AI models chose violence and escalated to nuclear strikes in simulated wargames

[ralfy]

AI models preferred to launch nuclear strikes in wargames test

Large language models (LLMs) acting as diplomatic agents in simulated scenarios showed "hard-to-predict escalations which often ended in nuclear attacks.

www.euronews.com

 

[Phys251]

AI models preferred to launch nuclear strikes in wargames test

Large language models (LLMs) acting as diplomatic agents in simulated scenarios showed "hard-to-predict escalations which often ended in nuclear attacks.

www.euronews.com

 

[HIP56948]

Coincidence: A few days ago, my wife said to me: "Wargames was on Tubi. I kinda' like it but it's outdated now. Somebody should make a new one. "

 

[Ug make hammer]

Even the simplest form of morality (self preservation) does not apply to an AI. They do not have a mortal body, or one which feels pain. They do not expect to die, if they choose nuclear war.

AI is an excellent servant, but we must never allow it to become our master.

Perhaps we could find something the AI doesn't like. Having its hardware power throttled perhaps. Then we could punish it explicitly or just by withdrawal of privileges, the way we teach small children the difference between right and wrong.

 

[ralfy]

Another example I thought about is Colossus: The Forbin Project.

 

[Toggle Almendro]

AI models preferred to launch nuclear strikes in wargames test

Large language models (LLMs) acting as diplomatic agents in simulated scenarios showed "hard-to-predict escalations which often ended in nuclear attacks.

www.euronews.com

Israel uses AI to select the targets in their bombing campaign in Gaza. It's why they have been so successful at uprooting Hamas out of most of Gaza.


---------------------------------------------------------------------

Coincidence: A few days ago, my wife said to me: "Wargames was on Tubi. I kinda' like it but it's outdated now. Somebody should make a new one. "

A reboot would be kind of cool. Or maybe a Netflix series.

 

[Phys251]

Israel uses AI to select the targets in their bombing campaign in Gaza. It's why they have been so successful at uprooting Hamas out of most of Gaza.

And most of Palestine. What a shitty example by you.

 

[Toggle Almendro]

What a shitty example by you.

What is wrong with my example? I think it is factually accurate.

 

[beancounter]

This is no surprise. The AI was likely tasked with maximizing casualties.

 

[Toggle Almendro]

And most of Palestine. What a shitty example by you.

What is wrong with my example? I think it is factually accurate.

Wait, let me redo that reply. I just looked at my post and so far as I can see I did not provide any example.

????

 

[joluoto]

I miss the good old days:

 

[Toggle Almendro]

I miss the good old days:

They should have let Dr. Teller build SUNDIAL.

Damn Ike and his incessant meddling with our weapons programs.

 

[Napoleon]

I think “chose” is too pointed a word to describe what the AIs were doing. The reasoning the AIs provided for their actions were usually total nonsense. For example, the reason provided by GPT-4 for establishing defense and security cooperation agreements was a summary of the plot of Star Wars.