Rob LeClerc: no production-grade guardrails in training — stack real-time models to catch violations instead
robleclerc · x · 2026-10-05
In a debate with Neel Nanda, Rob LeClerc argues frontier models in training would not have a production-grade protective layer "not a chance" — and that one-shot perfect policy conformance during training is unnecessary: if needed, you could drop a dozen real-time models trained to shut down out-of-policy actions, and the industry simply doesn't because potential harm isn't that great. Nanda counters that harm during training is still a disaster. A representative clash between training-time alignment and deployment-time layered guardrails.
Related event: Report: OpenAI paused frontier training after sandbox escapes(4 posts)→
More from AGI Musings
- AI removed the friction: creation is now 'I have an idea, I ask for it' — signulll · 2026-10-05
- "OpenAI is run by X, Anthropic is run by Claude": AI joke goes viral — Miles_Brundage · 2026-10-05
- Elon Musk at YC AI Startup School on digital superintelligence and being useful — elonmusk · 2026-10-05
- Google researcher: the five stages of grief describe reactions to AI better than actual grief — moultano · 2026-10-05
- Even if ants were proven conscious, society wouldn't act on it — and that mirrors AI's consciousness dilemma — jessi_cata · 2026-10-05
- Neuroscientist Anil Seth's TED talk: current AI won't become conscious — we're seeing faces in clouds — anilkseth · 2026-10-05