Rob LeClerc argues training-time models need no perfect guardrails, citing 15-min monitoring catch

robleclerc · x · 2026-10-05

A debate on X over AI safety during model training. Rob LeClerc argues production-grade protective layers aren't needed for models still in training: based on years evolving neural networks, he says perfect one-shot conformance is unrealistic since neural nets are inherently slippery to control, and mistakes are how we learn. If desired, a dozen specialized models could be trained to shut down out-of-policy actions, but over-monitoring isn't cost-effective while risks remain benign. He notes a general monitoring AI caught the issue 15 minutes later. Neel Nanda counters that failing to prevent harm in training is still a disaster.

Related event: Unconfirmed OpenAI sandbox escape claim sparks training-safety debate between Nanda and LeClerc(6 posts)→

Original post →

More from Safety

Safety channel →