Andy Halliday

Yeah.

So I want to talk about deep seek and in general, the Chinese open source models.

So deep seek is open source, open weights.

That means that anybody here in Silicon Valley who's building a startup can use this all at parity platform.

highly capable, very comparable to frontier models for basically for free for the cost of hosting themselves in their own infrastructure and running those models.

And so there's, there's a game at play here that goes beyond who has the smartest reasoning model.

The truth is here now that deep seek and also the Alibaba and you know, the Quinn models are,

Moonshot, Kimmy 2, all of these open source models that are coming out of China are really perfectly adequate to all of the application space that you might use these advanced reasoning models for.

But that doesn't change the fact that Gemini 3.0 is still at the state of the art.

It is measurably better than any one of the

Chinese open source models.

But it's so close now.

DeepSeq came out with 3.2.

And by the way, they also, last week, they released DeepSeq Math V2, which was important because there's this thing called the International Math Olympiad.

And humans score around 80% on International Math Olympiad.

I shouldn't say humans generally.

I mean, the very greatest mathematicians score around there.

And the model scored 118 out of 120.

on the putnam competition and solve five of the six international math olympiad problems for the 2025 exam uh hitting the gold standard so it's basically what's important about this is that it's as smart as any mathematician out there pretty much

And it also developed, in order to get to that level, it generated a new model, an algorithmic model for solving math problems called a generator verifier system.