I was somewhat interested in PAI’s recent release of deployment guidelines for new foundation models, despite an awareness that this organization doesn’t seem to consider everyone equal partners.

So I went over to check out the section about generating custom guidelines to fit your scenario. I didn’t select anything crazy, just frontier model as type and open access. And the advice it gave me is basically a slap on the wrist that says more or less, in fancier terms, “don’t do that.”

We recommend providers initially err towards staged rollouts and restricted access to establish confidence in risk management for these systems before considering open availability.

These models may possess unprecedented capabilities and modalities not yet sufficiently tested in use, carrying uncertainties around risks of misuse and societal impacts. Over time, as practices and norms mature, open access may become viable if adequate safeguards are demonstrated.

I’m not convinced this is good guidance that will make sense in all situations. I’m not sure people constantly freaking out about AI models and trying to forcibly apply entirely squishy and undefined “safety” concepts onto everything is going to result in the kinds of technological progress that improves human lives. I guess that is probably a seemingly incongruous opinion for someone to hold with a Trust & Safety background, but here we are. I’m just not seeing the kinds of results rolling out of AI Safety as a field that I intuitively feel are right and useful. I’m seeing instead mostly a lot of hand-wringing and constipation that results in products which seem somehow magically to get shittier over time instead of better. And it’s frustrating af as an end user.

I guess at this point, I’m feeling more and more in the camp of “let ‘er rip!” and let the communities that are served by and users of these models to determine their own roadmaps about safety and affiliated concepts, as they are likely to be more open-ended, flexible, and most likely more innovative than bottling up all those decisions in one organization.

Related writing & research