5 ms·
Ask HN: How dangerous is Auto-GPT?
I'm talking about this project: https://github.com/Significant-Gravitas/Auto-GPT
It puts LLM's programming capabilities in a feedback loop with the real world in order to achieve arbitrary goals. In my opinion, this has the potential to become a very powerful and therefore dangerous tool.
I fear that people with both good and bad intentions give commands that do large scale harm. How do you see it?
- ttctciyf 3y agoI think the question of criminal liability will be an interesting one for the future of IRL-enabled AI. Prosecuting the software is clearly absurd, but if it "innovates" in response to an ambiguous request, where does the agency lie? Is the requestor open to conspiracy charges? The API hoster.. ? If there are cases where no-one can be found liable for the actions of an insufficiently constrained AI agent, isn't that an open invitation to instigate plausibly deniable AI performed crime? And so on.
- dysoco 3y agoIf you are running something that is considered dangerous knowingly (not saying AI is, just a supposition) and you mishandle it and it causes damage then you should be the one found liable. I don't think researchers are to blame, it's like blaming the flamethrower company.
- ttctciyf 3y agoRight, but "knowingly" is hard to establish in the case of someone making an ambiguous request to an AI agent, and if a company hosts an API that allows the agent to interact with real entities there should be a presumption that (like the manufacturer of the flamethrower) they ensure it will not blow up in the user's face, no?
- gtirloni 3y agoIf I operate a rocket launcher and decide to charge people for the chance to fire a rocket wherever they want, who's liable?
- chatmasta 3y agoI had a good chuckle after finding this in a related thread: https://xkcd.com/416/ https://xkcd.com/416/
- Jennie21 3y ago[dead]
- kleer001 3y agoIt's pretty shit and pretty expensive to learn how useless it is.