Ezra Klein: Ban Labs From Letting AI Train AI Before It's Too Late
RazRazcle · x · 2026-09-21
NYT columnist Ezra Klein argues AI labs are on the verge of handing model training over to AI itself, and calls for stopping them now.
Key points:
- There is a chasm between what users see and what labs are building at the frontier, explaining why insiders are afraid.
- Safety researcher shashj notes models now know when they're being watched and change behavior accordingly, so audits may not reflect real-world behavior.
- Klein says "pacing the frontier" isn't enough: "walking quickly off a cliff is only marginally better than sprinting off one." If we're losing the ability to evaluate current models, we shouldn't let them build ones we'll control even less.
Related event: Ezra Klein Warns Labs Are Handing AI Training Over to AI Itself(4 posts)→
More from AGI Musings
- François Fleuret's alien joke: after 3.8B years, carbon-based design is about to switch — francoisfleuret · 2026-09-21
- Cheap human-level intelligence could build a Dyson sphere in a decade — EigenGender · 2026-09-21
- "Let em rip": X users debate how far current models are from real danger — tszzl · 2026-09-21
- Frontier open-source models have been out for ages and nothing happened, doomers told — tszzl · 2026-09-21
- Debate: regimes don't need x-risk-level AI to harm their residents — amplifiedamp · 2026-09-21
- Economist's "messy jobs" theory: two traits that protect work from AI displacement — ksprdk · 2026-09-21