Google has rebuilt the small team it set up this spring to catch Anthropic in AI coding, and the reshuffle lands just as the company tallies a steady loss of senior research talent to its rivals.
The team came together in April. Insiders call it a coding strike team, a focused group given one job: close Gemini's gap against Anthropic's Claude on agentic programming, the kind of work where a model writes code, runs it, reads the errors and fixes them with little human help. The Information first reported that Sebastian Borgeaud, a research engineer who used to run Gemini's pretraining, was leading it. About two months later Google reorganised the group and widened its remit to cover "midtraining," the stage between a model's broad initial training and the final tuning that teaches it to follow instructions.
The push is coming from the top. Sergey Brin, back in a hands-on role, told staff in an internal memo that the company has to move faster. "To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers" of finished code, Brin wrote, in a memo reported by TechRadar. DeepMind's chief technology officer, Koray Kavukcuoglu, has been involved as well.
Six departures in five months
The rebuild is happening on a thinning bench. Google DeepMind has lost six researchers in five months to Meta, OpenAI and Anthropic. Denny Zhou, who founded the reasoning research team inside Google Brain, went to Meta. Noam Shazeer, a Gemini co-lead and one of the authors of the original transformer paper, announced his move to OpenAI on June 18. Jonas Adler and Alexander Pritzel, two senior researchers tied to Gemini and AlphaFold, confirmed plans to join Anthropic, as did John Jumper, who shares the 2024 Nobel Prize in Chemistry for his work on AlphaFold. Dawn Song, one of the most cited researchers in computer security, is joining Meta's Superintelligence Lab.
What stings is that the talent is not just leaving, it is reinforcing the exact competitors Google is trying to catch. Several of the six went to Anthropic and Meta, the labs Brin's memo holds up as the benchmark.
Google is not short on evidence that coding matters. The company has said AI now writes a large share of its own code, and agentic coding has become the field where Claude has built the clearest lead. The pressure is showing on the product calendar too. Gemini 3.5 Pro, expected in June, slipped past the month without a public release, with Google pointing to further performance work before launch.
For now the strike team is Google's answer, broadened from a narrow coding squad into something closer to a rework of how Gemini is trained. Whether that closes the gap or merely slows the bleeding is the question the next model release will have to settle.
Commentarii · 0