Toby Ord: AI-built AI risks inheriting misalignment across model generations
tobyordoxford · x · 2026-09-14
Oxford researcher Toby Ord argues there are reasons to worry about recursive self-improvement beyond mere speedup of AI development: when an AI model builds the next AI model, any misalignment in one could be inherited by all of its successors, compounding the risk across generations.
More from AGI Musings
- Grok evaluates a BrainCode-vs-MindCode thesis: LLMs still differ from brains — wschroll · 2026-09-14
- Critics claim AI-doomer hysteria is funded by Anthropic investors in a tight-knit network — mjdramstead · 2026-09-14
- Bryan Johnson: AI Makes Death Optional, Humanity Is Birthing New Gods — mjdramstead · 2026-09-14
- China state newspaper blasts Anthropic's AI slowdown calls as Cold War tactic — Tolopono · 2026-09-14
- Commenter: 'Pacing the frontier' may be a better investor story than running out of compute — flavioAd · 2026-09-14
- After Gates, SBF, and Aschenbrenner, why 'give me power, I'm more moral' still fails — nptacek · 2026-09-14