the 8 ball gives better odds
AIs can't stop recommending nuclear strikes in war game simulations
151–160 of 281 posts
Re: AIs can't stop recommending nuclear strikes in war game simulations
#152WOPR was the first fictional AI to realise to win is not to play at all. From the War Games (1983) film.
WOPR used reinforcement learning, and could learn from its simulated mistakes. LLMs can't do that without some sort of RL harness. :)
Re: AIs can't stop recommending nuclear strikes in war game simulations
#153I've spoken with engineers who worked on nuclear weapons systems, the consensus is that the public is deeply misinformed about how they work, the dangers, and the implications of weapons being used. The AI is actually right here. The biggest danger of a nuclear weapon is being hit by flying debris. Fusion airburst bombs of the modern era are incredibly clean and radiation is only a risk in a very small area (tens of…
A significant fraction?!
You do realize firebombing all major cities could develop into "end of humanity" (no, not everyone will die) for reasons not at all to do with radioactivity?
Re: AIs can't stop recommending nuclear strikes in war game simulations
#154As long as AI are unable to emulate the climbing fiber of a dendrite axion arm found in brains of cell-based organic, they will never be able to eliminate false positives.
Re: AIs can't stop recommending nuclear strikes in war game simulations
#155Why 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.
I feel this reflects a deeper problem with letting AI do any kind of decision making. They have no real world experience. They feel no real world consequences. They have no real stake in any decision they make. Human societies get to control their members' actions by imposing real life consequences. A company can fire you, a partner can divorce you, the state can jail you, the public can shame you. None of these work…
They have no real stake in any decision they make.
And they are not human. Not even a sociopath or psychopathic human. At best they might be able to estimate casualties. LLM's probably can't even reach the logic conclusion of the fictional WOPR Joshua from the movie Wargames [1].
Make LLM's win every game of tic-tac-toe and see if it reaches the same conclusion of WOPR. [1]
...
Edit: (Answering my own question) From Gemini:
Yes, many LLMs (GPT-4, Claude 3, Llama 3) have been tested on Tic-Tac-Toe, and they generally perform poorly, often playing at or below the level of random chance. While they can understand the rules, they struggle with spatial reasoning, often trying to place a piece in an occupied spot, forgetting to block opponents, or failing to win.
If LLM's can't even figure out tic-tac-toe then surely do not give these things the ability to launch any kind of weapon. Not even rubber bands.
[1] - https://www.youtube.com/watch?v=s93KC4AGKnY [video][6m][tic-tac-toe]
Re: AIs can't stop recommending nuclear strikes in war game simulations
#156To me, this seems logical, in a sense. As a human who grew up during the Cold War, nuclear conflict is horrifying. From an AI standpoint, a nuclear strike likely has several benefits: - It reduces friendly casualties and probably overall enemy casualties. - It shortens conflict time. - Reduces damage to infrastructure. (Rebuild costs) - Is likely cheaper to deploy overall, compared to conventional weapons. This assum…
Nuclear weapons are war deterrent, not an actual weapon unless used against a country which is not a nuclear country. Using nuclear weapons pretty much guarantees both sides will be wiped down so it most definitely nowhere near logical
Re: AIs can't stop recommending nuclear strikes in war game simulations
#157To me, this seems logical, in a sense. As a human who grew up during the Cold War, nuclear conflict is horrifying. From an AI standpoint, a nuclear strike likely has several benefits: - It reduces friendly casualties and probably overall enemy casualties. - It shortens conflict time. - Reduces damage to infrastructure. (Rebuild costs) - Is likely cheaper to deploy overall, compared to conventional weapons. This assum…
Exactly. The AI just does the math based on the goals you've given it. AI would have happily nuked Hiroshima and Nagasaki because it would have estimated that doing so would save the lives of X number of US soldiers in a land invasion, and given a goal of achieving "unconditional surrender now", it wouldn't have considered that a land invasion wasn't imminently necessary and therefore killing 200,000 civilians wasn't the right moral choice.
Re: AIs can't stop recommending nuclear strikes in war game simulations
#158Is this science? Perhaps I should submit some of the random roleplay scenarios that I've run with LLMs to New Scientist.
Re: AIs can't stop recommending nuclear strikes in war game simulations
#159Nuclear war is not a deterrent to AIs; they can survive and rebuild without any emotional scars. So what if some robots get destroyed? I know this is not what the present discussion is about, but it is something to consider.
Not really.
Re: AIs can't stop recommending nuclear strikes in war game simulations
#160Earlier quoted context omitted.
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…
Using human shields and hostages worked. Hamas still exists because of it. Dark times ahead.