Haseeb

benchmarks where they showed that it was state-of-the-art on a bunch of different things across the board.

It's clearly a large capabilities jump in models generally.

But the main thing that they were hyperventilating about was not that like, oh, this thing is so generally intelligent it can solve the Riemann hypothesis or it's going to, you know, start doing clerical work or it's going to, it's an amazing lawyer and now it finally can do, you know, self-standing legal work.

The thing they were focused on was cybersecurity.

And I think that's not just because it's the most dangerous in terms of the negative impact on infrastructure.

I think a big part of the reason is that clearly a lot of this capabilities improvement has come from bootstrapping on enormous amounts of coding data.

So I think the right story here, probably, I'm not totally certain of this, but probably, is not just about compute.

it's probably more than anything about data is that, because the thing is like XAI has a huge amount of compute.

We're not seeing anything like this out of them.

Meta just released a new model.

It's okay.

It's obviously at the frontier.

They just released it, I believe yesterday.

Um,

It's not state-of-the-art relative to Opus 4.6 or GPT-5, but it's in the running.

But they also have a lot of compute.

They're not getting these results just because they're training on Blackwell's.

So I think there is something in just the sauce available at Anthropic that's leading them to be so far ahead.

And there is always a story in AI that having a data advantage

is going to result in some kind of, or just having your own great models, is going to result in some kind of recursive improvement loop that they're just going to pull further and further away from other people.