Geoffrey Irving: 'Slop proofs' may stump even AI, a risk for RSI-era alignment delegation
ronbodkin · x · 2026-11-20
DeepMind alignment researcher Geoffrey Irving warns that a mountain of 'slop proof' that technically checks out isn't just incomprehensible to humans — AIs may also struggle to build on it until the proof is simplified. He maps this onto AI companies trying to delegate alignment work to not-yet-superintelligent machines during recursive self-improvement: uninterpretable outputs could bottleneck further progress. Ron Bodkin adds that this is why human understanding remains critical as AI agents automate research, especially for safety.
More from AGI Musings
- Former Take-Two CEO Ben Feder calls GTA VI the 'last pre-AI blockbuster' — zephyr_z9 · 2026-11-20
- Hinton: new-language experiment shows LLMs truly understand, not just predict next word — coherence · 2026-11-20
- Tech founders surprisingly receptive to AI risk talks, says researcher — austinc3301 · 2026-11-20
- Fed paper: AI investment is booming, but productivity and labor effects haven't shown up yet — brucemacv · 2026-11-20
- Researchers push back on AI-generated papers: sharing results is part of science — KyleMorgenstein · 2026-11-20
- Beff Jezos slams Anthropic ahead of IPO: 'regulatory capture under the guise of safety' — dansitu · 2026-11-20