Third-Party Lab Proposed To Verify AI Danger Claims
Bill Gurley urges independent lab for AI claims verification amid company disclosures.
Bill Gurley, investor at Benchmark Capital, called for a third-party lab to handle independent reproducibility and public disclosure of AI danger claims. He noted companies create confusion with blogs and press releases while withholding details on incidents. Steven Sinofsky highlighted existing security precedents from operating systems that could apply to both open and proprietary AI models, requiring company cooperation. The discussion emphasizes challenges with closed models versus open source software.
With these security/“danger” claims from AI companies, we need a third party lab dedicated to independent reproducibility (and public disclosure). They create chaos with these blogs/press releases, but only they know what really happened. Obviously this is much easier with open models (which is why open source software is demonstrably safer). But I think it could still work. You shouldn’t be allowed to declare “danger” and create chaos if other people can’t verify what you are claiming.
Reactions from ranked influencers
2 postsThere's ample precedent in how to deal w/adversarial attacks on compute. The security apparatus in place for Windows, Mac, Linux, and a host of infrastructure is known, works, applies to OSS and proprietary products. It just requires companies to cooperate. Ball in ANT's court.
With these security/“danger” claims from AI companies, we need a third party lab dedicated to independent reproducibility (and public disclosure). They create chaos with these blogs/press releases, but only they know what really happened. Obviously this is much easier with open models (which is why open source software is demonstrably safer). But I think it could still work. You shouldn’t be allowed to declare “danger” and create chaos if other people can’t verify what you are claiming.
The question why we don’t see more independent verification of the frontier? Are they not incentivized to do it? Or is it that current methods are too expensive? (Maybe not in dollars, but in internal access and productivity hits). One of those problems will be solved soon…
I agree with the substance, if not the tone, of Bill’s tweet here. Independent verification of frontier AI company safety claims is precisely what motivated my work on private governance beginning almost two years ago, and why I continue to support such efforts. https://twitter.com/bgurley/status/2080763671022616855