Google DeepMind's sharpest week of talent losses in years — six named researchers dispersing to Meta, OpenAI, and Anthropic since February — has a structural explanation that goes beyond competitive poaching or equity upside. The lab made a decision about what kind of AI it wanted to build, and the researchers who disagreed voted with their feet.

The most consequential departure may also be the least reported. Denny Zhou, who founded Google Brain's reasoning research team and spent eight years building the technical foundations that made large language models capable of multi-step thinking, had already been at Meta for roughly four months when news of Noam Shazeer's jump to OpenAI landed on June 18, 2026. There was no announcement. No farewell post. Zhou's move surfaced only through a quiet update to his LinkedIn profile. His title: Research Scientist at Meta Superintelligence Lab.

The same week, Jonas Adler and Alexander Pritzel — two senior researchers identified with both Gemini and AlphaFold — confirmed plans to join Anthropic, extending to four the number of researchers leaving for the Claude maker within days. Nobel laureate John Jumper had announced his own move to Anthropic on June 19. Then, on June 25, Dawn Song, one of the most-cited researchers in computer security, announced she was joining Meta Superintelligence Lab as Vice President of AI Research — bringing the core founding team of her AI safety startup, Virtue AI, with her.

TechCrunch confirmed the Adler and Pritzel departures on June 24.

Who Left Google DeepMind and Where Did They Go

The six departures span a five-month window and three rival organizations.

To Meta Superintelligence Lab: Denny Zhou, who founded and led the reasoning research group within Google Brain — the team that produced the techniques now embedded in every major AI model. Also Dawn Song, the UC Berkeley professor and longtime AI security researcher who co-founded Virtue AI in 2024; her co-founders Bo Li and Sanmi Koyejo are joining Meta as well.

To Anthropic: John Jumper, a DeepMind vice president who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for AlphaFold, the AI system that predicted the structure of more than 200 million proteins and transformed structural biology. Jumper led the AlphaFold team starting just six months after finishing his PhD. Jonas Adler, who led Google's AI coding work inside DeepMind, and Alexander Pritzel, who worked on pretraining — the initial phase in which models learn from massive datasets — are also heading to Anthropic. Adler and Pritzel both contributed to AlphaFold research alongside Jumper.

To OpenAI: Noam Shazeer, a co-author of the landmark 2017 paper "Attention Is All You Need" and one of two co-leads of the Gemini project. Google had spent approximately $2.7 billion to effectively reacquire Shazeer — through a licensing deal with his company Character.AI — in 2024. He left less than two years later.

Denny Zhou and the Reasoning Legacy Google Quietly Walked Away From

Zhou's departure is different from the others in kind, not just timing.

Shazeer and Jumper were prominent enough that their departures triggered a single-day decline of roughly 5 percent in Alphabet shares, wiping out approximately $225 billion in market capitalization. Zhou attracted no such headlines. But among the researchers who understand what his work actually produced, his move to Meta may carry the deepest technical significance.

Zhou joined Google Brain in 2017 and founded the lab's reasoning research group. Over eight years, his team developed three methods that are now standard components in how large language models process difficult problems.

Chain-of-Thought prompting works by including intermediate reasoning steps in the examples shown to a model. Rather than asking it to jump directly to an answer, the prompt demonstrates how a human would think through a problem step by step. The 2022 paper showed that this simple change, applied to PaLM 540B with just eight examples, produced state-of-the-art accuracy on the GSM8K benchmark of math word problems — outperforming fine-tuned models trained on far larger labeled datasets.

Self-Consistency extends that insight by treating the model's own outputs as a population. Instead of accepting the first answer a model generates — which chain-of-thought makes more reliable but still imperfect — the method samples multiple reasoning paths and selects the answer that appears most frequently. Complex reasoning problems often admit several valid solution routes; the correct answer emerges as the consensus.

Least-to-Most Prompting decomposes a problem into a sequence of sub-problems, solves each in order, and uses each answer as context for the next. It addresses a specific failure mode in earlier prompting: models that could answer each component question individually but failed when the components required sequential dependency.

These three methods underpin modern AI reasoning pipelines in some form at every major lab. Zhou himself was referred to internally within Google's Gemini organization as the "king of reasoning." Google Scholar records more than 128,000 citations to his work. He is now building the next version of that capability inside Meta.

What Chain-of-Thought, Self-Consistency, and Least-to-Most Prompting Actually Do

Understanding why Zhou's departure matters requires understanding what these techniques are and are not.

Chain-of-Thought prompting does not modify a model's weights. It does not retrain anything. It operates entirely at inference time — the moment a user submits a prompt — by structuring the context in a way that elicits a different behavior from the same underlying model. This is what made it so powerful: the techniques revealed that reasoning capacity already existed in large models but was not being activated by direct-answer prompting.

Self-Consistency works because language models are not deterministic in the way calculators are. Given the same question, a model with sampling turned on produces different outputs across runs. These outputs are not random — they cluster around the model's actual knowledge — but some are wrong. Marginalizing across that distribution to find the most consistent answer is statistically more reliable than taking any single output.

Least-to-Most Prompting addresses what researchers call compositional generalization: the ability to solve problems that require combining known steps in new sequences. Prior models could answer sub-questions but failed at the combination. The decomposition strategy produces what Zhou's team described as "easy-to-hard generalization" — a capability that has since been extended in more complex forms across the field.

The researcher who built and refined all three of these methods is now at Meta. Zhou's group at Meta Superintelligence Lab is positioned to extend this work in an organization that has the compute and institutional focus to pursue it.

Meta's Talent Haul: Dawn Song, Denny Zhou, and the AI Safety Bet

On June 25, Dawn Song announced on X that she was joining Meta Superintelligence Lab as Vice President of AI Research.

Song co-founded Virtue AI in 2024 alongside Bo Li and Sanmi Koyejo, both leading figures in AI security research. The startup built automated red teaming tools, runtime guardrails, and AI governance benchmarks — security infrastructure designed to make AI agents safer as they become more autonomous. Li and Song will report to Nat Friedman within Meta's Superintelligence Lab; Koyejo will report to Rob Fergus, who heads Meta's fundamental AI research division.

Song spent 19 years as a professor at UC Berkeley, where her research spanned program analysis, binary security, encrypted search, and AI safety. Her work focused on making AI systems not just capable but trustworthy and verifiable — a set of problems that become more urgent as AI agents operate with less human oversight.

Axios reported the Virtue AI hire on June 25, citing an internal Meta memo.

The combined picture at Meta is notable: a lab that has acquired the researcher most responsible for LLM reasoning methodology and the researcher most responsible for enterprise AI security methodology within the same short window.

"As we ship AI products to billions of people and build increasingly capable agents, keeping those systems safe, reliable, and trustworthy is foundational," the Axios-reviewed memo stated.

The Coding Strike Team and the Midtraining Decision

The internal development that preceded the most visible departures began roughly two months earlier. In April 2026, Google formed what it called an AI Coding Strike Team — a focused group tasked with closing Gemini's gap against Anthropic's Claude and OpenAI's coding tools in agentic programming tasks.

The team was led by Sebastian Borgeaud, a research engineer who had previously run Gemini's pretraining. Sergey Brin, who has returned to an active hands-on role at Google, and DeepMind chief technology officer Koray Kavukcuoglu were also directly involved, Neowin reported on June 25, citing The Information.

The competitive pressure was explicit. Google CFO Anat Ashkenazi had acknowledged that AI currently writes approximately 50 percent of Google's code. Anthropic has stated publicly that Claude writes close to 100 percent of its code. Brin put the gap directly in an internal memo: "To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers" of final code.

By late June, the Strike Team's scope had expanded beyond agent-layer engineering into midtraining — a technically significant escalation.

Midtraining is the phase between pretraining and post-training in the standard pipeline for building large language models. Pretraining exposes a model to a massive, diverse corpus of text. Post-training — which includes supervised fine-tuning and reinforcement learning — adapts the model toward specific behaviors. Midtraining bridges those two phases by introducing higher-quality, domain-specific data — code, mathematics, instruction-formatted examples — into the training curriculum while pretraining is still underway. Research confirms that midtraining is most effective precisely in the domains where Google has the largest competitive gap: code and mathematics.

Unlike a prompt-engineering or interface improvement, a midtraining change reshapes what the model knows about code at the level of its weights — not just how it is asked about it. Google's decision to expand the Strike Team's mandate to midtraining means it is restructuring Gemini's fundamental training curriculum, not just its product surface.

Shortly before Shazeer announced his OpenAI move, computing resources dedicated to one of his projects were reassigned to a London-based DeepMind team, according to two people familiar with the matter cited by Bloomberg. The reallocation was framed internally as a collaboration and pretraining consolidation effort. It was also the proximate moment before Shazeer announced his departure.

The World Model Fracture: What Google Chose Not to Build

The Strike Team's mandate and the departures share a common context: a shift in research priority that many of the departing researchers found incompatible with why they joined DeepMind.

World models — AI systems that build an internal representation of an environment and predict how it changes over time — are considered by many of the field's most prominent researchers to be the most plausible path toward artificial general intelligence. Demis Hassabis, Meta's Yann LeCun, and Yoshua Bengio have all argued that world models are essential to building AI systems capable of genuine reasoning, planning, and causal understanding, as distinct from statistical pattern matching at scale.

Google DeepMind had long pursued this direction alongside its commercial work. According to sources familiar with the lab's internal direction, work on world models has been stalling and may be deprioritized as computing resources concentrate on coding performance. Internal tensions over this shift have deepened.

The internal fracture is an argument about what kind of AI Google is building. The Strike Team's expansion into midtraining is a commitment to optimizing the Transformer paradigm for a specific commercial application — code writing — rather than pursuing architectural research that might transcend that paradigm. The researchers who left are disproportionately those who built and believed in the Transformer's reasoning extensions, the longer-horizon AGI research path, or both.

Semafor spoke with Demis Hassabis at the Cannes Lions Festival on June 23. "There's a lot of talent movement between all the leading labs and we win our fair share of the top talent," Hassabis said. "We have by far the biggest and broadest research bench of any of the labs out there. It's a ferociously competitive market right now, the most ferociously competitive it's ever been in the tech industry."

Does Losing Researchers Hurt Gemini Right Now

Not immediately. The models the departing researchers built remain at Google. The weights are not transferable. The training runs are not undone. DeepMind retains deep research talent, enormous compute infrastructure, and billions of daily users.

Jefferies analyst Brent Thill called the departures "noise" in a note to clients, maintaining a Buy rating on Alphabet with a $445 price target. D.A. Davidson analyst Gil Luria offered a different framing. "The race at the frontier right now appears between Anthropic and OpenAI," he told investors in the days following the Shazeer and Jumper announcements.

The talent flow has a measurable directional quality. Engineers at DeepMind were nearly 11 times more likely to leave for Anthropic than the reverse, according to SignalFire's 2025 State of Talent Report. Anthropic's two-year retention rate leads every frontier lab. OpenAI's forthcoming IPO and Anthropic's $965 billion valuation following its most recent funding round provide pre-IPO equity that publicly traded Google cannot match structurally.

What Google can match is compute, infrastructure, and a user base that remains unrivaled. What it cannot easily replace is the institutional judgment that comes with decades of building foundational systems — the specific knowledge of where the Transformer architecture's real limits are, and what comes after.

Denny Zhou now works for Meta. Noam Shazeer, who knows the answer to that question as well as anyone alive, now works for OpenAI.

Frequently Asked Questions

Why are AI researchers leaving Google DeepMind in 2026?

The departures reflect a convergence of financial and strategic factors. Anthropic and OpenAI are both approaching initial public offerings, offering pre-IPO equity that publicly traded Alphabet cannot match structurally. Inside DeepMind, the expansion of the AI Coding Strike Team — including its recent broadening into midtraining — has shifted resource allocation and research priority toward commercial coding applications. Researchers whose work is aligned with longer-horizon AGI paths, such as world model research, have found that alignment weakened. Computing resources previously dedicated to Noam Shazeer's pretraining work were reallocated to other teams before he announced his departure for OpenAI.

What is the internal strategic conflict at Google DeepMind driving these departures?

The core dispute is between two theories of how to build artificial general intelligence. The Google Coding Strike Team's mandate — including midtraining Gemini on coding tasks — represents a commitment to optimizing the Transformer paradigm for high-value commercial applications. World model research, which Demis Hassabis, Meta's Yann LeCun, and others consider essential to genuine AGI, requires pursuing architectural questions that do not produce near-term commercial payoffs. Researchers who joined DeepMind to work on the world model path are encountering a lab that has shifted its resource allocation toward the coding path. The departures are the result.

What did Denny Zhou actually build at Google, and why does his departure to Meta matter?

Zhou founded Google Brain's reasoning research team and spent eight years developing the techniques that enable large language models to solve multi-step problems. His most important contributions — Chain-of-Thought prompting, Self-Consistency, and Least-to-Most Prompting — are now embedded in the reasoning pipelines of every major AI lab. Chain-of-Thought works by including intermediate reasoning steps in prompts, activating reasoning capabilities already present in large models. Self-Consistency samples multiple reasoning paths and selects the most consistent answer. Least-to-Most Prompting decomposes complex problems into sequential sub-problems. The researcher who built and extended all three is now building the next generation of that capability at Meta.

Does this talent exodus mean Gemini will get worse?

Not in the near term. The models the departing researchers built remain at Google, and their weights are not transferable. Google's compute infrastructure and user base remain formidable. What the departures signal is a longer-horizon question: the organizations now accumulating the researchers who built Gemini's foundational capabilities — its reasoning methods, its pretraining engineers, its science leadership — will be better positioned to outpace Gemini in the next generation. A $2.7 billion reacquisition of Noam Shazeer held him for less than two years. The structural challenge is not resources. It is research priority alignment.