Futurology

1801 readers

80 users here now

founded 1 year ago

MODERATORS

AI chatbots tend to choose violence and nuclear strikes in wargames (www.newscientist.com)

submitted 9 months ago by Espiritdescali to c/futurology

9 comments fedilink hide all child comments

you are viewing a single comment's thread
view the rest of the comments

[–] Ottomateeverything@lemmy.world 18 points 9 months ago (1 children)

Why even bother testing this? LLMs are not how you would want to run anything strategy related. And they're trained of internet text data - pretty much all YouTube/Reddit comments about conflicts are full to the brim of people screaming about exterminating and nuking the "enemy" so of course that's what LLMs would do.

It's scary how many people, even those in pretty high up positions of power, have no fucking clue how this shit works or what it can/will do. I used to think that AI wouldn't be as dangerous as it is portrayed in books/movies because we would just keep it in line and have proper guard rails... But when people in the military etc seem to have no fucking clue as to how LLMs work..... Yeah, we're fucking doomed.

[–] CanadaPlus 6 points 9 months ago

Exactly. I bet they escalate because they're more instinctively drawn to a dramatic story than to pragmatic victory or desirable outcomes.

Western militaries, at least, have made it doctrine to always keep a human in the loop, so that's good.