DeepMind talent exodus & the search-vs-scaling bet (Graepel)

status: open, opened 2026-09-07, last moved 2026-09-11. AI google-deepmind

• The thread: DeepMind's founding-tier departures, culminating (so far in our record) in Thore Graepel (AlphaGo/AlphaZero/MuZero, TrueSkill, ex-Altos SVP, DeepMind post-AGI team Nov 2025–Jul 2026) leaving to found an undisclosed venture applying explicit AlphaGo-style search (MuZero lineage: policy/value/learned world model) to robotics — "go back to the architecture and fundamentally redesign it so that it does proper reasoning." Cross-beat: an architecture thesis against the imitation-learning stack behind the humanoid beat.
• 2026-09-07: opened from the Sep 7 ai digest. Reported by Sifted (Sunday); his own site confirms; no name/investors/co-founders yet, first talk at LSE in Nov ("From AlphaGo to AGI"). Exodus context our archive had missed entirely: Silver → Ineffable Intelligence ($1.1B seed, Jan 2026), Jumper → Anthropic (Jun), Shazeer → OpenAI (Jun, ~7% Alphabet drop), Jeff Dean → Discovery Loop + Hassabis → Chairman/ Alphabet Chief Scientist (Aug 5, Kavukcuoglu day-to-day); Zeki Data: DeepMind EMEA hiring share 49% → 18.6% in three years.
• Watch for: venture name/funding, the LSE talk, whether DeepMind publicly counters the exodus narrative, next senior departure.
• 2026-09-11: Discovery Loop is back in market seeking ~$50B (Business Insider, 6:42 PM ET; Reuters carried it) — weeks after the $1B/~$10B round. The departure list is now a funding arms race: Dean's venture pricing at $50B without a product on our record, while consumer-agent Instinct seeks $1B at $10B (The Information) on the same capital. Note the shared structure with the robotics public-markets thread: valuation set by narrative before filings exist.
• 2026-09-08: second architecture bet this week — MIT TR profiles Danijar Hafner (Dreamer/world-models lineage) building a stealth startup on agents that plan ahead. MuZero-style search (Graepel) and learned world models (Hafner) are different answers to the same anti-token-stream thesis. Not yet a DeepMind-departure datapoint — watch whether he is.