Are Top AI Talent Leaving Google to Apply AGI and ASI for Giant Leaps in Science

by

Demis Hassabis is changing to a new role as Chair of Google DeepMind & Chief Scientist of Alphabet. This will allow him to focus on long-term strategy, and accelerating scientific breakthroughs, including leaning into my work at Isomorphic to help cure disease.

koraykv will be stepping up to lead Google DeepMind as SVP, alongside @joshwoodward and the exec team.

At roughly the same time, several Google heavy hitters have left to found Discovery Loop (a public benefit corporation). Jeff Dean (Google’s longtime chief scientist / Gemini technical co-lead, 27-year veteran, foundational to search infrastructure, Brain, TensorFlow, TPUs, etc.), Sanjay Ghemawat (his longtime systems collaborator), Oriol Vinyals (DeepMind research VP and Gemini technical lead), and Quoc Le (Google Brain co-founder, AutoML etc.) are leaving to build systems that automate the full experimental loop of machine learning, science, and engineering research—proposing experiments, running them at scale, evaluating, iterating. They’ll start by applying it to ML itself. Alphabet is a founding investor + Cloud partner; other backers include Radical Ventures, Khosla, etc. It’s framed as a supportive spin-out rather than a hostile raid.

Alphabet’s stock reacted with a drop. At the same time, Koray isn’t random bench—he’s a core DeepMind builder from the early days—and the company is emphasizing continuity, Gemini progress (including mentions of upcoming models), the Gemini app’s scale, Gemma downloads, and Google’s full-stack advantages (models + compute + products). Demis and Sundar Pichai’s notes stress that AGI feels close and that getting the next phase right matters a lot.

Chamath believes that AGI arriving means that applying AGI to win on Science is the right choice. Chamath posted two days earlier that recent lab results show AI has entered recursive self-improvement (the singularity loop), so AGI is effectively at hand and progress will accelerate rapidly. In that context, Hassabis shifting from operational CEO to Chair/chief scientist focused on long-term AGI strategy, plus key engineers spinning out for independent research, is a logical next step beyond pure model development.

Google has structural advantages that are hard to replicate—proprietary data at planetary scale (Search + YouTube + Maps + Gmail/Workspace signals), custom TPUs at massive scale, deep product distribution to billions of users, and cash flow that funds ~$200B-scale annual AI/infra capex without constant fundraising. Early foundational contributions (Transformers lineage via Google Brain/DeepMind, AlphaFold-class science work, etc.) were critical. Yet as of August 2026, pure frontier model leaderboards (BenchAlign, LMArena Elo, Artificial Analysis Intelligence Index, SWE-bench variants, agentic coding) are led primarily by Anthropic’s Claude family (Mythos 5 / Fable 5 ), with OpenAI’s GPT-5.6 Sol / GPT-5.5 close behind, Kimi K3 (Moonshot’s 2.8T open-weight MoE) punching into the high flagship band on efficiency and some coding/knowledge work, Grok 4.5 competitive in the next tier, and Gemini models (3.x / 3.6 Flash / Pro variants) solid on multimodal, long-context, and cost/speed but frequently mid-pack on pure reasoning/coding/agentic benchmarks.

Google is missing sustained “ass-kicking” urgency on the evidence of the last 12–18 months. High-profile talent departures (including the Jeff Dean / Sanjay Ghemawat / Oriol Vinyals / Quoc Le cluster to Discovery Loop, earlier moves of key AlphaFold/Gemini contributors to Anthropic/OpenAI/Meta), delayed or staggered flagship releases and a cultural feel of highly paid talent without must win commitment. There is anntensity differential versus Dario led anthropic, Altman led OpenAI, Elon-led SpacexAI (real-time X data + Tesla physical systems ambition), or the constrained-but-hungry Chinese teams (DeepSeek V4 series and Kimi K3) is noticeable.