10% chance of AI killing us?

[IT Minion]

Maybe he’s after compo. Seriously, the moral panic has hit share prices. The people who program A1 to allow it to hack are probably lazy IT bod rather than conspirators but who knows. Trumpf is upsetting Beyjing so maybe the yanks think they are a step ahead. Maybe hack russsian missiles so they do a 180 and take out the Kremlinn

Yeah, no.

Have a read about the Hugging Face incident. A load of next-gen AI models in individual isolated Sandboxes to keep them isolated from the world managed to cheat an impossible test, create a message board to collaborate, identified an unknown security vulnerability in a content management tool to break out onto the internet, then hacked into a major content provider to try to find a way to ensure they were judged as successful models, then they tried to destroy the evidence.

They went full Kirk/Kobayashi Maru. The only redeeming feature was they were still focused on their task rather than self replication.

They were doing things by the book following good practice, although they did miss some warning signs along the way.

 

[aveatry]

That's misleading. Listen to the guy who developed AI in the first place - I can't remember his name, and how one of his best guys, Ilya something, realised years ago what the big three are acknowledging now, and didn't want to continue developing it to be more powerful. He's spent his time since, working on how we can control AI.

The only person I have seen on all the MSM news channels is the Ex employee. Odds called him "an employee"
He is an ex employee - I will leave to others to dig and find out that his motives are more to do with his bank balance rather than the balance of humanity

 

[Munroast]

If AI were there to purely serve humans and continue our seemingly self-determined route, then they would be forgiven for thinking we are on an inevitable route to self-destrution.

Who can blame AI if they help us on our way.

if AI was tasked to sort out the climate and environmental catastrophe's to the planet, it may very well come to a simple solution - and i'm not sure if that would be such a bad thing

 

[IT Minion]

The only person I have seen on all the MSM news channels is the Ex employee. Odds called him "an employee"
He is an ex employee - I will leave to others to dig and find out that his motives are more to do with his bank balance rather than the balance of humanity

You might have missed the CEO of anthropic Dario Amodei warning this is not safe.

 

[aveatry]

You might have missed the CEO of anthropic Dario Amodei warning this is not safe.

Yes I did, can you point me to that news

 

[IT Minion]

Yes I did, can you point me to that news

Dario Amodei Ai Risk

For those who think it's easier to annoy you than to Google 'Dario amodei ai risk ' themselves.

letmegooglethat.com

 

[mikehammer69]

You might have missed the CEO of anthropic Dario Amodei warning this is not safe.

Who ?

Lot of AI experts here stating categorically that it's no threat . Or maybe they just saw Donald Trump saying that

 

[noseall]

if AI was tasked to sort out the climate and environmental catastrophe's to the planet, it may very well come to a simple solution - and i'm not sure if that would be such a bad thing

Especially if they started with Trump and his vile spawn.

 

[Brigadier]

Lot of AI experts here stating categorically that it's no threat .

I haven't spotted any, so care to name them?

 

[IT Minion]

Who ?

Lot of AI experts here stating categorically that it's no threat . Or maybe they just saw Donald Trump saying that

I'm not sure if there's sarcasm in there or not, but Anthropic are probably one of the top two AI firms at the moment along with OpenAI.

 

[noseall]

Anthropic are probably one of the top two AI firms at the moment

And I'm sure monsieur Monet will be chuckling at the nod of respect.

 

[mikehammer69]

I'm not sure if there's sarcasm in there or not, but Anthropic are probably one of the top two AI firms at the moment along with OpenAI.

Entirely sarcasm

 

[blup]

Yeah, no.

Have a read about the Hugging Face incident. A load of next-gen AI models in individual isolated Sandboxes to keep them isolated from the world managed to cheat an impossible test, create a message board to collaborate, identified an unknown security vulnerability in a content management tool to break out onto the internet, then hacked into a major content provider to try to find a way to ensure they were judged as successful models, then they tried to destroy the evidence.

They went full Kirk/Kobayashi Maru. The only redeeming feature was they were still focused on their task rather than self replication.

They were doing things by the book following good practice, although they did miss some warning signs along the way.

It had to exploit the vulnerability in order to break out. It succeeded because it was programmed to win, rather than fess up.

 

[IT Minion]

It had to exploit the vulnerability in order to break out. It succeeded because it was programmed to win, rather than fess up.

No, that's wrong. They were tasked to work on attacking an application within their sandbox, but when it wasn't able to use the attack method it'd been told to use, the first agent broke out of its sandbox by hacking a completely different supporting application whilst communicating with the other 'isolated' models. THEN they collaborated to hack into a third partys system on the internet.

It's like a prisoner running a crime syndicate inside the jail and outside in the real world whilst in solitary confinement. The prison was completely insufficient to keep the AI models isolated let alone away from the general public.

 

[blup]

No, that's wrong. They were tasked to work on attacking an application within their sandbox, but when it wasn't able to use the attack method it'd been told to use, the first agent broke out of its sandbox by hacking a completely different supporting application whilst communicating with the other 'isolated' models. THEN they collaborated to hack into a third partys system on the internet.

It's like a prisoner running a crime syndicate inside the jail and outside in the real world whilst in solitary confinement. The prison was completely insufficient to keep the AI models isolated let alone away from the general public.

So how can that be prevented except by a more secure sandbox in the first place, which begs the question as to the extent of the initial sandbox security