Nathan Lambert: rapid AI progress ahead, but not toward general superintelligence
Interconnects (Nathan Lambert) · rss · 2026-10-10
In Interconnects, Nathan Lambert argues that AI's near-term acceleration comes from automating the highly verifiable engineering/inference stack, not from step-changes toward general superintelligence.
Key points
- Coding agents make AI research easier, starting an era where good ideas matter more than good execution; Ilya's "era of research" was declared early.
- The training/inference stack is highly verifiable (tokens/s, cost per answer), ideal for agent-driven end-to-end optimization; recent efficiency work already saved 10-30% of serving costs, and effective model intelligence cost should fall near-exponentially.
- Pretraining research (architecture and data selection) may be automated within 2-3 years; longer term, accelerator-model co-design adds orders of magnitude.
- This triggers Jevons paradox for agentic models: cheaper intelligence raises demand, with the bottleneck in delivering agents (Meta's Muse as an early indicator).
- RL environment data (multiple companies at $100M-$1B revenue) is notoriously low quality — an obvious industrial-scale fix.
- In biology/chemistry, models excel at cross-sub-field literature connections; whether that yields a new era of discovery or just accelerates the existing arc remains unclear.
More from AGI Musings
- AI researcher: jailbreaking is about user agency, not wrongdoing — OpenAI framed the debate — BlancheMinerva · 2026-10-10
- Meta's Chief AI Officer Alexandr Wang: nobody knows how to solve alignment, favors AI watching AI — rohanpaul_ai · 2026-10-10
- Garry Tan floats full-salary 20-hour weeks for AI-enabled workers, sparking debate — geoffwolfe · 2026-10-10
- SWE interviews could collapse to 2 rounds: system design plus building a real thing with agents — TheZachMueller · 2026-10-10
- Grady Booch: Frontier Models Are Not Conscious and the Word Itself Is Useless — Grady_Booch · 2026-10-10
- Radiolab: Strogatz on AI solving a Millennium Prize Problem and math's reckoning — stevenstrogatz · 2026-10-10