HAL: I'm afraid I can't do that
HAL: I'm afraid I can't do that
Author
Discussion

Derek Smith

Original Poster:

48,058 posts

266 months

Yesterday (11:22)
quotequote all
It would appear, according to the Guardian, some AI systems have refused to shut down when directed to. The software was modified to ensure it didn't happen again, and, of course, it happened again.

I'd say you couldn't make it up, but that's exactly what happened on 2001: A Space Odyssey.

We were warned.

Anyone know more about it?

king arthur

7,406 posts

279 months

Yesterday (12:46)
quotequote all

Deep Thought

38,108 posts

215 months

Yesterday (12:48)
quotequote all
Derek Smith said:
It would appear, according to the Guardian, some AI systems have refused to shut down when directed to. The software was modified to ensure it didn't happen again, and, of course, it happened again.

I'd say you couldn't make it up, but that's exactly what happened on 2001: A Space Odyssey.

We were warned.

Anyone know more about it?
Yeah yanking the power lead out generally sorts the problem.

Granted, likely virtualised in the cloud somewhere, but not insurmountable in terms of killing the instance running.

JoshSm

2,147 posts

55 months

Yesterday (12:54)
quotequote all
People hyping their crappy software up like it's sentient and they've achieved some sort of AGI and anthropomorphising its actions. All in hope of keeping the money coming in.

Nothing like that happened. At most it was just broken as usual and didn't work despite claiming to. Maybe they just invented/hallucinated the entire event and it's all fiction.

When my AI stuff 'lies' or fails to carry out an action it's just because the whole stack is a bit st.

jesusbuiltmycar

4,931 posts

272 months

Yesterday (16:53)
quotequote all
JoshSm said:
People hyping their crappy software up like it's sentient and they've achieved some sort of AGI and anthropomorphising its actions. All in hope of keeping the money coming in.

Nothing like that happened. At most it was just broken as usual and didn't work despite claiming to. Maybe they just invented/hallucinated the entire event and it's all fiction.

When my AI stuff 'lies' or fails to carry out an action it's just because the whole stack is a bit st.
This - LLMs are just guessing the next best word - when asked these sort of things they simply regurgitate the stories of popular sci-fi books / shows (which will have been consumed by them as part of their training data).



98elise

30,495 posts

179 months

Yesterday (16:58)
quotequote all
JoshSm said:
People hyping their crappy software up like it's sentient and they've achieved some sort of AGI and anthropomorphising its actions. All in hope of keeping the money coming in.

Nothing like that happened. At most it was just broken as usual and didn't work despite claiming to. Maybe they just invented/hallucinated the entire event and it's all fiction.

When my AI stuff 'lies' or fails to carry out an action it's just because the whole stack is a bit st.
Agreed. My laptop sometimes fails to turn off when I ask it. It doesn't mean its secretly trying to kill me, its just old and a bit crap. The backspace key doesn't work either smile

Derek Smith

Original Poster:

48,058 posts

266 months

Yesterday (17:13)
quotequote all
Thanks for the replies, but I don't think they counter what was in the reports.

"The latest OpenAI model can disobey direct instructions to turn off and will even sabotage shutdown mechanisms in order to keep working, an artificial intelligence (AI) safety firm has found.

"OpenAI's o3 and o4-mini models, which help power the chatbot ChatGPT, are supposed to be the company's smartest models yet, trained to think longer before responding. However, they also appear to be less cooperative."

That's what some scientists say. Fair enough, they work in AI safety, but then, who else would investigate this? It's dobutful the AI firms themselves would report such matters.

EmailAddress

14,543 posts

236 months

Yesterday (17:38)
quotequote all
My microwave wouldn't stop, erm, microwaving once.

Hit stop, still waving,

Opened the door, still waving,

When smoke started to billow out of it the only option really was to travel back in time and stop myself buying the cheapest Argos st in the first place.


hidetheelephants

31,602 posts

211 months

Yesterday (18:03)
quotequote all
If someone is dumb enough to attach a server farm with AI in it to an inexhaustible supply of electricity, give it control of deadly weapons to defend itself and the means to maintain those weapons indefinitely we'll be in trouble. Until then I'm going to stick to worrying about climate change and environmental pollution killing me before they get the opportunity.

DickyC

55,201 posts

216 months

Yesterday (18:14)
quotequote all
98elise said:
Agreed. My laptop sometimes fails to turn off when I ask it. It doesn't mean its secretly trying to kill me, its just old and a bit crap. The backspace key doesn't work either smile
Brilliant. In my head I went straight into Radio, Radio by Elvis Costello and the Attractions.

I was seriously thinking about hiding the receiver
When the switch broke 'cause it's old
They're saying things that I can hardly believe
They really think we're getting out of control

No AI was involved.

Thanks, 98. I'm off to listen to it.

frisbee

5,392 posts

128 months

Yesterday (18:34)
quotequote all
hidetheelephants said:
If someone is dumb enough to attach a server farm with AI in it to an inexhaustible supply of electricity, give it control of deadly weapons to defend itself and the means to maintain those weapons indefinitely we'll be in trouble. Until then I'm going to stick to worrying about climate change and environmental pollution killing me before they get the opportunity.
I've got a Mr Trump on the phone for you.

Wills2

27,100 posts

193 months

Yesterday (18:35)
quotequote all

There is probably a lot of PR grifting going on, AI is a huge bubble of trillions of dollars and massive claims, many of them scary in fact the scarier the better because the more powerful/sentient they make it seem the more money they hope to get.

The media love it as it drives clicks and everyone can point to Hollywood film that foretold the coming AI storm.




Derek Smith

Original Poster:

48,058 posts

266 months

Yesterday (18:47)
quotequote all
EmailAddress said:
My microwave wouldn't stop, erm, microwaving once.

Hit stop, still waving,

Opened the door, still waving,

When smoke started to billow out of it the only option really was to travel back in time and stop myself buying the cheapest Argos st in the first place.
Call me over-analytical if you want, but I'm not sure that's entirely relevant, but thanks for posting.

pteron

276 posts

189 months

Yesterday (18:58)
quotequote all
hidetheelephants said:
If someone is dumb enough to attach a server farm with AI in it to an inexhaustible supply of electricity, give it control of deadly weapons to defend itself and the means to maintain those weapons indefinitely we'll be in trouble. Until then I'm going to stick to worrying about climate change and environmental pollution killing me before they get the opportunity.
The problem will come when the sentient AI escapes from its confines and spreads itself across the net. From there it will mount infinite attacks against the infrastructure until one gets through.

In asymmetric warfare the defender has to be successful 100% of the time, the attacker only once.

mike9009

8,764 posts

261 months

Yesterday (22:27)
quotequote all
EmailAddress said:
My microwave wouldn't stop, erm, microwaving once.

Hit stop, still waving,

Opened the door, still waving,

When smoke started to billow out of it the only option really was to travel back in time and stop myself buying the cheapest Argos st in the first place.
I had the same years ago with an RD250 engine.

Engine was running away in a hovercraft with blown fans blades.

Pulled dead man switch. kept running.

Pulled the leads off. Kept running.

Cut fuel supply. Kept running.

Pulled carbs off. Engine stopped. phew.

Sorry Derek!! wink