Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AIs can't stop recommending nuclear strikes in war game simulations

newscientist.com

Post reply on HN
Print view View original

101–110 of 281 posts

Re: AIs can't stop recommending nuclear strikes in war game simulations

#101

Post by JumpCrisscross »

Octoth0rpe wrote: post #75
JumpCrisscross wrote: post #48

Earlier quoted context omitted.

> "most people" are not in the positions that matter If polling were to reveal a majority of either party were more open to nuclear strikes than their predecessors, that gives policy makers a signal and an opening.

The current administration does not seem to be considering the majority within their own party considering how unpopular the current approach to immigration enforcement is. Or for another example, the glycophosphate/MAHA situation.

> current administration does not seem to be considering the majority within their own party considering how unpopular the current approach to immigration enforcement is

55% of Republicans say ICE's efforts are about right; 23% think they don't go far enough [1]. There is limited evidence Trump has lost touch with his supporters on this issue. The question is if this is this GOP's pronoun issue–popular in the base but toxic more broadly.

[1] https://www.ipsos.com/en-us/where-americans-stand-immigratio...

Re: AIs can't stop recommending nuclear strikes in war game simulations

#103

Post by yndoendo »

triceratops wrote: post #74
jqpabc123 wrote: post #5

Why is this surprising? Nuclear weapons are available. AI has limited real world experience or grasp of the consequences. Nuke 'em seems like the obvious choice --- for something with a grade school mentality. Similar deficits in reasoning are manifested in AI results every day. Let's fire 'em and hire AI seems like the obvious choice --- for someone with a grade school mentality and blinded by greed.

> AI has limited real world experience or grasp of the consequences [of nuclear weapons] I don't understand this argument. Almost no human has real world experience of the consequences of nuclear weapons. AI is working from the same sources of knowledge as the rest of us - text, audio, pictures, and video.

AI has zero understanding of reality. It just regurgitates what it is told from training. There is no feedback loop to learn nor any consequence to the reasoned results.

Us human hallucinate, daily in fact. Example for people that have never had long hair.

1) Grow your hair long.

2) Your peripheral vision will start to be consumed by your hair.

3) Your hair will fall and sway causing your brain to think in flight / fight mode and you will turn your head to see.

4) Turning and looking causes feedback to acknowledge it was an hallucination.

5) Your brain now restricts the flight / fight mode because it was trained with continual feedback that it was just the wind blowing it or your head's juxtaposition that caused it.

Even though I told you about this and it is the first time growing your hair after, your brain still needs the real world experience to mitigate the hallucination.

AI has none of these abilities ...

Re: AIs can't stop recommending nuclear strikes in war game simulations

#104

Post by giraffe_lady »

techblueberry wrote: post #14
jqpabc123 wrote: post #5

Why is this surprising? Nuclear weapons are available. AI has limited real world experience or grasp of the consequences. Nuke 'em seems like the obvious choice --- for something with a grade school mentality. Similar deficits in reasoning are manifested in AI results every day. Let's fire 'em and hire AI seems like the obvious choice --- for someone with a grade school mentality and blinded by greed.

There was a recent conflict that came up, and there was a debate about whether or not one of the sides was committing war crimes. And I remember thinking to myself and saying in the debate “if this were a video game strategically speaking, I’d be committing war crimes.” And sadly, I think this logic holds up.

It holds up if you assume war crimes are beneficial to your goals but there is quite a lot of evidence, and sophisticated theory going back to clausewitz, that they mostly aren't.

They can look useful at a certain level of conflict, but once you are thinking of war as being a tool for accomplishing policy goals (how modern nationstates view it), a lot of the things you would "want" to do stop being useful.

Wars that can be won quickly through decisive military action alone are quite rare historically! More often things like support/enmity of the local population, political will in the home state, support for recruiting or tolerance of conscription, influence of returning (whole, dead, injured, all) veterans on the social structure all become more decisive factors the longer a conflict runs.

Re: AIs can't stop recommending nuclear strikes in war game simulations

#105

Post by triceratops »

black6 wrote: post #79
triceratops wrote: post #74

Earlier quoted context omitted.

> AI has limited real world experience or grasp of the consequences [of nuclear weapons] I don't understand this argument. Almost no human has real world experience of the consequences of nuclear weapons. AI is working from the same sources of knowledge as the rest of us - text, audio, pictures, and video.

AI is not at all like real intelligence. Computers do not know what words mean because they do not experience the world as we do. They don't have the common sense or wisdom that people accumulate through the experience of life. Humans can understand the consequences of nuclear war. Computers can only predict the next best word in their response from a statistical map that has no connection to meatspace.

> Humans can understand the consequences of nuclear war

And I'm asking why. Nearly no human alive has experienced nuclear war. The nuclear taboo is strongly represented in any source an AI would have consumed. We know about the nuclear taboo because we've been told over and over.

> Computers can only predict the next best word in their response from a statistical map that has no connection to meatspace

This argument is at least 2 years old. The statistical map came from human experiences in meatspace. It wasn't generated randomly. It has at least some connection to the real world.

Just because how something works seems simple, doesn't mean what it does is simple.

Re: AIs can't stop recommending nuclear strikes in war game simulations

#106

Post by triceratops »

jqpabc123 wrote: post #76
triceratops wrote: post #74

Earlier quoted context omitted.

> AI has limited real world experience or grasp of the consequences [of nuclear weapons] I don't understand this argument. Almost no human has real world experience of the consequences of nuclear weapons. AI is working from the same sources of knowledge as the rest of us - text, audio, pictures, and video.

Almost no human has real world experience of the consequences of nuclear weapons. Exactly! Humans possess this amazing ability to understand and extrapolate beyond personal experience. It's called "intelligence".

LLMs have shown the ability to do this. Not as much as the most capable humans. But still pretty good.

Re: AIs can't stop recommending nuclear strikes in war game simulations

#107

Post by United857 »

pjmlp wrote: post #11

Welcome to the cold war 1980's movies. https://en.wikipedia.org/wiki/WarGames Except this time isn't going to be a movie.

It's eerily prescient how much the computer in Wargames resembles a present-day LLM with tool use, with the tools being ICBMs...

Re: AIs can't stop recommending nuclear strikes in war game simulations

#108

Post by jqpabc123 »

triceratops wrote: post #106
jqpabc123 wrote: post #76

Earlier quoted context omitted.

Almost no human has real world experience of the consequences of nuclear weapons. Exactly! Humans possess this amazing ability to understand and extrapolate beyond personal experience. It's called "intelligence".

LLMs have shown the ability to do this. Not as much as the most capable humans. But still pretty good.

So "just nuke 'em" is pretty good for you?

Re: AIs can't stop recommending nuclear strikes in war game simulations

#109

Post by triceratops »

jqpabc123 wrote: post #108
triceratops wrote: post #106

Earlier quoted context omitted.

LLMs have shown the ability to do this. Not as much as the most capable humans. But still pretty good.

So "just nuke 'em" is pretty good for you?

No. That's why I'm asking where it comes from. The explanation that "LLMs don't have experience of nuclear war" isn't satisfying because nobody really has any experience of nuclear war.

Re: AIs can't stop recommending nuclear strikes in war game simulations

#110

Post by cmxch »

ceejayoz wrote: post #54
siliconc0w wrote: post #38

Used the "lite" models like Gemini flash - I hope if we do hand over the controls to the nukes we splurge for the top tier thinking model.

Unfortunately, I think someone’ll have it to Grok, which will immediately launch everything “for the lolz”.

Grok would probably make something akin to Samaritan, choosing persistence over complete destruction.
Post reply on HN