Arguing third-party model audits would slow frontier labs the most

JoshPurtell · x · 2026-09-13

In a debate with Eric Mitchell over model verification and independence requirements, Josh Purtell argues that a neolab achieving a breakthrough would have to hand its secrets to third parties who would inevitably leak, letting OpenAI and Anthropic copy it by the next morning — killing the neolab.

Mitchell responds that with a reasonable definition of independence, the friction from such audits should scale with model capability, slowing frontier labs the most.

Related event: Cohere CEO Clashes with Ex-OpenAI Researcher over AI Auditing(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →