Mainland companies cut costs by building models trained on smaller amounts of data, requiring less computing power but optimized hardware, said Lee Kai-Fu, founder of 01.ai and former head of Google China.

According to the ranking recently announced by UC Berkeley SkyLab and LMSYS, the Yi-Lingtning model of startup 01.ai ranked third, tied with Grok-2 of x.AI, behind OpenAI and Google. This ranking is based on users' scores for query answers.

01.ai and DeepSeek are mainland AI companies that are adopting a strategy of focusing on smaller datasets to train models, while hiring cheap, highly skilled manpower.

The FT said Yi-Lightning's inference cost is 14 cents per million tokens, compared with 26 cents for OpenAI's GPT o1-mini. Meanwhile, GPT 4o costs up to $4.40 per million tokens. The number of tokens used to generate a response depends on the complexity of each query.

Yi-Lightning founder revealed that the company spent $3 million on “initial training,” before fine-tuning for different use cases. Lee said that their goal was “not to create the best model,” but to build a competing model that was “5-10 times cheaper.”

The method that 01.ai, DeepSeek, MiniMax, and Stepfun have applied is called “expert modeling” — simply combining multiple neural networks trained on domain-specific datasets.

Researchers see this approach as a key way to achieve the same level of intelligence as big data models but with less computing power. However, the difficulty with the approach is that engineers must orchestrate the training process with “multiple experts” instead of just one general model.

Due to difficulties in accessing high-end AI chips, Chinese companies have turned to developing high-quality data sets to use to train expert models, thereby competing with Western rivals.

Lee said 01.ai has non-traditional ways of collecting data, such as scanning books or collecting articles on the WeChat messaging app that are not accessible on the open web.

The founder believes that China is better positioned than the US, with a huge pool of cheap technical talent.

(According to FT, Bloomberg)

Meta Launches 'Self-Evaluating Learning' AI Model Social media giant Meta has just announced a new AI model with self-evaluating learning (STE) feature, which can narrow down human intervention in the AI ​​development process.