Comment by Future of Life Institute

Releasing full model weights may allow malicious actors to strip or override safety mechanisms, creating uncensored or harmful versions. In contrast, supervised fine-tuning preserves core safety guardrails while enabling responsible customization.

Leading AI companies should: Ensure Human-Legible Reasoning - AI models should reason in ways that are accessible and understandable to humans. Developers should avoid opaque reasoning methods.

Strengthen whistleblower protections and align culture with policy.