Scaling compute still fruitful in advancing AI, Google DeepMind scientist from China says

Amid heated discussions in the AI community about the future of scaling – the process of increasing computational resources and training data to develop better AI models – Yao said the method was still expected to yield results for at least a year “until we hit the hard boundary of data”.

“There are still many low-hanging fruits to be picked,” Yao said in an interview with the Post on Wednesday, shortly after AI pioneer and OpenAI co-founder Ilya Sutskever said on a podcast that the sector was returning to an “age of research” following “an age of scaling”.

Sutskever’s comments come amid concerns of a financial bubble in the AI industry, where US hyperscalers, including Google and Microsoft, have committed hundreds of billions of dollars to AI infrastructure to train the next generation of large language models, spurring demand for advanced AI chips from industry leader Nvidia.

However, the rise of Chinese AI start-ups such as DeepSeek and Moonshot AI has raised questions about such extensive spending. Faced with restrictions on access to advanced US chips, these companies have focused on algorithmic enhancements to improve AI models.

Yao believed there was “no reason” to choose between the scaling and research approaches, as the AI industry was “always engaged in both”.