Agentic AI & the Future of Science with Michael Bronstein & Nick Magnuson of Qlik – Tech Talks

We are back at Click Connect 2025, and I am reconnecting with one of the guests, one of the folks that I spoke to last year here at Click Connect. Uh, Mike Hornstein, you, you have ideas. They always get me thinking.

Nick, thank you for arranging this. Yep. Uh, first, I guess before we get started, uh, tell us a little bit about yourselves.

Uh, who are you? Why am I talking to you? Well, uh, probably you should ask why you, you're talking to me, But I am gonna ask myself that.

Yes. Uh, I'm Michael Bronson. I'm, uh, deep Mind professor of AI at the University of Oxford, computer Scientist, and I'm on the, um, ai, uh, consult, the advisory board of Qlik.

And I'm Nick Magnuson. I'm head of AI at Qlik. And so obviously very interested in this conversation.

Uh, absolutely. So I'm Stephen Foskett, the organizer of Tech Field Day. And, um, one of the things that I love about Click Connect is being able to speak with people who challenge me, who come up with different ideas.

And wouldn't you know it within two seconds. Michael, you challenged me. One of the things I hadn't thought about wa with agentic AI is the impact of agentic AI on science.

Uh, what is your, uh, thesis here in terms of the validity or the application of ag agentic AI to science? Yeah, so I think, uh, the way that we do science, um, has not changed, uh, since the time of Newton Oral, right? So the technology might have changed, uh, but the method itself, which is hypothesized predict and test, uh, this was the step of science, which in about four centuries has brought us to where we are now, right?

Quite amazing progress for human civilization and technology has been used for, uh, experimental testing of, uh, predictions for predictions themselves, right? So now do a lot in simulation, but the hypothesis generation has always been a human endeavor. So it's been always, whatever you call it, the, the eureka moment, the, the ingenuity, the, the, uh, the human genius, the, the inspiration, sometimes A guess And sometimes a guess, right?

So, and, um, now, uh, we see for the first time in the history of, uh, human civilization for the history of science, that, uh, the machine is, uh, increasingly taking part, not being only the mechanical amplification of the human brain, but part of the creative process. And, uh, for me, the, the, well probably for many others, that the real test for this will be whether we'll see an important discovery, scientific discovery, invention, uh, depending on how you, you look at it, that would, uh, be so major that if, for example, would be able to win a Nobel Prize. Well, we've already heard, uh, for example, of the application of AI to protein folding.

Uh, would you consider something like that to be a significant, uh, in invention or discovery, or is that just a brute force? So to some extent, it's good for us, I would say. So this is not a discovery, this is a technique, okay?

And there is this famous quote from Sidney Brenner. He was a novel, ate, uh, uh, in medicine. He was saying that the breakthrough in science comes, uh, first of all, from new techniques, new discoveries, and new ideas in this particular order, if you think of history, of science, invention of the microscope, right?

So this is what, uh, enabled more or less the, the modern biology. It was a technique. It led to discoveries.

It led to new ideas, right? So discovery, example of discovery is the structure of the DNA, right? That enabled modern genomics, uh, and idea is theory evolution, right?

That, again, enabled a lot of things that, that we, we have nowadays. So techniques always come first. AI in a sense, it clicks all the boxes.

So it is definitely a new technique. It's al it is already living, leading to new discoveries and the conjecture here that will lead to significant discoveries and possibly also new ideas. So he just said that it clicks the boxes.

So I just have to, I just have to bring it in. Um, on, on QQ, uh, Qlik is predominantly thought of, at least by me, as, uh, being relevant to enterprises, to businesses who are trying to, uh, you know, build modern, productive data-driven applications. Uh, but you're not just in business.

I mean, you guys are in science as a cycling team here. Uh, it it goes beyond, uh, the traditional business world, doesn't it? Yeah.

I mean, any of these applications require data and they require some level of analytics. So yeah, it, it can transcend almost any industry. That's why, uh, Qlik has so many customers across so many different industries, and all of 'em are interested in how they can employ agents and AI to make, uh, make things more productive or, uh, you know, solve new use cases that weren't previously solvable with traditional, uh, traditional technologies.

So you mentioned agents, um, you know, you're talking about using AI in science. Uh, again, one of the things I hadn't considered was ag agentic AI as an extension of the lab in, in a way, uh, how does, how, how does agentic AI work in science? Or how would it work in science?

Yeah, so when it comes to experimental science, the interaction with the physical world is complicated. It's messy, right? So you need to do experiments in the wet lab.

So a lot of things can go wrong. So this is probably one of the, the limitations. So we need to reinvent the lab to purpose it for, uh, for the machine, for the agents rather than for the humans.

My conjecture is actually that in formal fields of science, like mathematics, uh, we'll see these breakthrough, uh, faster, because software, it's all software, right? So software is soft, everything is possible. So, but seriously speaking, one of the, one of the things that is, uh, in these fields, when you, you think of theory improving, right?

So this is what mathematician is do. It is difficult to prove a theorem. It's very easy to verify whether the proof is correct, right?

So I can formally run it through some, uh, logical verifier and see whether it's solve, it answers the question, right? So same thing about software generation. I can run a piece of code that, that some generative agent produce, and I can see whether it does what it's supposed to do.

It's more difficult, uh, to do an experimental sciences where usually it'll require doing an experiment. Okay? So as you mentioned, math is essentially just software.

Just software. I don't mean that to, to diminish it. It it's software.

It's uh, it's, you know, programming. It was a mathematical, uh, discipline forever. Uh, what about in the lab?

What about biological science and physical science? Is, is AI gonna have an impact there? I think it's already having impact.

Sorry, what you mentioned alpha fold, right? So, so it's a new method that allows actually to ask and answer questions that biologists were even not thinking about as possible. So it changes the way that that work in this field is being done, which is, which is, uh, I think is more important transformation than a particular way or particular piece of equipment.

Uh, so it's change of the mindset. So it's a paradigm shift. Now, what, uh, my, in my opinion, what needs to happen is if we believe, again, in this vision and, uh, that, that, uh, science will be increasingly more AI driven, uh, in autonomous way by AI agents, the very, uh, way that we produce and consume data must be, uh, reimagined.

So now data is still human tension. So we produce data that is, uh, was developed so that experimental technologies were developed initially for humans. So in case of alcohol game, it was trained on PBB protein database.

So, so it's structural models, three dimensional models of proteins acquired by x-ray crystallography or cry and methods. We can imagine completely new data modalities that, uh, say nothing to a human. I don't see anything in this data.

It makes sense only in conjunction with appropriate machine learning model. So my view is that, uh, the experimental technology and the, uh, and the ML technology needs to be developed, uh, in tandem. So co-development of both the wet lab and the dry lab.

Yeah, that, that's amazing. It's amazing to think about that because as you say, if you look at the history of science, essentially, so much of science has been developed based on graph theory. Essentially.

I'm going to plot my points, I'm gonna draw a line, and then I'm gonna draw a conclusion based on my perception of the shape of that line. And yet, as you say, an AI doesn't need that crutch, it doesn't need that sort of translation into the real world. It could maybe handle data in a completely different way, maybe on, um, more dimensions as again, I mean, humans have trouble with, um, moving past a three dimensional, uh, image of data.

Um, and, and, and so, you know, how, how would you, where, where would you take this in the future? And do we have the systems that could even store and process, um, this new type of data? So I think we do have the systems of, uh, store and process this, uh, this type of data.

Uh, it is, um, really about picking the signal out of noise. So machine learning, not every problem is necessarily amenable to machine learning. So the problem, the problem selection is important in retrospective alpha fold.

Uh, protein folding was an easy problem, right? So I'm saying easy. Of course, it's unfair.

It won the noble prize. So it was a difficult problem, right? It was a big challenge for 50 years, but easy in the sense that it has some kind of internal structure that in retrospective, we can look at it and say, actually it was solvable by a mouth.

So this, uh, understanding, uh, which problems might have this structure, uh, on which we can apply. Machine learning is very important. Language is such a problem.

Mathematics might be such a problem, right? And it all boils down to some kind of what, what I can call degenerate solution space where we talk about proteins. So the, typically we'll hear some astronomical number, the number of proteins that might exist in nature.

We don't see that in nature, right? So proteins are result of evolution. So nature is lazy.

It likes to, to reuse bits and pieces that works well, maybe 3 billion of years ago in some organism, just reuse it somewhere else. This is also the case in language. This is also the case in, in software, right?

So we write software by reusing, uh, recycling bits and pieces of code that somebody else wrote, right? So this is often how software products are developed in mathematics as well. So we, we have some standard proof techniques and we try to reduce our problem to a particular pattern that, that, uh, successfully worked in the past.

And have you seen yet, uh, AI systems coming up with novel, uh, new approaches to mathematics, or, I think we start seeing it a little bit. It's probably too early, and I think it's promising. You know, are you worried about this?

Are you worried that he's gonna come to you and say, I need to store this completely different data set that you've never seen before? Oh, I think we're, we are starting to see that, we're starting to see a lot of the synthetic data that's being created for these types of use cases. And yeah, the, there is really no limitation in terms of how big that data can, can be.

And I think for a lot of applications, agents are necessary to explore a whole variety of different scenarios that would take a human far too long to do. Uh, it wasn't ever really reasonable to do in the past. Um, but you know, I'll take the challenge.

You know, it's funny that you mentioned this because, um, one of the big announcements this year from Qlik, and, uh, one of the big things that everyone in the industry is talking about is basically taking data lakes and making them useful in terms of analytics and querying. And that is a very hard problem. It's hard to get a lot of data into a data lake.

It's hard to process that data. It's hard to query that data, or at least it has been traditionally. And now all of the companies in this industry, especially Qlik, are working on solutions that would make a data lake a more approachable set of data.

Um, I imagine that if we move to an AI driven type of science, we will be generating a lot more data. One of the truisms that I found on the AI podcast that I've run is that the ability to process data has increased our appetite for data. And so essentially, especially, you know, with edge applications, for example, we deploy more sensors, we deploy more high resolution, we are collecting more data than ever before, simply because we can.

And I think that it would be the same in science, right? I mean, if, if I'm a lab, uh, you know, scientist deciding how my lab is gonna be, I'm probably going to, I dunno, go nuts on the data, you know, basically put all the sensors in there. 'cause maybe I'll find something, right?

Yeah. So multi multimodality is probably important consideration, not only in science, I guess in general. So we often have different data sources.

The, the quality of the data that, uh, the scale of the data, the type of the data, the structure of the data might be very different, right? So again, in biological applications, so you might have sequencing data for proteins, you might have data that is available only partially. So when you are designing a new drug, you might know, I dunno, the structure of your therapeutic targets, but not the, the full protein.

So you see maybe just some, some important part of it. So, so it's, uh, you don't know often a priori, uh, what kind of data you have, what is missing, uh, and you need to deal with this, uh, uh, missing bits and pieces, partial information, uh, uh, data coming from multiple sources with different level of reliability. Uh, so that's probably one of the, the challenges that we're gonna be seeing in terms of the, the data infrastructure that analyze this, uh, this future.

Well, I, I do worry about synthetic data generally. I mean, certainly we've been talking about training AI models with synthetic data, synthetic data in science. Is that a thing?

It is a thing. So, well, we use simulations all the time, right? Anywhere from molecular dynamics to weather forecasting.

So, uh, the issue is, again, when to trust the simulation. And in many complex systems, we have only very rough idea of how the system works, right? So we don't have an equation of a cell.

So we even at the very basic primitive of, uh, of, of life, we don't know how it works, right? So, so, uh, simulation works to some extent. What is important is to be able to verify whether the simulation produces right results.

And if we were to replace it with machine learning models, with generative models, uh, in case of images, right? So we have these famous examples of, uh, hands with six fingers, what is the analogy? The experimental, uh, life science analogy of this, right?

So if we're able to verify experimentally very quickly and very cheaply in that scale, uh, and discard some of the solutions of these generation, right? Or this, uh, this simulation and retain only those that make sense. So that, that's probably will be the key.

Wow. I, uh, Nick you have a a, an AI council. Yeah.

Do they, do they challenge you as much as they challenge me? Do you, do you come out of these meetings and say, we're gonna have to rethink things? Yeah, Well, all the time.

And it's not just on the scientific side. There's all the ethical side and the regulation side that, you know, we get a lot of really constructive input and yeah, it, it is an iterative process. We always go back and redesign and rethink our current, you know, our current thinking.

It, it, it's just incredible. I, I appreciate talking to you. I'm gonna say every year 'cause, 'cause we'll come back and talk to you next year.

Uh, thank you very much for including us. If you enjoyed this conversation, uh, we did post another discussion, uh, last year where we had an equally mind blowing time. So look in your favorite search engine for, uh, tech Field Day.

Uh, actually, you know what, we'll put a link in the, uh, in the show notes for this one. Also check out the tech field Day plus YouTube channel, where we're gonna be posting more interviews like this from click Connect and the regular tech field day website and YouTube channel where we're posting, uh, presentations from Click Connect this year where we dive deep in some of the new products and announcements. Thank you very much for joining us.

Thanks for your time. And, uh, catch some of the other videos.