Grant Harvey

Yeah.

And it's answered that coherently.

The model itself, this is where I think people get so confused.

No one ever interacts with the model.

They interact with an application layer on top of a model.

Yeah.

They interact with the product.

Claude is a product, Gemini is a product, and ChatGPT is a product.

So you're improving the model, but often a lot of what we're doing as well now is helping give information to improve the product as well as the base model.

People talk about the toxic garbage.

There's not much of that either.

There's just loads of irrelevant stuff.

There's just petabytes and petabytes of dross or random stuff that is adding no benefit.

So it's just weight.

And that's cost.

That's inference cost.

That's all these things.

Honestly, the open source thing, a lot of it is a cost play and a performance play.

Like, because if you just have less data and fewer hyperparameters, your model is more efficient and easier to interact with and easier to adjust as you need it.

But then the problem is that maybe sometimes because it's easier to adjust, you can adjust it in ways you don't expect.