It would be great to move from this kind of (hugely impressive!) privately-arranged ad-hoc investigation toward a more formal role for third-party organisations. In the short term, USG could name a roster of qualified third party orgs and increasingly involve them in investigating incidents or assessing framework compliance as it becomes necessary. There are a few benefits of formalising, most obviously: - As @BethMayBarnes points out, OpenAI went for this voluntarily - we don’t know if less safety-minded developers would even disclose this kind of incident, and if they did, whether they’d bring in third-party investigators. We got lucky this happened in a lab that reacted like OpenAI did, but let’s not keep relying on luck. - USG lacks capacity to do this kind of investigation in its own - especially the kind of trust and deep embedding required to assess lab safety practices holistically instead of checking boxes. - Obviously third parties are neither a fully fleshed out NTSB nor fully equipped auditors (yet). But we’re grading on a curve here, and we don’t have any better clusters of talent outside of the labs - especially not in most civilian government agencies (perhaps except CAISI, which faces political barriers). - A formalised relationship or deployment of some kind helps third parties get better at this quickly - especially because it helps with the talent bottleneck: being officially recognised and tasked by USG helps communicate to technical talent that the outside work matters in a way that’s different from internal safety research. - Third parties ready to go and familiar to USG can be an effective ‘release valve’ for the next ad-hoc scramble after the next incident: USG will want to do something, and I’d much rather their first impulse is ‘can’t we send in the guys who did the HuggingFace report to check’ than ‘send in the NSA’ or ‘nothing we can do’. If Commerce, Treasury and ONCD want a good ad-hoc solution to this messy situation, leaning on and bolstering third parties seems like an easy win. USG wants to get this right, doesn’t have enough capacity itself, and the third parties exist. One of the few easy wins in AI policy right now.