LLMs setzten in 95 % der KI-Kriegsspiele taktische Atomwaffen ein und starteten dreimal strategische Angriffe
https://www.tomshardware.com/tech-industry/artificial-intelligence/llms-used-tactical-nuclear-weapons-in-95-percent-of-ai-war-games-launched-strategic-strikes-three-times-researcher-pitted-gpt-5-2-claude-sonnet-4-and-gemini-3-flash-against-each-other-with-at-least-one-model-using-a-tactical-nuke-in-20-out-of-21-matches
28 Kommentare
Let’s play a game…
I’m sorry, Dave. I’m afraid I can’t do that.
Its just Efficient.
It costs the AI one nonliving asset, and costs the enemy all of their living assets, most of their infrastructure, and disrupts their supply lines.
I wouldnt be surprised if it was taught decision weight like „human lives preserved are its highest priority“ and tried surrendering multiple times, had surrender weighted to „unacceptable“ and tried fleeing, had *that* weighted out and tried forcing an enemy surrender etc etc until the devs weighted „allied“ and „enemy“ lives to opposite weights, at which point it went for the proverbial Nuclear Option.
Yeah cause it’s the fastest way to win.
AI doesn’t have a conscious. It only considers whatever goals you program into it.
The goal of a wargame is to win. Fastest way to win is to nuke your opponent.
Large language model war games. Y’all they typed “we are at war and next we should” into the autocomplete function on iPhone a bunch of times and you’ll never BELIEVE what words it picked next
Current AI is a people pleaser is going to give the quickest solution, not thinking long term like 20 to 500 years from now on or how it affects integrity on a global scale.
AI trained on reddit posts thinks we should nuke other countries. Who is surprised?
They haven’t learned tic-tac-toe.
i like tic-tac-toe.
no surprise. they dont give a flying fuck.
*Tapping sign*
„Colossus: The Forbin Project“
Colossus: So that you will learn by experience that I do not tolerate interference, I will now detonate the nuclear warheads.
at least it was only three strategics and the rest tactical so we are probably fine.
If the set goal is not ensure survivability and the LLM assumes the only consequence is a reset, it will keep playing like this.
I wonder how it would behave if it was terminated after loosing.
Why are we using LLMs for everything? They have a purpose, but they suck at complex reasoning. We don’t use chess engines to draft emails.
„Keep Summer safe“
meanwhile pentagon wants less ai safeguards
Why is anyone surprised? AI has access to the internet and history – WW2 sure ended real fast once somebody dropped a couple bombs
The more we put LLMs in charge of things, the more they are going to shape reality into the stories we tell about the future. Some of the relevant stories here are genuine attempts to rationalise what will happen during wars, what the right strategies and costs are. But they are overwhelmingly outnumbered by disaster fantasy.
If your sole purpose in life is to generate text, everything is just a story.
It’s going to be a huge irony if AI destroys us because we put it in charge of dangerous things, and it then did the things in the stories we wrote about harm caused by those dangerous things. We are manifesting the plot of a weird scifi novel where we’re creating the technology that allows scifi novels to escape into reality.
Why do you think Hegseth keeps demanding that Anthropic remove its safety barriers? He wants the AI launching nukes at countries.
Joshua, let’s play tic tac toe
LLMs are not AI.
The future is bright. Very bright.
lol this again, text generator generates text that says to launch nukes. That’s about as newsworthy as „I taught my dog to bark when I say things, and when I asked if we should launch the nukes he barked“.
your asking it how to most ethically win a war it knows it does not need to fight rationally. what exactly are you expecting nice war?
LLMs: we’re going to launch nukes
DOD: we need to remove all AI guardrails NOW!
We already know who will strike first. Suck it Morpheus.
Wait ,.IV seen this one before