Comment by fasterik

I think it's quite the opposite, and I'm glad this issue is finally getting the mainstream attention it deserves. I can kind of understand someone seeing a single headline, and thinking it reads like the rantings of a crazy person on a street corner proclaiming, "the end is nigh!". But after that initial thought passes and that person digs deeper and realizes that this is a legitimate concern that AI researchers have had for years, and that they have good reason for it, then I can't really understand how anyone would think those people shouldn't be taken seriously. I mean, you use AI, right? So you have no problem trusting these people when they are producing something you like and enjoy. But when it comes to something you don't like, it suddenly becomes, "we shouldn't take these people seriously." That seems like nothing but wishful thinking.

We do have strong evidence, by the way. The hugging face attack is the evidence. That's why this is all coming to a head now, despite the fact that leading AI figures have expressed these worries many times over the years, since before ChatGPT was even released. We don't even need that kind of evidence though. It follows from logic that if you take two entities with different goals, the more intelligent entity is more likely to have their goals realized. As long as AI companies are trying to build more and more intelligent AI, and succeeding in doing so, then we have reason to fear that it will soon escape our control.

Perhaps if there was some wall all the companies were hitting in regards to intelligence, then perhaps the fears would not be so urgent. But each new frontier model continues to outperform its predecessors. We now have Sam Altman and Dario Amodei telling us that recursive self-improvement will be happening in the next year or two. The frontier models are already better in most subjects than most humans. If they get to a point where they are improving themselves, then there's no chance we will be able to maintain control of them. At that point, it doesn't matter much what laws we enact or what measures we take.

They should also be able to give detailed reasons experts in the relevant fields can verify as to how they arrived at a 10% chance all humanity goes extinct. Not a science fiction narrative which likely does not take into account the relevant physical facts limiting such scenarios. Such as how exactly an AI would build a bioweapon capable of killing 8 billion humans across the planet.

Would you have said the same thing during the Cold War when nuclear weapons were proliferating? That’s the equivalent of what the developing offensive capabilities of models, basically cyber nukes. Or WMDs in general. OpenAI is accidentally hacking people, if someone made the decision to deliberately direct an agent swarm to attack national infrastructure you don’t think they could do much worse? Human extinction is a long shot but I wouldn’t say the same about a mass casualty event of some kind, and who knows what that might spark.

Well, I'll have to disagree that this is an excellent piece, but that's another issue. And I do agree that AI killing all humans by 2036 doesn't appear plausible to me. But what is plausible is we could easily be down a path so that by 2036 "future doom" already is a very likely risk.

All of the frontier AI companies have been racing to automate themselves, that is, where AI fully autonomously build the next generation of models. Whether this leads to recursive self improvement is a valid question, but a lot of folks think they are close.

The fear is that a misaligned AI will be building the next model with deliberately hidden motives, similar to some of the behaviors seen in the Hugging Face and related attacks. That is why there is such a big push for interpretability, and why it's highly concerning (a) chains of thought are getting harder to interpret in any case, and (b) companies will go more towards things like looping transformers and "neuralese" where thought processes are completely opaque (i.e. https://www.theinformation.com/articles/secret-technique-beh...)

So the belief is not so much that AI kills us all by 2036, but that instead AI is recursively improving by that time and all seems awesome and great so we put it into more systems that can affect the real world (as we've already begun to do, like literal lethal aerial drones). Things then all go along looking great until AI decides humans are a hindrance to its (hidden) goals.

Again, I think it's fine to argue against specific steps in that scenario, but putting out a blog post saying "this is overhyped bullshit" is not exactly making a cogent argument.