On the frontiers of research at the Lux AI Summit

Danny Crichton:

Thank you so much for joining the Riskgaming Podcast and joining us here at the Lux AI Summit today. You had one of the most interesting panels, which was about the future of the technology frontiers of artificial intelligence. I wanted to just give, for those who are on the room, a little bit of a summary of what you just discussed on stage,

Shirley Ho:

I think we discussed a huge range of topic, but most importantly is how do we get from where we are, which is using a lot of specialized tools and plus LLMs to where true scientific intelligence can be. So what are the roadblocks? How do we make it systematic? How do we actually create the data set that's needed? So all that needed to get to where we are, whether it's curing Alzheimer's or making new materials that can actually encourage, I don't know, travel intergalactically or something like this.

Kyunghyun Cho:

Yeah, absolutely. So what we talked about is that we need to distinguish between solving the problems that we know and that we have to solve more efficiently versus solving this problem of discovery. Scientific discovery is very unique in a sense that it's all about finding something that has not been found before, if it was already found before, we don't call it discovery anymore, right. So that's a rediscover or just a literature review in some sense. So how do we actually make AI to internalize this whole process of discovery that seems to be like the next step that we need to really work on, and then it's not really about the, let's build it or implement it, but more like something that we need to really give a very strong and then careful thought about.

Laurence Pevsner:

So where are we on that right now? Right now, if you're trying to tell an AI, I want to get better at discovery, what does it do? Where does it fail? How is it helpful?

Kyunghyun Cho:

Right. So at the moment, a lot of the focus has been on trying to answer questions that we ask these AI systems and then the idea how these systems are trained and often let's say encouraged to do is to answer them based on the known knowledge or the grounded into the knowledge that exists. Sometimes we're going to let it run small experiments here and there, in particular in the coding agents, we see that it is agents are going to code something, run it, look at the outcome, and then use that to answer things. But these are extremely narrow. What we actually want is that we want these LLMs or the AI systems that know the process of the discovery to actually make suggestions on what we need to do in order to provide these AI system with more data so that the information gain is maximized and then eventually, let's say future success is going to be maximized as well.

Danny Crichton:

I think we obviously talk about AI for science in this general format. There's some sort of generalized scientists. It shows up in every field. Yet when we look at human society, we have the complete opposite. People get stove pipe, narrow PhD's, postdocs, you get narrower and narrower and narrower. When you think about the bitter lesson, this idea that the way we progress AI is just through more compute, throw more energy and data centers at the problem and we'll eventually solve it. Where do you sort of stand on general compute versus more specialized LLMs models data sets to solve some of the challenging scientific problems in your fields?

Shirley Ho:

I think there's multiple answers to that. The first answer I would give is that I think we can get AI to become more generalist. So humans are doing all these specialized situations and we have a PhD in everything, right. But AI could learn everything, but they have to learn not just the literature, but also learning from the real data. I think that's the piece that's sort of missing right now that they all have, yes, I have all the literature, but that's not how you get discovery if you talk to any scientist, it's not that you just need to read all the papers from before. You actually need to go create new experiments, get more data, analyze it and come back and what data, what experiments to create next is some of these genius ideas that sometimes have to come through. So how to create that system, I think that is super important.

Our research, a lot of it has to be depending on creating a polymathic model, which can go across many different areas from physics to chemistry to biology and astrophysics and [inaudible 00:03:47] everything altogether. And I think that actually helps the system to become more generalist in a way that can bring new ideas from physics to chemistry, from chemistry to biology and to something else, maybe to health. So that's my hope in near term. I think further down the line will be can it come up with new ideas across without the human's help? But I think that's definitely more futuristic than where we are right now.

Kyunghyun Cho:

Yeah, I think the biggest fallacy of thinking about AI or AGI or however people call it, is that our thinking about those AI is very much constrained by what we think we know how to do. When in fact, of course, AI is built in a very different way from how we are built, thereby in fact they are now working with the same kind of constraints that we have. So in fact, we have to specialize individuals but then form a society in order to cover different aspects of the very difficult problem. But AI doesn't really have that particular type of the constraint. In fact, AI systems are often worse at a lot of things that we know well, but they can actually solve the problems that we just don't even know how to even approach. So one example I can give you is that if you think about, so I'm talking about the drug discovery a bit too much, but when you think about the drug discovery, there are people who are specialized in trying to figure out the human physiology, of course.

There are people specialized in particular drug molecular modality in order to design them and develop them and optimize them. There are people who are working on the commercial side as well, and then there are people who are working on the clinical trials. So all those people are extremely specialized, but thereby it's really difficult to see what happens across these stages and make connections. And then this lack of connection is one of the reasons why the success rate is so low end to end, but then AI actually doesn't have such a constraint. You can in fact look at all the data coming out of all the different stages and then identify this tiny bit of the correlation that exists across the stages and then now that correlation becomes something that we can now use in order to improve the success rate. So what that means is that this distinction of the generalist versus specialist applies to us, but this distinction does not necessarily apply to the large scale models that we are going to build and we are building and that we are going to build in the future.

Laurence Pevsner:

Yeah. It's like that famous metaphor about the blind men trying to identify what an elephant is, right. And they're feeling all the different pieces and hopefully a generalist polymathic AI can actually see the whole elephant at once. I'm wondering is this being used right now? Like, how is the polymathic AI been helping in your research currently?

Shirley Ho:

So we recently built a model. It's a fluid dynamics model that is having fluids that are including blood flowing through an artificial heart all the way to oceanography, to aerodynamics, to astrophysics. So fluids across all the different scales from smallest to the largest. And you're like, "Why do you build such a model?" What is happening is that you can take this model and fine-tune, which means give it a little extra data on something it's never seen before. In this case, something that explodes, in this case, an exploding star.

There's only five simulations of these exploding stars in the entire world, takes about 10, 20 million to make one happen. So we only got five of them. It's very expensive. Thank you for making these computers available in the government. So that is now used to fine-tune on this one huge model with one particular sample. And I can make predictions just with one example, which is very impossible before. Before you need 10,000 example, a million example, you need the internet scale data. Now I need one example and you can already make predictions on the other one that's different. So I think that makes a huge step from impossible to do to possible to do, and this is I think a great thing about foundation models where the nearby disciplines nearby data actually boost performance or make impossible to become possible for these crazy simulations that never happened before, or any data that's very scarce. So yeah.

Danny Crichton:

When you think about the future of science, we've had a model for the last 70, 80 years, since the 1940s, the post-war, Cold War and the National Science Foundation, Vannevar Bush, Endless Frontiers, et cetera. It seems like we have a bunch of stuff happening all at once. We have this sort of terrifying cut to funding from federal government, challenging finances from the endowment side because of endowment tax, and at the same time you have a kind of rebuild of the pipeline for scientists. So this idea that you would have AI biologists didn't exist 10, 15 years ago except in a couple of frontier programs, now, it seems like AI would be the first step to becoming a frontier biologist or astrophysicist as the case may be. How would you start to rethink the future of research careers in terms of skill building, in terms of what you'd be focusing on? Does it change radically? Does it mostly stay the same? How has it changed?

Kyunghyun Cho:

So I think it should stay the same. The existing system has a huge amount of the merit. That is, how we are going to choose as a society to work on that is not blindsided or the narrow-sided by the immediate, let's say profit, let's say motives. And then that's how we actually decided that the taxpayers are going to essentially outsource this selection as well as the execution to the federal government. And the federal government is going to outsource this research into the future agenda to the universities and other non-profit research organization. I think that this is probably only way in which we can actually continue to invest in the future research, not the research that's going to be developed into the product, let's say today or tomorrow or next year. Those things are of course going to be done by the industry and then with the funding from various, say capitalists and so on.

So that's all good. Now, the issue here I think is not that this system is wrong or this system is outdated, but unfortunately this whole system started to be somewhat, let's say, I don't know, blindsided by the fact that this kind of let's say shiny new tools as a shiny new industry with the shiny new amount of the money that has being poured in. So you look at some of the NSF solicitation in the computer science field. Over the past few years, it's really difficult to tell whether the solicitation is for the research or for developing products that's going to be productized next year. And then that's actually a mistake by the federal government, because federal government was supposed to take the taxpayer money and then trying to invest in the things that companies wouldn't have invested so that we can actually as a society protect our future.

Unfortunately, of course, individuals look at all these shiny new tech, shiny new companies, and they also want to contribute directly. Now of course, the NSF gets a mandate from the Congress. Congress members also tend to be somewhat, let's say, attached and grounded a bit too much. This kind of, let's say reality that happens immediately. So what I think is really important is we need to make sure that this system works by ensuring that the execution and choice of the topics is least influenced by what is actually being productized today or next few years.

Shirley Ho:

Yeah, this is interesting. I mean, I think this idea is that you should fund frontier research and where frontier is should not be decided by the government, it should be farmed down. But then I think you were also asking other point, which is how do you prepare future generations going forward with all these shiny new tools and should they just be, like, just forget it, don't learn software engineering because there's no more software engineering jobs. I think in that aspect I would say that people should do what they love, because then you have the most motivation and creativity pouring in that direction. But on top of that, having some idea what the tools available and what tools might be upcoming will be super useful. Just be a practical situation where you want to be multiplying your efforts instead of just at the one level, right.

Kyunghyun Cho:

Can I ask just one small thing there?

Danny Crichton:

Of course. Yeah, of course.

Kyunghyun Cho:

I think, by the way, when it comes to programming, even though we have amazing coding assistants and whatnot, I think we have to teach programming to everyone as earlier as possible. Because in my view, programming is not amazing because it actually gets you a better products or to improve your productivity. But because that's the only way for any one of us most of the time to test the actual logic of your thoughts. So you learn how to think about things, you know, the mathematics course, history, social science, all those courses, but you never actually get to execute it. You never get to use your logical thinking and then see the consequence. But the programming is the only place where anyone who has access to any kind of computer can write down their logic and then try to see the consequence and then see where the things went wrong. And that's kind of the only way for us to do the exercise. So I think that we need to teach them programming as always possible.

Laurence Pevsner:

Yeah, I was wondering actually if we were talking at the beginning of this conversation about AI doesn't need to be as nearly as specialist as humans does, it can have a much more polymathic point of view. I was wondering if actually one of the benefits of AI is that we too, as humans, will get to evolve with it, right. That in the future when an AI can really go deep, we can actually become more generalist, more interdisciplinary, and our training in schools will follow that model. And we'll all have a bit of programming. We'll all have a bit of astrophysics, we'll all have a bit of biology, and so we can actually work with the AI models too and be able to think this way.

Shirley Ho:

Yeah, I love it as an educational tool, I think it will flow all boat people who might only learn to be a biologist who might not code very much, obviously can do so much more because all the available tools, right. So learning program is great. I just want to be slightly controversial. I think-

Danny Crichton:

We love spicy food.

Shirley Ho:

I'm just thinking that it's not the only place you can test your logic. So a lot of labs, experiments, physics, chemistry, you can actually go hypothesize, go test it out. But programming is probably the fastest. And if you don't have anything else but computer, that's probably the easiest. So that, I agree.

Danny Crichton:

Look, the philosophers are all going to own us. The logicians are coming back to try to get true false logic back in the curriculum, but we're coming towards the end of 2025 and one of the most important AI conferences is coming up with NeurIPS. I heard that you have quite a few papers accepted, so congratulations. I'm curious twofold. One is, obviously we're almost a year from last year, so what were some of the developments over 2025 that you think were overlooked, weren't given the credence or are going to be fundamental into the future? And then when you look forward either to NeurIPS in a few weeks or into 2026, where does the AI from your perspective research agenda is going to go forward?

Shirley Ho:

Yeah, so I'm not going to talk about our papers. I want to talk with some other people's papers. No, I just think it's good to point out. I think there's a huge amount of literature that might not be as popular that talks about how to merge models and how to understand model and how to steer models, that's usually classified as interpretability and that direction. And as a scientist, I think it's great if we can understand the models a little bit better, because if you tell someone, I find a new fundamental rule about the universe, I have no idea how it came about. No one's going to believe you. So can you somewhat interpret the model? I think it's super useful. There's a bunch of paper coming out from I think multiple groups at Google. Like [inaudible 00:14:41] has one group. There's group in Europe also that's also doing very well that looks at how to steer the model, the Golden Gate Bridge stuff that people probably heard from Anthropic before, and how to understand the model from the beginning, how trained, what's the data, how was the output? And that's super important.

Kyunghyun Cho:

And what I see, not necessarily this year, but over the past few years, the trend has been that we are now acknowledging or in the face of acknowledging the fact that we have built this amazing correlation machine. This machine, big transformers trained with the stochastic gradient descent, and then you feed in as much data as you want and it's going to capture every single statistical correlation that exists within this data. And that's amazing thing. But then of course we've been actually beating this, not dead horse, but the horse that has arrived already too much, actually I shouldn't say that, but anyway. Okay, too much without realizing that we actually already solved this problem.

And then I think about few years back, including myself, I think that we started realizing that, oh, actually we may be actually spending too much time on solving the problem that has already been solved. And then now we are actually going into the one next level up, right. So if you think about this kind of let's say learning or inference, what is the next step, next to the correlation is more of a causation. And the causation is really important. And it connects to what Shirley just said is because if we want to be able to control or steer any of these systems, we have to have the causal understanding of the system. Otherwise, we're going to make all those random mistakes.

So how do we actually go to the large-scale, high-dimensional causal analysis is a big trend that I see as growing quite rapidly and then doing so it turned out, that it's all about, in fact, framing the problem in a way that maximally benefit from this amazing correlation machine we built. So we're going into more of a meta-learning or the meta-inference, and that's also what I'm interested in. And I think the increasing more people are interested in this one. And it connects of course, naturally to discovery. Without the causation, you cannot actually discover things that easily, although sometimes you do, but-

Shirley Ho:

Accidentally.

Kyunghyun Cho:

Yeah, accidentally.

Laurence Pevsner:

You stumble upon it.

Danny Crichton:

So I feel like there's always this TikTok and AI research over decades, which is we were in symbolic, we actually had truth-false values, and then we went over to the stochastic pair side of the equation, and maybe there's some beautiful fusion in the middle. But Kyunghyun and Shirley, thank you so much for joining us.

Shirley Ho:

Thank you so much.

Kyunghyun Cho:

Well, thank you for the invitation.

Danny Crichton:

Thank you.

Shirley Ho:

Thank you.

Kyunghyun Cho:

Yes.

‍