Mark Williams-Cook

They'd laugh me out the room.

Yeah.

So yeah, I'm, you know, I'm super excited about all of this, um, all this technology, but I just, um, I'm cautious through experience about how much of it as well by the, by the AI tech industry is, is posturing for, um, funding for, you know, various stakeholders that are invested, you know, it's why they call it, um, hallucinations, right?

Yeah.

Because it sounds nice in a border, and nobody wants to say, oh, yeah, it just makes up shit once in a while.

I mean, that's how I passed A-level biology on that exact principle.

Exactly.

I need to write something.

I was actually thinking about that paper ground when you said it, because I saw that, I think it was this morning or actually just yesterday, where they were like, oh, we figured it out.

But it's like, and as you rightly say, it's like the fitness function, whatever you want to call it, of what you're rewarding for.

But how on earth do you fix that if, you know, because you need to test it, it needs to have this and...

And people, you know, when I had conversations with them about, oh, they reward the correct, you know, correct answers, just correct answers.

And then the question comes up with how do you know it's correct?

Right.

Because you need pretty much a bank of human people to to verify that.

Because this is really, correctness is, for me, a really interesting topic from an SEO point of view, right?

Because Google has said before previously, like, we don't know what is correct or not.

And they rely very heavily on consensus, right?

Which I'm assuming is how a lot of the models decide whether they need to kind of ground answers or not.

So if you ask, you know, what do red blood cells do?