Europe is reliant on US digital services. There is a worry that the US government could use this to pressure us. This may be the first instance where Europe’s access is limited (or delayed) because of US government decisions.
Europe is reliant on US digital services. There is a worry that the US government could use this to pressure us. This may be the first instance where Europe’s access is limited (or delayed) because of US government decisions.
No big loss, it’s just another next version frontier model that’s being PR hyped as ‘dangerous’ (again, they do love to present LLMs as an existential threat, to distract from the environmental and societal threats and general scamminess of the industry).
Here’s a reasonably deep analysis of why it’s a nothingburger. TLDR:
It could actually be a good thing if this nothingburger gets some attention and scares some politicians. It could be yet another case which demonstrates that the US don’t want to be a partner of the EU anymore and that the EU must invest in their own solutions.
Broken clocks are right twice a day.
That’s PR hype. Look closer. It comes from some company that claims that it can do the same thing but much cheaper. They’re just trying to make some money before Mythos goes public.
Hmm, care to back that up? After having a poke around looks like an author and pretty bog standard security consultancy, nothing about automated exploit discovery.
The author does a deep dive into Anthropic’s own Project Glasswing announcement and IMO credibly refutes some of their PR Hype numbers. He references the work of ‘AISLE, is an AI-security startup’ but doesn’t seem affiliated AFAIK (perhaps this is where you got that idea?). He notes that all of the ‘evidence’ for Anthropic’s claims comes from three self referencing self authored articles. He goes further in depth with benchmarks and so on, provides references, makes sense. Seems solid enough to me.
Right. So what do you need me to back up?
Really? That the author has anything to do with AISLE (you claim PR Hype, remember?) beyond referencing it (‘comes from some company that claims that it can do the same thing but much cheaper’), and that the arguments are in any way fallacious (which makes profit motive moot).
Pretty sure I’m being trolled.
I did not claim that.
You quoted some PR hype. I pointed it out.
You’re welcome.
I thought the source was a random researcher
The quoted bit comes from a cybersecurity company. I don’t know if that blogger has a conflict of interest. The post is pretty bad, though.
The one I read was pretty good.
https://www.flyingpenguin.com/the-boy-that-cried-mythos-verification-is-collapsing-trust-in-anthropic/
After I saw how credulous the author was, I didn’t read the rest very carefully. I think it was mostly about there not being enough detail to judge, which didn’t stop him from jumping to conclusions. Again, not something that indicates that the article is worth my time. Details about unpatched potential vulnerabilities is not something you’d expect to be published. We’ll just have to wait a bit. So far, it sure doesn’t look like they will be vindicated.