Less than two months after establishing a specialized "AI Code Rapid Response Team," Google has quickly initiated team restructuring, aiming to close the gap with Anthropic in the highly commercially valuable generative AI programming field. According to information, the team responsible for Gemini development at Google DeepMind has adjusted its responsibilities, shifting focus from single-code tool development and intelligent agent optimization to more core model "mid-training" stages.

The so-called mid-training refers to the process of using high-quality, carefully selected data to "supplement" the model after it completes large-scale pre-training but before formal instruction alignment and fine-tuning. Industry research shows that this strategy significantly improves the model's reasoning capabilities, especially for highly structured abilities such as programming and mathematics. Google's move marks a shift in strategy — no longer merely optimizing prompts or product interfaces, but deeply digging into the "basics" of underlying model capabilities.

In April of this year, Google formed this rapid response team led by DeepMind research engineer Sebastian Borgeaud, aiming to tackle large and complex programming tasks. The deep involvement of company co-founder Sergey Brin and DeepMind Chief Technology Officer Koray Kavukcuoglu demonstrates Google's management's determination to regain dominance in the code field. Within DeepMind, Anthropic has been viewed as a leader in this field due to its Claude series models and outstanding performance in code tools.

Besides competition on the product development front, Google is currently facing severe talent competition pressures. Noam Shazeer, co-lead of the Gemini project, recently left to join OpenAI, and two other key researchers are reportedly joining Anthropic. The continuous flow of core talent to competitors means that while Google tries to narrow the technological gap, it must also deal with challenges related to team stability.