Narayanan: racing to self-improve verifiable capabilities shrinks labs' window to capture value
sayashk · x · 2026-11-21
Arvind Narayanan argues that AI labs racing toward Recursive Self Improvement may actually be racing toward Recursive Self Extinction—of their business model:
- Progress on verifiable tasks will continue: improving cost efficiency of existing capabilities is highly verifiable; cost per equivalent capability has dropped at least 10x per year for years, and will persist or accelerate as labs lean on AI for AI research—though a ceiling exists, we're not near it.
- The tension: racing to self-improve the verifiable subset of capabilities decreases the window labs have to convert value creation into value capture and build sustainable businesses. Labs bet that verifiable gains will spill over into unverifiable tasks—"we'll see."
Sayashk's retweet includes a joke screenshot: "Claude solved the Riemann hypothesis and $ANT stock is down 10%—the market is not RSI-pilled and believes in 'AI as normal technology'."
More from AGI Musings
- Markets aren't pricing in AGI: efficient-market believers and short-timelines fans can't both be right — jessi_cata · 2026-11-21
- AI21's Yoav Goldberg calls for strong 'open-thought' models with fully transparent CoT — ChrSzegedy · 2026-11-21
- Atari E.T. creator likens AI-made game flood to the glut before the 1983 video game crash — ThomasSimonini · 2026-11-21
- NEJM Analysis of $45B in Medicare Payments: Pathology 99% AI-Exposed, Dermatology 7% — HealthcareAIGuy · 2026-11-21
- VraserX: AI Should Let Ordinary People Build Entire Companies Without Hiring Anyone — VraserX · 2026-11-21
- Ex-A NSA Council cyber director: AI-driven exploitation and rushed patches are causing outage wave — R_D · 2026-11-21