Steve Brown

All the major models.

They've all clearly, and you can test this, they've all clearly been trained on a vast corpus of medical literature and medical information.

So they've all recognized that medicine is a major use case.

And probably, you know, chat GPT, you know, watching what everyone's typing in, they're seeing a lot of people are using it for medical questions.

Big part of the GPG presentation.

But, you know, I found that actually GPT-5 was less accurate than GPT-4 in my case.

Really?

Yeah.

Yeah, I mean, I think it may be over-trained.

It may be going too deep in rabbit holes, but I don't know the cause.

I mean, these LLMs are a black box.

We've compressed all this knowledge in there, but no one really knows how to go explore it, other than doing the approach that I'm doing is like, well, let's create multiple agents to explore multiple different pathways.

But it may be more accurate in some things and less accurate in other things.

But I'm using all the major models and all the major versions of the models.

And I can tell from this that they all have been trained on medical knowledge.

That doesn't mean they give the same answers.

They give different answers.

So if I say I've got 10 different major models that I'm using and 36 different agents trained on different points of view, that's 360 different answers, potentially.

But when you go synthesize those and have them kind of...

cross-check each other, you find areas of convergence.