Google's DeepMind Forms Coding SWAT Team to Catch Up with Anthropic, Bloem Leads the Charge

According to 動察 Beating monitoring, The Information cited three sources to report that Google's DeepMind has assembled a task force composed of researchers and engineers to specifically enhance its in-house coding model. The task force is led by DeepMind Research Engineer Sebastian Borgeaud, who previously led DeepMind's pre-training efforts; co-founder Sergey Brin and DeepMind's Chief Technology Officer Koray Kavukcuoglu are directly involved.

The direct trigger for team formation was Anthropic's recent model release. Internal DeepMind researchers believe that Anthropic's coding tools have surpassed Gemini's code generation capabilities. In a recent memo, Brin wrote that the team must "urgently close the gap in agent execution capabilities, turning the model into the primary code creator" to win this final sprint. An agent refers to AI that can handle multi-step tasks.

The gap has specific figures. Boris Cherny, head of Claude Code at Anthropic, stated in January that "almost 100%" of the company's code is written by AI; Google CFO Anat Ashkenazi mentioned during the February earnings call that coding agents are only responsible for about 50% of the internal Google code work.

The task force focuses on long-cycle coding tasks, such as writing new software from scratch, which require the model to read through multiple files, understand user intent, and are currently the most challenging part of AI coding tools. The training corpus is also being adjusted: Google has begun training models on its proprietary codebase because of significant differences between internal and public code, where general coding models do not perform well on internal projects. These internally trained models cannot be publicly released but can, in turn, help iterate on versions that can be made public.

Internally, Google has set up an internal coding tool usage leaderboard called Jetski; some teams outside of DeepMind have started organizing mandatory AI training. In the memo, Brin requested that every Gemini engineer must use internal agents when performing complex multi-step tasks.

The longer-term goal is the "AI takeoff" mentioned by Brin, which involves AI that can self-improve. He has repeatedly told employees that enhancing coding capabilities is key to reaching this milestone; with AI that can do math and run experiments, theoretically automating the work of AI researchers and engineers on a large scale is possible. OpenAI already has similar internal tools to help researchers generate experimental code more quickly.