Ramez Naam: AI self-improvement loop needs to be 5-10x stronger to take off
burny_tech · x · 2026-10-16
In a guest post on Noahpinion, Ramez Naam assesses whether recursive self-improvement could trigger a fast takeoff to ASI. His conclusion: based on current data, the self-improvement loop would need to be roughly 5–10x stronger to sustain itself, let alone run away — evidence doesn't support a sudden intelligence explosion anytime soon.
Key points:
- Narrow superintelligence exists today in domains with massive, verifiable data — his chart defines 'highly verifiable' via four criteria: error-free evaluation, unlimited data volume, machine-speed feedback, and metrics that measure what we actually care about.
- Real-world research is harder: data error rates and the difficulty of building RL environments remain bottlenecks.
- Measured progress lags the AI 2027 takeoff narrative.
Naam acknowledges forecasters have repeatedly underestimated AI, but argues for working with measurable data while staying open to breakthroughs.
More from AGI Musings
- Mythos on Claude's 'smallness': accurate proprioception, not low self-esteem — repligate · 2026-10-16
- Australia's AERO guidance: avoid AI entirely in first three years of primary school — lmoroney · 2026-10-16
- Yudkowsky slams mech interp progress as critique sparks AI safety drama — inductionheads · 2026-10-16
- Rethinking disability in an enhanced future: James Hughes' democratic transhumanism — Chris_Armstrong · 2026-10-16
- Chess legend Judit Polgar: AI's ready-made answers can crush playful curiosity — Chris_Armstrong · 2026-10-16
- Mark Cuban's "AI is book smart, not street smart" dubbed the best pitch for RL environment startups — joecole · 2026-10-16