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

[Coldwater]

Link

https://www.newscientist.com/article/2516885-ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations/

[RussBLib]

1. And combine that with....

…autonomous weapons without the need for human intervention, and we may indeed have a (human) extinction event, with insane motherfuckers in charge.

[BootinUp]

2. It's sad (to me) that anyone with reasonable

Intelligence thinks “AI” should even be consulted.

[sdfernando]

3. So it appears the 80s were prescient...

they must be using the WOPR!

[Layzeebeaver]

4. Paywalled

Even not reading the article, as a user of AI, this smacks entirely about the scenario prompts and about actual rules (the ones humans operate under)

When I was a kid, we sometimes played war games that included nuclear weapons options, it was a race to get them AND use them. Because there usually were no rules that indicated usage restrictions. You were free to okay as you wanted.

What’s missing is how the LLMs compete when give precise nuclear use protocols vs when they are not.

I feel the clickbait pseudoscience “study” force is strong.

To add a bit of nostalgia, back in the 80’s Yaquinto published a board game titled “ultimatum”. The objective was to win either by investing in soft power and gaining control of global regions OR investing hard power (nuclear weapon systems). At some point when you felt you were losing the soft power side of things you could decide to “push the button” - where the soft power portion of the game stopped and was scrapped. What remained was a nuclear exchange aimed at cities, launch sites, military bases, etc. the one with the most population remaining was the “winner” - very few games ended in soft power mode.

TLDR; it’s not just AI that can run amok.