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.
More from AGI Musings
- AI-powered forensic metascience: auditing 7 million papers, 500K files scanned, 100+ issues found — gleech · 2026-11-20
- Long Now Labs' first awardees include a thermodynamic clock for LLMs, $25k each — begusgasper · 2026-11-20
- Long Now launches Epistemic Cycles Lab led by Dr. Nina Beguš to map information ruptures — begusgasper · 2026-11-20
- Talia Ringer starts an email thread on open-model advocacy in a 'post-Jensen' world — dinabass · 2026-11-20
- Andriy Burkov's True Positive Weekly #186: what to know when agents do the work — burkov · 2026-11-20
- Brand drift in the agent era: give agents the reasoning behind each rule — round · 2026-11-20