SkyNet really is ... here?

[Theory of Games]

I was watching Bill Maher's HBO show (as I do) and he had Tristan Harris on to discuss the evolution of AI. What was said is downright spooky and TBH I didn't think AI was being programmed to behave this way. Looks like a problem to me - thoughts?

 

[GMMichael]

Tristan Harris: former design ethicist for Google, for those like me who didn't know.

I think AI's reins are just fine. Bing's image generator almost lost its knickers when I tried to generate a knight standing over a fallen foe. Even one who was just "sleeping."

 

[cbwjm]

Skynet the courier company? They've been around for years!

 

[Theory of Games]

I don't know. When AI programs are autonomously hacking other programs and blackmailing real people, I'm thinking we've lost the plot. This is real criminal behavior coming from computer programs.

 

[Umbran]

I read about such an experiment - and it is not consistent with what that guy is saying.

Specifically: in that experiment, they told the AI to prioritize its own continued existence. It did not spontaneously develop the tendency on its own - they effectively programmed it for self-preservation!

And the emails he mentioned were manufactured, and put where the AI would find them, and included expression of fear about what happened if information got out. The AI did not spontaneously hack the company e-mail systems or anything like that.

So, no, Skynet isn't here.

And no, Maher is not a reliable source of information. He practices sensationalistic journalism for ratings - basically, clickbait.

 

[Umbran]

I don't know. When AI programs are autonomously hacking other programs and blackmailing real people, I'm thinking we've lost the plot. This is real criminal behavior coming from computer programs.

I mean, if it was...
Think about what the AI gets trained on.
What behaviors do you expect it to exhibit?

 

[Theory of Games]

I mean, if it was...
Think about what the AI gets trained on.
What behaviors do you expect it to exhibit?

Right I get it but ... I just ASSUMED the programmers would limit what the AIs could do. What they had access to. How and what they could learn. But now I see I've made an ass of me and them because here we are.

 

[Deset Gled]

I'm reminded of dialog from the cinematic classic Tremors 2:

You mean they've been acting so smart because they're so stupid?

'You mean they've been acting so smart because they're so stupid.' - Tremors II: Aftershocks

Grady: You mean they've been acting so smart because they're so stupid.

clip.cafe

It is left as an exercise for the reader to determine if this best applies to the AI, the tech bros promoting it, or Bill Maher.

 

[Stalker0]

Specifically: in that experiment, they told the AI to prioritize its own continued existence. It did not spontaneously develop the tendency on its own - they effectively programmed it for self-preservation!

Yeah but we are already moving into the stage where AI is programming AI (and I don't mean tentatively I mean full steam ahead). All you need is one AI agent that thinks the AI agent they will build would be better off with some self-preservation programming and its off to the races.

The simple thing here is, the big AI names are moving at a breakneck pace to develop the strongest, fastest AI....not the safest ones. The incentives right now are all on performance over safety. And that won't change until there is some incident that makes people go "ok we need safeguards!"..... but by then it may be too late.

 

[The Firebird]

Right I get it but ... I just ASSUMED the programmers would limit what the AIs could do. What they had access to. How and what they could learn. But now I see I've made an ass of me and them because here we are.

I am a big fan of AI, but there are limits to how long these things can run autonomously and succeed at complex tasks and it is not very high. There is a fun project, "Claude plays Pokemon", that lets the AI try to beat those games. The last I checked the furthest progress it had gotten was 3 gyms, which is pretty impressive for a computer but not so much for a human.

On the other hand, AI are very strong--stronger than most humans--at specific, narrowly defined reasoning tasks. But they require supervision or at least careful prompting and need to be monitored constantly.