3 ms·
> I very much doubt ChatGPT has any ability to shove me in a pen. An LLM with the ability to call arbitrary APIs, and to write and run code, absolutely has the
by Blahah 3y ago
> I very much doubt ChatGPT has any ability to shove me in a pen.
An LLM with the ability to call arbitrary APIs, and to write and run code, absolutely has the ability to shove you in a pen. Not directly of course, but by incriminating you or in some other way making you the target of humans who can physically shove you in a pen.
Langchain agents (for example) have all the necessary tools already. It doesn't require using words to manipulate you, and doesn't require you to perform real world actions on the behalf of the LLM. Using just words, an LLM can autonomously act in the real world.
- c_crank 3y agoThe LLM with the ability to write infinitely powerful malware to hijack anyone's computer (still unproven) has the ability to commit crimes to try and frame humans so that other humans would then arrest them. This is the sort of thing that would make people more concerned with the arbitrary arrests done by the state than the mischievous AI hacker program.
- mrguyorama 3y agoALL AI doomerism rests on this magical claim that "superintelligence (which is inevitable for some reason) means infinite power". It's not even trying to wrap it up in reasonable philosophy or an actual argument, they just claim it as a given and try to rush you past it before you can argue.
- ben_w 3y agoHow is "idiot uses AI to design virus, requests lab prints virus genome, has virus delivered, pandemic happens" any of those criticisms? AI doom doesn't require more intelligence than whatever the dumbest way to go is. Super-intelligent AIs just have the extra problem that, if they exist, you can't out-smart them. But you don't need them for AI to be a problem.
- mrguyorama 3y agoWhat stops that person now?
- ben_w 3y agoChatGPT being made to not do that. RHLF? Filtering? I don't know how. Other AI being harder for idiots to use. Gene printing labs looking out for the smallpox genome, or so I hear. But I don't know how they do that, so I can't say if I trust it.
- Blahah 3y agoRemarkably, almost nothing. In my comment above I wasn't saying that I think it's likely that chatgpt would have a person imprisoned, just that it's technically achievable. There are paid services for swatting. People do that to each other. An LLM could also do it with just a few lines of code. I don't think it's doomerism to say that people have made it possible to harm one another by sending messages and calling APIs, so an LLM could do it too. Humans are the problem, not AI.
- ben_w 3y agoWe don't disagree much, FWIW; for me: > Humans are the problem, not AI. is > Humans are the problem, as well as AI. And in no small part because we make buggy AI, and that seems likely to still be true as capabilities go up. But we're still a bunch of monkeys at the keyboards, almost literally and not just metaphorically.