Further, even if you are right, so what. Is that a reason to just accept bad public policy? That’s like saying, anyone can learn how to make smallpox at home with a basic lab set up so we should just ignore any safety measures.
Not necessarily, but it should probably inform that public policy. I think the problem is no one knows what the public policy should be assuming that scenario is true. Even if you, somehow, regulate away massive GPU cluster training making such future training impossible, existing models are already here. Further already training smaller models for things like images, speech, and other specialties is cheaper than the bigger models.
I agree that we need some regulations like everything else, but it’s not clear to me what the right policy should be. I think the European ai act is a fine start, but it’s clearly not enough nor does it necessarily limits the training portion just the application portion. Not to mention that the requirements there can be summarized into something like “you have to be careful, and show evidence you tried to be careful”.
That sounds reasonable. If applied to OpenAI and their agents multiple times breaking out of bad secured sandboxes, it should be enough.
But limiting the training?
There really is china and they have a different approach I suppose. But it is possible to talk with them.
Does it? To me it seems reasonable for OpenAI to argue they did try to be careful evident by the sandbox, they just made a mistake. Almost every 0day is categorized by something like that. We haven’t had a long history of establishing a negligence charge to security bugs. Could you be sued because you didn’t demonstrate “carefulness” and used Linux which is not written in a memory safe language and has had multiple CVEs before? How complicated should the chain of an exploit be to demonstrate “carefulness” to the courts?
> training
OP was the one suggesting that training could be controlled because massive gpu clusters could be regulated the way a nuclear power plant could. If you assume training costs won’t drop, then it’s feasible I guess. However, unlike a nuclear reactor, the final training result isn’t a radio active material, but rather an ordinary file that anyone can load and use for inference.
We used to run Microsoft Word and other popular applications with 8 MB RAM and it worked fine.
I am working on reducing ram requirements to run models but the weights are already compressed to the edge of the Shannon entropy boundary and might resist further compression.
Computing always gets cheaper and faster over time. We can argue about the exact rate of improvement but the results are inevitable and uncontrollable.
'When disagreeing, reply to the argument instead of calling names. "That is idiotic; 1 + 1 is 2, not 3" can be shortened to "1 + 1 is 2, not 3." '
Also LLM's have something to do with smallpox as a unrestricted LLM will happily guide any wannabe terrorist in how to make them.