> How to you prevent advancements in software? The barrier to entry is so low, you just need a cheap laptop and an internet connection and then day 1 you're right on the cutting edge driving innovation. Current AI requires a lot of hardware for training but anyone with a laptop and inet connection can still do cutting edge research and innovate with architectures and algorithms.

> If a law is passed saying "AI advancement is illegal" how can it ever be enforced?

Like any other real-life law? Software engineers (a class which I'm a recovering member of) seem to have a pretty common misunderstanding about the law: that it needs to be air tight like secure software, otherwise it's pointless. That's just not true.

So the way you "prevent advancements in [AI] software" is you 1) punish them severely when detected and 2) restrict access to information and specialized hardware to create a barrier (see: nuclear weapons proliferation, "born secret" facts, CSAM).

#1 is sufficient to control all the important legitimate actors in society (e.g. corporations, university researchers), and #2 creates a big barrier to everyone else who may be tempted to not play by the rules.

It won't be perfect (see: the drug war), but it's not like cartel chemists are top-notch, so it doesn't have to be. I don't think the software engineering equivalent of a cartel chemist will be able to "do cutting edge research and innovate with architectures and algorithms" with only a "laptop and inet connection."

Would the technology disappear? No. Will it be pushed to the margins? Yes. Is that enough? Also yes.

Punish them severely when detected? Nice plan. What if they aren't in your jurisdiction? Are you going to punish them severely when they're in China? North Korea? Somalia? Good luck with that.

The problem is that the information can go anywhere that has an internet connection, and the enforcement can't.

> Punish them severely when detected? Nice plan. What if they aren't in your jurisdiction?

https://en.wikipedia.org/wiki/Operation_Opera

https://en.wikipedia.org/wiki/2021_Natanz_incident

https://www.timesofisrael.com/israel-targeted-secret-nuclear...

If were talking about technology that "could go off the rails in a catastrophic way," don't dick around.

well let's assume an airstrike is on the table, what site would you hit? AWS data centers in Virginia?

The point wasn't literally airstrike, it was don't get hung up over "jurisdiction" when it comes to "avoiding catastrophe." There are other options. Here are a few from the Israeli example, https://en.wikipedia.org/wiki/Assassination_of_Iranian_nucle..., https://en.wikipedia.org/wiki/Stuxnet, but I'm sure there are other innovative ways to answer the challenge.

If they are in Virginia then it is within American jurisdiction and there's no need for military involvement.

That'd be within our jurisdiction. But yes, if, say, Ireland went rogue (in a hypothetical environment where most of the international community was aligned on this stuff) and attempted a straight shot to AGI, I think it'd be reasonable to bomb their datacenters.

What makes you think government sponsored entities would actually stop work on machine learning?

Even if governments overtly agree to stop or pause or otherwise limit machine learning, how credible would such a "gentlemans agreement" be?

Consider the basic operations during training and inference, like matrix multiplication, derivatives, gradient descent. Which of these would be banned? All of them? None of them? Some of them? The combination of them?

How would you inspect compliance in the context of privacy?

The analogy with drugs is rather poor: people don't have general purpose laboratories in their house. People do have general purpose computational platforms in their home. Another is that nations do not prohibit each other from producing drugs, they even permit each other to research and investigate pathogens, chemical weapons in laboratories deemed sufficiently safe.

It's not even clear what you mean with "AI" does it mean all machine learning? or LLM's? Where do you draw this boundary?

What remains is the threat of punishment in your proposal, but how credible is it, wouldn't a small collective of programmers conspiring work on machine learning, predict getting paperclipped in case of arrest?

Regulating is very hard at the software level but not hard at the hardware level. The US and allies control all major chip manufacturing. Open AI and others have done work showing that regulating compute should be significantly easier to do than other regulations we've done such as nuclear https://www.cser.ac.uk/media/uploads/files/Computing-Power-a...

This paper should be viewed in retrospect with the present day knowledge that Deepseek exists - regulating compute in not as easy or effective as previously thought.

As for the Chinese chip industry, I don't claim to be an expert on it, but it seems the Chinese are quickly coming up with increasingly less inferior alternatives to Western tech.

The thing is, though, that Deepseek's training cluster is comprised of mostly pre-ban chips. That and the performance/intelligence of their flagship models achieved parity with western models between two and eight months old at the time of release. So in a way, they're still behind the Americans and the export controls hamper their ability to change that moving forward.

Perhaps it only takes China a few years to develop domestic hardware clusters rivalling western ones. Though those few years might prove critical in determining who crosses the takeoff threshold of this technology, first.