Lee Cronin
speaker
545 appearances
2 recordings
1 series
first heard Dec 2023
last heard Jun 2024
Lee Cronin’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
I agree, but the super intelligence I want, I want to be able to have a discussion with it about... coming up with fundamental new ideas that generate knowledge. And if the superintelligence we generate can mine novelty from the future that I didn't see in its training set in the past, I would agree that something really interesting is coming on. I'll say that again.
If the intelligence system, be it a human being, a chatbot, something else, is able to produce something truly novel that I could not predict, even having full audit trail from the past, They're not going to be sold.
But you're saying truly novel. I think they are in the training set. I think everything it produces comes from a training set. There's a difference between novelty and interpolation. We do not understand where these leaps come from yet. That is what intelligence is, I would argue. Those leaps.
And some people say, no, it's actually just what will happen if you just do cross-domain training and all that stuff. And that may be true. And I may be completely wrong. But right now, the human mind is able to mine novelty. in a way that artificial intelligence systems cannot. And this is why we still have a job and we're still doing stuff. And, you know, I use chat GPT for a few weeks.
Oh, this is cool. And then it took me too. I had to. Well, what happened is it took me too much time to correct it. Then it got really good. And now they've they've done something to it. It's not actually that good.
And the unit future is bigger than the present, which is why human beings are quite good at generating novelty because we have to expand our data set and to cope with unexpected things in our environment. Our environment throws them all at us. Again, we have to survive in that environment. And I mean, I never say never.
I would be very interested in how we can get cross-domain training cheaply in chemical systems, because I'm a chemist, and the only thing I know of is the human brain, but maybe that's just me being boring and predictable and not novel.
Yeah. I mean, this is one that goes with my team. I try and do things that are obvious but non-obvious in certain areas. And one of the things I was always asking about in chemistry, people like to represent molecules as graphs. And it's quite difficult. It's really hard. If you're doing AI in chemistry, you really want to basically have good representations. You can generate new molecules.
They're interesting. And I was thinking, well, molecules aren't really graphs. And they're not continuously differentiable.
could i do something that was continuously differentiable i was like well molecules are actually made up of electron density so they got thinking say well okay could there be a way where we could just basically take a um take a database of readily solved electron densities for millions of molecules so we took the electron density for millions of molecules and just train the model to put to learn what electron density is
And so what we built was a system that you literally could give it a... Let's say you could take a protein that has a particular active site or a cup with a certain hole in it. You pour noise into it. And with a GPT, you turn the noise into electron density. And then, in this case, it hallucinates like all of them do.
But the hallucinations are good because it means I don't have to train on such a large, such a huge data set. Because these data sets are very expensive. Because how do you produce it? So... So go back a step. So you've got all these molecules in this data set, but what you've literally done is a quantum mechanical calculation where you produce electron densities for each molecule.
So you say, oh, this representation of this molecule has these electron densities associated with it. So you know what the representation is and you train the neural network to know what electron density is. So then you give it an unknown pocket. You pour in noise and you say, right, produce me electron density. It produces electron density that doesn't look ridiculous.
And what we did in this case is we produced electron density that maximizes the electrostatic potential, so the stickiness, but minimizes what we call the steric hindrance, so the overlaps that's repulsive. So, you know, make the perfect fit. And then we then used a kind of like a chat GPT type thing to turn that electron density into what's called a smile.
A smile string is a way of representing a molecule in letters. and then we can then so it just generates them just generates them and then the other thing is then we bung that into the computer and then it just makes it
The robot that we've got that can basically just do chemistry. Yeah. So kind of, we've kind of got this end to end drug discovery machine where you can say, oh, you want to bind to this active site. Here you go. I mean, it's a bit leaky and things kind of break, but it's the, it's a proof of principle.
Well, the hallucinations are really great in this case, because in the case of a large language model, the hallucinations just like just make everything up to when it doesn't just make everything up, but it gives you an output that you're plausibly comfortable with and thinks you're doing probabilistically.
The problem on these electron density models is it's very expensive to solve a Schrodinger equation going up to many heavy atoms and large molecules. And so we wondered if we trained the system on up to nine heavy atoms, whether it would go beyond nine. And it did. It started to generate molecules of 12. No problem. They look pretty good.
And I was like, well, this hallucination I will take for free. Thank you very much. Because it just basically, this is a case where interpolation, extrapolation worked relatively well. And we were able to generate the really good molecules. And then what we were able to do here is... And this is a really good point, what I was trying to say earlier, that we were able to generate new molecules...
from the known data set that would bind to the host. So a new guest would bind. Were these truly novel? Not really because they were constrained by the host. Were they new to us? Yes. So I do understand, I can concede that machine learning systems, artificial intelligence systems can generate new entities, but how novel are they? It remains to be seen.
Showing 481–500 of 545 · page 25 of 28
← Previous
Next →