Eine Klasse von KI-Modellen, die als beängstigend leistungsfähig angepriesen wurden, hat der Regierung offenbar zu große Angst gemacht und ist nun außer Gefecht gesetzt
https://gizmodo.com/anthropics-models-hyped-as-scarily-powerful-apparently-scared-the-government-too-much-and-now-theyre-disabled-2000770341
21 Kommentare
Smells like bullshit to gin up investors and other assorted rubes.
didnt a version of one of those get stolen and is out in the wild anyway?
its also hard to tell if thats legitimately a threat or if this is a corrupt play at juicing the stock price when the IPO happens
There is a new space race right now between China and the USA. Whoever wins gets access to a lethal set of cyber tools. While most people think this type of stuff is benign and used for spying, it’s actually a lot more malicious because all nation states basically have Trojan backdoors and malware embedded into each other’s critical civilian infrastructure.
This means with a push of a button, you can disable another country’s power grids, water treatment plants, oil/gas pipelines, etc.
PR bullshit
It only happened because they refused to work on army projects so Donnie dumpo decided to screw them with export controls
Amazon CEO jassy was the one to snitch them out, he told Bessant that his team jailbroke it and they banned it.
>One of the tougher guardrails, designed to quietly punish users who abused the model, was even deemed ill-conceived, prompting Anthropic to apologize.
I think this undersells just how bad this was and in the context of this article misrepresents the motivation by conflating it with safety.
Rather than being some vague abuse criteria it was specifically detecting use of the model for development of LLMs, distributed training systems etc. Not just distillation attempts but explicitly targeting any potential competitors using it as a tool.
The punishment was also silently sabotaging their work. Once they flipped it to fail fast and loud it turns out this was tripping on all sorts of stuff from basic research to open source projects and ML work unrelated to LLMs.
Imagine if a business like Microsoft decided they’d quietly introduce bugs for anybody doing OS development or building an IDE in Visual Studio.
It’s weird that the gov wants the ability to tell Ai companies they can’t release a new model
And at the same time they want no regulations on ai
“This tool is too scary and powerful for release.
It could be yours for the low low introductory price of $50 a month”
I wonder how much these AIs know about the Epstein files and other revealing data files, since they’ve scanned so much of the Internet files. Did AI reveal something that scared them?
Doesnt the US have an „invention secrecy act“ which hides this level of technological breakthroughs from being public in the first place?
Edit: https://en.wikipedia.org/wiki/Invention_Secrecy_Act
Task failed successfully!
I’m worried about restrictions and government meddling with open weighted models. „You can’t have this on your personal machine because its too dangerous and you could do something bad!“… This isn’t something we should allow to be encouraged.
We’re going to blame AI for the US government being disabled? As if Trump was actually functional before AI scared him?
It’s so super special they can’t tell you about it, it’s that super special.
Once the check clears they’ll be back.
I don’t like AI but it’s actually insane the federal government without any warrants or congress or anything can just tell private companies what to do.
The restriction only prevents Anthropic from allowing its use by noncitizens. A market still exists if Anthropic wants to build the tooling to comply.
How does Anthropic comply with only allowing US citizrns to use its product? Why, ID verification, of course.
Don’t be fooled. This has nothing to do with the PR hype of Anthropic’s models or national security. This is simply yet another prong in the coordinated push for a deanonymized Internet wrapped in a fearmongering excuse.
Anthropic hasn’t taken the bait immediately. It’s unclear whether that’s because they are principled, or because they need time to build the verification infrastructure. My chips are on the latter.
Part of me thinks this is all part of the administration’s vendetta against Anthropic, but Anthropic kind of brought it upon themselves by publicly stating how dangerous mythos is.
Stupid ridiculous, practically false headline.
The actual reason for banning this AI model from public access was the ability to jailbreak them, and then violate export restrictions. „Fear“ played no factor in this decision. The models being „too powerful“ played no factor in this decision , at all.
Whatever.
Sounds like more fucking AI bullshit to me