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.

Broaden dangerous-capability evaluations and link thresholds to binding mitigations. Evaluations have gaping holes (no AI R&D data), and no procedure connects threshold breaches to deployment decisions, making thresholds effectively non-binding.