Andy Halliday

And this is where one model proposes a proof to the math problem, and the other one critiques that, and it goes back and forth in a process like that in order to arrive at the correct conclusion.

So it does that on a stepwise basis, meaning it's going to give it a confidence score at each step of the solution in the proofing process for a math problem.

Now, that's indicative of the ways that reasoning advancements are happening now at a more... at a non-scalar level.

Like, it's not that they're just giving it more examples and the large language model that's doing this kind of math work is just getting more and more experience.

They have new methods of reasoning that are being applied algorithmically in order to get to that performance level.

So...

Take away too long, don't read.

It's basically Chinese open models are pretty much as good, you know, for every practical purpose.

And when I say open source, open weight, it means you don't have to pay on an ongoing basis as you do with all of the closed models that are at the state of the art, which is just, you know,

small measurement ahead of where the Chinese open source models are.

And they are achieving this without access to the very latest and greatest chips that are used in the training runs for the closed source models, because those are subject to an export embargo by our government.

I think that's right.

That's the point is that if you're going to use one of these models on a daily basis for something practical like shopping or, you know, just general information retrieval,

It's useful to have it accumulate memory about that so it understands the context and the areas of interest that you have.

And that makes memory in Perplexity more important than memory inside Claude Opus 4.5 for me, because I don't want it to memorize everything that it's looking at in terms of the coding work that I'm doing with that, right?

So Perplexity, I think, is a good, good example of where memory will be very useful.

And I think for the 800 million weekly users who are using ChatGPT, memory in ChatGPT is very useful because those 800 million users aren't doing complex work or multi-customer work like you are, Brian.

They're basically oral.

Yeah.

Yes.