Why DeepMind Veterans Are Leaving London: The Ongoing AI Talent Drain at Google

According to Bloomberg, leading AI researchers Jonas Adler and Alexander Pritzel announced they will leave Google to join Anthropic, and DeepMind alignment researcher Arthur Conmy also plans the same move. Both Adler and Pritzel were regarded internally as key contributors to the Gemini model—Adler to AI coding work and Pritzel to the model‑training pipeline.

AI coding becomes the primary commercial battleground

The article notes that over the past year AI coding has turned into the main arena for large‑model commercialization. Products such as Claude Code, Codex, Cursor and Devin have pushed large models from chat tools into real‑world productivity scenarios, creating a high‑frequency, high‑demand, and easily quantifiable market. Google offers a plethora of tools—Jules, Antigravity, Firebase Studio, AI Studio, Gemini CLI, and Workspace Gemini—but the abundance of options leaves developers uncertain about which to adopt.

Training pipeline expertise is the hidden gold

Pritzel’s work on the training system, while less publicly celebrated than "Nobel"‑level breakthroughs, is described as the real "hard currency" inside frontier‑model companies. The article enumerates the nuanced decisions that define successful training: selecting and discarding data, timing model scaling, balancing compute ratios, diagnosing loss‑curve issues (data vs. optimization vs. architecture), distinguishing genuine benchmark gains from over‑fitting, and choosing the next pre‑training direction. Such tacit knowledge rarely appears in papers and is difficult to reproduce without insider experience.

From research lab to model factory

DeepMind’s Gemini era marks a transition from a purely academic research institution to a "model factory" that competes directly with OpenAI and Anthropic and integrates model capabilities into Google’s core product ecosystem. In this factory logic, control over compute resources, training schedules, and model roadmaps translates into real decision‑making power.

Geographic shift and cultural impact

Several of the departing figures were linked to DeepMind’s London‑based team. Observers suggest that pre‑training focus is gradually shifting from London to Mountain View, implying not just a change of office location but a deeper reallocation of internal power, resources, and cultural emphasis.

Talent dynamics in a giant versus focused AI firms

Google’s massive scale brings both advantages and burdens: its search, advertising, cloud, Android, YouTube, Workspace, TPU, and DeepMind divisions provide a full‑stack AI capability, yet they also require constant coordination across product lines and resource schedules. In contrast, Anthropic and OpenAI operate as narrowly focused model companies where model iteration is the central mission, making them highly attractive to top researchers who seek faster impact and fewer bureaucratic hurdles.

Implications for Google’s role in the AI race

While Google still possesses unparalleled AI infrastructure—custom chips, cloud services, foundational model teams, and product distribution channels—its recent talent outflow raises the question of whether it will continue to lead the next AI revolution or become a primary talent pipeline for rivals. The article concludes that the ongoing departures give Anthropic another win in the current talent battle.

References: Bloomberg report; TechCrunch article (2026‑06‑24); related X posts.