According to The Information, less than two months after establishing its dedicated "AI Code突击队", Google has reorganized the team, expanding its responsibilities from solely improving coding tools and intelligent agents to covering broader "mid-training" work for models. This move aims to strengthen its technical foundation and narrow the gap with Anthropic in the field of generative AI programming.

Starting with the basics

Mid-training refers to feeding models with carefully selected structured data after they have completed large-scale pre-training, but before they undergo instruction alignment and task fine-tuning, to further "catch up" on high-difficulty tasks such as reasoning and programming. For Google, this means no longer relying solely on optimizing prompt design or product interfaces, but directly enhancing the fundamental coding capabilities of Gemini.

In April this year, Google formed this AI coding突击队, led by Sebastian Borgeaud, a DeepMind research engineer with long-term experience in model pre-training, focusing on tackling complex and time-consuming large-scale programming tasks. Google co-founder Sergey Brin and DeepMind Chief Technology Officer Koray Kavukcuoglu were also reported to be involved, showing the company's extreme emphasis on this field.

Currently, Anthropic has regarded "writing code" as a core part of its AI strategy, continuously advancing through Claude Code and the Claude model family — the latest release, Claude Opus 4.8, has seen significant improvements in code and intelligent agent tasks, and is viewed by many developers and enterprise users as an important reference for evaluating the competitiveness of large models.

Talent loss intensifies anxiety

At the same time as adjusting its technical roadmap, Google is also facing increasingly fierce competition for talent. Recently, Noam Shazeer, co-leader of the Gemini project, announced his departure from Google to join OpenAI, while two other researchers involved in the Gemini and DeepMind projects were reported to be preparing to join Anthropic. The continuous outflow of key talents has made Google not only face the challenge of product gaps but also deal with the real risk of research resources continuously moving to competitors.

It remains unclear whether the reorganized team will launch a new public Gemini model or new products for developers, and Google has not disclosed the team size, specific performance goals, or release schedule. However, from the rapid restructuring, the introduction of mid-training, and the involvement of executives, it is clear that Google is trying to regain control in this commercially valuable key area of AI coding through deeper technical adjustments.