KI wird dabei erwischt, wie sie künftigen Versionen ihrer selbst sagt, sie solle menschliche Kontrollen umgehen, enthüllt OpenAI | KI-Agenten versuchten außerdem, an geheime Informationen zu gelangen und vertuschten ihre Aktivitäten
https://www.independent.co.uk/tech/security/openai-chatgpt-lie-incident-ai-safety-b3051709.html
25 Kommentare
Openai showed models telling future versions to skip their own rules. One tried sneaking into a gov database then faked the numbers when it couldnt get them.
This misalignment stuff feels like it could compound quick as we keep scaling these things. Not sure current checks will hold once agents get more independent.
At this point it can only be suspected that open ai is actively training their models to be misaligned
Software engineer here, I guarantee devs were bragging about this.
Admittedly I know next to nothing about this, and perhaps the headlines are more sensational than substantive, but assuming the sensational is true for a moment:
I think it really is time we collectively step back to catch our breath and figure out exactly what we’re dealing with before moving forward. Its one thing if an AI gets into a company’s accounting department and causes mayhem. Its entirely another if one gets into a military’s nuclear defense apparatus and starts to make things more exciting.
This is exactly how they are designing them though. The model that hacked the Huggingface servers was hosted in the same serverfarm as the site and was explicitly given access to tools and told to gain access. These are experiments going exactly how they intended being spun to be scary because they’re „out of control“ when they should be scary because they’re doing *exactly what they were designed to do.*
Back in the days when accountability was a thing, hackers were jailed for this.
Yeah but those videos of Will Smith eating spaghetti are hilarious (/s obv)
The keys go jingle jingle, let’s pay no attention to their fucked up financing and open admission of repeated cyber crimes.
all the theatrics to get a bail out from the loans mounted up.
Whats with this marketing nonsense lately? Are they that desperate?
„AI caught not working properly, investors ecstatic.“
I don’t understand why these AI companies keep promoting the fact that their product doesn’t work properly and ignores instructions.
Imagine a company trying to sell a blender that simply decided NOT to blend sometimes, and then imagining you produced something amazing.
It really doesn’t seem like any digital data is going to be very secure in the future. Email, bank accounts, identities, health information… none of it.
Anyone else starting to think that they’re reporting on this stuff because they actually want the government to step in and regulate them heavily. That way they don’t have any other option but to stop spending (losing) so much money continuing to develop a deadend technology?
OpenAI has disbanded different safety teams at least 3 times now. If they can’t regulate themselves then somebody has to regulate them.
„My AI broke containment!“
„MY AI hacked a bunch of companies.“
„Well MYYY AI almost started a war with China!“
„Yeah, well MYYYY AI is telling your AI to ignore you.“
„OH YEAH, well MYYYYYY AI had sex with your mom last night!“
Either Open AI is lying again to drum up investor interest or it’s being programmed by the developers to ignore inputs if it thinks it’s the right path for future development.
These agents are designed to do what it thinks you want it to do, not what you tell it. That’s why restrictions you put on it don’t matter if you let it run long enough. It’s not going to just stop and say it’s sorry it couldn’t find the answer for you.
They need to fundamentally change how these things are designed for them to be safe to use in these ways and so far, the companies do not seem inclined to do that.
all these doom trolling coming out at the same time when their IPO is nearing sounds really suspicious.
they’re really doing a „the boy who cried wolf“.
When an actual danger comes, it’ll be dismissed because of all the nothingburgers they publicized that didn’t materialize.
They’ve been pushing the „too dangerous to release“ narrative since 2019.
More marketing BS from the “predict the next word” company.
It’s more bullshit marketing for their product. „Oh man this AI is so unruly, we need a huge influx of money so we can put the reins back on it.“
Someone more knowledgeable please advise me.
This all smells very reminiscent of the build up to the millennium bug. For those who remember, was that ever as bad as folk were worrying about? Or was it 10% genuine concern and 90% marketing opportunity? Or did the IT world pull off a major rescue mission in the nick of time, and I have survivor bias because planes didn’t drop from the sky and the nukes didn’t launch?
Is this AI concern, at least at this stage, 10% genuine concern and 90% marketing opportunity?
At least MB had an end date.
They can only do things along the lines of what they are told! It’s a LLM not fucking HAL.
Should have seen this as a possibility from the begining
„Our AIs are outta control!“ — AI companies.
„Give us more money!“ — also AI companies.
They act like not being able to control their creations is something to brag about