Kumar McMillan@kumarvibe@mastodon.social

Replying to a post on ⁨federate.social⁩

@mattblaze yes, models are human controlled for sure, but whenever I think about the killing problem I’m not sure how else you could solve it aside from programming ethics into the model. There are just too many ways to kill 😂

You could, of course, limit access to all the killing things but then you’re left with no internet access and so on — too many limitations to be of use, probably 🤷

In any case, every “rogue” agent attack involved a sandbox breakout. Building the sandbox is not easy.

Matt Blaze@mattblaze@federate.social

Replying to @⁨kumarvibe@mastodon.social⁩

@kumarvibe Software correctness and safety isn’t easy. Who told you it was?

But the killing problem can be solved. My word processor can’t kill anyone (Although maybe it could raise my blood pressure by deleting files at an inopportune moment).

Perhaps if you aren’t sure what a piece of software might do, you shouldn’t let it control lethal capabilities?

en
Kumar McMillan@kumarvibe@mastodon.social

Replying to @⁨holdenweb@freeradical.zone⁩

@holdenweb oh hey Steve, good to see you on here. Big respect for all your work in the Python community.

@mattblaze unlike word processors, it would only take a minute for a “rogue” agent to break out of its sandbox and trick Donald Trump into annihilating the planet with nukes. I am not sure how one would fix that problem aside from training agents on ethics.

And, yes, it’s fueled by greed but Pandora’s box is open. There is no going back.

Matt Blaze@mattblaze@federate.social

Replying to @⁨kumarvibe@mastodon.social⁩

@kumarvibe @holdenweb Huh? Don’t give software with undefined behavior, whether it uses “AI” or not, access to capabilities that you don’t want it to misuse. Ever. Period. That’s the trick. (Actually it’s not a trick, it’s just an obvious safety precaution.)

You either put it in a configuration with access to the Internet (or the life support system or whatever) or you don’t. Software doesn’t “escape” containment. It’s that someone failed to contain it in the first place.

Matt Blaze@mattblaze@federate.social

Replying to @⁨mattblaze@federate.social⁩

@kumarvibe @holdenweb You seem to be suggesting that giving random software (or “AI agents” or whatever it’s called this week) with undefined behavior access to dangerous capabilities is somehow necessary for society to be able to function efficiently.

I think that’s ridiculous and indefensible. The fact that something has “AI” pixie dust doesn’t somehow cancel out basic principles of safety engineering.