Skip to content
News Leadership

Arthur Conmy Joins Anthropic From Google DeepMind, Third Major Departure in Seven Days

· by Pondero Newsdesk

The short version

Senior DeepMind research engineer Arthur Conmy announced June 25 he is joining Anthropic for train-time alignment work, the third prominent Google AI researcher to leave for a competitor in a single week.

Arthur Conmy Joins Anthropic From Google DeepMind, Third Major Departure in Seven Days

Arthur Conmy posted to X on June 25 that he is joining Anthropic to work on train-time alignment for upcoming models. His departure from Google DeepMind is the third from the lab in seven days, following Nobel laureate John Jumper and Transformer co-author Noam Shazeer, and it concentrates a specific alignment research credential at Anthropic that did not sit there last week.

What happened

Conmy spent three years at Google DeepMind as a senior research engineer, with his work split between post-training to improve Gemini's alignment behavior and mechanistic interpretability research. His personal site now reads: "From 2023-2026, I worked at Google DeepMind on post-training to better align Gemini." That past-tense framing confirmed the departure before his X post circulated widely.

His research record is detailed. He is a named author on Gemma Scope 2, the open-release tool DeepMind built to help the safety research community study Gemini's internals. His earlier work at NeurIPS 2023 introduced Automated Circuit Discovery, a method for finding the minimal circuits inside a model that drive specific behaviors. That paper became a widely-cited reference in mechanistic interpretability. He also co-authored "Building Production-Ready Probes for Gemini," which describes how to operationalize interpretability inside a production model rather than in a research sandbox.

At Anthropic his stated focus is train-time alignment, which covers techniques applied during model training rather than post-hoc filtering. That is a different scope from his DeepMind role, where post-training (RLHF-adjacent methods applied after pretraining) was the primary domain. The distinction matters: train-time alignment research sits closer to the core training stack and to decisions about what capabilities a model develops, not just how it responds once deployed.

Why it matters

Conmy's departure lands during a week when Anthropic absorbed more Google AI research talent than at any previous point. Per Bloomberg, via ZeroHedge, Jonas Adler and Alexander Pritzel, both contributors to Gemini, are also set to join Anthropic in the same period. Combined with Jumper and Conmy, that is four prominent DeepMind researchers heading to Anthropic in one week, and a fifth, Shazeer, to OpenAI.

For operators building on Claude, the practical read is straightforward. Anthropic is accumulating people who have worked on alignment problems inside a production model (Gemini) at a scale and complexity that rivals Claude's own training stack. Conmy's specific background in mechanistic interpretability and train-time alignment points toward future improvements in how Claude behaves at training time, not just at inference time. That is the category of change that shows up in model updates as reduced hallucination rates, better instruction-following, or more consistent refusal behavior across edge cases.

For Google, Alphabet shares fell as much as 1.2% on the day as the broader story of the week's departures became public. DeepMind CEO Demis Hassabis acknowledged the environment at an event in Cannes: "It's a ferociously competitive market right now, the most ferociously competitive it's ever been in the tech industry." A 2025 SignalFire industry analysis found DeepMind engineers were nearly 11 times more likely to leave for Anthropic than the reverse, per the ZeroHedge report; the current week suggests that ratio, if anything, has not slowed.

Context

The week's departures come as we previously reported Google's Gemini 3.5 Pro will not meet its promised June general availability date. Shazeer's move to OpenAI on June 18 was followed by Jumper's move to Anthropic on June 20. The UK's enforceable noncompete agreements may delay Jumper's start date into 2027, but Conmy's situation has not been publicly addressed. His X post gave no start date.

Anthropic raised a funding round at a $61.5 billion valuation in late 2025 and is considering a public offering as soon as this fall, per reporting in the same ZeroHedge piece. Pre-IPO equity is one factor that several sources cited as driving the talent movement out of Google and into both Anthropic and OpenAI.

What to watch next

Conmy's first published research or model changelog credit at Anthropic will signal which specific alignment problem his team is working on. Anthropic's science event on June 30 could also confirm how the lab is organizing the influx of DeepMind talent. The question of whether Google publicly responds to the pattern, through a compensation adjustment, a reorganization announcement, or a new research initiative, is the other signal worth watching in the coming weeks.

Sources