No, because if you forbid the dissemination of nuclear research and materials then it's very difficult even for highly-resourced organizations (e.g. government of Iran) to produce nuclear devices, but if you would tomorrow forbid the dissemination of all future "AI research" then there are probably a million of people worldwide who would each be able to recreate the "dangerous capabilities" based on what they already have in their heads and current laptops, as long as someone gives them a few million $ worth of commodity (and thus not restrictable) computing hardware.

Nuclear technology is tricky to get working even if you have decent scientists and all the published research about basic principles. For current ML tech, a big point is that the core techniques are really simple, the simple textbook methods work at large scale, you don't necessarily need any special sauce which could be restricted. Yes, there is a lot of engineering work that goes into a full system to scale it efficiently, but all of that can be replicated by any decent bunch of engineers, unlike nuclear weapons.

The cat is out of the bag; you can't put the genie back in the bottle, all the "sensitive knowledge" you'd need to restrict AI is already known by at least a few CS undergrads in every single university in the world.

For nuclear research you could be right, but not so much with bioweapons. Already 20 years ago you could manufacture the polio virus from scratch just using publicly available data. https://www.science.org/content/article/poliovirus-baked-scr...

Those can be regulated because they require a level of physical access to materials. Centrifuges and viral samples and whatnot.

A given AI, once created, is "just" data. It can be sent around the world and run on commodity hardware. Nothing special needed beyond the expertise that created it in the first place.

Think of how relatively easy it is to clamp down on bioweapons, and compare that with our ability to stop malware.

The same is largely true of bio weapons. The difficult but is R&D, but a small group with access to a decent collage lab could recreate some dangerous shit if the people using it knew what to do.

Training new SOTA (state-of-the-art) models does require a large amount of compute resources. So we could at least pause here and not make more powerful models publicly available until we understand the implications of wtf actually just happened over the last 6 months.

Certainly the models and weights that have been released publicly are out in the wild and near impossible to retract. That damage is done.

Yes but those have higher barrier to entry, MAD dynamics and the fact that you can't clone a nuclear weapon off Github.

Can you clone ChatGPT's weights off of GitHub?