My thoughts on the regulation of AI advancement

What would slowing down AI actually look like?

Dario Amodei is calling for slower development of the most powerful AI models, and Sam Altman has backed him. I think this is a good idea, But we should be able to see what is actually happening at each company.

Dario’s proposal includes independent evaluators. If I were Dario, I would propose having a clear consequence, if a model fails agreed safety tests, the lab pauses further capability training and wider deployment until outside reviewers verify the fixes. That should apply to internal models too. Delaying a launch while continuing the same work behind closed doors wouldn’t address the problem.

I think a good place to start is with a 90-day trial across participating labs, including shared testing standards, and public summaries of what was delayed and why. Passing a test wouldn’t prove a model is safe, but failing one should be publicly released.

Governments would need to make those requirements binding, so the company taking precautions isn’t simply handing its competitors an advantage. Getting countries to agree would be harder, but common testing and reporting standards seem like a place to start.

I want better AI (as I consider myself an AI poweruser). But I also want “we should slow down” to come with a clear answer about when they actually would.