Investor Ramez Naam: AI self-improvement loop is 5-10x too weak for takeoff
jessi_cata · x · 2026-10-11
Ramez Naam, founder of Planetary VC, argues with data that AI's self-improvement loop is far too weak to trigger an intelligence explosion: even with no humans in the loop, it is 5-10x too weak to sustain itself. Citing OpenAI's own numbers, researchers used 124x more tokens per person and produced 7x more code, yet ran only 1.6x more experiments — a severe mismatch between compute growth and output.
He contends the labs' own data contradicts what their leaders claim, and discusses what a truly self-sustaining improvement loop would actually require. The original poster notes this is one of the most important open questions right now and welcomes the numbers-based analysis, even though confident conclusions remain hard.
More from AGI Musings
- Miles Brundage: 'Superintelligence' as a term won't help public adoption — it could backfire — Miles_Brundage · 2026-10-11
- David Sacks rips AI alignment field: 'spectacularly unsuccessful' with nothing to show in 10 years — AndyMasley · 2026-10-11
- Terence Tao's 27-slide deck: math is entering an era of proof abundance — ns123abc · 2026-10-11
- Engineer argues AI isn't special: classic systems-engineering controls still apply — AlexTensor · 2026-10-11
- Satya Nadella says assume all AI models are 'compromised' and calls for verifiable, contained AI systems — The Verge AI · 2026-10-11
- Mathematician Lior Pachter publicly rebukes Terence Tao's view of pure vs applied math — lpachter · 2026-10-11