Chinese AI Rival Claims to Beat Industry Leader

The key takeaway is that Chinese AI company iFlytek has developed a new AI model called Spark v3.5 that its creators claim can beat OpenAI's industry-leading GPT-4 Turbo model in several key areas like language, math, and coding capabilities.

Summary

• Spark v3.5 beats GPT-4 Turbo in language capabilities, math and coding tasks according to its creators. It is also marginally better at multimodal tasks like ingesting text and producing images.
• The model was trained on a less advanced computing platform than GPT-4 due to US export restrictions on AI technology to China.
• Spark v3.5 can synthesize human speech conveying different emotions and tones. Its voice recognition capabilities also exceed OpenAI's Whisper model in 37 languages.
• There is no standardized benchmark for comparing AI models. Companies use dozens of academic benchmarks to test performance in different areas.
• Spark v3.5 has been integrated into various smart devices and apps in China, including school blackboards, tablets and a mobile app that transcribes phone calls.
• The model was trained over 90 days on a computing platform called "Feixing No. 1" which lacks Nvidia's state-of-the-art GPUs that were used to train GPT-4 and ChatGPT.