I can't think of great interventions for ensuring third-party model access.

A safety researcher at LessWrong argues that external AI safety auditors and researchers are severely hampered by their limited access to cutting-edge AI models. While AI labs give their employees the latest tools, third parties working on safety often lag behind by two to sixty times in capability—during the exact months when that research matters most. Author Cleo Nardo sketches potential interventions: direct advocacy to sympathetic lab employees, framing the issue as a productivity question, and building internal support. But Nardo doesn't think any solution is bulletproof, mostly because they'll likely be overtaken by changes in the strategic landscape. The core argument remains: if outsiders are to spend billions making AI development safer, they'll need near-parity access to the models insiders use daily. According to LessWrong, that's not where we stand today.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton