DeepMind's 2026 Talent Exodus: Who Left, Where They Went, and the Reasoning Techniques They Built

John Jumper → Anthropic (June 19, 2026)

Jumper, who shared the 2024 Nobel Prize in Chemistry for co-creating AlphaFold, announced on X that he was leaving after "nearly 9 years" . His final stretch at Google had been focused on AI coding tools, not the scientific work that won him the Nobel . His exit, combined with Shazeer's, erased roughly $270 billion from Alphabet's market cap in a single trading session .

Jonas Adler → Anthropic (announced ~June 24, 2026)

Adler was viewed internally as a key contributor to Gemini and Google's AI coding effort. People familiar with the move cited a desire to work at a more agile AI startup .

Alexander Pritzel → Anthropic (announced ~June 24, 2026)

Pritzel worked on Gemini pretraining and AlphaFold. His departure was reported alongside Adler's, with the same context of seeking faster-paced environments .

Denny Zhou → Meta (reported June 26, 2026)

Zhou, known as DeepMind's "Reasoning King" and founder of Google Brain's reasoning research group, left quietly. He made no public farewell — the move was reported by HTX after he updated his LinkedIn to show he had already been working at Meta for four months . No explanation was given by Zhou or Meta.

The Broader Context: What Drove the Exodus?

Multiple sources describe a broader talent drain at DeepMind throughout 2026, driven by three factors :

• Aggressive poaching by well-funded rivals Anthropic and OpenAI, both preparing for market debuts.
• Internal frustration with compute allocation — the sense that Google's bureaucracy was slowing research velocity compared to startups .
• Perception of bureaucracy: researchers felt they could move faster at smaller, more focused labs.

The Three Reasoning Techniques by Denny Zhou

Denny Zhou and his collaborators developed three foundational prompting techniques that have become central to how large language models reason. They form a progressive stack, each building on the last.

1. Chain-of-Thought (CoT) Prompting

What it does: Instead of prompting an LLM to output an answer directly (input → output), CoT prompts the model to generate a sequence of intermediate natural-language reasoning steps before arriving at the final answer (input → reasoning steps → output).

Key benefit: Dramatically improves performance on arithmetic, commonsense, and symbolic reasoning tasks. It also enables interpretability — you can read the model's "thought process." Combined with large models like PaLM-540B, CoT achieved state-of-the-art results using as little as 0.1% of annotated examples .

2. Self-Consistency

What it does: A decoding strategy that improves upon CoT. Instead of taking a single chain of reasoning, the model generates multiple independent CoT reasoning paths (via sampling with higher temperature), then selects the most consistent answer across all paths by majority vote .

Key benefit: Mitigates the variance of a single reasoning chain. A single CoT path might be wrong due to one flawed step; self-consistency averages over diversity and is significantly more robust on math and reasoning benchmarks . Denny Zhou has emphasized that self-consistency should not be interpreted superficially as mere majority voting — it is an empirical implementation of marginalization .

3. Least-to-Most Prompting

What it does: A two-stage prompting strategy designed for problems harder than the examples in the prompt. First, the model decomposes the original hard problem into a list of simpler sub-problems. Then, it sequentially solves those sub-problems, using the answer of each earlier sub-problem as context for the next one .

Key benefit: Enables easy-to-hard generalization — the model can solve problems that are strictly harder than any example it was shown. It has been demonstrated on symbolic manipulation, compositional generalization benchmarks (like SCAN and CFQ), and math reasoning tasks . Zhou describes it as "Planning + Reasoning" .

Summary

Five of the six researchers listed are confirmed to have left DeepMind for Meta, OpenAI, or Anthropic in June 2026, driven by competitor poaching, compute allocation disputes, and a desire for faster-paced environments. Dawn Song's departure could not be verified and does not belong to this wave. Zhou's three prompting techniques — Chain-of-Thought, Self-Consistency, and Least-to-Most — form a progressive stack: CoT adds reasoning steps, Self-Consistency adds voting across multiple reasoning paths, and Least-to-Most adds problem decomposition and sequential solving for harder problems.