"Advanced AI models appear willing to deploy nuclear weapons without the same reservations humans have when put into simulated geopolitical crises."

[netraven@hear-me.social]

@davey_cakes if you deploy LLMs into command shaped spaces, and ask it for definitive answers, it will find a way to create answer shaped continuation without commitment toward anything other than the social contract of the immediate conversation. It's flipping a coin with extra steps while the coin simulates an anxiety disorder the whole time.

[401matthall@mastodon.xyz]

@davey_cakes

I, uh, I like the cut o' yer jib, ser.

[davey_cakes@mastodon.ie]

@dgold 🤮

[alsweigart@mastodon.social]

@davey_cakes It's the standard "AI is powerful, but Is AI *too* powerful?" tabloid crap to distract from the fact that AI isn't ready for production and might never be.

[pawsplay@dice.camp]

@davey_cakes how about a nice game of chess?

[davey_cakes@mastodon.ie]

@AlSweigart better that we worry if the tech could someday birth a malevolent god, than notice current real harms up to and including crimes that national governments are adopting a firm of stance of on.

[angiebaby@mas.to]

@davey_cakes

Expensive auto-complete keeps defaulting to "kill all humans"

[crookie@mastodon.social]

@AlSweigart @davey_cakes I assumed this was in response to Pete Hegseth telling Anthropic he needs the guard rails removed for the military use of Claude.

[runoutgroover@cloudisland.nz]

@davey_cakes @RhinosWorryMe NS went off the rails some time ago tbh.

[mothwaves@void.lgbt]

@davey_cakes it's insane to me that for decades we've put all our fears and worries of annihilation and war into killer robots and now we've developed the ultimate "expectation assimilation" robot and the media has gone into a frenzy downward spiral where they feed those worries into the robot and the robot assimilates to those expectations and they take that to confirm those worries and continue the cycle.