How is AI changing science for good?

Chris Smith put this question to materials scientist Anna Ploszajski, astrobiologist Lewis Dartnell, mathematician Ems Lord, and materials scientist Anna Ploszajski...

Xander - Absolutely. I mean, astronomy as a very kind of data-driven field, we sort of have more data than we know what to do with. And machine learning in particular, that particular discipline of AI, is especially useful for picking out the interesting patterns in all of the noise, finding the useful bits of information through these swarms of data that we're getting.

Chris - And EMS is the fact that the government appreciate we're going to need a lot of people who are very good at making, using and deploying this sort of software, the reason they're so interested in maths, because you can't write computer code if you're not at least reasonably versed in maths.

Ems - I think that's part of it. But when we're looking at these chatbots and you can put in your questions, they actually can be incredibly valuable.

Chris - Have you tried solving any maths equations with one?

Ems - Well, what I've done is I've been putting in NRICH questions and I've put in some questions for primary aged children. I've put in questions for five year olds just to see what it would come out with.

Chris - I'd not thought of that. What does it do?

Ems - I mean, it's learning as it goes along. So to begin with, I can ask a question. We've got one called Eggs and Baskets, and you get clues about how many eggs in each basket. You've got to work it out for five year olds. The answer it gained the first time was comparable to that GCSE question. Lots of Xs and Ys and Zs. And I put in, 'well, you know, thank you. However, could you answer it in the style of a five year old?' And then it started using words rather than algebra. And then because it seemed wordy, and these are five year olds, I said, 'well, what sort of diagram might you draw? What kind of sketch?' And so it's that little skill about knowing what questions to ask. So it's not just typing something in, but it's asking the question that will be helpful. So, not just what's the answer, but how did you get it? What strategies might I use?

Chris - Did it reveal it's working? Did it show it's working in, in sort of exam speak? Because that's one of the criticisms that many individuals level at AI that they're not explainable. They can't tell you how they got to the answer they did unless of course they've gone and got it from somewhere else on the internet, seen your question that you've thrown at it before, and they've just dredged up that bit of information and parroted it.

Ems - They did come up with quite a good reasoning explanation, which wasn't the same as any that we'd got on the website already. What would concern me is the lack of visualisation at the moment, but that's an area they're working on as well. So I really value diagrams, visuals. They're great for explaining to people the thinking behind the problem. And also, if I'm looking at something a bit more academic, I'm not getting the references at the moment where they've been looking. So do they mine our own questions? Where have they got those from? But the actual range of strategies, you can encourage it to suggest different ways. That's really valuable for a student who's stuck and just doesn't get one way they can get another. And the first few lines can act as a hint. So if you're learning to problem solve, if you use it the right way, it can be really useful.

Chris - And in astrobiology Lewis, useful?

Lewis - Yeah for similar reasons as Xander was saying with astrophysics. Churning through the data and looking for patterns. But, my sort of university post actually did a lot of lecturing and marking essays from students, and we've basically now passed a threshold where these language models are good enough that it's practically impossible to tell if a student has cheated. If they've plagiarised and they've obviously cut and pasted a paragraph off Wikipedia, you can only find that very easily, but prove that it is the case. If they've used a language model to create an essay, you can't prove that has happened, because you can run the same language model on the same prompt a second time and it comes up with something completely different.

Chris - You're obviously much better informed and well versed in your subject than the lecturers who were teaching Michael Crichton, who wrote Jurassic Park, because in his autobiography he says, when he was at Harvard, he originally started studying English before he switched to medicine. And he said he got so dispirited and disillusioned with his teachers. In the end he started handing in tracks of famous poets and famous writers, <laugh>, and they wouldn mark that. He said, 'when Blake got a C, I realised English was not for me', at least at Harvard, and switched to medicine. So it was medicine. And ultimately Hollywood that he, he did that. Material science must be a massive beneficiary of this sort of machine learning technology, Anna, for sort of doing the engineering with atoms and working out recipes for novel materials and arrangements of atoms that your subject involves.

Anna - Absolutely. In fact, I can remember during my PhD, a lot of material science is to do with, as you say, the arrangements of atoms inside materials and machine learning can help us to predict where those atoms will sit because there's a lot of very distinct symmetry and the sizes of atoms, for example. So these machine learning algorithms can really help us to predict where the atoms are going to sit if you give them these different parameters, for example. And, I use those during my PhD to solve new crystalline structures that we didn't know about before.