Anthropic's new model is game-changing. At least, that's what they say in their advertisement designed to make money. Claude Mythos has apparently found hundreds of vulnerabilities in every major browser and operating system, including a remote code execution in FreeBSD and an out-of-bounds write in FFmpeg. They outline this in their now-public system card, which I’d be tempted to trust if AI companies hadn’t been saying this for years now. Remember when GPT-2 was “too dangerous to release”, but then turned out to be a worthless slop factory? Anthropic’s recent behaviour seems suspiciously similar, and I’m hesitant to believe in a product which literally has “myth” in the name.

The new model has apparently been in use internally for several weeks. This is probably true: for one, it explains the recent collapse in the quality of their app. In my opinion, an endorsement from the lab that leaked its own source code last week should be anything but a selling point. Perhaps Anthropic isn’t gatekeeping the model to protect humanity (as they ostentatiously claim), but rather to avoid having to prove the benchmarks they’re so eager to share. Although, come to think of it, this could actually be a blessing for the open-source projects who won’t have to deal with a deluge of superintelligent slop.

Whether or not this is all a stunt, it seems to have worked wonders for their marketing. Social media is drowning in hype from techbros who seem very quick to put an ADVERTISEMENT designed to MAKE MONEY before Anthropic’s history of exaggerating Claude’s capabilities (read: lying).

Maybe I’m wrong. Maybe in 10 years I’ll curse myself for not foreseeing the end of software as we know it. Or maybe I’ll wonder how Anthropic managed to trick select software maintainers into sharing critical security details in exchange for a product that never materialized. Either way, I think it’ll be fun to watch.