Will Douglas Heaven

They're not really going in and doing maths on it.

They've built a kind of software tool

that allows them to go back to what we were saying earlier.

We all know what it's like to type a bit of text and get an answer out.

They will be able to peer inside the model, if you like, and follow a little bit of the trace of what happens inside the model between that input and that output.

And it's really early days, so they're only getting tiny glimpses yet.

but they're starting to see some of the inner processes, the pipelines, the pathways that data takes from beginning to end.

So I think that's the sense that we're talking about.

We're doing early biology on these things.

We're studying them under a microscope, really.

Yeah, yeah.

I mean, I think that's one thing, maybe the main thing that makes this technology so powerful.

Rather than having to code algorithms by hand to do all the things that we know LLMs can do, those algorithms are sort of learned by the system itself, like why you train it.

So we all know now that you throw a lot of data at...

at a model and it learns how to do stuff with that data.

But the magic there is that people don't have to sort of code it themselves.

But that then gets at why it's probably better to think of these things being grown rather than built because nobody has actually gone in piece by piece and said, oh, then this bit goes there, that bit goes there.

A better way of thinking of it is if, I don't know, for any gardeners out there, you know, maybe you have a trellis or a frame in your garden.

So you can set up the scaffolding.

You can sort of say, I want it to go, I want my plant to go in this direction.