Narayanan: labs racing toward Recursive Self-Improvement may hit Recursive Self-Extinction of their business model
random_walker · x · 2026-11-02
Princeton's Arvind Narayanan argues that the race toward Recursive Self-Improvement may actually be a race toward Recursive Self-Extinction — of AI labs' own business models:
- Verifiable tasks keep improving: cost per equivalent capability has dropped at least 10x per year for years, and AI-accelerated AI R&D will sustain or accelerate this far from any ceiling.
- Unverifiable tasks remain hard: AI inherits the same limits (ambiguous success criteria, slow, noisy feedback) that constrain humans.
An accompanying diagram contrasts the slow 'intelligence loop' (deployment → real-world feedback → revenue → compute/talent → better models) with the fast, verifiable 'efficiency loop' (AI does AI R&D → cheaper inference). The paradox: falling prices undercut revenue growth — the more efficiently labs improve themselves, the faster they erode their own monetization.
More from AGI Musings
- Ex-Adobe Firefly platform lead cites $5b AI-influenced ARR in AI jobs spat — ericelliott_ · 2026-11-02
- AI reply bots are now so good even a dev can't tell humans from bots — xiaohu · 2026-11-02
- Revisiting 'Machine Behavior': the 2019 Nature paper that predicted today's AI science — iyadrahwan · 2026-11-02
- Most Teams Use AI Like 100 Clones of One Employee — Adam Smith Explains Why That Fails — sebkrier · 2026-11-02
- Self-described early agent engineering pioneer says he's interviewed for hundreds of roles — ericelliott_ · 2026-11-02
- World Bank webinar on Nov 3 unpacks AI readiness as Africa growth rises to 4.3% — ChinasaTOkolo · 2026-11-02