Models & Researchgoogleai researchai developer toolsmidtraining

Google Reorganizes Coding Strike Team Around Midtraining

||By LDS Team
6.7
Relevance Score
Google Reorganizes Coding Strike Team Around Midtraining

For practitioners tracking where coding-model quality actually comes from, Google's latest move is a tell: the company is reorganizing its AI coding strike team to add a dedicated midtraining phase, rather than betting only on better tools and agents. The Information reported that Google DeepMind is expanding the team, formed roughly two months ago to close the gap with Anthropic, to work on midtraining, the stage between broad pretraining and final instruction tuning where a model is exposed to carefully selected data. Prior research suggests midtraining is especially effective for code and math, where models must move from general language ability to structured problem solving. The reorganization, reportedly involving Sergey Brin and DeepMind CTO Koray Kavukcuoglu, follows a string of senior departures to Anthropic and OpenAI. The signal for builders: leading labs increasingly believe coding capability is won in the training pipeline, not just in the agent harness layered on top.

Why it matters

The interesting claim is about where coding ability is created. By moving the strike team toward midtraining, Google is signaling that it does not believe better scaffolding, prompts, or agent loops alone will close its gap with Anthropic on code. The capability has to be built into the model earlier in the pipeline.

What was reported

According to The Information, Google DeepMind is reorganizing the AI coding strike team it created about two months ago, expanding its focus beyond coding tools and agents to include midtraining. Midtraining sits after a model's broad initial pretraining but before the final stages that tune it to follow instructions. In practice it gives engineers another chance to expose the model to carefully selected, domain-specific data, which prior research has found particularly effective for code and mathematics. The reporting names Sergey Brin, now in a hands-on role, and DeepMind CTO Koray Kavukcuoglu as directly involved, and cites an internal memo in which Brin framed the goal as bridging the gap in agentic execution and turning the models into primary developers of final code.

The talent context

The pivot lands amid a run of senior exits. Reporting ties the reorganization to the planned departures of researchers working on Google's coding and training efforts to Anthropic, alongside other high-profile losses to Anthropic and OpenAI. For a strike team chartered to catch a faster-moving rival, losing pipeline expertise while changing strategy is a meaningful execution risk.

One open question

The Information notes the midtraining work reportedly draws on Google's own proprietary codebase, which could complicate releasing those gains in public models. That tension, between training on internal code to win on capability and shipping models broadly, is one practitioners should watch as the second-half 2026 coding race plays out.

  • Google is betting coding capability must be built in training, not just tooling.
  • Midtraining targets code and math via curated data inserted after pretraining.
  • Strategy shift plus senior departures raises real execution risk for the team.

Key Points

  • 1Google DeepMind is reorganizing its AI coding strike team to add a dedicated midtraining phase aimed at improving coding capability.
  • 2Leading labs increasingly believe coding skill is built in the training pipeline, not just in the agent tooling layered on top.
  • 3The pivot, made amid senior departures to rivals, signals Google sees catching Anthropic on code as a training problem.

Scoring Rationale

This reveals how a frontier lab thinks coding capability is actually produced, shifting from agent tooling to training-pipeline interventions, which is directly relevant to anyone evaluating coding models. It matters because it signals where Gemini's coding trajectory may go and reflects intense competitive pressure from Anthropic.

Practice with real Ad Tech data

90 SQL & Python problems · 15 industry datasets

250 free problems · No credit card

See all Ad Tech problems