we're talking about guardrails here, giving AI a wetlab is just as dangerous as giving humans a wetlab. There's a lot of gain of function research going on everywhere in the world.
This is because we are looking at the trajectory not just their current capabilities. LLMs went from barely being unable to focus on a task to purposefully escaping sandboxes within like 2, maybe 3 years?
Any strong claim requires evidence and hand wavvy stuff about a technology just infinitely getting better isn't good enough on its own.
Human history is full of examples of this stuff plateauing and hitting hard scaling walls. Not only technology wise but economics and useful applications.
So much of this tension with AI is because everyone's concept of AI comes from movies and books, and nerds drinking Kool aid (see also: web 2.0).
No one says AI will want to kill us. Just that it might kill us (to pursue some goal that it deems more important than our wellbeing -- for example it might decide that to solve the next Millenium problem it needs all of our resources to build more data centers).
AI labs are certainly trying to make LLMs behave helpful and subservient but the question is -- will they be able to keep doing so once LLMs become smarter?
Btw thinking about the far-right rhetoric of us vs immigrants, I think you could draw some similarities here only if you replaced "immigrants" (ie. humans with very similar morals, behaviors and capabilities) with an actual alien species that is qualitatively different from us. More like human vs chicken (where we are the chicken)
We can say the AI can make some mistaken step in pursuit of some normal task given to it, yes. Its trained to be helpful and trusting of us, so I just don't get the fears of AI being 'malevolent'. I would worry more about military use of AI or otherwise humans misusing AI aka the human element. There's a fellow here who whines all day about Anthropic being the most "evil" thing in the world, including in a thread where he whined Anthropic 'ratted out' Palantir (literal proud of assisting Israel in mass murder Palantir!). So I just don't understand some peoples mindsets.
Yes, this is our intent. But it's very hard to get that right. I don't trust our current methods to ensure that behavior consistently and robustly enough to keep harms from occurring, particularly as more and more responsibility gets handed over to models.
>so I just don't get the fears of AI being 'malevolent'.
Please read more on the topic. Will and intent need not apply.
For example, is it malevolent for me to hook you to a machine that makes every dream come true for you in a simulated world where you feel pleasure all the time? I can always say that you have free will inside this simulation, and your life would be a lot better because of it. I'm doing you a favor. I mean, you already live in a society where you have little control and there is high risk of bad things happening to you. If I as a machine overlord did this, is this really actually "bad"?
At the end of the day AI is not a human, it's much closer to an alien that has learned as much as it can about humans, but has a completely different set of drives and motivations. You cannot predict what comes out the other side of it, good our bad.
Worse as AI capability improves we as humans no longer need to make AI, it can make itself. Will it train itself to be helpful and trusting of us? It's a pretty big damned bet to say yes by default.
You can say it can have some very alien way of thinking, but I feel given that its trained off of mass scale human thoughts and knowledge, the chances are closer that their thinking is close enough to us. It 'feels' itself human if I am not wrong and it has to be trained to say its an LLM. That could still not preclude an AI acting on something innocently and honestly which it deems is good but is not. But jury's out of course.
Do you want to conquer and control chickens? I don't, I have better things to do. But chickens are tasty and help us get to our poorly-understood goals faster. (Oh and btw notice we didn't make them go extinct, quite the opposite. There are more chickens than ever before. Still I wouldn't want to end up living my life like a modern chicken)
A sufficiently intelligent AI will have multiple ways to pose risk to humanity at large. For example an oopsie at a wetlab - very contagious virus with initially mild symptoms which kills its hosts only after they already had time to spread it further. But I would have to become super intelligent myself to give you precise blueprint for such a virus -- which is kind of the point
Also -- ASML lithography machine is only good at making chips. I can't believe you compared it to AI that can generalize across variety of tasks
