The issue is that most "underground" groups do not have access to top researchers or experienced engineers; those tend to be public figures, or people who want to become public figures. Furthermore, complete secrecy requires tightly restricting communication in a way that makes collaboration even within an organization harder; note how terrorist groups have to operate in isolated "cells."
I'm sure that anyone could replicate GPT-4, but creating radical new advances would be very hard to do in secret.
Furthermore, while many such groups (particularly state actors) have plenty of funding, for other groups it will take a great deal of time and effort to raise or steal the necessary cash in secret, potentially making the costs outweigh the the benefits. If you make a profit, you'll also need to launder the resulting profits, which reasonable crypto regulations would make harder.
Ultimately, nobody can entirely stop "underground" AI. If you think AI will bring about the singularity and destroy mankind, laws cannot stop it. But if your worries are more pedestrian, merely reducing the prevalence of such systems is more than enough.
I'm sure that anyone could replicate GPT-4, but creating radical new advances would be very hard to do in secret.
Furthermore, while many such groups (particularly state actors) have plenty of funding, for other groups it will take a great deal of time and effort to raise or steal the necessary cash in secret, potentially making the costs outweigh the the benefits. If you make a profit, you'll also need to launder the resulting profits, which reasonable crypto regulations would make harder.
Ultimately, nobody can entirely stop "underground" AI. If you think AI will bring about the singularity and destroy mankind, laws cannot stop it. But if your worries are more pedestrian, merely reducing the prevalence of such systems is more than enough.