Lee Cronin: observation

Transcript context

…Yes. So novelty, it is interesting whether these large language models will ever be able to come up with something truly novel. I can show on the back of a piece of paper why that’s impossible. And it’s like the problem is that… And again these are domain experts kind of bullshitting each other. The term generative, right. Average person say, oh, it’s no, no, no. Look, if I take the numbers between zero and 1000 and I train a model to pick out the prime numbers by giving all the prime numbers between zero and a thousand, it doesn’t know what prime number is. Occasionally if I can cheat a bit, it will start to guess. It never will produce anything out with the dataset because you mine the past. The thing that I’m getting to is I think that actually current machine learning technologies might actually help reveal why time is fundamental. It’s like kind of insane. Because they tell you about what’s happened in the past, but they can never help you understand what’s happening in the future without training examples. Sure, if that thing happens again. So let’s think about what large language models are doing. We have all the internet as we know it, language, but also they’re doing something else. We having human beings correcting it all the time. Those models are being corrected, Steered.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.