Adam Kucharski

So I think we almost, by skirting that too much,

we actually avoid the problems that people sometimes care about.

And I think a lot of the nice work that's been going on in causal inference is trying to get people to confront this more head on rather than say, okay, you can just stay in this prediction world and that's fine.

And then just later, maybe make a policy suggestion off the back of it.

I think that's a really good point.

And it kind of shows almost a lot of the transition I think we're going through currently.

And I think particularly for things like, you know, things like smoking cancer, where it's very hard to run a trial, you can't make people randomly take up smoking.

having those additional pieces of evidence, whether it's an analogy with a similar carcinogen, whether it's a biological mechanism, can help almost give you more supports for that argument that there's a cause and effect going on.

But I think what I found quite striking, and I realized actually that it's something that had kind of bothered me a bit, and I'd be interested to hear whether it bothers you, but

With the emergence of AI, it's almost a bit of the kind of the loss of scientific satisfaction.

I think, you know, you kind of grow up with learning about how the world works and why this is doing what it's doing.

And I talked, for example, of some of the people involved with AlphaFold and some of the subsequent work in installing those predictions about structures.

And they'd almost kind of made peace with it, which I found kind of interesting because I think they started off being a bit uncomfortable.

You've got these remarkable AI models making these predictions, but we don't understand still biologically what's happening here.

But I think they'd just settled and say, well, biology is really complex on some of these problems.

And if we can have a tool that can give us this extremely valuable information, maybe that's okay.

And it was just interesting that they'd really kind of gone through that process

kind of process, which I think a lot of people are still grappling with and that almost that discomfort of using AI and what's going to convince you that that's a useful, reliable prediction, whether it's something like predicting protein folding or getting in a self-driving car.

What's the evidence you need to kind of convince you that's reliable?

And this, this was a fascinating story because this, Kirk Goodell, who's like one of the greatest logical minds of the 20th century and did a lot of work, particularly in the early 20th century around system of rules, particularly things like mathematics and, and whether they can ever, uh,