Chinese scientists claim they have built an artificial intelligence (AI) system that can beat one of the world's most widely used large language models (LLMs).
iFlytek's Spark v3.5 outpaces OpenAI's GPT-4 Turbo in language capabilities, math and coding, and is a close second to the American AI system in several other areas, according to Chinese-government-affiliated media outlet Shine.
Spark v3.5 is also marginally better than GPT-4 Turbo at multimodal tasks, iFlytek's chairman Liu Qingfeng said Jan. 29 at a company conference. This means it's more proficient at understanding one kind of input and providing a different form of output as an answer — such as ingesting a text prompt and producing an image.
GPT-4 Turbo is an upgraded version of GPT-4, which is used to power ChatGPT, and is widely considered to be one of the most powerful AI tools since it rolled out in November 2023.
There is no single, standardized method for comparing LLMs with each other, nor is there a publicly available database comparing different proprietary AI systems. Rather, companies use dozens of different benchmarks to generate performance scores in different areas. AI companies often use them to compare their own models with industry-leading variants.
Google, for example, revealed in December 2023 that its new Gemini LLM beat the standard version of GPT-4 and other leading models in 30 of 32 academic benchmarks used in AI research and development. These included high school exams and tests on morality.
Reporting by the state-owned China Global Television Network (CGTN) revealed Spark v.3.5 can synthesize speech that conveys different emotions, tones and speech patterns. Its voice recognition also outperforms OpenAI's Whisper in 37 languages including English, Chinese, French and Russian, CGTN reported.
iFlytek has integrated Spark into including smart devices, school blackboards and tablets, according to Shine. The company also launched a voice-to-text mobile app in partnership with China Mobile on Jan. 29, which uses Spark v3.5 to transcribe phone calls and highlight key information conveyed in the conversation.
The AI tool was trained for 90 days on a computing platform called "Feixing No. 1." Due to restrictions on AI-related exports to Chinese companies, imposed by the U.S. government, the AI company wasn't able to train the model using state-of-the-art components. These are namely graphics processing units (GPU) built by Nvidia — including the A100, which was used to train ChatGPT, as well as the H100 and H200 chips.
Sign up for the Live Science daily newsletter now
Get the world’s most fascinating discoveries delivered straight to your inbox.
Keumars is the technology editor at Live Science. He has written for a variety of publications including ITPro, The Week Digital, ComputerActive, The Independent, The Observer, Metro and TechRadar Pro. He has worked as a technology journalist for more than five years, having previously held the role of features editor with ITPro. He is an NCTJ-qualified journalist and has a degree in biomedical sciences from Queen Mary, University of London. He's also registered as a foundational chartered manager with the Chartered Management Institute (CMI), having qualified as a Level 3 Team leader with distinction in 2023.