Post

I lost faith in all these companies; for me, unless proven otherwise, they are driven by money and market share. As far as I know, they might have made a deal with Anthropic. Putting my tinfoil hat on.

Yeah, fair enough. But can you assure me that the AI hackathon to reach those numbers wasn't planned to increase the visibility of Anthropic's new model? I'm obviously going overboard with the conspiracy theory, but these AI companies are dirty af.

I tend to believe that Mythos could indeed be very capable. There are some cynics out there who think it's just a marketing stunt. To that, I'd say: is that something you want to be wrong about? It'd be much better to find out "false alarm" than to realize "oh no, they weren't kidding"

I miss times where programming was about programming, not about reading AI-generated PRs and doing security patches 3 times per week

in theory this graph should drop quite sharply once the security bugs are fixed and better systems are in place to prevent new ones, right

Several groups confirmed the same vulnerabilities can be found with Opus. So other than marketing stunt, by Anthropic giving away millions in compute for free in exchange for companies pumping it's "scary to release" model, what was it? Why didn't companies do this with Opus already?

That doesn't really mean anything - if you point Opus at the relevant files and ask to find the bug it'll find them, of course. Same for a human. It's conceivable that the model is particularly good by sheer virtue of not needing to know which files and interaction must be scrutinized.

"of course" is very load bearing there. Did you read the bug reports? These are complex chains that span huge codebases and are difficult to reason about. These are not "find the bug in these 10 files" kind of bugs.

And Opus was able to do the same thing as Mythos - it was pointed at a project and asked to find vulnerabilities, while given an absolutely massive compute budget. Because that's the marketing trick - giving unlimited compute until it finds a bunch of holes. Existing models do just fine at that.