[revel]

The real problem with AI, as I see it, is that we are not ready to give the average person the capabilities they will soon have access to. The knowledge necessary to build nuclear weapons has been readily available for decades, but the constraints of the physical world put practical limits on who can do what. We are also able to limit who can do what by requiring formal education and licenses for all kinds of professions. This is no longer going to be strictly true; or at least not in the same way. If we impose no restrictions, literally everyone on the planet will eventually be able to do anything, provided they have the resources.

The fact is that AIs are more likely to be accomplices than masterminds, at least for the foreseeable future. What we are afraid of is that they will end up being just as terrible and flawed as we are, but it's more likely that they will do what they do today: what we ask them to do. The greater risk is therefore from malicious internal users rather than from the technology itself.

Perhaps more to the point, there is no effective way to limit the output of an LLM or place restraints on how they work. I think it's foolish to even attempt that -- they just don't work that way. The better way to regulate this kind of technology is to focus on the people. Some data classification and licensing program seems a lot easier to implement than the road we currently seem to be going down; which is either no regulation or insanely restrictive regulation.

[avgcorrection]

The planet has been ruined (climate change) by elite-domination since the Industrial Revolution. Those elites have almost annhilated us with nuclear weapons. And here you are navel-gazing about Joe Beergut being “capable” of doing… unclear, no examples are even given.

[erksa]

> We are also able to limit who can do what by requiring formal education and licenses for all kinds of professions.

I agree with your take for the most part, however cautiously more optimistic about this. Removing the barrier of entry for most thing will lead to more people be able to contribute in professions that they otherwise would not have the know how to do.

This will lead to more good things than bad things I think, but we seemingly only focus on the bad extremes and assuming everything else is the same.

Modern Javascript has reduced the barrier of entry to software development. It has had an massive impact on the industry and with an influx of people which leads to the industry seeing more and more creative solutions, as well as some really wild ones.

Maybe I'm terribly naive, but I think we seriously underestimates how many people wants change -- just missing the push to spring into action. For some the moment is about to arrive, for others it will come later.

[revel]

I think you're getting to the heart of what a smart approach to regulation would look like (for me at least). Try as I might, I can't imagine a situation where someone is able to subvert an LLM to come up with a world-ending scenario using its knowledge of onion-based recipes or by painting fingernails. There is no point to regulating every single potential use of an AI: to do so is to cede the future to a less scrupulous nation. We already know the kind of data and functionality that will likely lead to a dystopian hellscape because we already do those things and they are already regulated.

Paradoxically, a heavy-handed approach to regulation could guarantee a bleak future for humanity. We risk increasing the incentives for malicious use at the same time as we make legitimate use prohibitively expensive / regulated out of existance. If the war and crime machine is cheap and easy to monetize, don't be surprised when that's what gets built.

The future could be unimaginably better for everyone, but we won't realize that future unless we get the regulation right.

[Der_Einzige]

Hey, I wrote a whole paper on how to "limit the output" or "place restraints" on an LLM!

https://paperswithcode.com/paper/most-language-models-can-be...