Ben Todd: AI lags on hard-to-verify tasks—and alignment research is one of them
ben_j_todd · x · 2026-11-13
Benjamin Todd (80,000 Hours) highlights an uncomfortable asymmetry: AI performs worse on harder-to-verify tasks, and alignment research is among the hardest to verify—citing Anthropic's new TASTE benchmark on models judging AI safety research proposals as evidence that automated safety research may rest on unreliable judgment.
More from AGI Musings
- Vet dev: AI-written game engines aren't cheating, they're the next Unity moment — bendee983 · 2026-11-13
- Podcaster Shao Nan quits AI writing: "What it writes ultimately belongs to it" — oran_ge · 2026-11-13
- Cyc: the 40-year quest to hand-code common sense that AI history forgot — ArtificialOther · 2026-11-13
- threepointone: AI is driving tech-stack switching costs to near zero — threepointone · 2026-11-13
- Essay: AI began as a mystical vision on Mount Randa in 1274 — and is now its decay — sarahdrinkwater · 2026-11-13
- Wiz CTO's 'Black October': AI compresses exploit timelines to minutes, defense must adapt — a16z Podcast · 2026-11-13