The Strike Team is causing the departures it was built to prevent

Google DeepMind expanded its AI Coding Strike Team into midtraining this week, one day after two more senior researchers walked to Anthropic and eight weeks into a run of six senior exits. The interesting thing is not that Google is losing people. It is that the intervention meant to stop the losses is the mechanism generating them, and that the technical diagnosis underneath it, Anthropic's coding lead is a midtraining problem, not a post-training one, is a more consequential admission than the org chart it produced.

The Strike Team, briefly. Formed in April 2026 under Sebastian Borgeaud, with Sergey Brin and Chief Scientist Koray Kavukcuoglu overseeing. The mandate was agentic coding — the multi-step, autonomous workflows where Anthropic's Claude models had opened a visible gap. On 29 June, per The Decoder and AI News Weekly, the team's scope expanded into midtraining: the curated middle layer of the training pipeline that sits between raw pretraining and post-training fine-tuning (RLHF and its cousins). In the same week, Jonas Adler and Alexander Pritzel, a coding contributor and a model training contributor, both left for Anthropic.

The midtraining admission

Start with what the reorg is actually saying. If you believe your competitor's coding advantage is a post-training problem, you fix it with more reinforcement learning from human feedback (RLHF — the fine-tuning stage where human preferences shape model behaviour), better preference data, more iteration. Fast, cheap, contained. You do not expand a strike team into midtraining to do that.

Midtraining is where you shape capability before fine-tuning begins: curated data mixes, synthetic data generation, domain-specific weighting. Intervening there is slow, compute-heavy, and requires re-running expensive stages of the pipeline. You go there when you have concluded that the gap is baked in earlier than fine-tuning can reach.

That is the real news in this week's reporting. Not the org chart. The diagnosis.

There is a fair counter, and it deserves stating cleanly. Midtraining is not a novel idea invented for this moment; curated data mixes have been part of frontier training pipelines since at least the GPT-3 era. It is possible the "expansion" is a rebadging of work already in flight, and the new framing is mostly internal signalling. That reading is available. But rebadging does not usually require a Sergey Brin-overseen strike team and public reporting; the specificity of the intervention suggests something structural is being asserted, not just relabelled.

The reorg is the mechanism

Now the feedback loop. The Information's reporting, echoed by TechTimes, links the Strike Team's expanding compute footprint to tension with Noam Shazeer's pretraining work. Shazeer, who had returned to Google in the 2024 Character.AI deal, left for OpenAI. The causal chain (Strike Team scope creep → compute reallocation → Shazeer exit) is sourced to people familiar rather than to Shazeer himself, and it should be held at that confidence level. But the shape of the loop is worth taking seriously even before the specific link is confirmed.

A strike team is a compute claim as much as an org claim. When a named-executive-overseen unit expands scope, other teams inside the same GPU budget lose allocation. Researchers whose work depends on that allocation notice. Some of them have offers.

The reorganisation intended to stabilise the coding effort is the same reorganisation that keeps producing the departures the effort needs to prevent.

The sequence, laid out, is uncomfortable. Strike Team forms in April to close the Anthropic gap. Compute reallocates. Shazeer, whose pretraining work is one of the affected areas, leaves for OpenAI. The team expands into midtraining to compensate for what has been lost and for the deepening diagnosis of the gap. Adler and Pritzel — exactly the kind of midtraining-adjacent contributors the expanded team would need — walk to Anthropic in the same week the expansion is announced.

Whether or not each individual link holds, the pattern is legible. Structural interventions in research organisations tend to accelerate the churn they are designed to arrest, because the intervention itself changes the internal price of staying.

The floor nobody is talking about

Under all of it sits the compensation math. Alphabet is a public company. It cannot issue pre-IPO equity grants. Anthropic and OpenAI can, and at the senior researcher level the delta is not marginal — it is potentially eight-figure equity upside on a plausible liquidity event within a few years.

This is not a Google-specific problem. It is a structural feature of being a listed hyperscaler competing with pre-IPO labs for the same twenty people. Meta has responded by paying cash at levels that make the equity gap irrelevant. Alphabet has, to date, responded with mandates, oversight, and named executives on strike teams.

Managerial responses to financial-structural problems have a poor track record. The Strike Team can offer a researcher Sergey Brin's attention, a compute allocation, and a mission. Anthropic can offer all three of those things plus a strike price on shares that have compounded meaningfully since the last funding round. The comparison is not close.

None of this means Google is doomed at recruiting. It means the retention lever the Strike Team represents, elevated status, focused mandate, executive proximity, is working against a headwind it cannot itself neutralise. Every reorg that does not address the equity floor is fighting the wrong war.

The counter-signal: they still shipped

Here is the fact that complicates the narrative. On 30 June — the day after the Strike Team story broke, the day after Adler and Pritzel were reported walking — Google shipped Gemini 3.5 Flash and a preview of Gemini Omni Flash. The API changelog describes Omni Flash as multimodal with improved coding and reasoning.

An organisation losing six senior researchers in eight weeks still moved product on schedule. That matters. The "Google is losing" frame is a talent story; the shipping cadence is a capability story; the two are not the same metric, and they can diverge for longer than the talent narrative usually assumes.

DeepMind is thousands of researchers. Six named exits over eight weeks, from three destinations, is legible as a pattern in reporting but is not obviously abnormal for senior AI research attrition in mid-2026. The clustering may be as much a reporting artefact, journalists notice when three of the departures land at the same lab, as a structural signal.

What I would not conclude from the shipping cadence, though. That the midtraining diagnosis is wrong, or that the equity floor does not matter. Product can ship on the strength of work done six to twelve months ago. The teams that shipped Omni Flash preview are, in significant part, not the teams whose composition next week's launches will depend on. Shipping now is compatible with a weaker bench for what ships in Q1 2027.

What is actually being contested

If the internal diagnosis is right — Anthropic's coding advantage is a midtraining data-curation advantage, not an RLHF advantage — then the competitive layer of the training stack has moved. Not to raw pretraining scale, where Alphabet's compute position is genuinely strong. Not to post-training RL, where startups iterate faster than hyperscalers. To the curated middle, where the moat is a combination of data mixes, synthetic data pipelines, and taste about what a model needs to see and in what proportion during the shaping phase.

This is a more interesting claim than the strike team story that carries it. It suggests the next few years of frontier competition will be fought in a layer that is harder to see from the outside, you cannot benchmark a midtraining data mix from the API, and where the advantages compound quietly. If that reading is right, Google's problem is not that it is losing researchers. It is that it is losing the specific researchers who know how to run the middle layer, at the moment the middle layer becomes the contested one.

What I would watch, then. Not the next headcount story. Two things: whether Gemini's coding-benchmark trajectory in the second half of 2026 shows the discontinuity you would expect from a genuine midtraining intervention landing, and whether Anthropic's next model generation extends the coding gap or plateaus. The first tells you whether the Strike Team's diagnosis translates into results the market can see. The second tells you whether the diagnosis was correct in the first place.

The answer to whether Google is losing is neither yes nor no. It is that Google has correctly identified a technical problem it has structural difficulty solving, has responded with an intervention whose mechanism partially undermines the response, and can still ship product in the meantime. All three things at once. Which is, as these things go, a more honest picture than either side of the narrative is currently selling.

Footnotes

¶