Arrow Forum 2026 keynote, Germany (~40 min). The main-stage version of the bubble-pop argument. Stress-tests the assumptions behind the next five years of work: AI capabilities keep climbing, but frontier access is now a government-and-bubble variable, so the only guaranteed floor is Sonnet-class open-weight models (GLM-5.2). Diffusion is slow, but forward-deployed engineers close the gap — and Palantir, OpenAI, Anthropic, AWS, and Microsoft are pouring billions into them; the move is to do the FDE's job in-house (reach the invisible 80% and make it visible). Walks the seven-stage roadmap, the three worker types, the per-worker token ladder (100K to 10B/day), and an audit of every EUC primitive translated into its cognitive-era successor. Reconstructed from slides; not recorded.
Discussion about this post
No posts