AI Researchers and Executives Continue to Underestimate the Near-Future Risks of Open Models
Argues that open-weight models undermine the defenses proposed for near-future AI risks, since guardrails can be fine-tuned away, use cannot be monitored, and bad actors have access by default.