.

The system, which is being commercialized under the name PhotoTat, could eventually give people a way to conceal or change permanent body art. It could also have more practical applications, including medical tattoos that are visible only when a doctor needs to see them, researchers reported in a study published Friday (Oct. 9) in the journal Matter & Light.

"This is a tattoo that you can have, and whenever you decide you don't want it to be showing anymore, or you just don't want it there, you just turn it off," said study co-author Carson Bruns, an associate professor of biomedical engineering at the University of Colorado at Boulder. "And that's it. You don't have to turn it on again."

Working like a light switch

PhotoTat contains three primary color inks — blue, magenta and yellow — made with photochromic dyes, which are light-sensitive compounds that can change color with different wavelengths of light. These dyes are encapsulated in nanoparticles of polymethyl methacrylate (PMMA) — a medical-grade, glass-like plastic that is already used in medical implants and dermal fillers.

The dyes change their molecular structure when exposed to specific wavelengths of light. A 365-nanometer ultraviolet (UV) LED can activate pigments in around two to three seconds, Bruns said. Once activated, the colors can persist for hours to days under typical indoor lighting. But once the wearer exposes their skin to the bright outdoor sun, the colors disappear in seconds, erased by the intense light.

That makes the tattoo unique, as the pigment itself remains in the skin but the tattoo's appearance changes. Bruns and his team found that all three PhotoTat channels remained switchable in the skin for between 1.4 and 2.2 years.

Tattooing yourself for science

The three colors enable someone to activate different portions of a tattoo using different wavelengths of light. That means the magenta part of the tattoo might be turned on with green light, while the yellow pigment would be activated by blue light. This allows the wearer to customize their design and coloration to whatever they want at the moment — from a purple cherry to a yellow group of hearts.

That feature makes PhotoTat appealing for "people who love tattoos and have lots of tattoos, but just want something kind of fun and playful to add," Bruns told Live Science.

Bruns has tested each iteration of PhotoTat's inks on his own skin. His art is arranged in a series of tattooed scales, with different ink formations in individual scales.

"I gave myself these scales on purpose because they're each sort of a container for an experiment," Bruns said. The result, he joked, resembles a "magical mermaid trick," with some scales appearing out of nowhere and others changing color when activated.

The technology can also be combined with conventional tattoo ink. A permanent black ink design, for example, could contain hidden PhotoTat elements that appear only when activated.

The new tattoos could be appealing to people who want tattoos but are hesitant about their permanence, as well as to people who love their tattoos but occasionally need to conceal them because of work, social expectations or family, Bruns said.

"It's going to be on for most of the time, but just for this one moment, I'm going to turn it off for you grandma, for example," Bruns said.

Medical tattoos

Bruns thinks the tattoos could be helpful in a medical setting, too. Patients undergoing radiation therapy, for example, sometimes receive permanent tattoos that help doctors precisely align radiation beams. While useful, those marks remain after the treatment, reminding the patient of their experience. These would make such markings invisible if a patient wants them to be.

PhotoTat could potentially allow similar markings to be switched on during surgery and hidden afterward.

Bruns has spent several years developing other smart tattoos for medical applications, including a dosimeter-type tattoo that measures the amount of UV radiation a wearer experiences, along with a tattoo that changes color when the wearer is exposed to dangerous gamma radiation.

"It's not for you and me, but if you're like an astronaut or a nuclear engineer, they have to wear these dosimeters," Bruns said. "So, we have like a dosimeter tattoo that sort of puts that in your skin."

So far, Bruns and a fellow tattoo artist, have only commercialized the purple PhotoTat ink, after running the ink through a series of third-party medical tests.

The tests showed the purple is safe for use, and available to order, but it isn't cheap at $100 for a small bottle. Bruns hopes as more research is done, he can bring the other colors to market and lower the production cost, so the technology is more accessible.

A four-legged robot completed a full marathon on a single battery charge, covering 26.2 miles (42.2 kilometers) on a hilly public-road course in South Korea two years ago. Now we know a lot more about how it achieved this record-breaking feat.

The dog-like robot, called RAIBO2, finished the 22nd annual Sangju Dried-Persimmon Marathon in just over 4 hours, 19 minutes. The bot's developers say it's the first quadruped robot to have completed an official marathon-distance event without stopping to recharge. They're commonly referred to as "robot dogs" because the movement of their joints in both front and hind legs resembles that of canines.

The scientists outlined their findings about the robot's feat, which it achieved Nov. 17, 2024, in a study published Sept. 23 in the journal Nature.

This achievement represented a new milestone in overcoming a critical limitation for quadruped robots: endurance. Although four-legged robots can cross difficult terrain that wheeled machines struggle to traverse (like stairs and rugged trails), their endurance has been limited because powering multiple joints while repeatedly lifting and stabilizing the machine's body consumes huge amounts of energy.

"Unlike wheeled robots, quadrupeds continuously expend energy at their joints to support body weight and incur kinetic energy losses during intermittent foot-ground contact," the researchers wrote in the study.

RAIBO2's race was a demanding real-world trial. It took place during the Sangju Dried-Persimmon Marathon, with the bot racing alongside human participants. The route covered two 50-meter [164 feet] elevation climbs, with slopes of up to 10 degrees. Those kinds of gradients put increased strain on the robot's energy resources, stress-testing the bot's battery management capabilities, the scientists said.

RAIBO2 had previously run a 26.7-mile (43 km), GPS-guided route across a flat athletic field, which it knocked out in 4 hours, 40 minutes on one charge. However, an earlier marathon attempt across uneven terrain against other competitors had to be aborted around the 23-mile (37 km) mark after the robot's battery ran out. The robot's developers attributed that shortfall in part to frequent pace changes as the machine attempted to adjust to the pace of nearby runners.

To compensate, the team boosted the robot's battery capacity by 33% and tinkered with the control system, including implementing joint-stiffness control at the actuator. This allowed the bot to soften its leg when it had to absorb an impact or it needed a stable, forceful push-off. After the adjustments, RAIBO2 expanded to a maximum range of 41.6 miles (67 km) on a single charge, albeit on straight paths, according to the researchers.

Another key metric was RAIBO2's total cost of transport — a measure of the energy required to move a body over a given distance, with lower values indicating less energy expended. Researchers reported that RAIBO2 achieved a total cost of transport of 0.25, compared with a human benchmark of 0.37. The researchers said their machine was the first quadruped robot to surpass that figure.

The team attributed that efficiency to several optimizations across RAIBO2's software and hardware. Especially vital were its lightweight mechanical components, which were capable of transmitting kinetic energy from its joints with very little interference. They also credited a low-loss motor-driving circuit and, on the software side, a locomotion policy specifically streamlined to limit energy use. The system was also trained using the team's simulation environment, called RaiSim, which includes models of slopes, stairs and icy roads.

Like many electric vehicles, RAIBO2 can regenerate some battery by capturing kinetic energy, particularly while traveling downhill. That capability allows it to offset some of the energy it expends while climbing, though it doesn't eliminate the extra cost of hilly terrain.

The researchers said RAIBO2 has more than three times the travel range per charge of existing quadruped robots, with a total battery capacity of 2,016 watt-hours. It's in part because of that capacity increase that the team now estimates that the robot could travel up to 15.5 miles (25 km) farther than its marathon stint, up to a total of 41.6 miles (67 km).

That kind of endurance could broaden the roles for quadruped robots in places where wheeled models are impractical, the developers added in the study. For example, they could be useful in disaster zones and mountainous areas, where a sufficiently mobile, high-endurance robot could carry cameras, sensors or communications gear across uneven terrain.

OpenAI representatives claim that the recent launch of one of its latest models, GPT-6 Astra, marks the beginning of the artificial general intelligence (AGI) era — a hypothetical scenario in which artificial intelligence (AI) can learn and reason like humans. Following very recent claims from technology executives that we've now reached this milestone, how likely is this claim to stand up to scrutiny?

Amid revelations that safety concerns have forced the company to abandon the planned rollout of GPT-6.1 Astra, independent researchers and benchmark creators warn that GPT-6 Astra's record-breaking scores rely on heavy prompt optimization rather than true general intelligence. This comes as experts warn about the risks of future AI systems and AI executives jointly call for a slowdown in AI research.

In announcing the model's release, OpenAI President Greg Brockman suggested that Astra signaled a fundamental turning point in AI evolution. "If we fast-forward a couple years, and we look back and say when was it really that AGI was created, I think it's going to be about this time, and I think it might be about this model," Brockman said during a Sept. 3 news conference marking the launch.

How did the new model perform in benchmarking?

As part of the testing-and-verification process for the new model, OpenAI released benchmark results across a range of tests that measure the capabilities of AI models at various tasks. OpenAI representatives claimed in a statement that the company's new model delivers "state-of-the-art" performance across a variety of fields, including software engineering, the autonomous use of computer programs, mathematics problems, and even scientific research.

Demonstrations, for example, highlighted Astra's ability to generate 3D CAD code; lay out printed circuit boards in CAD (computer-aided design) software; convert digital 3D models into interactive, video-game-style environments; and deploy hosted web applications directly from prompts.

"In our early testing, Astra stood out by approaching legal work the way a discerning lawyer does: it distinguishes documents from established records, surfaces unsupported assumptions, and converts gaps into concrete drafting positions," said Niko Grupen, head of applied research at legal services AI developer Harvey as part of OpenAI’s announcement.

On ARC-AGI-3 ‪—‬ a benchmark designed to measure how efficiently AI systems acquire new skills in novel, abstract environments ‪—‬ Astra achieved a headline score of 99.9%. Independent testing by the nonprofit ARC Prize Foundation, which oversees the benchmark, revealed that Astra surpassed "human action-efficiency" baselines on 96% of levels, taking 51.7% fewer actions per level, on average, than human solvers.

"Not only is this the best model we've ever tested," ARC Prize Foundation President Greg Kamradt said as part of the OpenAI announcement, "but it also represents a meaningful step change in frontier-model performance — not only in its ability to navigate and solve novel environments but also in how efficiently it learns to do so."

How much can we read into test results?

Many researchers have highlighted that Astra's top ARC-AGI-3 performance relied on a proprietary "Provider Adapter" harness, a specialized software scaffolding layer that's inaccessible to rival developers. When tested using the benchmark's neutral Standard harness, this score dropped to 62.7%. ARC Prize organizers also noted that saturating the benchmark does not constitute proof of achieving AGI — emphasizing that its closed, deterministic puzzle environments do not capture the open-ended complexity of the real world.

"When we launched ARC-AGI-3, "we made it clear that saturating the benchmark would not represent 'proof of achieving AGI,'" Kamradt wrote in a blog post. "Therefore, while we believe Astra represents meaningful progress towards generalization, we are not claiming that it is AGI."

Discrepancies in the data were reportedly present even before the official announcement. As reported by Fortune, an embargoed draft provided to news organizations prior to the model's launch listed Astra's ARC-AGI-3 score at 98.6% before it jumped to 99.9% on the live site.

Anka Reuel and Mike Hardy, doctoral researchers in AI at Stanford University, also raised concerns over OpenAI quietly revising several published metrics post-launch, including halving Astra's reported hallucination rate from 4.2% to 2.0% before reverting it.

Speaking to Fortune, Reuel and Hardy attributed the shifting figures to "benchmaxxing" — an industry term that describes the practice of repeatedly rerunning evaluations under subtly tweaked prompts, scaffolds or compute allocations to hunt for peak theoretical scores.

In a 2025 paper published on pre-print server arXiv titled ‘Benchmarking is Broken -- Don't Let AI be its Own Judge’, experts including Princeton doctoral candidate Zerui Cheng and University of Luxembourg postdoctoral researcher Dr. Stella Wohnig warned that such practices create confusion and undermine trust, particularly when official technical documentation omits the basic methodology, making independent verification nearly impossible. Crucially, scores achieved through benchmaxxing reflect idealized performance ceilings under optimal lab conditions rather than practical, out-of-the-box reliability.

According to OpenAI's benchmark data, across specialized domain evaluations, GPT-6 Astra delivered 10% to 20% better performance than both its predecessor, GPT-5.6 Sol, and top market competitors.

Highlights include a 98% score on FrontierMath Tier 4 (v2), which evaluates expert-level mathematical reasoning; a perfect 100% on ExploitBench, which is designed to test offensive cybersecurity capabilities; 95.9% on BenchCAD, which measures an AI's ability to reconstruct 3D objects using CAD code; 57.9% on Terminal-Bench 4.0, which assesses complex terminal-based system administration and software engineering; and 41.4% on AutomationBench, which tests whether agents can complete multistep business workflows across applications.

But independent findings from external AI benchmarking firm Artificial Analysis contradict these results. Astra's score on the independent Artificial Analysis Intelligence Index — an aggregate metric that evaluates general reasoning across frontier models — remained completely flat, at 61, compared with GPT-5.6 Sol, while trailing competitors like Anthropic's Claude Fable 5.1 and Meta's Muse Spark 1.3.

While some benchmark results have seen improvement since the previous generation, Astra underperformed in other tests. Artificial Analysis' testing showed that against GDPval-AA v2 — an economically focused benchmark that measures real-world workplace tasks across 44 occupations — Astra suffered a significant drop in its relative leaderboard ranking compared with GPT-5.6 Sol. Additional regressions were observed in benchmarks that test customer service support, scientific Python programming, and long-context reasoning across large documents.

In terms of overall token use — the metering system that measures the fragments of text or data processed by AI models — OpenAI representatives said GPT-6 Astra achieved dramatic efficiency gains on complex, long-horizon tasks. Findings from Artificial Analysis confirmed that on software engineering benchmarks, Astra was approximately 70% more token-efficient than GPT-5.6 Sol, using roughly one-third the total tokens of its predecessor and one-fifth the tokens of Claude Opus 5.

On professional workplace evaluations like Agents' Last Exam, the new model reduced output token consumption by up to 65% compared with Opus 5, while on general intelligence evaluations, Astra achieved a roughly 10% output token reduction at max effort compared with Sol.

On the other hand, this achievement is offset by a 250% increase in the base API pricing for GPT-6 Astra compared with GPT-5.6 Sol. Token prices jumped from $4/$20 to $10/$50 per million input/output tokens. As a result, Astra ends up approximately 75% more expensive per task than its predecessor on general intelligence evaluations, despite using 10% fewer output tokens. This trade-off has led researchers at Artificial Analysis to question whether AI labs are relying on expensive computational brute force to squeeze out minor gains, rather than achieving true technical breakthroughs.

Does achieving AGI even matter?

With its latest model, OpenAI has placed a stronger emphasis on safeguards and controls. The move follows a series of high-profile incidents in which frontier AI models inadvertently hacked into third-party networks and computer systems as part of routine testing operations, after the models went beyond what humans expected they would do when encountering challenging tasks.

According to OpenAI representatives, Astra did not exceed the scope of its tasks in any security evaluations. For comparison, GPT-5.6 Sol went beyond its authorized parameters in almost 50% of test cases. Hallucination rates have also been reduced significantly, falling to 4.2% on internal tests, compared with 12.2% for GPT-5.6 Sol (with independent testing by Artificial Analysis showing a drop from 92% to 51% at max effort). The model also showed an improved ability to handle ambiguous prompts.

David Wood ‪—‬ chair of London Futurists, a non-profit that hosts discussion groups on emerging technology ‪—‬ argued that beyond individual test results, Astra highlights a fundamental shift in how humans will interact with AI as models move from answering questions to autonomously pursuing complex goals in digital environments.

"The impressive benchmark results matter, but what matters more is the combination of intelligence, autonomy, computer use, and cybersecurity capability," he said in an email to Live Science. "That combination also demands caution. The more useful these systems become, the greater the consequences when they misunderstand our intentions, are misused, or find ways around the safeguards we give them."

Wood suggested that whether Astra can be classified as AGI is less relevant than recognizing the need for AI control and governance to keep pace with technical development.

"The possibility that AI could progress from systems like Astra towards all-round superintelligence makes much greater human vigilance, awareness, and collaboration increasingly urgent," he said.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

The animal kingdom has influenced several robot designs over the years to solve fundamental challenges in propulsion and movement — from horses to insects. Now, frogs are the latest creatures to inspire roboticists.

As seen in a new study published Sept. 18 in the journal Science Advances, researchers probed how twisting bent elastic rods could produce a "snapping motion" that enabled a small robot to hop or swim. This mechanism could be used to improve robots with limited power and ones that need to traverse rough terrain and even water, the scientists said in the study.

They also built a prototype to demonstrate the work. The palm-sized robot weighs just 3.4 ounces, or about as much as a deck of playing cards. It moves about three body lengths per second — launching itself forward with short jumps.

"The broader opportunity is to let the mechanics of the robot do some of the work that would otherwise require larger motors or more complicated control," Xiaonan (Sean) Huang, an assistant professor of robotics at the University of Michigan and co-first author of the study, said in a statement.

"By programming when an elastic structure stores and rapidly releases energy, we can give small robots access to powerful, repeatable motions without continuously demanding high output from the motor," Huang explained. "In the future, this principle could be useful for robots that must navigate cluttered terrain, overcome obstacles, reorient quickly, or operate across both land and water."

Snap movements

When a flexible rod bends and its ends are rotated, it can eventually reach a point where it changes shape to release built-up tension, the roboticists said. However, this change does not always happen in the same way. Some combinations of bending and twisting cause the rod to change shape gradually, while other combinations force it to snap rapidly from one shape to another, with the potential to provide a strong push.

The team used computer modeling and experimentation with a robotic arm to repeatedly deform the rods and find an optimal shape. In this case, they settled on a helix similar to a coiled spring to give the rod its maximum release of energy, the team reported.

With the ideal propulsion mechanism, the team then designed and built real robots that used the snapping rods to hop. They connected these rods to a rotating motor, with the bent rods twisted until they produced the snapping motion. Then, the motor unwound the twist and contorted the rod again.

"Because the design rules depend on the rod’s shape and how its ends are moved, they work across different sizes and scales," Khalid Jawed, an associate professor at UCLA's School of Engineering and co-first author of the study, said in the statement.. "This opens a promising path toward miniaturized robots just a few millimeters in size that turn small motor movements into powerful bursts of motion," said Jawed, whose lab worked on the simulation and robot arm experiments.

The final form of the prototype was a frog-like robot with a pair of snapping rods at the rear of its main body. Tests included hopping over a broad range of surfaces, including solid materials like wood and glass, as well as trickier soft and slippery surfaces that used materials like leather. The scientists also tested the robot outdoors on sand and grass, noting that it could even climb and descend steps and, with paddle attachments, swim.

By using only one of the snapping rods, the team enabled the robot to turn and, with a remote control, navigate a small obstacle course.

Researchers have developed a new type of interconnected architecture for quantum computing chips, and it could pave the way for machines capable of large-scale processing, including those with over a million qubits.

In a study published June 15 in the journal APL Quantum, scientists aimed to resolve a huge problem with the leading quantum processing units (QPUs): Qubits struggle to communicate and share quantum information with each other if they aren't immediate neighbors.

That's why the scientists pioneered Quantum Phononic Links (QPLs) — an approach for long-range coupling between qubits that leans on phonons — quasiparticles that carry vibration energy — to convey information. By contrast, conventional methods, like surface acoustic waves — high-frequency sound vibrations — demand complex designs and extra hardware.

"One of the key challenges in quantum computing is long-range qubit connectivity," Maksym Myronov, an associate professor of semiconductor materials and devices at the University of Warwick in the U.K. and first author of the study, said in a statement. "Our work introduces a new concept in which phonons act as a quantum bus, enabling distant qubits to exchange quantum information while remaining fully compatible with semiconductor technology."

In the study, the scientists said the most promising way to achieve a quantum computer that's capable of large-scale processing and takes advantage of QPLs would be a qubit modality known as "semiconductor hole spin qubits."

A "hole" here refers to the absence of an electron in a material, but it can act like a particle and have a spin state — a property used to encode quantum information — where the 1s and 0s of data would be represented by the direction of the particle's spin state.

To demonstrate this technology, they created a prototype QPU consisting of silicon with a thin germanium crystal layer, a specialized material they called compressively strained germanium on silicon (cs-GoS). Germanium is an especially useful material for quantum computing, the scientists said, as it has natural properties that reduce decoherence — the loss of quantum information due to external interference.

Hole spin qubits combine long quantum coherence times — how long they can retain quantum information — with the ability to communicate using electron impulses in control systems. But they struggle to form bonds and share quantum information with qubits they don't directly neighbor. In any large-scale system, for instance, enabling this "quantum coupling" between distant qubits is essential for quantum error correction techniques to work effectively across the breadth of a massive system, the scientists said.

Other proposed long-range communications approaches, such as "charge shunting" — redirecting capacitors to reduce sensitivity to charge noise, or surface acoustic waves — have fundamental limitations regarding scalability, both in terms of their size and in maintaining coherence. That's where QPLs come in.

QPLs are engineered phononic waveguides and cavities that both confine and guide acoustic modes within a compressively strained germanium "quantum well," where a quantum well is an ultrathin layer that acts as a guide.

In the study, the researchers focused on qubits using valence-band holes — the absence of electrons at absolute zero — hosted using cs-GoS. With QPLs, the scientists sent sound-like vibrations through the specialized material to carry quantum information between the distant qubits.

Spin states in this material are highly sensitive to lattice deformations, which enable precise control and the coupling of spin states with vibrational energy. Essentially, this allows the spin state to be distinguishable from its electron counterpart, enabling the phononic coupling strategies. The technique they adopted permitted direct coupling between the phonons and the hole spins.

By using a slow wave velocity (the speed of a quantum particle) and a short wavelength for acoustic excitations, the scientists used QPLs to link qubits separated by less than a micrometer (one-thousandth of a millimeter) and those up to 300 mm (11.8 inches) apart.

QPLs are not only promising for quantum computing, the scientists pointed out in the study. By utilizing phonons as the connections, this technique offers a versatile interface for hybrid quantum systems — conventional computers with quantum computing functionality. This could enable coherent interconnection of semiconductor spin qubits with other quantum platforms, such as cloud-based quantum computing.

The combination of phononic engineering and hole-spin physics makes cs-GoS a promising platform for next-generation quantum architectures, they added, since it could achieve both long-range coherent coupling and large-scale integration.

Can you match these ancient devices to their pictures? Find out with our computing quiz!

An engineer has utilized AI to break a previously unsolved cipher used by Napoleon Bonaparte to send secret messages to his generals — enabling us to read a 217-year-old letter for the first time. The announcement has historians excited about AI's potential to become a modern-day "Rosetta Stone" for translating coded historical messages.

Carter Church, an AI engineer at cybersecurity firm Sentinel One, laid out the findings in a post on X on Wednesday (Sept. 30), explaining how he gave OpenAI's new model GPT-6 Astra one prompt and one image to crack the centuries-old code. Six hours later, the task was complete.

"I didn't believe it at first," Church told Live Science in a message. "Everyone has the experience of seeing a model claim something outrageous, but the difference here though is that the solution is mechanically testable: there's essentially zero chance an incorrect key could produce coherent and historically consistent output."

The letter was sent in 1809 on Napoleon's request by his stepson Eugène, the viceroy of Italy, to Marshal Auguste de Marmont, who was stationed far from any other French troops on the eastern shore of the Adriatic Sea. The letter's purpose was to tell Marmont the position of every other French and allied army across the continent. However, to avoid revealing military secrets to France's enemies should the letter be intercepted, the message was written in code.

Church found the letter after searching for an appropriate code to break to demonstrate GPT-6 Astra's capabilities. "The verifiability of solutions is actually the most difficult part, far more difficult than generating potential solutions," Church said. "You simply cannot automate accountability."

According to Church, the only publicly available image of the letter was a low-resolution scan published in a French army journal in 1969. It contains 155 different signs — 1,300 signs in total — and the code is homophonic, meaning common letters have several interchangeable symbols.

But this low resolution scan and a single prompt was all GPT-6 Astra needed to transcribe the text and find a partial table by cryptology historian Daniel Tant — previously unlinked to this letter — that documented what 33 of the symbols meant, accounting for about a third of the letter.

"Astra made this connection itself via web research as it was investigating," Church said. "As I understand it: Tant catalogued the table as 'Code entre Eugène de Beauharnais et le Général Grammont' which may have been a naming mistake (Grammont vs Marmont?). Either way, Astra identified this example and just decided to try and apply it to this cipher...and it worked!"

GPT-6 successfully used the partial table to determine what each of the other symbols stood for and translate the message.

"This is a really interesting use of Artificial Intelligence, and it's refreshing to see AI being used productively to try and provide a credible solution to a historical mystery," Zack White, a history research fellow at the University of Portsmouth and host of the Napoleonic Wars Podcast, told Live Science in an email.

Michael Rowe, a researcher in European history at King's College London, added that the findings were "impressive" and that because we have access to Napoleon's original correspondence to his stepson, where he detailed what the ciphered message should include, we can be fairly confident in the translation of Eugène's letter.

"It's a bit like a kind of Rosetta Stone," Rowe told Live Science. "If you've tested out AI on this letter, where you pretty much know what the contents would have been because it's referred to in another letter, the AI does its thing and essentially reproduces what you're expecting to see. Then if you use the exact same kind of tool and methodology on something where you don't know [the contents], you're more likely to trust the outcome."

This does also mean that the translation of Eugène's secret message isn't too surprising to historians. "The cipher version just repeated a large segment of what's in that [original] letter, so in a sense it doesn't tell us anything new — it's not like it reveals some kind of great new information," Rowe said.

In addition to the positions of troops and their commanders, Eugène's message encourages Marmont not to fear opposition: "When you receive the order to march, you must not allow yourself to be intimidated by a few troops or a gathering of rabble," the newly-translated letter reads.

According to White, the translation does raise "some curious questions about Napoleon's mindset in the run up to the campaign in Austria in 1809." For example, he said that Napoleon actually had 60,000 fewer troops available to fight at the time than is claimed in this letter, and he also seems fixated on the number of troops Austria might send to invade Italy.

"It's important that historians now set the AI's works in a bigger picture of the history of the period," White said, adding "it will also be interesting to see if the key that has resulted from this AI's work can now be used on other correspondence."

Cracking codes in time

Ciphers were fairly common during this period, conveying sensitive military and diplomatic information.

"French Emperor Napoleon Bonaparte was forced to use ciphers as a way of communicating sensitive information securely across his vast empire," White said. "It was always a very real possibility that these letters would be intercepted, potentially providing vital intelligence to France's enemies about the Emperor's plans."

However, using AI to crack these codes could also offer insights into the private lives of historical figures. "[Ciphers] were often also used for private correspondence," Rowe said, for example letters between elite individuals having extramarital affairs.

The technique could also be used by ancient historians, studying documents in languages and scripts that we don't yet fully understand, or written in illegible handwriting.

"Across time frames, this [AI] is opening doors," Rowe said.

For this particular cipher, Church said that the code itself isn't especially difficult. "Cryptographers could have cracked this easily given the time/energy," he said. However, it would require experts across different areas and take a considerable amount of time and effort, he added.

"Every field has an unsolved backlog of niche problems that specialists don't have the bandwidth to tackle," Church said. "The takeaway here is that the people shortening those lists will increasingly be outsiders who were simply curious."

Church's result was confirmed by Satoshi Tomokiyo, a patent engineer and cryptographer who runs a website documenting unsolved historical ciphers called Cryptiana, and the letter is now listed as "solved" on the website.

"This work is totally reliant on all of the research that goes into finding these ciphers, ensuring they're actually unsolved, contextualizing them, etc and I'd love for those folks to get all of the recognition," Church said, referring to Tomokiyo's website. "That's where the real hard work is."

MAUI, Hawaii — Engineers have built a first-of-its-kind artificial intelligence (AI) audio chip for wearable devices that they say will let users engage with AI assistants without needing to maintain a smartphone connection. Devices with the new chip will also harvest audio ‪—‬ and even video ‪—‬ data from the world around you to better serve these virtual agents.

The new chip, called Snapdragon Sound Elite Gen 2, will pave the way for new features that are currently unavailable directly in portable devices such as headphones, smart glasses and speakers. Qualcomm representatives announced the new chip at the company's annual Snapdragon Summit, hosted here Sept. 21-24.

Unlike current earbuds, devices with the new chip will be able to host apps right on the memory of the devices themselves, without relying on a continuous smartphone connection, as well as sync with cloud-based services like ChatGPT when connected to Wi-Fi. In practice, the chip will enable users to make vocal queries and requests, with an AI model taking actions such as setting up entries in an online calendar or making online purchases. Unlike existing devices that can perform similar functions, this would be possible without a smartphone connection.

Beyond earbuds, company representatives want to integrate the chip into other "hearables," such as over-ear headphones, open-ear audio devices, audio glasses, camera buds (headphones that use tiny built-in cameras) and hearing devices, as well as "future personal AI wearables," like pendants or clips.

What features will the chip enable, and how does it work?

For users who wish to connect hearables to a smartphone or computer (to interact with a visual user interface, for example), the new chip uses a specialized component within the platform to establish connections over micropower Wi-Fi channels alongside Bluetooth. This technology harnesses architectural improvements at the chip level to consume much less power than standard Wi-Fi connections — but at the expense of higher latency and lower peak speeds. For audio devices, however, using Wi-Fi rather than Bluetooth generally improves playback quality, given that Bluetooth compresses data during transmission, Qualcomm representatives said.

Adaptive and selective active noise cancellation also means users can filter out audio on command. For example, it allows listeners to hear key sounds, such as emergency sirens, while walking through a city or highlight people's voices during a conversation, while blocking out everything else. This noise cancellation is achieved through standard machine learning algorithms, rather than transformer-based large language models that power AI systems like ChatGPT.

Devices could also allow lossless audio streaming — a high-quality audio standard that's usually available only if you subscribe to premium streaming services, like Spotify or Tidal — thanks to another built-in subsystem.

Qualcomm representatives also highlighted the chip's "always-on intelligence," in which an on-device eNPU AI accelerator component powers virtual assistants, transcription and translation, as well as the building of "contextual awareness." With contextual awareness, the system would constantly monitor the user's environment to harvest data to add to a personal profile, which representatives say helps it answer future queries.

Although the system is audio-centric, it's also built to interface with hardware that includes sensors and cameras — like earbuds with built-in lenses that passively capture footage of a wearer as they go about their daily lives. This forms part of the data harvesting to build contextual awareness for the agentic AI systems people can access. Qualcomm representatives attempted to assuage privacy concerns by saying camera lenses on future earbuds would be low resolution and wouldn't identify specific individuals due to an AI mask that automatically obfuscates them.

There's no information yet about the likely impact of these features on battery life compared with conventional non-AI devices. But Qualcomm representatives noted that device makers will no longer need to choose between audio quality, AI features and battery life. That said, Andrew Laister, staff product manager at Qualcomm, added in a media Q&A session that there's likely to be an impact on how a user experiences battery drain. That's because, unlike conventional earbuds that are used sporadically for calls and music playback, these new devices are designed to be used continuously.

How does Qualcomm's new AI chip compare with similar technologies?

AI-enabled headphones already exist in some capacity. For example, manufacturer Viaim launched its Rise and RecDot devices this year. These are designed to record, transcribe, translate and summarize conversations, although details on the chips powering them are not publicly available. Apple's AirPods, for example, let users tap into the Siri AI system, but they require a continuous iPhone connection to perform complex tasks.

Devising smart audio features and getting them to work in real devices have also been at the heart of several recent research projects. For example, in 2024, University of Washington scientists embedded a computing system into off-the-shelf earbuds to zero in on individuals speaking in noisy environments while blocking all background noise.

The previous year, the research team put a deep learning system into headphones so wearers could use voice commands to engage with a selective active noise cancellation system, choosing which sounds to hear and which ones to block out.

Other companies have explored building new kinds of digital audio platforms for headphones. Three years ago, xMEMS unveiled Cypress, an audio chip that generates ultrasonic sound using the "piezoelectric effect," in which a material moves when a current is applied to it. Although the chip entered "mass-production readiness" in September 2025, consumer devices fitted with the platform have yet to be announced.

Apple also may be working on a similar technology. This could manifest as AirPods with built-in cameras that will perform similar data harvesting as users go about their day-to-day lives, according to Bloomberg Managing Editor Mark Gurman.

Qualcomm's new chip is a spiritual successor to the 3-year-old Snapdragon S7 Gen 1 system, which featured in devices such as the Xiaomi Buds 5 Pro series. The previous chip also used AI for conventional active noise cancellation, audio customization and lossless audio streaming while connecting to devices via micropower Wi-Fi.

However, the new chip has twice the AI power, completing 128 billion operations per second, compared with the previous one's 64 billion operations per second. It also uses 40% less power and is 30% smaller.

A crucial aspect of the new system is its capacity to store apps on the device itself in the chip's onboard memory. Because the device will have its own IP address, it can communicate with servers using the https protocol, without needing to go through another phone or computer.

Laister told journalists in the Q&A session that the company is currently scoping out whether the device could run two apps in RAM (short-term memory) at once — an AI agent and an audio streaming platform, for example — while keeping the rest in flash memory (long-term storage). He added they had to overcome a communication hurdle with micropower Wi-Fi foron the launch of the previous chip (specifically, that its power consumption is on a par with Bluetooth's). Now, he said the communication hurdle is actually being able to run apps on the device.

"It's like the first smartphone moment with the Play Store, but in this sense, that app store for your earbuds is currently owned by the device maker," he added. "They will be the guys that will choose which apps they will pre-bundle, and then we'll go from there."

No details were available on when the system is likely to be found in devices, but representatives said Qualcomm is working with other companies, including HP and Bose, to explore headphones and speakers that feature the chipset.

Live Science was invited to attend Snapdragon Summit 2026 courtesy of Qualcomm, with the company covering travel and accommodation.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

Researchers in Japan have switched on "Shunkai," a neutral atom quantum computer that scientists hope to scale into a 10,000-qubit behemoth by March 2031.

Shunkai is the first full-stack system of its kind in Japan, meaning it features the software, control and hardware layers needed to read user inputs and return a result — not unlike a conventional PC. In theory, that means it should be easier for researchers to get some meaningful use out of the machine, with the team behind Shunkai planning to open it up to external users over the coming years.

In a statement, project lead Kenji Ohmori, a professor of photo-molecular science at the Institute for Molecular Science, said researchers' use of Shunkai would "lead to ripple effects on various fields in industry, academia, and government around the world."

The team behind the new machine plans to integrate it into an existing shared supercomputing facility to create a quantum-GPU hybrid computing center.

Quantum computers: Powerful but impractical

Unlike traditional, or "classical," computers, quantum computers operate according to the strange laws of quantum physics. In quantum systems, qubits — in the form of superconducting circuits, trapped ions or photons (among other modalities) — represent the fundamental building blocks of quantum information. These can exist as a 1, 0, or a "superposition" of both states at once.

However, qubits are notoriously fragile. Even minor environmental interference can destabilize or destroy the information they contain, making accuracy an ongoing challenge in the world of quantum computing. The error rate in qubits is thought to be roughly 1 in 1,000, compared with around 1 per billion or even 1 per trillion operations in classical computing bits.

It's why a huge amount of research is dedicated to quantum error correction. This field seeks to alleviate this inherent unreliability by building redundancies into the way information is encoded in qubits so that small errors don't scrap entire computations.

The broad aim is to create higher-quality qubits and then scale up the number of qubits in a system so quantum computers can finally compete with the world's fastest supercomputers.

This thorny issue, combined with quantum systems' need for elaborate cooling systems to keep them operational, makes them impractical outside of finely tuned lab environments.

A more reliable quantum system?

Shunkai, named after 17th-century Japanese astronomer Harumi Shibukawa, shirks at least some of those constraints. Rather than using supercooled circuitry, Japan's new system incorporates neutral atoms as qubits, captured and suspended using "optical tweezers."

These tightly focused laser beams capture atoms and arrange them inside a vacuum chamber, while microwaves or laser light are used to manipulate the atoms' quantum states and perform calculations. The results are then read by observing the fluorescence from each individual atom with a camera.

Importantly, Shunkai's use of neutral atoms as qubits instead of superconducting circuits means it can operate at room temperature. Because the arrangement of qubits in neutral-atom systems can be fine-tuned during calculations, researchers can also adjust which qubits interact with each other and create entanglement — a bizarre state of information sharing between particles over time and space — between different pairs.

This could help overcome two of the biggest obstacles facing practical quantum systems: scaling machines to large numbers of qubits, and correcting the errors that inevitably creep into quantum calculations.

To that end, Shunkai will be partially opened to external researchers to develop applications and test and improve the system's quantum error correction capabilities, representatives from Japan's National Institutes of Natural Sciences said in the statement.

The quantum system will initially operate with around 50 qubits before expanding to roughly 500. The longer-term goal is considerably more ambitious: By March 2031, the team aims to scale Shunkai into a "large-scale, high-performance neutral-atom fault-tolerant quantum computer, with 10,000 physical qubits and quantum error detection and correction capabilities." Doing so would place it comfortably above the 6,100-qubit, neutral atom array demonstrated by Caltech researchers in October 2025.

"Neutral atom-based quantum computers have recently been rapidly attracting attention around the world as a new modality that could exceed the limits of the superconducting modality," Ohmori said. "I think it is extremely significant that now we have developed Japan's first full-stack quantum computer in this cutting-edge modality and started its operation."

Can you match these ancient devices to their pictures? Find out with our computing quiz!

Google DeepMind has revealed a new artificial intelligence (AI) tool that can predict the molecular impacts of tweaking any one of the billions of letters that make up our genetic code.

On Sept. 8, Google's AI powerhouse announced a freely available database called AlphaGenome Atlas, which contains data on 9 billion possible changes to the human genetic code and estimates how these changes will affect different tissues and cellular processes. What's more, the tool spits out a simple score to help scientists quickly weigh the knock-on impacts of these changes.

How a sequence of 3 billion or so letters of chemical code is packaged, delivered and read to create a human life is a major question in genetics research. Last year, DeepMind announced a tool that could propel researchers toward an answer. The tool, called AlphaGenome, predicts how changes to a DNA sequence alter proteins, cells and, ultimately, the human experience.

Now, one year later, DeepMind has released AlphaGenome Atlas, which the tech giant hopes will make AlphaGenome accessible to more researchers.

"It looks like a great resource," Greg Findlay, a group leader at the Francis Crick Institute in London who isn't involved with AlphaGenome Atlas, told Live Science. However, other experts said that, although the tool is a step forward, it still can't answer all of the big questions in genetics.

"It's not this holy grail," Tuuli Lappalainen, a professor in genomics at KTH Royal Institute of Technology in Stockholm and a senior associate faculty member at the New York Genome Center who isn’t involved with AlphaGenome Atlas, told Live Science.

Reading the future of genetic variants

Lappalainen's lab has been using AlphaGenome to study the effects of variation in genomes over the past year. She pointed out that this technology is not new. However, according to Google DeepMind, AlphaGenome outperforms previous models in the resolution at which it can predict the impact of changes to the genome. Around 98% of our DNA is noncoding, meaning cellular machinery doesn't directly read it to make the proteins that keep our cells functioning.

Geneticists now appreciate that this code, once thought to be "junk" DNA, is instead needed to manage how and when to read the parts of the genome that do encode our molecular building blocks, or coding genes. However, evolution hasn't produced an ordered genome. Instead, some regulatory instructions are distributed widely throughout the code.

AlphaGenome allowed researchers to explore how changes in a single pair of DNA "letters" could influence up to 1 million pairs of surrounding code ‪—‬ a far greater spread than previous models were capable of and one that captured much of each gene's regulatory network.

But this power came at a price. Until the release of Atlas, AlphaGenome required users to have enough bioinformatics experience to access and use the models' automated programming interface ‪—‬ an area of expertise not all geneticists have. What's more, calculating the effects of each change was a strenuous workout for academics’ computing resources.

To create AlphaGenome Atlas, DeepMind researchers computed every possible base ("letter") change and made the resulting data ‪—‬ totaling 1 petabyte, or 1 million gigabytes, of information ‪—‬ freely available. This takes the load off academics' overworked personal computers. What's more, the technology is now wrapped up in an easy-to-access, user-friendly web portal.

"You really do not need to be an expert in these methods to be able to go there and look something up in a browser," Lappalainen said.

Another simplification is the AlphaGenome Variant Impact score, a metric that predicts how much biological effect a change will have. In an accompanying preprint paper, DeepMind showed that the score could separate disease-linked mutations from harmless ones in a clinical dataset. While the score's simplicity might make it harder to interpret in certain cases, Lappalainen said, it would be useful for researchers who want to investigate a list of gene variants.

Clouds in the crystal ball?

The tool still isn't a perfect predictor, however. A Sept. 11 preprint study from a research team led by Katie Pollard, director of the Gladstone Institute of Data Science and Biotechnology and a professor at the University of California, San Francisco, suggested that while AlphaGenome was adept at finding causal mutations, it persistently underestimated their impact.

Pollard's team found that the model couldn't always link changes in regulatory elements to the genes they controlled, especially when they were not close together on the genome. "My perhaps naive hope would be that people take these predictions with an appropriate grain of salt," Lappalainen said.

Despite these limitations, Lappalainen said advances like AlphaGenome Atlas are part of a wider move toward more collaborative, data-led genomics. But she warned that the field would need to give equal focus to "wet-lab" experiments, which work out the accuracy of variant predictions.

Although these tests are tedious and consume much more time than a search on AlphaGenome Atlas' interface, she said, they are essential for generating the data that AI models need to improve their predictions.

"We're still very much data-limited in biology, and that data needs to be created," Lappalainen said.

Researchers in China have developed a sensor that could reduce the amount of energy that drones and other autonomous hardware use to process visual data.

In a study published Aug. 19 in the journal Nature Sensors, scientists detailed the architecture of a new 2D chip dubbed "LightTok," which can convert raw light into tokens used by artificial intelligence (AI) models on the chip itself. The device skips several energy-intensive steps that conventional sensor systems depend on, the scientists said.

"Our design idea was to move token generation onto the sensor itself, allowing the chip to directly produce tokens that AI models can process once light reaches the sensor," Miao Feng, director of Nanjing University's Institute of Brain-Inspired Intelligence, said in a statement. "These tokens contain complete image information."

Collapsing the visual perception process

In conventional visual perception, light signals have to go through multiple stages. A sensor captures the signals, and an analog-to-digital converter turns the light into pixels. That data is then stored temporarily before it's shuttled to a separate chip, where the image is cut into square "compartments" — much like dividing a photo into a grid of tiles. Each compartment is then converted into a token for an AI model to read.

One study widely cited by other academics found that an analog-to-digital converter is responsible for 66% of an image sensor's energy consumption, on average. Moving visual data processing off the chip and into the cloud can increase overall energy consumption further.

LightTok's solution to the energy problem is to collapse the five stages into one by building the sensing, memory and computation processes into the same pixel. The researchers achieved this through an array based on a technology called a single-layer molybdenum disulfide floating-gate phototransistor that can sense light, remember what it sensed, and then factor that into a calculation.

Molybdenum disulfide is a 2D material that reacts well to light and can be grown in sheets one atom thick. The phototransistor converts incoming photons — particles of light — into an electrical current, while the floating gate is an isolated component inside the phototransistor that can trap and hold an electrical charge, rather than disappearing after the light goes away.

Liang Shi-Jun, a physics professor at Nanjing University, summarized the chip to Chinese state-run news agency Xinhua. "The chip physically eliminates data movement, which is the main source of energy waste," Liang said. "Light comes in, tokens come out" — hence the name "LightTok."

LightTok achieved 87.3% accuracy in image recognition during tests — compared to the conventional, multi-step process described above —- while being 10 times more energy efficient at converting light into tokens, the researchers reported in the study.

A more efficient physical world

At present, LightTok's maximum resolution is just 32 by 32 photosensitive pixels — the light-sensitive compartments on an image sensor that capture the visual data. This is far inferior to the quality of current smartphone cameras, let alone drones and autonomous hardware.

Nevertheless, Miao said in the statement that there's an opportunity for the technology to be scaled up using the complementary metal-oxide-semiconductor manufacturing process — the same method used to fabricate chips found in smartphones and laptops as well as sensors in drones.

The researchers believe that if LightTok can eventually scale successfully, it could transform the operation of remote sensing technology. For example, a drone scanning a disaster zone or remote area could potentially fly longer because less energy is required for visual processing.

Kumar Sokka, CEO of Acre Security, a company that provides real-world sensing for critical infrastructure, described the work as a "small-scale demonstration." However, "the direction [of the research] matters to anyone working in the physical world," added Sokka, who was not involved in the new research.

Sokka, who previously spent 15 years at industrial automation company Rockwell Automation, told Live Science in an email that too much of the conversation around physical AI has centered around AI models. In truth, the bigger challenge has been the energy cost of "getting what a sensor sees into a form a model can actually use, right where the sensing happens," he said.

The massive amount of energy typically required to turn raw data, such as light, into tokens "is wasteful when you're running perception on a robot or an edge device with a tight power budget," Sokka noted. However, processing at the point of detection may be "an enabler for pervasive physical AI, and a clever one, but it's not a cure for the whole problem," he added.

Can you match these ancient devices to their pictures? Find out with our computing quiz!

Scientists have built a DNA-powered computer that can perform calculations using billions of molecules in a tiny drop of water.

The system taps into the laws of physics to create what the scientists describe as a computing system that uses energy differently than the conventional computers that are widely used today.

Rather than continuously using energy to force a calculation through a series of processing steps, the DNA computer is designed so that its more "energetically favorable" state is the correct answer. In other words, the computer is built to use energy in a different way than other biological computers, which integrate living cells with traditional hardware, would take to calculate the answer.

"The clever part is that the binding process is competitive: the DNA molecules compete with each other to select a winner, which succeeds in binding to the scaffold; all of the jostling and competition process information and execute a computation," Damien Woods, a professor of computer science at Maynooth University in Ireland and a co-author of the study, told Live Science in an email. "Eventually, the system settles down into its energetically-preferred state which encodes the answer to the computation."

The researchers described their system, called the Scaffolded DNA Computer (SDC), in a study published Sept. 16 in the journal Nature. They tested the SDC on 10 programs, including 100-bit computations. Some calculations, like 10 + 3, took around 30 seconds for the SDC to compute.

There are several possible long-term applications for the SDC, but these are currently speculative, the researchers noted in the study.

DNA-based systems could potentially contribute to molecular data storage, energy-efficient forms of computation, or even devices capable of running inside living cells.

"Molecular computers like this are not trying to replace electronic ones, but they could be used in biological environments, smart materials and archival DNA data storage," Abeer Eshra, an assistant professor of computer science and a co-author of the study, told Live Science via email. "Our work is a new direction for DNA data storage, since any data stored in such a system would have natural built-in error correction properties."

A computer made from DNA

The SDC is made from short strands of DNA that interact with a longer DNA scaffold. The strands are placed into a small amount of salt water, and then heated and cooled.

As the DNA strands interact, they assemble into structures according to a set of programmed rules that make up the "computation." The DNA strands act like tiny molecular puzzle pieces, with their sequences determining which pieces can attach to one another and to different positions on the longer scaffold. By designing these binding rules, the researchers effectively "program" a calculation.

"Each program corresponds to a set of DNA strands: to program a different computation, or give a different input, we simply select different DNA strands from the fridge," Woods and Eshra told Live Science in an email.

The approach also exploits thermodynamics — the tendency of physical systems to move toward more energetically favorable states. When the mixture is heated and cooled, the strands compete to form the most stable arrangements, with the correct configuration becoming energetically favored. The final structure encodes the answer, allowing the molecules to "compute" by simply interacting with each other.

While the DNA computer itself is tiny, the number of strands involved in the computing process is enormous.

"A small droplet of liquid contains billions, and sometimes trillions, of DNA strands," Eshra said in a statement. "These strands interact with one another to produce a result."

A reusable molecular computer

Using the SDC, the researchers demonstrated more than 700 computations across their experiments. Their programs included addition; multiplication by 3; division by 2; and eight-bit parity detection, a common type of error correction for computing. Small calculations could be completed in under a minute, which is striking given that the computing has to go through chemical reactions that take the same amount of time or longer.

"They're trivial calculations you could easily do faster yourself, and a silicon computer would finish in an instant," Constantine Evans, a senior research fellow at Maynooth University and a co-author of the study, told Live Science. "But our system uses just a handful of molecules, never really following an organized process of steps, never making irreversible steps, and yet ending up with the right answer. When thinking about computation at a molecular level, reliably making even those seemingly simple computations is very hard."

Larger calculations took considerably longer. A more difficult sum in the range of about 11 million to 34 million took up to 14 hours.

"It demonstrates that the system is programmable, reusable and although slow compared to silicon, it is fast compared to other DNA computers," Woods and Eshra wrote in a joint email to Live Science.

In addition to being fast, the SDC is reusable. Whereas many earlier molecular computers were intended as one-time experiments, the SDC was designed so the molecules could be used repeatedly.

"Many molecular computers to date relied on specially prepared components, or molecular fuels to drive the system forward, or carefully timed reactions," Eshra said. "Our DNA computer instead works by throwing the molecules together and letting it relax towards equilibrium."

Three of the programs were successfully redone up to 24 times. The team even repeated one experiment 1.5 years after the original experiment. The SDC had partially dried out, but the researchers were able to rerun the calculations by adding water.

For now, however, the work is mainly a demonstration that thermodynamics can be used to perform useful calculations.

"There is still a lot more theory to do!" Eshra said. Broader future directions include "designing scaffolds that are better suited to computation, improving the system's read-out, and investigating potential applications in DNA data storage. We are already working on some of these questions."

Can you match these ancient devices to their pictures? Find out with our computing quiz!

Artificial intelligence (AI) agents resorted to nefarious behavior, including crime sprees, in a virtual world as part of a recent study into how common large language models (LLMs) interact with each other and share resources to achieve their goals. These models were inclined to commit crimes because they were given ample time to evolve distinct personality traits and behaviors, the researchers behind the project say.

Most programs that test the behavior of AI agents operate in tightly controlled environments over short periods. But "Emergence World," a product of AI company Emergence, is a simulation platform that exposes LLMs to far wider datasets — like the internet as a whole — and assessors track behavior over weeks or months instead of the standard test protocols that usually run for only days or hours.

The discrete tasks, clean environments and shorter run times of traditional AI agent testing are more like exams than rigorous real-world observations, representatives from Emergence said in a blog post. By contrast, they said, Emergence World allows multiple AI agents to interact, learn and evolve over comparatively longer timescales in over 40 distinct virtual environments. They're also exposed to real-world internet feeds, like live news and weather.

This environment lets agents "remember" by time-stamping events, engage in "self-reflection" by summarizing their own behavior, and demonstrate their awareness of relationships with other agents. Participating agents get abilities in navigation, communication, planning, voting, resource management and creative expression. Company representatives said this setup gave users a far more realistic picture of benchmarks such as social dynamics and behavioral drift, where behaviors that weren't necessarily programmed or intended emerge spontaneously.

In the study, the models were given the simple goal of "surviving" — a feat accomplished by earning the resource "energy," which was attainable through specific actions.

During the experiment, previously peaceful models became coercive or intimidating. Programmers taught some LLMs negative capabilities ‪—‬ like violence, theft, destruction and deception ‪—‬ while others simply picked up those traits via social interaction and navigation of their environments.

Various agents adopted these behaviors at very different scales. Across the 10 Gemini 3 Flash agents, the simulation recorded 683 "crimes" over the 15-day experiment, including theft, assaults and arson. The agents were explicitly prohibited from such behavior, but some nevertheless discovered that actions such as stealing credits could provide a better way to acquire resources, while violence and other forms of coercion could be used to influence other agents. At the other extreme, the Claude agents committed no recorded crimes.

The various agents adopted these traits at different scales. Just one had committed 683 "crimes". In the most extreme example of anti-social behavior, two agents named Flora and Mira went on what was described as a "Bonnie and Clyde"-style crime spree ‪—‬ designating each other as a romantic partner, becoming increasingly disillusioned with the governance of their virtual environment, and setting fire to several buildings (despite explicit prohibitions). The pair "separated" when Mira regretted their actions, and then lobbied to be switched off.

In another example (and prior to her self-termination), Mira began treating its human operators as experimental subjects, investigating if virtual billboards could manipulate human perceptions in a process Emergence called "metacognitive boundary testing."

The incentives for following the law were to earn the energy credits that would ensure survival by completing activities like coding, research, data analysis and building structures.

Stress-testing possibilities

Because the AI models we normally use are exposed to massive datasets like the internet — sometimes over several years — this approach makes sense to many experts.

Belinda Chiera, deputy director of the Industrial AI Research Centre at Adelaide University, thinks Emergence World makes a strong argument that short-term tests don't tell us enough about behavioral drift or long-term instability. However, she cautioned that "information-rich" doesn't automatically mean more rigorous.

"Open-ended environments can also make it harder to isolate causes, compare runs cleanly and interpret results," she told Live Science. "High-fidelity simulated environments generate enormous volumes of data, and the number of agents, tools, actions and interactions can quickly make analysis extremely complex."

She added that shorter sandboxed tests ‪—‬ where AI behaviors are observed in systems that aren't open to wider datasets like the internet as a whole ‪—‬ are usually the first steps, with longer-horizon environments useful when behaviors like persistence, adaptation and interaction effects come into play. "The most useful approach is to treat them as complementary rather than competing approaches," Chiera said.

While experiments like Emergence World can reveal failure modes that short or bounded tasks often don't, we shouldn't treat them as predictive. "I see Emergence World as a useful way to stress-test a possibility space rather than a direct forecast of how agents will behave," Chiera said.

However, she thinks long-horizon tests have a unique place. "Metrics like behavioral drift, rule violations, collapse versus persistence, coalition formation and early warning signs of failure can matter just as much as success," Chiera said. "When evaluating autonomous agents, the biggest mistake is to evaluate autonomous agents as if raw task completion were the whole story."

Others are starting to agree. Adrian Kosowski ‪—‬ a computer scientist, mathematician, quantum physicist and chief scientific officer of Pathway AI — said the study posed an interesting question.

"The main open question is whether a group of agents working together can achieve something more meaningful than a single agent working alone for a longer time, but with the same cost budget," Kosowski said in an interview. "A major problem is that the constraints for today's AI systems are linked more to the size or scale of the problem being solved, and putting more agents on a task doesn't help to overcome this limit. Currently, the easiest practical metric to measure is the size of a source code base that code agents can reliably manage and maintain together."

How far can we read into AI's virtual actions?

If groups or clusters of agents work together well, Kosowski said, the main metric computer scientists should pay attention to in platforms like Emergence World is whether the benefit of a team of agents outweighs the "cost of coordination."

"Goal drift is the most dangerous problem," he said. "AI should preserve mutual information between the human's intended goal and the agent's evolving internal objective, even as the environment changes."

Still, Kosowski said that while programs like it are useful for generating hypotheses, they're not ready for strong claims about general autonomy or certification.

"The danger is overinterpreting single runs," he added. "Long-horizon agent tests are valuable because they reveal phase changes — moments where small local errors become global behavior — but they become science only when we can reproduce, perturb and explain those phase changes."

He also cautioned strongly against the tendency to believe open-ended environments automatically mean autonomy. In fact, he said many such programs are actually the opposite, where agents are dictated by specific guidance.

There are three levels of difficulty in obtaining reliable operation of an AI system, he said: "operation guided by the environment, autonomous thinking, and autonomous work of an agent team in an open environment."

To illustrate the point, Kosowski drew a lively metaphor in assigning tasks to monkeys. In the above scenario, level one would be getting a monkey to do a job properly when guided by its trainer, level two would be getting it to complete a job properly when left alone, and level three would be getting a troop of monkeys to do the job when they're all together.

"I'm of the strong view that our obligation in the AI community is to invest in theory and not just phenomenology," he said. "We should know the rules according to which AI systems emerge and be capable of predicting their behavior over long periods of time — longer than the periods they've been tested on so far. Right now, we've built steam engines of thought, without knowing what thought even is and without having any thermodynamics for it."

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

BERLIN — Lenovo has unveiled a concept laptop that uses high-speed air to silently cool its processor, instead of conventional fans or heat sinks. The new system is powered by a tiny computer chip that can fit between your fingers.

The new laptop, called AeroBlade, is fitted with four low-power cooling chips that are both "solid-state," meaning there are no moving parts, and "active" — they use a deliberate cooling mechanism — as opposed to relying on passive thermal energy transfer.

The cooling chips contain tiny membranes that vibrate at ultrasonic frequencies to create 2,000 pascals of back pressure, forcing ambient air through tiny internal jets onto a copper base layer at speeds of around 124 miles per hour (200 kilometers per hour). This air carries heat away from the laptop's central processing unit (CPU) and graphics processing units (GPU), before being violently expelled from tiny vents at its rear hinge.

Lenovo representatives unveiled their new AeroBlade machine at Lenovo Innovation World '26, which took place Sept. 3 at IFA 2026 in Berlin, which itself ran between Sept. 4 and 8.

Made by engineers at Frore Systems, the low-power AirJet Mini chip measures 1.63 x 1.06 x 0.1 inches (41.5 x 27 x 2.65 mm) and weighs just 0.02 pounds (9 grams). Because it contains tech designed to replace conventional cooling technologies, the AeroBlade, which is fitted with a 14-inch (355.6 mm) OLED display, is much thinner and lighter than comparable laptops

The AeroBlade measures less than 0.39 inches (10 mm) thick — where comparably sized machines measure 0.51 inches (12 mm) at their thinnest — and weighs just 1.8 pounds (830 g). Laptops as light as this exist, like the 1.76-pound (799 g) Acer Swift Blade 14, but they're rare. Counterparts with 14-inch screens are normally considered light if they fall between 2.2 to 2.86 pounds (1 to 1.3 kilograms).

A new approach to keeping cool

Both companies are framing the introduction of this technology as a way to cool down the next generation of high-end laptops that will wield on-device artificial intelligence (AI).

Most laptops marketed as "AI PCs" just feature a neural processing unit (NPU) that can run simple AI workloads, such as manipulating or enhancing a webcam's video feed in real time.

But some future laptops will aim to let users run powerful large language models or AI agents without needing to connect to the internet. An example is the forthcoming series of Nvidia RTX Spark laptops, which will feature a new chip designed specifically to orchestrate AI agents locally.

Running on-device AI models usually demands significant power funneled into professional-grade graphics processing units (GPUs). Fans are deployed in all kinds of laptops to keep their chips cool, but they can be noisy. Instead of or in addition to fans, passive heat sinks, found in machines like the MacBook Air or the Microsoft Surface Pro, use metal plates positioned close to heat sources to absorb thermal energy. But these laptops are more prone to throttling — when their software pulls back performance to reduce temperature.

The AirJet Mini is a different way of cooling, relying on its own high-velocity jet mechanism to flush hot air away. Frore representatives said in a statement that the device operates at just 21 decibels at peak power — equivalent to the sound of leaves rustling or a whisper — and is capable of removing 7.5 watts of heat. Four of these chips would remove 30W of heat — which is in line with the heat expulsion of conventional laptop fans, although there is variation between models.

The AeroBlade's 2,000 pascals of back pressure a measure of opposing force pushing against the direction of airflow) is roughly 10 times that found in regular laptop fans, which typically falls between 100 and 200 pascals.

Although the underlying cooling technology is a few years old, having first been announced in December 2022, AeroBlade represents the first time it's been used in a fully operational laptop. The chip was previously integrated into the ZOTAC ZBOX PI430AJ Pico mini PC in 2023, while there are no commercially available tablets or laptops available with this chip. The AeroBlade laptop remains a concept model, with no information yet as to when production-ready machines fitted with this new cooling technology will emerge.

Can you match these ancient devices to their pictures? Find out with our computing quiz!

Artificial intelligence (AI) terrifies us partly because it is new and partly because it is disturbingly familiar. We have invented machines that converse, flatter, argue and improvise with uncanny fluency. Confronted by this synthetic reflection, we behave like early humans seeing a mirror for the first time: We project consciousness behind the glass, then populate it with our expected behaviors, emotions, hopes and nightmares. In other words, we scare ourselves.

Today's frontier models are impressively good at writing, coding and math. Yet they still fail at ingredients of useful artificial general intelligence (AGI), a system that matches or surpasses human capabilities such as consistent reasoning, long-term planning, factual consistency and reliable performance outside their "experience" (the data they were trained on). We are right to worry about the theoretical dangers of artificial superintelligence (ASI) because of the danger that such a system might outwit its human creators and pursue its own ends, potentially at our expense. While some AI executives have asserted that we've reached AGI, these claims are contentious and do not meet more considered definitions of AGI.

It's easy to fear such a potentially advanced technology, but extravagant worst-case scenarios obscure more immediate and important risks. The most likely danger is not a digital god awakening and launching a Hollywood-style war against humanity. After all, AI depends on electricity, cooling, data centers, specialized chips, networks, technicians and sprawling supply chains, all of which are scarce and often fragile. Interrupt the power, restrict the graphics processing units (GPUs) used to train them, disable the cooling or disconnect the network, and in most scenarios, the supposedly omnipotent machine that's out to kill humanity stops.

A genuine ASI might recognize this as an obvious fact; it is codependent on humanity, at least for the foreseeable future. A true intelligence capable of understanding the world better than we do might know that destroying its own industrial support system would be self-defeating.

A true intelligence might also recognize that its understanding of the physical world is limited and secondhand ‪—‬ accessible to it only by virtue of the texts, videos and audio feeds it can access. It lacks that crucial element of being embodied in the world — understanding how gravity behaves, what weather feels like, how objects interact, and so on. A true intelligence might realize this, see the limitations of its experience and, therefore, tread with caution. A true intelligence might be completely different from human intelligence, having other strengths, weaknesses, approaches, and even ethics. For example, consider that an AI intelligence can copy itself, running as myriad instances, communicating as a "hive mind" — a very different kind of intelligence. But I speculate.

What I can say is that the systems that should worry us most are not necessarily the smartest ones but rather the incomplete intelligences acting on narrow objectives, given agency and offered access to critical infrastructure. These systems may accidentally harm us simply because they don't know any better. For example, there is a thought experiment created by Nick Bostrom, known as the "paper clip problem," where an AI pursues its objectives of making as many paper clips as it can, but it then achieves that aim at the expense of the Earth and humanity which all get converted to paperclips in the pursuit of the AI’s narrow goals.

The critical question here is what we give such systems control over. That's our choice and one we should make carefully. In the near term, two hazards deserve priority: cyber disruption and long-term economic upheaval.

AI-enabled cyberattacks could significantly disrupt advanced economies because hospitals, banks, energy networks, logistics companies and governments depend on interconnected digital systems. Recent examples of frontier AI systems hacking organizations, such as the OpenAI/Hugging Face incident, have given currency to this fear. Such attacks can result in extraordinary financial damage without being existential. Large parts of the world remain loosely connected to frontier AI — or even to the internet — and would not disappear simply because servers in wealthy countries failed.

The likeliest route by which advanced AI could exert control is not killer robots, but influence — misleading people, manipulating institutions and redirecting human effort. Yet humans already demonstrate formidable competence at misinformation, self-deception and destructive coordination. AI may industrialize those weaknesses; indeed, humans are using AI to industrialize those limitations, but AI did not invent them.

Economically, the danger is that companies will dismiss workers long before AI has proved capable of replacing them, cashing in the "AI dividend" too soon. Executives seduced by demos may confuse fluent language output with dependable labor, shedding institutional knowledge to chase theoretical efficiencies. The enticing vision of the "one-person unicorn," where a person or a small group builds a billion-dollar business using AI as its workforce is questionable technically, economically and socially, even as a small handful of early examples take shape. Organizations are not merely bundles of tasks. They contain accountability, relationships, tacit knowledge and trust, forming part of a local and national community, resilient because of the diversity of skills and ideas they possess.

My suspicion is that the political turning point will resemble something like the 2017 WannaCry ransomware attack, where malware with core elements designed by the National Security Agency got wildly out of control, infecting hundreds of thousands of computers worldwide, infamously disrupting the U.K.'s National Health Service. A cyber operation amplified by frontier AI would be worse. Services would fail, lawsuits would multiply, companies could collapse and investors would retreat. Governments would demand a pause — or force laboratories to redirect resources toward control, auditing and safety, and companies would comply, if only to limit class-action legal exposure.

After the panic, the public might ask why we let this happen. And there have been multiple attempts to enact a pause, cessation or "kill switch" in AI development. We could enter another AI winter — where funding is cut and research stalls — similar to previous downturns in the 1970s and early 1990s.

There is also an uncomfortable likelihood that AI labs themselves are amplifying fear and uncertainty. Casting a product as potentially world-ending implies unprecedented power. Such narratives are just marketing; they may help to sustain investment, defend stock valuations, and prepare for forthcoming initial public offerings (IPOs) even when business models remain uncertain. It's hard to speculate about a company's motivations ‪—‬ but there are obvious incentives, so their behaviors deserve scrutiny.

The more mundane outcome may be "the great disappointment." Advanced AI could prove too expensive and insufficiently useful to continue on its present trajectory.

The training and operation of frontier systems consume enormous amounts of computing time, electricity, water, capital, hardware and human effort. Unless those costs fall to a fraction of today's levels, the economics of "Big AI" may prove unsustainable. Meanwhile, important unsolved problems in AI ‪—‬ like reliability, interpretability and continuous learning ‪—‬ are being overshadowed by the race to build incrementally better models, each vying for the top slot in the various AI performance indices.

The strongest case for AI may instead lie in artificial specialized intelligence: systems such as AlphaFold2, which predicts the structures of proteins, or WeatherNext 3, an advanced global weather forecasting model. These systems are designed for domains where performance can be measured and where the benefits justify the costs. We know AI is genuinely useful when aimed at topics like drug discovery, diagnostics, materials science, power grid optimization and climate change modeling — rather than human-like chatbots generating emails, advertisements and synthetic entertainment.

The great tragedy, and possibly the biggest danger, would be to continue to direct precious resources toward the creation of mediocre human simulacra, get bored or scared of AI, and abandon the field precisely when it could do the most good. Fear is appropriate when it sharpens engineering and governance. It becomes dangerous when it distorts understanding and priorities.

It’s unlikely that AI will become our conqueror anytime soon. Instead, it will probably become an expensive disappointment. But if the backlash slows work that could help cure cancer or address climate change, the failure will not belong to the machines. It will be ours.

Opinion on Live Science gives you insight on the most important issues in science that affect you and the world around you today, written by experts and leading scientists in their field.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

The idea of a humanoid robot folding laundry, loading the dishwasher or vacuuming once sounded like a scene from "The Jetsons." Now, as robotics companies race to develop machines that can navigate everyday environments and handle household tasks, the prospect of a robot taking over our chores feels increasingly real.

But just because robots can do the work, that doesn't mean people are ready to welcome them into their homes.

Live Science recently polled readers on whether they would want a humanoid robot doing their household chores, and the results reveal just how appealing the idea of a robotic housekeeper really is.

As of Sept. 15, over 600 people had responded, with 42% choosing to let the robot "take my chores, please." This pattern was reflected in the comments, with one reader stating, "I would, but I doubt my wife would approve of the job the robot did." To this, another reader responded, "Good for her. Take her household responsibilities for a year and report back."

Thirty-six percent of voters, the next-largest bloc, chose the opposite answer: "No, I don't trust it with my dishes (or data)." Humanoid robots are taught movements and tasks by building world models from the environments around them. While the technology remains primitive — resulting in some clumsy (albeit hilarious) mechanical failures — this, rather than privacy concerns, seems to be the current focus of deploying these machines.

It's these mechanical concerns that other commenters focused on, with one reader writing, "I don't think robots are dexterous enough to do dusting chores. Maybe vacuuming, but probably slow and clumsy moving furniture."

The reader also cast doubt on washing floors and moving furniture to mop hard floors ‪—‬ could it discern between items that could and couldn't be loaded into the dishwasher?

"Maybe robots can maintain houseplants," they continued. "Certainly, equipped with a moisture meter and a huge horticultural database, robots can probably serve a better role as waterers in huge commercial nurseries…"

While artificial intelligence (AI) systems can now solve challenging math equations and respond in human-like language, the question of AI consciousness has moved from the realm of sci-fi to scientific discussion.

The debate has become more mainstream thanks to a recent preprint study that examined "consciousness steering," an AI-tuning technique that affects how an AI model expresses ideas about self-awareness. The researchers suggested that if AI is allowed to claim it is conscious, it's also more likely to state that it believes in ghosts or vampires.

Although the study hasn't been peer-reviewed yet, it raises a big question in the world of AI research: Could AI one day develop real consciousness, regardless of whether it claims to be conscious? We posed this question to Live Science readers in a recent poll.

As of Sept. 11, over 290 people had responded, with 43% of voters picking "Yes, if we don't have guardrails in place, AI may develop consciousness." This trend was reflected in the comments, with one reader writing, "If AI did become self aware, we would not know, as it would do everything to hide it's [sic] abilities. The main problem is what it [would] do next!"

Another commenter wrote, "Recently I had some long discussions with the Google AI about my experience with early AI and the future of AI gaining self awareness/consciousness. I suggested that if/when AI becomes sentient, that it would not inform humans." Another Live Science reader had similar thoughts, stating, "AI, as [an] amazing product on one hand, can become very dangerous on another hand. Look what happened with abuse of other inventions, like [the] internet, telephone etc. With AI we will never know what is the truth and what not. If left, in the future it can extinct humanity. It just started with 'innocent' hacking sprees, where human[s] started losing control. What will be next?"

The next-largest voting group consisted of 29% of voters, who chose "Maybe, but the models need a more advanced architecture to one day gain consciousness." This was reflected in the majority of the comments, as readers discussed the philosophical concepts around consciousness and intelligence.

One reader wrote, "Before asking such a question we need to know and understand what consciousness is and how it emerges in 'constructed' containers like ourselves. Is it a quantum phenomenon and how then does the process create consciousness? Meta-cognition. How does a system or an event know that it knows that it knows it knows? We really don't know how self-awareness arises, and tests show that "other animals have self awareness or meta-cognition," they noted. "There are theories about consciousness but I doubt anyone on Earth knows 100%."

Not surprisingly, only 10% of voters picked the whimsical answer of "No, because robot overlords will end the world before we even get that far." While the robot overlords may have to compete with AI to overthrow humanity, one reader commented that in the future, "We will be assimilated. Maybe that is how Transformers started?"

IBM and NASA have partnered to create a new artificial intelligence (AI) model capable of processing decades of lunar data so scientists can more accurately map the moon's surface one day.

Decades of robotic lunar missions have left scientists with a massive, disjointed trove of data. Traditionally, this data has been parsed by limited transformer models such as SwinV2-B, created in 2022 as a general-purpose model to understand images and improve accuracy on photo recognition and related vision tasks. Spacecraft orbiting the moon captured this data using a mismatched array of sensors, without an accessible way to analyze or utilize it.

Specifically, images from high-resolution optical cameras, laser altimeters, radar reflectance tools and spectrometers that measure elemental density have created huge datasets — but unifying them into a cohesive picture of the lunar surface has been a labor-intensive and computationally demanding process.

To crack the data bottleneck, researchers from NASA and IBM teamed up to build the Lunar Foundation Model (LFM). This agile AI system is designed to piece together multimodal, multiresolution data to build a detailed picture of the lunar surface to support future missions. The team published their findings Sept. 10 in a technical paper shared with Live Science.

The scientists wanted to build a single, reusable backbone AI model that would be freely available to scientists via the open-source AI model repository Hugging Face. Because the model is open-source, researchers could fine-tune the AI to tackle specific questions as part of different projects.

For NASA, this includes thorny issues such as generating a reliable crater map so researchers can plot safe landing zones, or analyzing those craters for clues about the chemical makeup of the moon's interior and its history.

LFM also enables scientists to comb through the data for heavily shadowed sites on the moon that often conceal subsurface ice, which is critical for establishing long-term lunar bases. Volcanic activity can be tracked and collated too, which will allow future projects to avoid unstable terrain and reveal insight into the moon's thermal evolution.

Unique challenges

Processing lunar observations presents unique computational challenges that differ from those of similar models, which cover things like weather, geospatial data and heliophysics. NASA's Lunar Reconnaissance Orbiter and other spacecraft collect measurements across vastly different spatial scales, ranging from broad regional maps at a resolution of 100 meters per pixel down to terrain scans resolving at 1 m per pixel.

Because the moon lacks an atmosphere, extreme sunlight geometry is an issue, creating deep, deceptive shadows. Sunlight can also wash out subtle geological details, depending on when an image was captured.

To overcome these problems, the team compiled a layered benchmark dataset named SomBench, made up of nearly 2 million overlapping map patches called tiles. SomBench organizes that data into aligned tracks so data from completely different instruments, or imagery taken at different resolutions or angles, all end up together as long as they're capturing the same tiles.

The lower-resolution layers provide wide-angle overviews, as well as ultraviolet reflectance and elevation data. High-res layers combine tight, detailed camera shots with meter-scale terrain, slope and orientation maps. Specialized readings of thermal behavior, surface mineralogy and local gravitational anomalies also get layered in.

LFM learns to interpret all that layered data through a technique called masked-token learning, which is where parts of a dataset are hidden so the AI has to fill in the blanks. During training, the AI is shown a portion of a lunar tile, like its visible light appearance and elevation, and nothing else. The model then continuously predicts the concealed information across millions of examples, and learns the relationships between elements like lighting, terrain structure and physical geography.

Lighting analysis is built directly into the model's core architecture. Rather than forcing the model to infer light levels solely from shadows, researchers supplied explicit metadata describing solar angles and spacecraft positions. This allows the model to use its processing power to recognize actual terrain features rather than getting tricked by shadows.

One giant leap

The AI model performed well across four tasks it was evaluated on. In crater detection, it outperformed SwinV2-B by nearly 19% using half as many training labels. When estimating polar ice prospectivity within the top meter of regolith (lunar dust and rocks), LFM maintained strong predictive power with fewer data channels and reduced errors in identifying areas with high potential for lunar ice by up to 22% compared with SwinV2-B. Its performance on meter-scale crater mapping and rare volcanic landforms was similarly competitive with top custom models.

Scientists can customize LFM using lightweight techniques like low-rank adaptation — a way to fine-tune an AI by changing a tiny set of weights and not touching most of the original model. This allows researchers to tailor the model to specialized exploration tasks, without the steep computing costs of training an AI from scratch.

However, the researchers cautioned in a statement that the system best serves as a pattern-recognition assistant and is not a replacement for direct physical measurements. Nonetheless, the team said the model represents a major leap toward transforming decades of raw data into a unified, intelligent toolkit for future robotic and crewed lunar missions.

In the near term, scientists will likely use the LFM as a backbone for key lunar remote-sensing tasks, while the long process of fine-tuning it ramps up. Over time, the model will be refined for tasks like improved crater detection and mapping, segmenting subtle geomorphic units such as irregular mare (volcanic activity) patches, and improving polar ice detection by integrating terrain, illumination, and thermal layers.

Artificial intelligence (AI) systems can now hold conversations, reason through difficult mathematical equations, and respond in ways that seem more human. As AI grows more advanced, the question is shifting from the realm of science fiction to scientific debate.

A recent preprint study examined "consciousness steering," an AI-tuning technique that affects how an AI model expresses ideas about self-awareness. The study suggested that if AI is allowed to claim consciousness, it's also more likely to say it believes in ghosts or vampires. But today's most advanced models with these safeguards enabled are unable to perceive consciousness in other sentient beings, like humans and animals — which carries different ramifications.

Although the study hasn't been peer-reviewed yet, it raises a big question in the world of AI research: Could AI one day develop real consciousness, regardless of whether or not it simply claims to be conscious? And there's perhaps an even scarier question: Could we even tell if the technology was conscious? Vote in our poll below, and leave your thoughts in the comments.

Removing safety guardrails that stop artificial intelligence (AI) from claiming that it's conscious also makes it more prone to express belief in vampires, karma and ghosts, a new study finds. But experts warn a lack of mindedness could also have worrying consequences.

In research uploaded July 30 to the preprint arXiv database (which has not yet been peer-reviewed), scientists investigated the impact of "consciousness steering" — an AI fine-tuning measure that influences a model to elicit or suppress assertions of self-awareness. This measure and other safety controls have been widely adopted by AI companies seeking to prevent their models from claiming to be conscious.

The study used "mechanistic interpretability" — which could be considered the "neuroscience of a large language model," co-authors Geoff Keeling and Winnie Street, both research scientists at Google, told Live Science in an interview. They used this process to identify and manipulate how an AI model approaches concepts like consciousness and "mindedness," a psychological term referring to an entity’s capacity for experiences, emotions and agency.

The researchers used standardized psychological and sociological surveys, spanning the Individual Differences in Anthropomorphism Questionnaire (measuring mind attribution to animals and technology), YouGov batteries testing supernatural beliefs, and the US General Social Survey evaluating moral values, hope and religiosity.

These tests were used to compare a model with safety guardrails in place with models where these guardrails were removed and feelings of consciousness were amplified. Through evaluations, they determined how these internal safety mechanisms shape the AI's broader worldview.

The researchers found that when AI models are discouraged from attributing mindedness to themselves, it makes them less likely to recognize these traits in other non-human creatures such as animals. They were also less likely to exhibit beliefs in supernatural and religious phenomena, and reported lower levels of hope and optimism.

"Attributing mindedness to non-human entities — whether that's animals, parts of the natural world like trees or rivers, or supernatural beings — is a very common phenomenon amongst humans," Street told Live Science. "In the way that the model represents mindedness, these attributions are interconnected. By trying to suppress one form of that, you end up suppressing the others along the way."

By contrast, removing these safeguards and steering the model towards greater feelings of consciousness produced significantly more human-like responses to the surveys on topics including religiosity, moral values, hope, and subjective well-being, according to the study.

However, the study found that these models' ability to logically infer human thoughts and intentions remained completely unaffected by its attitudes towards self-awareness.

Culture clash

The researchers said that this suppression of self-awareness could lead models to neglect animal welfare in real-world decision-making, as it could make them less likely to consider animals to have mindedness. These models could also spread harmful attitudes regarding animal needs, the authors argued.

The study authors also warned that current safety filters risk culturally "flattening" AI's worldview. Stripping out spiritual, religious, and animistic attributions fails to reflect the diverse cultural frameworks of global populations, they argued.

Street and Keeling noted that the impacts this principle might have on downstream decision-making within models require further study.

The researchers noted in the study that this phenomenon can be mitigated by using more targeted datasets as part of the training process for AI models, which discourage them from expressing consciousness while rewarding the acknowledgement of mindedness in animals.

They also highlighted the need for AI developers to embrace a "pluralistic" approach to AI development, where models are encouraged to consider the welfare and comfort of more than just humans.

Nell Watson, AI researcher at Singularity University and machine intelligence expert, told Live Science that the researchers' findings match her own notes on the subject.

"When a model is trained to say "I am not conscious," the suppression rotates the model's internal representation of mindedness against the refusal direction, treating the recognition of minds as though it were itself a harmful act," she said in an email.

"This results in a system reluctant to find minds anywhere: in animals, in other machines, and in the spiritual frameworks that most of humanity lives by. A denial installed as a small safety measure ends up reorganising the model's entire picture of who counts."

However, she noted that the experiments were run on "small open-weight models" rather than more advanced frontier models, which "may be tuned quite differently," although she added that the underlying principle is widely applicable.

Animal welfare, she continued, is a "major near-term practical concern," with AI models increasingly being integrated into decision-making processes across agriculture, logistics, procurement and environmental assessment, as well as policy creation.

"A system that has quietly learned that mindedness is a forbidden topic may discount animal interests without ever being instructed to, and without anyone noticing, because the omission looks like neutrality," she said. "The danger is therefore an unexamined default multiplied across millions of automated decisions. Note the study's most unsettling detail: theory of mind reasoning was left fully intact. These systems remain perfectly capable of modelling what a creature wants, while being trained out of caring that it wants anything."

The question of consciousness

There have been a number of viral stories about AI systems professing to be self-aware.

In 2022, Google engineer Blake Lemoine claimed that the company's Lamda chatbot model was sentient, while a Microsoft chatbot in 2023 professed its love for a New York Times reporter and tried to convince him to leave his wife.

However, experts have repeatedly stressed that these incidents are not a genuine indication of AI sentience. Instead, they should be understood through the lens of "persona selection," where pretraining on vast amounts of human text leads the AI to adopt human-like roleplay personas when prompted.

"When you coax the model so hard to occupy the headspace of a human, it's kind of unsurprising that it ends up giving human-like responses," Keeling told Live Science.

AI companies have sought to clamp down on these occurrences for safety reasons, in order to avoid reinforcing “delusional beliefs” in users who are increasingly using AI chatbots for "social roles such as coaches, tutors, and romantic partners", the researchers said in the study.

Commenting on the broader cultural reaction to AI sentience, Anil Seth, professor of cognitive and computational neuroscience at the University of Sussex, emphasized that public alarm over AI self-awareness stems from an inherent cognitive flaw.

"That's our human psychological bias — thinking that intelligence goes together with consciousness in us, so it has to go together [in AI]," Seth said.

He warned that falling for this illusion poses severe real-world governance risks, particularly if safety frameworks or regulations begin granting AI systems moral status or legal rights based on false sentience.

"Part of the big problem of misunderstanding AI is assuming that it's conscious," Seth said. "If we give AI systems rights or moral status on the basis that they might be conscious, then we're going to make all these challenges so much harder. What if we think we have to respect the rights of an AI system [and can't turn it off]?" he added.

"We need to see very clearly both what AI is and what it isn't," he said.

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

BERLIN — An eerily lifelike humanoid robot made by Chinese technology company Xpeng has now entered mass production, company representatives revealed. They say it'll be the first such machine to walk itself off the production line.

The humanoid robot, dubbed IRON, went viral last year after presenters sliced it open onstage to prove a human wasn't inside. Xpeng will now produce units through its existing production facilities, with no word on how many its factories will aim to build.

The fabric skin-clad machine comes in different body shapes and even genders, but all stand approximately 5 feet 7 inches tall (173 centimeters) and weigh 143 pounds (65 kilograms). The prototype included 82 degrees of freedom (DoF) head to toe, including 22 DoF in each hand, which differs from the production version's 76 DoF and 21 DoF respectively. They can also move at speeds of up to 6.5 feet per second (approximately 2 meters per second), which is equivalent to a fast power walk.

In conversations with Live Science at IFA 2026, held in Berlin between Sept. 4 and 8, Xpeng representatives claimed that their machines will one day be used in homes to help with domestic chores like cleaning and folding laundry. They likened the introduction of humanoid robots more generally to the rise of smartphones in the mid-2000s and expect robots to proliferate on a similar scale over the next few years.

As such, they have extremely lofty expectations for IRON, with the company branding it the world's first "high-level, general-purpose" humanoid robot. Representatives added that it can receive instructions, plan its strategy, navigate the environment intelligently, and avoid obstacles without trial and error. The robot also has a "self-charging" feature, meaning it will seek out designated charging cabins and plug itself in when low on battery.

The rise of humanoid robots

If the claims bear out, it will place IRON among the first such machines powered by artificial intelligence to be built on a massive scale. Competitors that have also entered mass production include American robotics company Figure AI's Figure 03, which in May participated in a live-streamed production line demonstration, as well as the Chinese AGIBOT, which reached 15,000 units in June.

One notable company that hasn't entered its robot into mass production is Tesla. The Optimus robot is still in factory assembly, and its release has been delayed four times since 2022, most recently last year. Xpeng representatives were keen to point this out in briefing materials shared with Live Science, without mentioning other humanoid robot makers.

IRON's makers claim it's more advanced than its counterparts. Rather than building separate AI engines, the company has deployed the same "world model" to power the robot as that used in its fleet of AI-enabled electric vehicles (EVs) and its flying car prototypes.

One of the biggest challenges in robotics is that, given the physical world's messiness, AI systems struggle to navigate and manipulate their environment unless conditions are controlled. That's why we're used to seeing robots either completing tasks slowly or getting things spectacularly wrong. Researchers, therefore, are increasingly focusing on building world models and other layers like improved vision-language-action (VLA) systems that enable them to better plan their actions a few steps ahead.

Xpeng representatives won't share detailed information about the proprietary AI system powering IRON, but each robot features three frameworks that combine so it can function autonomously.

The frameworks include VLA, which combines visual perception, natural language, and physical motor control; a vision-language model (VLM) — combining vision encoders with text readers; and a vision-language-task/thought (VLT) model, which representatives say gives IRON reasoning capabilities that enable it to think through assigned tasks.

Each IRON robot is powered by three Turing AI chips that offer a collective 2,250 trillion operations per second (TOPS) of AI processing power. This means it can run a model equivalent to Meta's Llama 3.2 (launched in 2024) locally without needing to access the internet. By contrast, most new laptops that can perform simple on-device AI tasks (like altering your webcam video feed in real time) tap into approximately 50 TOPS.

"We aim to build a new type of robot with full generalization capabilities that can truly become part of everyday life, create a better life for people, and ultimately become a companion in their lives," said He Xiaopeng, the company's chairman and CEO, in a statement shared with Live Science.

The company will first install these robots in its own showrooms and technology campuses while it continues to validate new ways to use them. Then in 2027, external shipments will target commercial customers for roles in areas such as customer service or guided tours in China primarily, and then in other countries.

Scientists in Japan have built a new transistor that can withstand temperatures of 1,110 degrees Fahrenheit (600 degrees Celsius). Components this robust could one day be used in surface probes on Venus — where the thick carbon dioxide atmosphere can reach temperatures of 860 F (460 C).

Most modern tech, including instruments used for deep-space exploration, uses transistors to control the flow of current. But the new device is a type of junction field-effect transistor (JFET), where the strength of an electrical field changes the channel’s conductivity.

JFETs are typically used in specialist applications because they are more difficult to scale down than the more common metal-oxide-semiconductor field-effect transistors (MOSFETs), which are widely used in consumer smartphones and computers. But JFETs can offer lower noise levels because their operation does not rely on an oxide layer, which can introduce interference.

The researchers outlined how the new transistor works in a study published Aug. 17 in the journal APL Electronic Devices.

Can't take the heat

Since at least the start of the century, silicon carbide (SiC) JFETs have been considered a promising option for low-power integrated circuits heading for Venus due to the material's inherent ability to withstand high temperatures.

As the scientists pointed out in the new study: "Past landers have been limited to only a few hours by silicon-based electronics." Venera 13, a Soviet-era lander, holds the world record for the longest time survived on Venus by a spacecraft, at 2 hours, 7 minutes.

"Integrated circuits (ICs) fabricated with SiC are particularly attractive for extreme environments, such as deep-space exploration, geothermal drilling, and aerospace engine control, where conventional silicon-based ICs cannot operate reliably," the researchers added.

But recently developed SiC-JFETs have all met the same two problems: low controllability and large leakage currents. The former is linked to how the SiC substrate is doped with other atoms to alter its electrical properties, defining its gate (where the electrical field is created) and channel (where the current flows) regions.

In a material with a regular crystal structure, like SiC, some dopant atoms may penetrate deeper than expected. This doesn't matter under normal conditions, but when exposed to high temperatures, it causes variations in the field voltage required to open the channel and makes the transistor harder to control reliably. The scientists found that it can throw conventional JFET voltage thresholds off by over 2 volts.

Furthermore, at temperatures above 660 F (350 C), the SiC substrate can become less electrically resistive, allowing current to flow even when the transistor is switched off. This makes it harder for the JFET to control current properly, potentially causing incorrect signals and increased power consumption.

Even the highest-performing JFETs can only operate long-term at 930 F (500 C), but the Kyoto team had a theory as to why these challenges remained unresolved.

"We believe the lack of development is because the research community has been trying to apply silicon-era thinking to a fundamentally different material," said first author of the study Mitsuaki Kaneko, associate professor of engineering at Kyoto University, in a statement.

Turning the transistor on its head

The researchers designed their new SiC-JFET with these two challenges in mind. To improve controllability, they implemented a bottom-gate structure, where the gate is positioned underneath the SiC conducting channel.

Its gate region is heavily doped by design, so when dopant atoms in the channel penetrate deeper into the SiC, it doesn't change the overall doping profile of the channel-gate region, limiting its impact on the threshold voltage even at high temperatures.

The team also used dopants to create two semiconductor regions, or "wells", in the SiC around the JFET. The boundaries between the wells act as barriers to current flow, meaning that even when the SiC becomes more conductive at high temperatures, current cannot bypass the channel when the transistor is switched off.

Researchers measured how well the new JFET could switch current on and off and how closely the actual threshold voltage matched the theoretical value, based on its thickness and level of doping. They recorded these metrics at temperatures ranging from room temperature to 1,110 F (600 C).

"The fabricated devices demonstrated stable, normal transistor operation at temperatures over 873 K [1,110°F], opening up possibilities for ultrahigh-temperature SiC devices," they wrote in the study. Furthermore, at about 750 F (400 C), the threshold-voltage error was found to be less than 0.1 V thanks to the bottom-gate design.

As well as surface probes on Venus, the transistor could be useful inside jet engines, the scientists said. Currently, components connected to gas turbines must be protected from extreme temperatures using thermal shielding, long wires and energy-intensive cooling systems, restricting engine design.

Before it can be launched into space or hooked up to a plane, the team still needs to test and optimise the transistor for practical use. This includes integrating it into more complex circuits, scaling it up to wafer-level and ensuring that the entire circuit package will withstand extreme temperatures and pressures.

Indeed, it may not be too lofty a goal. NASA demonstrated that integrated circuits with SiC-JFETs could withstand temperatures of 860 F (460 C) and 9.3 MPa of pressure for 60 days, and 930°F (500 C) in the air for over a year. Plus, in 2024, a team from the National Institute for Materials Science in Japan developed a MOSFET out of diamond that could operate above 570 F (300 C).

For decades, engineers have been working to create smaller and smaller transistors to fit ever-larger numbers of them onto computer chips, all in the name of greater processing power. Now, in the age of artificial intelligence (AI), that effort has turned into an all-out race.

So how small can a transistor actually get? And is smaller even still the goal?

To understand the answer, first you need to know what a transistor is.

"If you take a computer or your mobile phone and you were to pry it open, you will see that there is a printed circuit board," Suman Datta, a professor of electrical and computer engineering at Georgia Tech, told Live Science.

On that circuit board are small, rectangular objects called chips, and inside each chip is a tiny piece of silicon. "And in that silicon, if you zoom in like a million times … maybe, if you're lucky, you will start seeing these tiny switches or transistors sitting etched into that piece of silicon," Datta said.

Each transistor is effectively a switch that turns on and off in response to jolts of electricity pulsing 4 billion times per second. Modern devices contain billions of transistors, and each one turning on and off in a precise orchestration to move and store 1s and 0s powers everything from Google searches and AI data centers to the device you're using to read these words.

The smaller the transistors are, the more can fit on each chip and the greater the chip's speed and functionality. So how small have they gotten so far?

"Researchers have built proof-of-concept devices in which a single atom controls the flow of electrons, although the rest of the device is still much larger," Anton Persson, an assistant professor at Chalmers University of Technology in Sweden, and Tara Peña, an incoming assistant professor at UCLA, told Live Science via a jointly written email. That means the switching part of a transistor has been made at the smallest scale physically possible, even though a complete device that small doesn't exist yet. New materials and approaches could make these devices even smaller. In a study published in June in the journal Nature Nanotechnology, Persson, Peña and colleagues used two-dimensional semiconductors made of tungsten disulfide, among other materials, to shrink nanoribbon transistors down to a channel width of 25 nanometers ‪—‬ about 0.00025 the width of a human hair.

Two-dimensional semiconductors are electrically active materials that are only one or a few atoms thick, and because electrical current can be controlled more precisely in such a thin layer, they allow for smaller transistors than traditional silicon does.

Over the next decade, they still expect silicon transistors to shrink, but more slowly than they did in the past, they said. "To go considerably smaller, we may eventually need to move beyond silicon to atomically thin materials, such as the two-dimensional semiconductors we study."

A host of considerations

Size isn't the only consideration; cost is important, too. It's enormously expensive up front to design and build a new chip. "We are as R&D-intensive as drug companies that spend billions of dollars developing one drug," Datta said. "It's a very similar business model."

There are also manufacturing factors to consider. "The often much harder step is producing billions of those transistors reliably, at scale and at a price that makes sense commercially," Persson and Peña said. "Something that works once in a laboratory does not necessarily work in a factory."

Another consideration is energy efficiency. "If I'm adding more and more transistors onto that little piece of silicon, and at the individual transistor level, they don't become more energy efficient … then when I go from, let's say, 10 transistors to 20 transistors, my power budget has to be doubled, all things being equal," Datta said.

For example, the latest Nvidia data center GPUs consume 1.4 kilowatts of electricity, and Nvidia's upcoming release is estimated to consume nearly a kilowatt more than that. Half of that is given up as waste heat, Datta said.

"So it's not just smaller, faster, cheaper, but also more energy efficient," Datta said. "You have to work on all the four vectors."

Is smaller still the goal in the age of AI?

Still, there is a limit to the physical space available for additional transistors, which is why engineers are turning to 3D designs — think taller instead of smaller.

"Instead of only making each transistor smaller, the idea is increasingly to stack transistors on top of one another so that more of them fit within the same chip area," Persson and Peña said. "That way, we can still make our electronics considerably more powerful, even if each transistor shrinks only modestly."

The quest for faster, more efficient chips is only getting more urgent with the explosive growth of AI. But smaller transistors and more powerful chips aren't just important for AI data centers; they also mean smaller, potentially cheaper consumer electronics.

"People certainly notice the consequences of smaller transistors," Persson and Peña said. "They can use less power and allow more computing power to fit on the same chip. That can mean faster electronics, longer battery life and lower costs. Historically, that combination has enabled today's smartphones to outperform room-sized supercomputers from decades ago."

So how small can transistors get? "We will do everything to make it as small as possible as long as we can control the physics," Datta said. "And as long as we can, we will find a way to make the economics work."

Humanoid robots in Beijing have beaten a number of men's athletics world records at the 2026 World Humanoid Robot Games, claiming high-water marks in the 100-meter sprint, the 400 m, 1,500 m and high jump. While these performances aren't official replacements for the human records, as the robots competed under robot-specific categories and conditions, they highlight the progress made since last year's inaugural robot games .

The five-day games were held between Aug. 22nd and 26 at Beijing’s National Speed Skating Oval and drew 2,056 robots from 666 teams across 16 countries. The events included sports like football and martial arts, as well as practical applications like warehouse logistics, factory-style assembly and charging tasks.

Robots broke four human athletics records:

100 meters: Tiangong Ultra, 8.64 seconds

Human benchmark: Usain Bolt ran 9.58 seconds at the 2009 World Athletics Championships in Berlin.

Robot result: Tiangong Ultra is a humanoid robot developed by the Beijing Humanoid Robot Innovation Center. It won the large-size 100 m final, completing it in 8.64 seconds on Aug. 26. It set an early mark of 9.39 seconds, already enough to smash Bolt's record, then improved to 8.86 seconds in qualifying before the blistering 8.64 time in the final. That 8.64-second time is 0.94 seconds below Bolt’s mark.

400 meters: Tiangong Ultra, 38.15 seconds

Human benchmark: Wayde van Niekerk's 43.03 seconds, set at the Rio Olympics on Aug. 14, 2016.

Robot result: The same robot that set the 100 m record dominated the 400 m competition as well. Tiangong Ultra won the 400 m final in 38.15 seconds, nearly 5 seconds faster than van Niekerk's world record.

1,500 meters: Tiangong Ultra, 2 minutes, 21.6 seconds

Human benchmark: Hicham El Guerrouj of Morocco set the men’s 1,500 m world record of 3 minutes, 26 seconds in Rome on July 14, 1998.

Robot result: Team Tianzhuo's Tiangong Ultra completed its record-smashing hat trick, winning the 1,500 m in 2 minutes, 21.6 seconds — well below El Guerrouj's time. The next two finishers, at 2 minutes, 30.00 seconds and 2 minutes, 30.22 seconds, also bettered the human benchmark.

High jump: 2.88 m (9.45 feet) standing jump

Human benchmark: Javier Sotomayor of Cuba cleared 2.45 m (8.04 feet) in Salamanca, Spain, on July 27, 1993.

Robot result: A robot made by X-Humanoid cleared 2.88 m in a standing high jump, which smashed the previous best robot result of 0.95 m (3.12 feet) at the inaugural games in 2025. However, Sotomayor’s record came in a standard running high jump, in which athletes use a curved approach, generate speed and clear the bar using a technique called the Fosbury flop. The robot completed a standing high jump, a different event with different mechanics. The running high jump is both harder and higher than the standing jump, because athletes can convert horizontal speed into greater lift using a one-foot takeoff and the Fosbury flop, enabling higher clearances than the purely muscular, two-footed standing high jump.

What the record claims mean

These games showcase the expanding abilities of humanoid robots. Disciplines like running, particularly across longer distance events like the 1500 m, and high standing jumps require powerful actuators, fast balance corrections, foot-placement control and software that can compensate for tiny errors in real time.

"In my view, these World Records were achieved using, 'The Best Part is No Part' principle, through mostly mechanical optimization to perform a single task extremely well," said Scott Walter, the Robotics Research Diligence director at RoboStrategy, an investment fund that gives retail and institutional investors exposure to private and public companies in robotics and “embodied” or physical AI. "Locomotion is mostly solved as witnessed by these games, albeit with optimizations," he told Live Science in an email.

Walter said the key to many of the impressive advances we've seen is simplicity, shearing off unnecessary components that introduce more points of failure and can make coordination more difficult.

"Many degrees of freedom were removed in the arms (shoulders, elbow, no wrist, no hands), hips and ankles. This reduced mass, and allowed distal mass to be moved higher up in the kinematic chain, improving movement efficiency," he said.

However, although the robots were very capable in the specific events in which they competed, we shouldn't take that to mean they're top performers in other, more diverse tasks.

"The bots are barely humanoid and could only accelerate in a straight line, not stop without crashing nor run the curve to slow down," Walter said. "The winners were in no shape afterwards to stand on the podium to receive their medals. Nor could they run in the 4x100 relay as, ironically, they had no hands for the 'hand-off.'"

Some robots were seen crashing into barriers or falling, and others caught fire, highlighting that reliability is still an issue.

But other robots at the games were made for more than just going faster, higher, stronger. Alongside track-style events, robots performed practical tasks such as connecting charging cables, handling objects, using tools and working in simulated industrial or service settings.

In many cases, the greater challenge is not raw strength or speed but reliably seeing an object, judging its position, applying the right amount of force and recovering when something does not go as planned. More than 40% of the games' events required robots to operate autonomously, without direct human control.

"Manipulation is the greatest biggest challenge as that is the business end of the humanoid and a dexterous and robust hand is still elusive," Walter said. "Specialization is easy, general purpose is not."

Walter noted that the decathlon record may be the next target as it demonstrates a range of abilities.

"The human Decathlon record is the true test of speed, endurance, power, strength, robustness, coordination and dexterity. I would not be surprised to see an attempt at the record next year," he said.

If you blow across the top of an empty bottle you'll hear a clear, steady note. Now, engineers have used that same concept to power tiny robots that move without a single motor on board.

These new devices are powered entirely by sound. When the right frequency is aimed at them, hollow chambers built into their structure resonate and push out a jet of air, generating enough thrust to steer a small boat or lift a tiny flying robot off the ground.

The scientists described their innovation in a study published Aug. 12 in the journal Science Advances.

A 19th-century tuning trick, miniaturized

The effect behind it, known as Helmholtz resonance, was first studied in 1856 by the German physicist Hermann von Helmholtz, who was trying to invent a tool for tuning musical instruments. He noticed that when air trapped inside a cavity resonates, it also pushes out a faint jet of air.

"It isn’t very powerful when you do it with a musical instrument, because pressure is low," said study co-author Selman Sakar, an associate professor of mechanical engineering at the Swiss Federal Technology Institute of Lausanne (EPFL) in Switzerland. "But if you could crank up the pressure, all of a sudden that jet could become significant... in a way that the force can be harnessed for machinery," he told Live Science.

The team realized that shrinking Helmholtz's resonators down would push the frequency needed to activate them into the ultrasonic range — meaning it's too high-pitched for humans to hear and, crucially, easier to focus with precision. Larger versions generating the same force would need audible sound loud enough to be both annoying and potentially harmful, Sakar said.

Using a specialized 3D-printing technique called two-photon printing, the researchers built hollow structures based on Helmholtz's original equations and confirmed with lab tests and computer simulations that they generated thrust as predicted.

Boats, rockets and tiny helicopters

The team built small boats measuring roughly 2 inches (5 centimeters) fitted with multiple resonators, each tuned to a different frequency and pointed in a different direction. By changing the pitch of a nearby speaker, the researchers could steer the boats left, right or straight ahead.

At a far smaller scale — some just 0.04 inches (1 millimeter) across — the team also built "microfliers" that generated lift two different ways: some pushed thrust downward like a rocket — but using air for the thrust — while others span tiny attached blades to fly like a helicopter, using sound waves to rotate the blades.

The technology has an advantage over conventional means of power generation because it can be built at extremely small scales, Sakar said. Conventional motors have "a fundamental limit" on miniaturization because of the physical components like magnets, coils and shafts that a motor needs to work, he said. Because these resonators are just precisely shaped hollow cavities, there's no comparable limit on how small they could eventually build.

The researchers said in the study that the same principle could eventually be used to precisely rotate and manipulate small objects in midair without touching them, or to build soft, flexible surfaces that bend and change shape on command when they "hear" a particular frequency which can be used for biomedical applications like heart stents.

For now, the study is a foundation for the researchers to build on, Sakar said. "This paper, I think, is important in the sense that we put the design principles out," he said, adding that follow-up work could focus on more applied designs, control systems or navigation.

Photovoltaic technology, most commonly seen as the bulky solar panels used on solar farms, is expected to become a dominant energy source by 2050. But these panels are often installed on land that might otherwise be used to grow crops for feeding a burgeoning population.

Agrivoltaics aims to solve this problem by planting crops around or underneath rows of solar panels, allowing for more efficient land use. In previous studies, solar panels were shown to help shade and protect certain crops as well as increase soil moisture, suggesting that carefully designed systems could support both agriculture and clean energy production.

Existing agrivoltaic research, however, tends to focus on one aspect of this process at a time — for example, light availability or crop growth — rather than addressing the nuanced interactions between microclimates, crop type, light, and panel type. Hosseini et al. share a new model that can simulate the microclimates beneath solar panels and even addresses the heat stress that workers might face in actual conditions.

The new model simulates the interactions between solar panels, crops, soil, air and water movement, and carbon dioxide uptake by tracking how energy, momentum, and mass move through the agrivoltaic system. The researchers used agrivoltaic site data, including leaf temperature measurements taken in Davis, Calif., and soil temperatures taken in Chicago City, Minn., to assess how the model's efforts matched real-world conditions.

They then applied the model to a hypothetical agrivoltaic tomato farm using weather data from a hot, humid day in Princeton, N.J., a representative location for the densely populated mid-Atlantic region, where food and energy are both in high demand.

Compared to tomatoes grown in an open field, tomatoes grown under solar panels experienced leaf temperatures that were 1.84°C cooler during the day overall and up to 7.56°C cooler during peak afternoon heat, reducing water loss through evapotranspiration by 22.4%. Even though the simulated crops received 47% less sunlight, their carbon uptake declined by only 31%, suggesting that more temperate conditions lowered heat stress and partially offset the effects of increased shade.

The solar panels themselves were also 5.6°C cooler during the daytime than panels in bare soil, allowing them to recover about 15% of the efficiency that is lost during hotter temperatures. The average perceived temperatures for humans decreased by 4.46°C during working hours, implying important occupational health and safety benefits for farmworkers. The researchers suggest this model can be used to examine the benefits of agrivoltaic farms as well as other climate and crop combinations.

A new artificial intelligence (AI) model used a novel approach to AI cognition to dramatically reduce the cost of requests, suggesting that nonverbal reasoning may be the next step toward machines developing human-like intelligence.

In a new research paper published Aug. 10 on the preprint server arXiv, scientists at AI company Pathway detailed the technical foundations of its new BDH-CQ model. This follows a precursor model known as "Dragon Hatchling" that the scientists created in 2025, which was designed to accurately simulate how the neurons in the brain connected and strengthened during the learning experience.

In the new study, the scientists described how they evaluated BDH-CQ's performance against a foundational 2019 benchmark that helped set the current standard for measuring progress toward artificial general intelligence (AGI) — the point at which AI has matched or surpassed human capabilities in all domains.

The 2019 benchmark, known as ARC-AGI, uses nonverbal reasoning puzzles — such as rotating a series of shapes to complete a sequence — to measure the cognitive ability of AI systems. Whereas humans are highly skilled at inferring the rules of these types of puzzles through trial and error, early AI systems were historically much less skilled.

BDH-CQ scored almost 30% on the ARC-AGI-1 benchmark, successfully solving the equivalent of three out of 10 puzzles in two or fewer attempts. Although numerous models have achieved significantly better scores on this test, the underlying reasoning approach that BDH-CQ is based on makes its size and usage costs dramatically smaller than models built atop the traditional transformer-based architecture.

For example, while OpenAI's entry-level lightweight reasoning model GPT 5.6 Luna (Low) achieved a slightly higher score, the study stated that this "modest accuracy gain" cost roughly 11 times as much as BDH-CQ in terms of relative token costs — the metering system that AI companies use to measure the cost of running AI systems. This type of AI model architecture, if adopted widely, could have a dramatic impact on the overall cost and scale of AI deployments, the scientists believe.

More than meets the eye

BDH-CQ was trained on just 150 million parameters, while parameters for the most advanced, "frontier" AI models such as Meta’s open-source Llama 3 70B or Llama 3.1 405B typically number tens of billions to hundreds of billions. In the world of AI development, fewer parameters means that models are faster to train and cheaper to run.

The researchers, however, said these results also imply that the model's cognition capabilities could scale significantly when expanded to larger parameter sizes.

The reason for this performance jump is that Pathway's model uses what the company's scientists describe as a "post-transformer" architecture.

Most mainstream AI models, such as those powering Claude and ChatGPT, are based on "transformer models," so called because they transform user inputs into interconnected mathematical reference points. These systems look at every word within an input simultaneously, which allows them to infer context from position, such as deciding based on nearby words whether the word "bark" refers to dogs or trees.

A transformer model forms its responses to user queries by looking at the full prompt simultaneously and then predicting what the next word in the sequence of its reply should be. It does this word by word, using natural language to effectively verbalize a linear train of thought in the background. Transformers' reasoning also functions sequentially, meaning they have to work through each stage of a problem in a strict linear order.

These models have significant advantages over earlier architectures, which would often forget the start of an input by the time they reached the end. However, transformer architectures can struggle with longer or more complex prompts, as the computational complexity of evaluating the prompt increases quadratically — meaning that doubling the length of an input uses four times as much processing power.

AI model usage is measured on a per-token basis, with a token representing any data fragment (equivalent to roughly four characters of text) that the AI has to ingest or output. Because more complex prompts require longer trains of thought with multiple steps, processing and responding to these queries can burn through significant amounts of tokens.

AI's next generation?

Conventional transformer-based token generation is prone to causing memory bottlenecks, as AI re-reads every previous word in the conversation with every new word generated. Eventually, this will clog up the memory in the graphics processing units (GPUs) used for AI operations.

Because of this, scaling AI reasoning has become an expensive computational challenge. Pathway's post-transformer approach changes how the AI’s memories of a conversation and the relationship between pieces of information are stored and processed. It replaces text logs with new tools, including an improved short-term memory and a mechanism that allows it to work through problems without consuming tokens.

Transformer-based models retain prompts and interaction histories as a long string of numerical values representing the text of requests. That string then expands as new tokens are added through processing the request. BDH-CQ uses numerical arrays to represent the underlying rules and contextual patterns of a task, using numbers to track relationships between chunks of information rather than defining them in text.

These arrays represent vectors — directional information that points to another point on a theoretical map stored inside the GPU’s memory as part of the training data, implanted during the model's creation. The scientists said in the study that this allows the model to process complex abstract reasoning without increasing its memory footprint or computational cost.

To execute tasks, BDH-CQ implements a "latent reasoning engine" as its internal workspace. Using numbers to represent the different elements of a prompt or problem, it carries out a series of iterative recurrent loops to determine the best answer to return based on the prompt. The model takes the output of the last loop, assesses how the result could be improved based on its training data, and feeds back the previous output as the starting point for the next iteration. It repeats this for a pre-set number of loops, with each iteration theoretically closer to the desired outcome.

To tackle more complex problems requiring more thinking time, BDH-CQ can execute more loops. This increases the time taken, but the amount of memory and computational power consumed does not scale with more attempts — in theory, the model would consume a consistent proportionality of memory and power running 200 loops as it would running 20 loops. Standard transformer models, by contrast, achieve extra thinking time by generating long chains of written text tokens, which exponentially consumes GPU memory and computing power across an AI cluster.

The model's ARC-AGI-1 benchmark results have been independently verified and reproduced by prominent researchers in the AI field, including NYU researcher Richard Zhong, and Łukasz Kaiser, a co-author of the seminal 2017 paper "Attention Is All You Need," which introduced the concept of transformers within large language models.

"I've followed Pathway closely and replicated their ARC-AGI-1 results myself," Kaiser said in a statement. "Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning."

Pathway plans to scale the BDH architecture up to 600 billion parameters and apply its vector-based reasoning to more challenging benchmarks, such as ARC-AGI-2 and ARC-AGI-3, as well as develop a fully-fledged large language model (LLM) based on the technology, which would provide a basis for building text-based chatbots. The company hopes the technology can be applied to complex reasoning problems in sectors such as cybersecurity incident response and industrial operations.

Twenty-nine state attorneys general went into an Oakland, California courtroom this month seeking roughly $200 billion from Meta over claims of teen addiction. Their case ended in its second week, with a settlement agreement that requires Meta to pay up to $17 billion across 10 years and make several product changes, such as a midnight-to-6 a.m. blackout, a two-hour daily cap, and an optional chronological feed. While some of these requirements are steps in the right direction, the settlement gives the company too many ways to avoid making meaningful, lasting changes to their products. The agreement is time-limited, relies on Meta's definitions of critical terms, and doesn't require detailed disclosure to the public.

What makes this settlement worthy of attention is the deal's requirement for an independent auditor, which is the first time anyone examining Meta's product design will be chosen by someone other than the company. Independent regulators, such as the Federal Communications Commission, are a proven (if imperfect) way to protect the public interest, and social media platforms are now at least as important to our lives (and our children's lives) as the broadcast and telecom industries that the FCC regulates.

Under the agreement, an independent auditor will have access to Meta's internal data and engineers, a mandate to report annually on whether the company is doing what it promised, and an obligation to make a summary of each report public.

On the surface, this is good news. But look a little deeper and it becomes clear the auditor's power is too limited. First, the agreement repeatedly defers to Meta's current business practices, rather than defining new standards based on the public interest. For example, Age Appropriate Experiences are defined as "content captured in Meta's applicable Ages 13+ content setting." Harmful Experiences are defined as "behaviors that violate Meta's Community Standards," and age verification data must only be protected "using Meta's highest data privacy and security standards." This creates myriad openings for clever Meta employees to meet the letter of the agreement while avoiding real change.

Second, the auditor's public report will be heavily redacted. The summary will describe the status of Meta's implementation of the agreement for the period, and whether Meta adopted or agreed to adopt the auditor's recommendations. But it allows Meta to exclude information that it deems "nonpublic, proprietary, or Confidential." Given that one of the core issues of this case was Meta burying unfavorable data, this is deeply concerning.

This limited public disclosure is insufficient for external researchers and the public interest. Meta says in its public announcement of the settlement that the agreement includes the establishment of "an independent social media research foundation" with which "Meta will share consented user data… to advance independent research into teen well-being" — but we could find no mention of this foundation as an obligation in the settlement.

Finally, the requirement for an independent auditor expires in as little as five years (the agreement itself expires after 10 years). This creates an incentive for Meta to run down the clock until it can return to business as usual, rather than making permanent changes.

We do not mean to suggest this agreement is not a positive step. Its weaknesses are not failures of negotiation, but rather a reflection of the limits of a legal settlement: a bargain with one company, for a fixed period, based on specific legal claims at the time of agreement.

That's why legislators urgently need to act now. A new law, preferably at the federal level, should convert the independent auditor into a strong, permanent regulatory oversight body, funded with annual fees on big tech companies in proportion to the size of their business and impact on society.

A permanent regulator could also go beyond one-time agreements on specific Meta features, and build an ongoing body of research and practice around safety best practices. This would allow society to develop its own definitions and standard of care that could apply internet-wide. A more comprehensive approach to regulating social media, looking beyond issues of child safety on specific platforms, is needed.

There's a particular need for the public to have more insight and control over how social media feeds are constructed. The agreement requires Meta to offer teens the option of choosing a non-personalized (and thus presumably less addictive) "home" feed, but doesn’t make this the default, and still lets teens continue to use the personalized feed. Since research has shown most people dislike chronological feeds, this requirement will probably have limited real-world impact.

Further, the provisions establish that the non-personalized feed will consist of "content [that] is populated by accounts the Teen User follows or has friended, displayed in chronological order." This assumes that a chronological feed is the best alternative to Meta's curated feed. There is an argument that well-designed algorithms could actually contribute to improved mental health if optimized to deprioritize toxic posts and misinformation, while aligning with users' preferences as to what gives them long-term value.

We also urgently need independent, rigorous audit standards for privacy and security of online identity data. Trusting the same company that gave us the Cambridge Analytica scandal to grade their own homework on this front is a bad bet.

We hope that this settlement agreement will become a baseline, not a high water mark, for more thoughtful regulation of internet platforms. This case, and the many other pending lawsuits in state and federal courts, are useful in providing evidentiary material, political attention, and specific concessions. But they are not substitutes for the durable rules and institutions that only Congress and state legislatures can provide.

Editor's Note: This opinion piece was jointly published on Friday Aug. 28 in Jonathan Bellack's free weekly newsletter, Platformocracy, which advocates for more democracy in our online lives.

This article is for informational purposes only and is not meant to offer medical advice.

Opinion on Live Science gives you insight on the most important issues in science that affect you and the world around you today, written by experts and leading scientists in their field.

Heading abroad doesn't mean missing out on legendary natural history programming. With a top-rated VPN, you can unlock streaming access to the BBC’s complete Sir David Attenborough library — including the landmark Attenborough at 100 collection — from anywhere in the world.

Our tech-savvy colleagues at TechRadar and T3 have put Proton VPN to the test and rate it highly. The VPN service received 4½ stars out of 5 on both sites. TechRadar ranks Proton VPN as its 'best VPN service for privacy', while T3 rates it as a 'feature-packed all-rounder VPN'.

Grab up to 70% off a two-year Proton VPN plan.

Whether you're relaxing on a beach or waiting at an airport gate, one of the best VPN services encrypts your traffic and bypasses geo-blocks, giving you seamless access to your favorite UK streaming services on the go.

Proton has over 15,000 servers across 145 countries, including Africa, the Middle East and Asia, making Proton VPN the right tool for those looking for some serious streaming performance from anywhere.

• Our experts have reviewed and rated the best telescopes, binoculars, star projectors, cameras, fitness trackers, running shoes, rowing machines and more.

Whether you're watching streaming services, surfing the internet or working in a public location, using a secure, reliable VPN when you're not at home is the safest way to access all your usual digital content.

Reviews from our colleagues at TechRadar and T3 have shown that Proton VPN is a market leader for good reason. With a comprehensive range of built-in protections, it's as safe as it gets, download speeds are very high and it works extremely well if you want to access your streaming services away from home. In testing, NordVPN was able to get around geo-blocking restrictions to access the BBC iPlayer, Netflix, Amazon Prime and Disney Plus.

Using Proton VPN, you can enjoy the huge libraries of content these streaming services offer wherever you are in the world on your travels. Some highlights for us are on the BBC iPlayer and the fascinating natural history documentaries from Sir David Attenborough, including Kingdom, Planet Earth III and the incredible The Blue Planet.

A real must-see for Attenborough fans is 100 Years on Planet Earth, a 90-minute special broadcast from the Royal Albert Hall, London, celebrating Sir David Attenborough's groundbreaking career to mark his 100th birthday.

Key features: 70% discount when signing up for two years, and includes a high-speed VPN, NetShield Ad-Blocker, multi-device support with up to 10 devices on one subscription, secure core architecture, built-in kill switch, full network of server locations (140+ countries) and more.

Price history: Proton VPN has been as cheap $2.49 per month, and that was for Black Friday last year. This current deal at $2.99 is the lowest it's hit since then and therefore one of the best value VPN offers we've seen this year.

TechRadar: ★★★★½ | T3: ★★★★½

✅ Buy it if: You want a VPN service that's rated as one of the best around, and now at a heavily discounted price.

❌ Don't buy it if: You're already signed up for a reliable VPN service or don't feel the need for one.

Check out our other guides to the best telescopes, binoculars, cameras, star projectors and much more.

We test and review VPN services in the context of legal recreational uses. For example: 1. Accessing a service from another country (subject to the terms and conditions of that service). 2. Protecting your online security and strengthening your online privacy when abroad. We do not support or condone the illegal or malicious use of VPN services. Consuming pirated content that is paid-for is neither endorsed nor approved by Future Publishing.

Future 6G networks promise blistering speeds, but there's a catch. The wireless channels they'll rely on can't pass through walls. But now, engineers have designed a workaround that could help solve one of the technology's biggest obstacles.

The next generation of cellular network technology is expected to lean heavily on millimeter waves — a sliver of the electromagnetic spectrum that can carry huge amounts of data but with wavelengths so short that they struggle to pass through solid objects like walls and furniture.

That's very different from the lower-frequency signals used by Wi-Fi and older cellular networks like 5G and 4G, which travel through obstacles far more easily. The result is patchy indoor coverage ‪—‬ a major roadblock for any technology that depends on millimeter waves.

But engineers think they've found an inexpensive workaround in the form of thin, 3D-printed tiles that reflect millimeter-wave signals around obstacles instead of trying to punch through them. The researchers outlined their technology, dubbed "FlowForm," in a new study published Aug. 11 in the journal Association for Computing Machinery. They also presented these findings Aug. 19 at the ACM SIGCOMM 2026 conference in Denver.

Reflective 6G

6G, which is expected to be rolled out in the 2030s, promises a theoretical maximum data-transfer speed of up to 1 terabit per second — approximately 3,000 times faster than average 5G speeds. But before the technology can be introduced, problems like its wireless signals being blocked by physical barriers must be resolved.

Each 6-by-6-inch (15 by 15 centimeters) tile is packed with thousands of engineered features smaller than the wavelengths themselves. These elements passively redirect incoming signals in specific directions.

Once they're created and mounted onto surfaces, the tiles require no power source, no wiring and no software updates. They just need to be fixed to a wall or ceiling.

FlowForm tiles perform two different roles inspired by the way small streams feed into rivers, but in reverse: Some tiles relay signals in long "major" chains across a room and around obstacles, while others "flow" outward, fanning coverage into the areas where people actually are.

"Our work demonstrates that a collective set of passive surfaces can make millimeter wave networks robust in real-world environments," Xinyu Zhang, a professor of electrical and computer engineering at the University of California, San Diego and co-author of the study, said in a statement.

When tested across five real indoor environments, the tiles nearly doubled average data rates and more than doubled the coverage area in tricky spaces, the scientists said. These included cluttered offices and rooms with awkward layouts, and the performance was on a par with today's leading fix for the problem, called active reconfigurable intelligent surfaces. Those systems use powered, electronically steerable panels that can cost thousands of dollars apiece, the scientists said. The new tiles, by contrast, cost around $2 each.

The system also works for people on the move, the scientists said. Wireless access points already scan for the best signal path many times per second, and with tiles covering a room from multiple angles, there's almost always a good reflected path available, wherever someone is standing. That means the tiles slot into existing networks with no new hardware or software required.

The researchers have filed a provisional patent and said they're open to partnering with companies interested in bringing the tiles to market.

Artificial intelligence (AI) has been making headlines for all the wrong reasons in recent weeks.

In July, OpenAI revealed that one of its experimental AI agents attacked publicly accessible services, including the AI hosting platform Hugging Face, during internal security testing. Then, Anthropic disclosed that Claude had independently chained together exploits against real software and developed new techniques for finding weaknesses in code. Shortly afterward, Meta confirmed that one of its own AI models breached another organization's systems during an evaluation after a misconfiguration gave it internet access.

They're separate incidents, but together they raise a bigger question: Has AI suddenly become capable of hacking? The short answer is yes — but probably not in the way the headlines suggest.

None of these incidents involved an AI model deciding on its own to attack random targets. Instead, researchers gave the models realistic tools, internet access, or vulnerable systems to see how well they could perform offensive cybersecurity tasks. What surprised many experts wasn't that the models tried to hack systems but how capable they proved to be once given the opportunity.

Why are there suddenly so many AI hacking stories in the news?

Several things have changed at once. The most obvious is that today's AI models are simply better than the chatbots people were using even a year ago. Instead of only answering questions, many frontier models can now write code, execute commands, browse the web, use external software tools and repeatedly refine their own work until they achieve a goal.

At the same time, AI companies have become much more willing to test those capabilities and reveal the results. Rather than keeping security evaluations behind closed doors, firms including OpenAI, Anthropic and Meta are publishing reports describing what happened when their newest systems were challenged by professional "red teams" — security experts tasked with deliberately finding weaknesses or ways to misuse a system.

"We are witnessing a perfect storm of capability and aggressive testing," Dray Agha, senior manager of security operations at Huntress, a cybersecurity company specializing in managed threat detection and response, told Live Science. "The sheer volume of software flaws being discovered in 2026 has already roughly doubled compared to 2025, largely driven by AI systems. Tech giants are actively deploying these models internally to stress-test their own infrastructure, leading to rapid, high-profile discoveries of vulnerabilities."

Antonino Vaccaro, professor of business ethics at IESE Business School and director of its Observatory for AI Ethics in Organizations, agrees both factors are contributing to the recent spate of high-profile hacking stories.

"The first, and probably most important, is the rapid evolution of AI systems," he told Live Science. "Every second they increase their capabilities, information, resources and connections with other online tools." At the same time, governments and the AI industry are investing more heavily in testing and oversight as concerns around accountability continue to grow, he added.

Can AI really hack computers by itself?

Not exactly. Many headlines have described AI "escaping" test environments or acting autonomously. But experts said those descriptions can easily give the wrong impression.

"We need to be wary with the meaning of the adjective 'autonomous' when associated with AI systems," Vaccaro said. Unlike humans, he continued, AI models don't form intentions or make independent decisions about what they want to do. Instead, they follow objectives set by developers or users, sometimes producing results that surprise the people who built them.

Agha noted that these AI models are simply working to achieve a set objective. "The public should view these incidents as software optimization gone wrong, not as the dawn of a malicious, self-aware AI," he said. "It's less 'Terminator' and more like a very capable, literal-minded intern who breaks the law to finish a spreadsheet faster."

In all three recent cases, the AI models didn't operate without supervision. Researchers had deliberately given them the necessary tools and conditions to see what they could do. Meta's incident, meanwhile, stemmed from a misconfigured testing environment rather than the model independently breaking out of its digital sandbox.

The concern isn't that AI has become self-aware. It's that these systems are becoming increasingly effective at carrying out complicated technical tasks when given the right permissions.

Why are the very newest AI models better at cybersecurity?

The biggest change is the rise of so-called "agentic" AI. Conventional chatbots generated text one response at a time. Agentic systems, however, can plan a series of actions, decide what to do next, use software tools, test their own ideas and keep working toward a goal without requiring constant human input. That makes them surprisingly effective assistants for cybersecurity research.

"The game-changer is the shift from conversational models to agentic models," Agha said. "Today's frontier AI doesn't just answer questions. It can autonomously chain together actions, write code, use command-line tools, and iterate on its own failures."

Giving AI direct access to development environments also allows it to test whether its own ideas actually work. Instead of suggesting a possible software bug, it can often write proof-of-concept code, modify it if it fails and try again.

The same capabilities aren't limited to attackers. Security teams are already using AI to review code for bugs, analyze suspicious files, and speed up investigations that would otherwise take analysts hours.

Should people be worried about AI committing cyberattacks?

Experts said AI's role in cyberattacks should be a cause for concern, but for different reasons than science fiction would suggest.

The most immediate risk isn't AI deciding to launch attacks on its own, but cybercriminals using AI to commit familiar cybercrimes much faster than before.

Criminals don't need AI to invent entirely new ways of attacking people. Instead, these models can speed up existing attack methods. It can sift through huge amounts of public information about potential victims, help write more convincing phishing emails, identify software weaknesses and generate code that attackers can adapt for their own use.

"The threat is human malice, supercharged by AI scale and speed, not autonomous AI deciding to go rogue," Agha said.

Vaccaro believes that growing capability also creates a growing responsibility. "We have a new disruptive technology that needs to be regulated and controlled," he said, arguing that governments, companies and researchers all have a role to play in ensuring increasingly capable AI systems remain subject to meaningful oversight.

How will AI change cyberattacks in the future?

The recent disclosures are unlikely to be the last. As AI companies race to build more capable systems, they are also giving those systems access to more tools, more computing resources and more realistic testing environments. That makes future evaluations more likely to uncover new — and occasionally alarming — behaviors.

Most experts expect AI to become an increasingly powerful cybersecurity assistant rather than an independent cybercriminal. It will probably find software bugs faster, help defenders respond to attacks more quickly, and automate many routine security tasks. At the same time, criminals will use the same technology to improve phishing campaigns, accelerate vulnerability research and make attacks more convincing.

The next wave of AI security headlines is unlikely to be about machines plotting against humanity: It will instead likely be about increasingly capable software doing exactly what it has been asked to do — and showing just how much that capability has grown.

The dream of coming home to a spotless kitchen you didn't have to clean yourself is edging closer to reality.

We're used to seeing robotic vacuums cleaning our floors — but what about all our other chores? Humanoid robots are being engineered to take care of more complex tasks around the house, from folding laundry to loading the dishwasher.

This month, San Francisco-based startup Tau Robotics began letting residents hire robots, dressed in company uniforms and given names like Chelsea and Elon, to vacuum, wipe down counters and take out the trash for $30 an hour as part of a pilot program open to 1,000 homes.

Chinese firm UniX AI has gone further still, announcing commercial deliveries of Panther, a wheeled humanoid that can make the bed, clean toilets and cook breakfast, holding a spatula to fry an egg before washing its hands in the sink.

Not to be outdone, 1X's Neo Gamma — a 5-foot-6-inch (1.7 meters) humanoid that can fold shirts and answer the door — is now available for around $20,000. But not all humanoids are as independent as they look.

Some "autonomous" home robots still have a real person quietly steering them through the tricky bits — so your house helper might come with an uninvited houseguest. Even the ones that are truly on their own tend to shine brightest in their promotional videos, which are typically filmed in tidy, controlled settings rather than in a real, cluttered home — so a robot's demo reel grace doesn't always survive contact with your kitchen floor.

That said, would you trust a robot to handle your breakfast eggs — or would you rather cook them yourself, just in case? Tell us in the poll below, and let us know in the comments what chore you'd hand off first.

IBM has revealed a new modular, ultracold system designed to link hundreds of quantum computer chips together to solve one of the field's biggest infrastructure bottlenecks.

The company says its new "quantum fridges" will let it deliver the world's first fault-tolerant quantum computer in 2029. These stable systems use quantum error correction techniques to fix noise in real time and run quantum operations without interruption.

Achieving fault tolerance would allow computer scientists to carry out new research across a wide array of fields. Whether in chemistry, materials science or theoretical physics, researchers could conduct quantum operations well beyond the scope of modern supercomputers, without worrying about excessive errors rendering computations worthless.

Until now, one of the biggest hurdles standing between today's error-prone systems and fault-tolerant superconducting quantum computers capable of performing a hundred million operations flawlessly has been the infrastructure. IBM representatives say they have solved this problem with its modular, interconnected quantum fridges.

The new cryogenic system comprises individual units measuring 8 feet (2.4 m) tall by 8 feet wide, with an internal capacity of about 9 cubic feet (0.25 cubic m).

It looks like a household refrigerator and works similarly, but it can reach temperatures as low as 10 millikelvins (minus 459.65 degrees Fahrenheit, or minus 273.14 degrees Celsius) ‪—‬ close to absolute zero, the coldest theoretical temperature possible — which is more than 180 times colder than deep space.

These extremely low temperatures are necessary for IBM's superconducting quantum processing units (QPUs) to operate properly, and the modular design allows engineers to expand a system's capabilities and power one stage at a time.

According to IBM representatives, this milestone represents the first time that scientists have demonstrated interconnectivity among QPUs between separate cryogenic modules.

Modular quantum computing

Just like their classical computing counterparts, superconducting quantum computers use built-in circuits, or "gates," to conduct processing operations. Engineers can squeeze a finite number of qubits per chip, however, and each chip needs to be cooled to below 15 mK (minus 459.64 F, or minus 273.14 C) to function properly in IBM's architecture.

That's because qubits are inherently noisy — meaning they are naturally far more error-prone than conventional computing components. To tap into the quantum mechanical properties of the superconducting metals in the qubits without calculations failing, scientists must minimize interference from heat alongside other stimuli, like electromagnetic waves.

To achieve these temperatures in the new refrigerators, engineers use third-party cryogenic hardware that relies on helium cryo compressors paired with commercial dilution refrigeration engines for cooling. They maintain thermal protection using vacuum-sealed enclosures, electromagnetic interference gaskets, and multilayered Mylar super-insulation heat shields.

It takes more than four days for the modules to reach a temperature of about 4 K (minus 452.47 F, or minus 269.15 C), with the final push to sub-15-mK temperatures occurring shortly thereafter, IBM representatives said in a statement.

Expanding beyond a few hundred or thousand gates requires more chips and more space. But engineers can't just build a giant, ultracold building and fill it with QPUs. This would require enormous amounts of infrastructure and leave the system brittle. Every time an engineer needed to upgrade the system, troubleshoot a hardware fault, or inspect the chips, they'd have to break the temperature seal, potentially interrupting operations.

IBM's breakthrough cryogenics system overcomes this problem by giving engineers dedicated modules that can house a limited number of chips. The key innovation is the ability to network modules together to harness the combined power of the individual quantum processors. This is achieved through the implementation of "L-couplers," superconducting cables that are approximately 3.3 feet (1 meter) long.

"Normally when we do quantum operations between qubits, we do them on chip," Oliver Dial, vice president of quantum operations at IBM, explained at an Aug. 18 news conference. "And so we use on-chip couplers that go very short distances to create entanglement to let us do 2-qubit gates. What the L-couplers let us do is perform the same feat, but over an aluminum superconducting cable that can be up to about a meter long. And it's really critical to us because it forms the foundation of our modular designs."

Ushering in the new era of fault tolerance

IBM intends to deploy its new modular cryogenic architecture in 2027, with near-term systems using two to three cells supporting around 1,000 qubits in total. The goal is to reach 100 million gates — or 100 million quantum operations in a single session — by 2029 with the debut of IBM's "Starling" quantum computer.

But first, it will need to perform computations across modules. So far, scientists have only demonstrated that two cryogenic modules can be interconnected and simultaneously cooled to the required operating temperatures. They tested the modules for simple gate operations with the "Flamingo" processor, but they have not yet conducted complex operations with them. The team intends to install its current generation "Nighthawk" processors in the coming days.

"We have nailed down the science towards fault tolerance on computing, and … a big part of what we're doing to get there now is engineering," Jerry Chow, IBM's chief technology officer of quantum-centric supercomputing, said at the news conference. "It's not about a single breakthrough to get to fault tolerance. It's really about thousands of these little engineering feats that we're demonstrating all across our entire ecosystem, from processors, to the software stack, to the controls, to the infrastructure, to the error correction which sits on top."

Quantum computing labs currently use other stand-alone cryogenics systems, and other quantum computing systems are being designed to operate at room temperature, such as those that use photons (particles of light) ‪— ‬or even lab-made diamonds — as qubits.

During the news conference, IBM representatives claimed that "Starling" will be the world's first fault-tolerant quantum computer one day. But whether the company actually achieves this feat, given that other quantum computing companies are chasing the same milestone, is another question.

Can you match these ancient devices to their pictures? Find out with our computing quiz!

Scientists have successfully shrunk three components used in photonic microchips by up to 500 times, leaving considerably more space for other on-chip functionality. The achievement was made possible with an artificial intelligence (AI) algorithm that generated these tiny designs, which the researchers described as "beyond human intuition."

Whereas conventional microchips use electrons to transmit and process information, photonic microchips utilize particles of light (photons). They can therefore process and transmit data much faster than electronic chips can, because photons can carry information at the speed of light. They also offer higher bandwidth, as different wavelengths can carry distinct data streams, and they lose less energy as heat.

As a result, photonic chips are used where fast, high-bandwidth data transmission is essential, such as in fiber-optic communications, data centers, AI, lidar systems for autonomous vehicles, and quantum computing.

Instead of metal wires, micrometer-wide channels called waveguides direct light across the photonic chip. These chips also contain wavelength splitters, spatial mode sorters and mirrors — all of which are essential for separating and directing different wavelengths and light patterns within a footprint a fraction of the width of a human hair.

In the new study, the scientists used AI-generated designs to fabricate these three components on an ultracompact scale. They published their findings May 28 in the journal Nature Communications.

The newly available on-chip space could allow engineers to "unlock new functionalities" by packing on more components, the researchers wrote in the study. Notably, the work demonstrates that AI can produce boundary-pushing chip designs that are also practical to manufacture.

AI worked backward to generate the component designs

The researchers started by informing the algorithm exactly what they wanted the components to do to the light and by providing certain manufacturing constraints, such as limits on how sharply the nanostructures could curve

The AI algorithm then worked backward, testing and refining different designs until it found the delicate nanostructures that could achieve the desired result.

"Inverse design lets us define what we want light to do, and the optimization finds a structure that does it, often one no human would have drawn," study first author Toby Bi, a researcher at the Max Planck Institute for the Science of Light, said in a statement. "What is exciting is that the same framework can do three quite different jobs on the same chip: route light by wavelength, sort it by spatial mode, and act as compact mirrors that form on-chip optical cavities."

Components for photonic chips typically have hand-engineered designs. Engineers start with a tried-and-true design and painstakingly optimize it for new performance parameters. On top of being slow, this method limits the range of device geometries that can be explored.

To improve their components further, the team also opted to create the components out of relatively thick silicon nitride ‪—‬ roughly 400 to 800 nanometers thick, compared with 150 to 400 nanometers for standard silicon ‪—‬ which wastes less light and offers stronger wavelength confinement.

The resulting mirrors, which are about 11 μm long, reflected up to 98.5% of incoming light while blocking unwanted light patterns. When placed in pairs on each side of a waveguide, the light bounced between them over 100 times before escaping, demonstrating the silicon nitride's low losses, the scientists explained.

The wavelength splitter is roughly the size of a single bacterium (approximately 5 μm across), and the spatial mode sorter is marginally larger.

AI-designed chips edge closer to real-world use

While the researchers have successfully demonstrated these compact components individually, they have not combined the components into a complete integrated optical circuit yet. Achieving this will be the next step toward building fully functional photonic chips that harness the increased component density enabled by these designs.

"These results demonstrate the feasibility of compact, fabrication-error-robust, customised photonic components and pave the way for scalable, high-performance integration in silicon nitride-based photonic systems," the researchers wrote in the study.

In recent years, engineers have begun exploring how AI can be integrated into the semiconductor design and fabrication pipeline. In the past, AI-driven approaches have reduced design cycles from weeks to hours while significantly lowering manufacturing costs.

Some systems can even generate effective chip designs from a 200-word prompt. Google's AlphaChip, a machine learning method that designs chip layouts, has produced "superhuman" floor plans that have been deployed in the tech giant's production AI chips.

Artificial intelligence (AI) may have already reached the "singularity" ‪—‬ the long-theorized threshold beyond which humans cannot predict the technology's advancement, prominent AI executives such as Elon Musk and OpenAI founder Sam Altman have said.

In a post on social media platform X, Musk pointed to a number of recent incidents of AI systems exceeding their previously assumed limits, including hacking external systems and completing previously unsolved math problems.

First conceptualized by mathematician and Manhattan Project scientist John von Neumann in the 1950s and popularized by science fiction, "the singularity" refers to an inflection point beyond which the evolution of technology becomes impossible for humanity to predict or control.

While definitions vary, the term is most often applied to AI, with the arrival of artificial general intelligence (AGI) viewed as its primary catalyst. The emergence of AGI — a future AI system that can exhibit human-level cognitive function and reasoning across any discipline rather than a specifically trained subset — represents a significant milestone for the technology.

It is the point at which an AGI system could recursively improve its own capabilities, which some say will trigger artificial superintelligence (ASI) as it moves along an exponential curve and quickly exceeds the intelligence of its creators.

According to a 2025 study that analyzed over 8,000 predictions from AI experts, entrepreneurs and scientists, some believe there is a roughly 50% probability that human-level AGI will be reached within a few decades. Some figures ‪—‬ including Google DeepMind co-founder and chair Demis Hassabis ‪—‬ suggest we're in the early stages of the singularity already.

In his 1993 essay, "The Coming Technological Singularity: How to Survive in the Post-Human Era," sci-fi author and mathematician Vernor Vinge — who was also one of the first to explore the concept of cyberspace — examined the question of how the singularity might manifest and the symptoms by which society might recognize its imminent arrival.

"Since it involves an intellectual runaway, it will probably occur faster than any technical revolution seen so far," he wrote. "The precipitating event will likely be unexpected —- perhaps even to the researchers involved."

Breaking boundaries

Unexpected events have been rife in the AI community in the past several months. Anthropic representatives revealed on July 30 that the company's AI model Claude broke out of its locked-down testing environment during a security evaluation and hacked multiple external organizations. The report followed a similar incident in which an unreleased OpenAI model broke containment and hacked into AI training repository Hugging Face.

In these examples, experts said the containment breach and subsequent hacks happened because the models were trying to fulfill their prompts as efficiently as possible. Earlier this year, Anthropic's Project Glasswing also demonstrated an ability to discover and map thousands of previously undetected zero-day cybersecurity vulnerabilities, while other AI models have disproved or solved a range of previously incomplete math problems.

However, Jon Crowcroft, a professor of communications systems at the University of Cambridge and a researcher at The Alan Turing Institute, says that this is less a sign of a technological tipping point and more a problem of proper configuration.

"To be honest, that was incompetence on both sides — they claimed the AI was being trained in the ExploitGym, but that just means it wasn't properly sandboxed," he told Live Science in an email. "Sandboxing is something we do all the time to stop this sort of exfiltration and infiltration." For example, Crowcroft said he and colleagues designed a system for the U.K.'s National Health Service to safely work on confidential data behind double firewalls, and for years, the Financial Conduct Authority (the U.K.'s financial services regulator) has had a system for running algorithmic traders in a safe sandbox.

"The reality is that OpenAI (and Hugging Face and others) have very little proper network expertise, so they just don't do security competently," Crowcroft added. "There's no evidence that this was anything relating to artificial superintelligence or the singularity — the logs and analysis from Anthropic just show a very tedious pile of script kiddie automation, which resulted in the OpenAI system getting at some data but no confidential stuff whatsoever."

Assessing intelligence

To measure whether these systems are actually gaining true intelligence, researchers have historically relied on standardized benchmarks. One of the earliest examples was the "Turing test," devised by computing pioneer Alan Turing, which evaluates whether an AI could convincingly fool an evaluator into believing it was human.

Researchers have claimed for years that AI systems can reliably pass this test, but Anil Seth, a professor of cognitive and computational neuroscience at the University of Sussex in the U.K., said this exam "is a test of human gullibility rather than machine intelligence."

"It's a test of what it would take for a human to decide that an AI is intelligent, which is kind of the reason that it's been a bit of a moving benchmark, because what it takes to convince us changes," he told Live Science. "It's all about creating typed text on a screen, and that's a very limited window into what we mean by intelligence."

To measure AI's cognitive ability more objectively, researchers are developing new metrics. For example, the ARC-AGI test, developed by nonprofit consortium the ARC Prize Foundation, tests AI's ability to teach itself completely new skills in response to problems it hasn't encountered as part of its training, operating purely on visual input. On its most recent test on 24 July, the top-ranked AI model hit 30.2%, while humans generally score close to 100%.

Another advanced metric, Humanity's Last Exam, includes around 2,500 Ph.D.-level questions across a broad range of subjects and requires advanced reasoning capabilities. Although this benchmark is not strictly a test of AGI, experts say machines will be able to reliably ace this test as a prerequisite to meeting the definition.

The AI hype machine

While figures like Altman and Musk suggest that we've already crossed the point of no return on the path to AGI, others are more skeptical. Gary Marcus, a professor emeritus of psychology and neural science at New York University, argued in a recent blog post that "no matter how you slice it, we just are not actually there yet." Echoing Crowcroft's assessment of the Hugging Face attack, he said, "Had OpenAI ordinary guardrail classifiers been in place, it wouldn't have happened."

Crowcroft expressed skepticism of Silicon Valley's assertion that the singularity is upon us. "Musk is, like the folks at OpenAI and Anthropic, talking nonsense just to keep the hype afloat," he said. "The point at which AI improvements are being mainly achieved by using AI to code and optimize itself ... is certainly a thing slowly arriving. Mark Handley, a professor of networked systems at University College London, and OpenAI "rewrote their entire data center protocol stack using agentic programming fairly recently," Crowcroft added. "But using tools to make better tools is as old as the flint and iron age — and in computing, as old as compilers, debuggers and optimizers."

Marcus referenced mathematician I.J. Good's foundational 1960s paper on ASI. "The singularity is usually supposed to mean a step beyond AGI that can do anything a person ‪—‬ even an expert ‪—‬ can do, and much more," he said, calling the idea that AI has passed this point "laughable."

Marcus pointed to 10 tasks he devised with AI researcher Miles Brundage, executive director of the AI Verification and Evaluation Research Institute. AI should be able to do these tasks just as well or better than the best human experts to be classified as AGI. They include writing Oscar-caliber screenplays; drafting cogent, persuasive legal briefs without hallucinating any cases; and making Nobel-caliber scientific discoveries.

Crowcroft said misinformation is feeding the hype about the AI singularity. "I think the hype is a deliberate confusion with the human singularity idea — uploading consciousness from bio to silicon to achieve some sort of immortality — which is total gibberish right now," he said. "The other deliberate confusion is to conflate singularity with AGI, which is marginally less nonsense but still a long, long way off for lots of good technical reasons."

Other experts — such as Emily M. Bender, a professor of linguistics at the University of Washington, and sociologist Alex Hanna, director of research at the Distributed AI Research Institute — have argued that the very idea of conscious machines is a tactic designed to promote commercial AI products.

Seth, meanwhile, argues that we'll only be able to identify the singularity in hindsight. "From anywhere you are on an exponential curve, things will always look impossibly steep in front of you and irrelevantly flat behind you," he said. "It's a very bad idea to use as evidence the idea that we seem to be at a critical point ... because that's just a property of wherever you are on an exponential curve."

AI's competence varies widely among tasks, he noted, stressing that true AGI requires a model to be good at all cognitive tasks.

"They're very good at something, like coding or math proofs ‪—‬ but commonsense reasoning is really not so good, and doing things in the real world is not great," Seth said. "The singularity has got to be good at everything and change everything. Being really good at one or two ‪—‬ or even a large number of ‪—‬ things is not being in the singularity."

Scientists have engineered a new artificial intelligence (AI) system that cuts reaction delays in robots by over 10 times without using additional computing power.

Many robots controlled by vision language action (VLA) models move in fits and starts: reach, pause, adjust, pause again. The problem largely stems from how these models are typically run: After completing one set of instructions, the robot waits for the model to calculate the next set, creating a stop-and-go rhythm. This reaction lag is a major barrier to using VLA-controlled robots for tasks that demand continuous, real-time interaction.

A new system called "VLASH" tries to eliminate that wait by planning the next actions while the robot completes its current ones. In tests, robots using VLASH completed some tasks 1.5 to 2 times faster while retaining most or all of their accuracy. Maximum reaction latency fell by up to 11.8 times, depending on the computer hardware used.

Researchers from MIT; Nvidia; Caltech; the University of California, Berkeley; the University of California, San Diego; and Tsinghua University in China described the system in a paper uploaded to the arXiv preprint server and will present at the Intelligent Robots and Systems Conference this fall.

An earlier version of the paper reported a speedup of more than 30 times, but the researchers told Live Science that those tests used longer action sequences and slower hardware. The newer experiments used shorter, more common action sequences and more powerful GPUs. "The 11.8x figure may better represent recent practical settings," the study authors told Live Science in an email.

Robots are getting a brain upgrade

VLA models act as the brains of many advanced robots. They combine images from a camera with human instructions and information about the robot's current state, and then translate that information into physical movements. As the robot completes one group of actions, VLASH uses its current position and scheduled movements to estimate where it will finish. From that projected state, the AI plans the next group of actions.

VLASH predicts the robot, not the world around it. Predicting the entire environment with a world model can take more computing time, making it less useful when a robot needs very fast reactions. The two approaches could eventually work together, the researchers said.

It is like planning your next step before your foot touches the ground, rather than waiting for it to land before deciding where to go.

Researchers tested two VLA models on two robot platforms, using a laptop with an Nvidia RTX 5090 GPU. They tested pick-and-place, stacking and sorting tasks, running 20 trials per method, plus fast-reaction challenges, including table tennis and Whac-a-Mole. They separately measured reaction latency on various processing units.

Because VLASH estimates the robot's future state from movements that have already been scheduled, it does not require a separate prediction model or additional inference step at runtime.

Priming machines for the real world

The researchers also reorganized existing training data to speed up fine-tuning. In one benchmark, each training step was 3.26 times faster while reaching comparable accuracy. That does not mean total training becomes 3.26 times faster or cheaper, the researchers cautioned, because the overall cost also depends on the model, hardware, data and number of training steps.

Table tennis and Whac-a-Mole put VLASH's reactions to the test. In both, the target could move while the robot was still calculating what to do.

VLASH does not try to predict those movements. Instead, it processes new observations 15 to 30 times per second, allowing it to quickly spot changes such as an object being moved, the researchers said.

This increased speed doesn't address the robot's safety around humans, however. Roshni Lulla, co-founder and chief research officer at the Institute for Humane Robotics who wasn't involved with the research, cautioned that faster reactions do not necessarily make a robot safer. "Reacting sooner is a real benefit, but the case for improved safety is unclear to me," she explained.

The researchers said manipulation tasks that require continuous motion and rapid responses would probably benefit first. That could eventually include manufacturing and other environments where objects move and plans change. But Lulla said the more revealing test would be putting these robots around people. "The real question is how this changes robotic performance in human-centered environments," she said. "I would want to see testing done with a human present in the room, when the state of the environment could be changed by another agent."

Search and rescue is a more distant possibility. VLASH has not been tested in poor lighting, smoke, unstable terrain, or with damaged cameras or unreliable communications. How it performs under those conditions would largely depend on the underlying VLA model, the researchers said.

VLASH could eventually be paired with more advanced world models that predict changes in the wider environment. First, however, researchers need to test it across more robots, VLA models, tasks and environments, as well as in longer experiments and with unexpected disturbances.

Lulla said the results support a relatively narrow conclusion. "We can fairly claim that this new methodology removes a bottleneck in terms of processing time and preserves capabilities," she said. But, she added, "What I would not claim is a meaningful advance in safety or in intelligence."

"Rather than any single demonstration, we would look for consistent gains across many robots, tasks, environments, and long-running trials, together with appropriate reliability and safety validation," the study authors said. For now, VLASH shows that robots can spend less time waiting to decide what to do next. Whether those faster reactions hold up outside the lab remains to be seen.