AI’s Next Frontier with Dr. Kyunghyun Cho

I was having a beer with Richard Bonneau, Rich Bonneau, who is the co-founder of the Prescient Design and who is now the VP, uh, at Genentech. And then he started to tell me about the problems that he's interested in in the molecular biology. And as I listened to his description, I started to see this kind of parallel. So, at the end of the day, what we care about in natural language processing is the meaning. Can this system extract the meaning out of the text and then try to

produce a text that's going to convey this extracted out meaning. That's it. Everything about machine learning from natural language processing, and then text is going to be nothing, but at the surface level, there's a sequence of strings. Now, the interesting thing is that in biology, in molecular biology, as well, if I just abstract out every possible detail. There are really important, yes, absolutely.

But that is not really critical for our discussion, is that we have a string of base pairs, and then what we really care about is this very high-level physiological phenomena that we see. And then you can think of this physiological phenomena as a meaning that is encoded by this string of the discrete symbols. Then it's exactly same as the natural language processing, or at least the structure behind these problems are same.

So, we started to actually work on it together with Vlad who is yet another, so, the third, let's say, co-founder of Prescient Design, who was back then a senior scientist at Flatiron Institute where Rich Bonneau was leading the Center for Computational Biology.

And then we really, literally started to add in all our knowledge that we gained from the natural language processing machine translation research into this kind of protein model, and we start with the protein function prediction and eventually protein design as well. So, that's how we actually, I, I got into this kind of AI for biology or drug discovery. So, I would just like to say anytime you'd like to have a beer with me, I'd be open to that.

Hi, and welcome to another episode of NEJM AI Grand Rounds. I'm Raj Manrai. I'm here with my co-host Andy Beam, and today we are absolutely delighted to bring you our conversation with Dr. Kyunghyun Cho, uh, who is the Glen de Vries Professor of Health Statistics and a Professor of Computer Science and Data Science at New York University. Andy, I think this was, I know we've said this before, once or twice at least, but this was really one of the most wide-ranging conversations we've

ever had. KC has been absolutely instrumental in everything from machine translation to protein engineering to some of the recent very impressive work on clinical AI. And he took us through, I think, some of the things I hadn't heard about in the stories behind not only those papers, but also his very interesting start in, in ML and AI and it's just a lot of fun. And I gotta say also just a really, really nice guy. Yeah. I couldn't agree more.

Raj, like, I think that you and listeners of the podcast will know that I'm a history of AI nerd and so KC was at Mila with Yoshua Bengio at the beginning of our current last decade of AI. Some really awesome stories that he talks about what was happening then. And like you said the body of work that he has put together is almost, like, on anyone else that we've had on the podcast. Machine translation, protein engineering. He started a protein engineering company that got acquired by Genentech.

Oh, by the way, some of the early models for language models in health care. And again, just super impressive. And like you said, like maybe one of the nicest, most affable, humble people that we've had on the podcast. So, some of the things that he's done with prize money that he's won is just laudable and admirable. And again, just all-around world-class researcher and like A+ human also. So, I just really, really enjoyed the conversation immensely.

The NEJM AI Grand Rounds podcast is brought to you by Microsoft, Viz.ai, Lyric, and Elevance Health. We thank them for their support. And with that, we bring you our conversation with Kyunghyun Cho on AI Grand Rounds. KC, thanks for joining us on the podcast today. We're excited to have you. Well, thanks for the invitation. I'm very excited to be here myself. KC, it's great to have you on the podcast. So, this is a question that we always get started with.

Could you tell us about the training procedure for your own neural network? How did you get interested in AI and what data and experiences led you to where you are today? So, it was actually purely accidental how I got into AI. What happened is that I was an undergrad computer science major back in Korea studying in KAIST, which is one of the premier, let's say engineering, schools in Korea.

And I was taking a lot of time 'cause I was taking this semester off, that semester off, just hanging out with my friends. You had to drinking, talking, thinking about the future, and all those random stuffs. So, it was final semester and I really had no idea what I was going to do. I was just a computer science undergrad. Knew a bit about software engineering, but nothing else.

And this friend of mine who was taking the course together with me brought this ugliest possible brochure you can ever imagine. It was a yellow, bright yellow brochure. And then the brochure was about the International Master's Program in Machine Learning and Data Mining, what they refer to as a macademia program in Helsinki University of Technology. That became shortly afterward, Aalto University. And then that really hit me. Yes, there are universities in Europe.

I probably should go there and then see how it feels like to study and live in Europe. So, I applied, got in, and then moved to Helsinki in 2009. Of course, the worst time to actually go to Europe and try to get a job eventually. That was right after the financial crisis. So, Nokia was effectively down, the whole Finnish economy was in shambles. But then of course, the university or the department wanted to support the international students like myself.

So, they created this program where the students are going to be unilaterally or randomly assigned to the research groups within the department. And then they're going to study or work on the research one day a week. And thereby they receive stipend and that's going to help the student to stay there and then study. And I was randomly assigned to a neural net group. I did not choose it. I actually did not know about the artificial neuron nets or anything like that. I was just assigned.

So, the first day I showed up at the lab and then the lab was actually shrinking as, unsurprisingly, it was 2009. There were two postdocs, one halftime professor, my supervisor back then, and then another PhD student. So it was, that's it. And I was the one master's student there. And I was told to look into this paper by Geoff Hinton and Roland Memisevic on how to train a third order restricted voice machine in order to learn the transformation or learning the style rather than the content.

I tried to read it. They asked me to reproduce it. I could not even read, finish reading the first paragraph of the introduction. I just had absolutely no background, necessary notes to read the paper. So, I went back, they gave me this book by David MacKay, late David MacKay on the Information Theory, and they asked me to read it, I think it was a chapter 16 or 17 on the whole field network and Bozeman machines. I read it. It was awesome.

I loved the topic, and then I thought, okay, I'm going to implement it myself. And then I was very, you know, fluent in let's say C, C++, and Java, but I've never dealt with this kind of matrix multiplication before. So, I opened up the MATLAB because it says math there. So, I thought that's the right thing to use. And I implemented the whole thing. It was so slow.

I brought it back to the postdoc, Alexander Ilin and Tapani Raiko, and I show them, okay, it showed them my code and then told them that I think it works, but it's so slow. I don't think we can actually work with anything more than 10 data points. And then they looked at the code and then kind of let say, told me how to write this code, or in particular, how to do the matrix multiplication on MATLAB, because I implemented by having three, four loops.

So, I was actually, literally, going through individual element trying to recompute the each and every element of the matrix after the multiplication. Of course, now I know that all we need is just a one-line, A times B, that's it. But I didn't know about it. And then that's when I realized that in fact this is really awesome because there are so many things involved in machine learning that involves the, everything from the hardware all the way to the applications.

And then being able to know, or the AI being able to come up with the algorithm is a superpower that we can actually have. And that's how I got into AI. In fact, it was a purely by chance or in fact I didn't choose, I was actually kind of volunteer to work on this. Yeah, we all do the looped matrix multiplication once before someone tells us, hey, there's this thing called blast that you can use. Absolutely. Um, so super, a great origin story there.

I think a big theme of folks on this podcast is like a moment of serendipity that gets people interested in what becomes their life's passion. Could you talk a, continue the thread? So, you did a postdoc, I think with Yoshua Bengio at Mila and now professor at NYU. Yeah. Take us through that and then we'll hop into some of your work. Yeah, absolutely. So, as I said earlier, the lab was very small.

The deep learning net lab or the neuro net lab that I was there as a master's student, first I decided to stay there as a PhD student getting funded by the academia of Finland. So, I got the funding from Finnish government. So, that was great except for the fact that the whole university decided to ensure that they, because it was economic hardship, was a major theme there. So, they wanted to ensure that they can actually have a sustainable future.

And thereby they decided to choose and focus on a smaller subset of the disciplines when it comes to research. And it was about 20, I think they 11 or 12 when the department decided that, yes, we are doing amazingly well on machine learning, but we are still too spread out. So, we need to actually choose a smaller subset of the areas and focus on it. And then they decided to focus on Baysian machine learning, which I think is a great idea. Kernel method, which I was like, I'm not entirely sure.

And then machine learning applications. And then they decided to ax neural net research in 2011, 12. So, the group actually effectively disappeared. That group doesn't exist anymore, by the way. Yeah. So, it eventually disappeared fully, but effectively back then.

So, I was in a kind of limbo, so my advisor was halftime professor anyway, the postdocs who were actually advising me Tapani Raiko and Alexander Ilin, their contracts were effectively terminated and they had to leave the university by the end of 2011, or 12, or something along the line. And I was like, what am I going to do? But I have a full funding for four or five years, and it has a funding, not only for my stipend, but for travel, almost unlimited conference travels that I could go to.

So, 2013, I went to very first ICLR conference that happened in Scottsdale, Arizona. I never been to Arizona until then. And then it was the beginning of May. It was the hottest summer I've ever experienced. And then the dryest possible summer as well. The skin just hurt. But anyways, so we're put in a hotel, very nice hotel in Scottsdale. I don't know if you've been to Scottsdale, if you are staying in a hotel. And then if you don't have a car. You got nothing to do.

You have to just stay there. And then that's exactly what we did. If I recall correctly, there were about 60 to 70 people at the whole conference, ICLR 2013. And we were having breakfast together, lunch together, dinner together, drink together, listening to talk together, giving a talk together, doing a poster session altogether, for the three to four days nonstop. And then one of the breakfasts, if I recall correctly, I sat next to

Yoshua Bengio. And then I asked Yoshua, so Yoshua, it turned out that I actually lost my supervisors and advisors in Helsinki. Can I actually come visit you as a research student? I have funding. You don't have to provide me with anything. I'll just be there and then do some research together with you and your students. Back then Mila, which was called Lisa back then, had about 30 or so students only. Now Mila had, uh, has about 1500 people. So, it was a completely different time.

And then Yoshua said, yeah, of course, why not? So, I literally just showed up four months later in Montreal and I sat down at the lab. And then one day Yoshua just showed up and then told me, okay, what do you want to work on? And then I said, what do you have in your mind? First, you can continue to work on what you were working on

before, that is a BMO machines, or you can work on a slightly better version of the VMO machine called the Generative Stochastic Networks. No one remembers them anymore, so we don't really have to talk about it. But I thought, okay, but these two are more or less the same thing. And then the third option he gave me was machine translation. I never had any kind of training about machine translation nor natural language processing. And then I ask Yoshua.

But Yoshua, I don't think, you know, machine translation, either. Is this a good idea? And then Yoshua told me that he felt that this is the right problem to solve at that point, like true visionary. And then he really knew what to solve. So, I was like, sure, sounds good. I started to read some of the textbooks and the existing literatures, but then I started to also think about how to do it from scratch. Can we do it from scratch?

If I did not know anything about machine translation and only about how to train neural network, can I still build a machine translation system? That's what I spent about two years since then as a postdoc. Now, the interesting thing is I wasn't a postdoc back then yet, right? I was a research visiting student. And after about five, six months, I realized that this is amazing environment.

There were, I could not think of any other environments where I can actually do machine learning research that matters with the people who actually care about it deeply. So, one day I knocked on the door of the Yoshua's office and then I asked Yoshua, Yoshua, this is great. Can I come back as a postdoc? I'll go back to Helsinki and then try to defend my dissertation and come back as quickly as possible. And you show was like, wait, you're not a postdoc?

Well, that's how I actually showed up, uh, about two months later to become a postdoc officially. And I spent another, let's say year and a half there. After that, thanks to the machine translation research that I was able to do under the Yoshua's advisement, we came up with a lot of interesting ideas. It was really fun time. Every month we could actually, literally, wake up and then think of some ideas, could implement it, and then those things, those ideas, about half of them really worked.

Which is an amazing rate, right? Usually research ideas. My experience is that they, you know, one out of 20 kind of works only, but back then it was like the right time. So, like, one out of two ideas would actually do something that was better than the idea what we were doing before. So, it was really fun. And then because that worked well, I decided to. Uh, go to academia 'cause I was like, it, it's really interesting. And also—. Yeah, could, could I hop in here and just add. Absolutely.

Please. A little context, like amazing stories from the history of AI, which is probably like one of the, like, most exciting decades ever. For folks who don't know ICLR as the International Conference on Learning Representations, when you were there for the inaugural, uh, meeting, there was 30 or 40 people. I think it's like tens of thousands of people now. They get tens of thousands of submissions and it's now like one of the biggest AI conferences on the planet.

Yoshua, your mentor is one, I just crossed a million citations, if I remember correctly. Yes. Making him, like, the most cited scientist in the history of science. And your papers are a non-trivial fraction of those citations. Probably like 10% of those million citations are some of this work that you're talking about now when you were a grad student. So, like, just an amazing set of stories there. Thanks for sharing those.

No, no. Uh, it was really, everything was quite accidental and coincidental. And then I gotta say, I don't know in a sense that this is something that I learned is that they really, you know, we all try our best, but there are definitely a lot of uncertainty in the whole life and the world and how things work to the point that they, sometimes you get, sometimes you don't get lucky.

And in this particular case, for that particular moment, starting from let's say 2012 to, let's say 2017, that was the area when, not let's say large, but reasonable size group of people, including myself, were just, you know, benefiting so much from the picking up of the whole, let's say AI and also by the environments. So, everyone wanted to invest into AI for, let's say, developing something or they're doing some kind of research, but there were only few people who were trained to do so.

Thereby we were just swimming in this kind of, let's say, uh, support that we are receiving. Probably not the bigger support than we actually deserved, but, you know, it was a fun time probably, thereby yes. I'm skipping ahead a little bit, but do you think that we reentered that period with large language models, if you fast forward a decade later, or not quite the same as back when it was in 2012 and for the few years after? Oh, I mean, it is very different.

So here, here's some thought about your idea. What happened, uh, from let's say 2012 or 11 until, let's say 2020, is that the, all these ideas that we are enjoying nowadays, all those coding agents or the LLM for the conversations and so on. All those things are the things that were created out of the research labs, both in academia as well as the big companies.

And then these ideas were in fact immediately actionable, but there was no one who actually have act on it to build any products, and there are no products that people are using. What that means is that we are effectively doing a very early stage product development at universities, and because it was done in universities, we

could do everything in open. And that's very different from the kind of pure product development where everything is somewhat, let's say opaque, because you need to have the trade secret in order to stay ahead of the competition. So, it was that duration where we could actually really build the products that are going to touch upon the billions of people now.

But in the open so that everyone can benefit from, and at the same time, we could still get the benefit both in terms of the finance and reputation and whatnot. So, it's, it was a very unique era. Now, everyone knows how impactful as a product these technologies are. Thereby everyone is building their own product. They are not sharing all the things. And a lot of things are not done in universities, but in companies. So, in that sense it's a very different era.

Although financially, I think everyone is doing pretty well. So, I'm happy for my friends and myself as well. Yes. Yeah, I think that this is like a natural transition point to your research. I wanna just make something kind of explicit that's been implicit. So, this machine translation project that you've been talking about is kind of the early seedlings of what we now think of as large language models. Back in the day, we used to call them sequence-to-sequence models.

And so, an LLM is just a specific kind of sequence-to-sequence model. But you and Yoshua really were kind of like breaking new ground. I think it's easy to go back up the family tree and point to this work as part of an ancestor of what we think about with large language models now. A couple things that you said just really resonated with me that, like, the neural net department got shut down when I was doing, I did a PhD in neural nets around the same time. Mm-hmm.

My advisors always told me that like, neural nets were the second best way to do almost anything. And so therefore, like never the best way to do anything. So, your point about the product aspect of this is interesting because I would say that it really was research because it was already sort of decided by the community to not be valuable. Neural nets were not the way to do anything. Machine translations were all Ingram-based.

They were all phrase-based. Look up things and you were doing something that was actually kind of heretical. Like, it's not, it was not the obvious thing to do that you could stick English sentences, one side of the neural net German sentences and the other side of the neural net. It would learn to map between those two things. Yes. So I think it really was, I think research in, in the like truest sense of the word because you didn't know if it was gonna work out. Yeah, absolutely.

Uh, we didn't know, um, at least I didn't know they was going to work out this amazingly well. Although, later on I actually found out that there were some earlier studies. One from CMU and then a couple them from Spain. Are, are you gonna say Schmidhuber? I feel like you're—. No, actually, uh, interestingly, Jürgen wasn't the first one to do this. Uh, this is one of those few things where Jürgen did not actually spend too much time on during nineties.

There were a couple of the papers from the early nineties from CMU and then a couple of the other groups in Spain, in fact, worked on you training a neural net, nothing else, just a neural net to map from one sentence in one language to the corresponding translation in target language. And I actually found that out a bit later.

And then this is one thing that I tell my students as well is that the, my students are obsessed these days to catch up with every paper that is being put on Archive or that that is being, let's say, uh, posted on Twitter and whatnot, or the X now. And what Goeff Hinton used to say, I don't know if he still believes it this or not, is that the, we shouldn't read too many papers.

The reason is papers tell us about what others have done based on the circumstances they were in, and also limited by the situations they were in, as well as their own kind of thinking process. And what that means is that even if the idea looks exactly same, that doesn't mean that that idea was executed in the same manner, and then the outcome may actually change.

So, if I actually had read these papers from nineties as well as the late eighties early on, and took them, or took their result at, at the face value, then I wouldn't have actually tried this out because they said that yes, it works, but it cannot be scaled up. They couldn't do it.

And then there were issues with the generalization or the lack thereof, but because I didn't know about them, but I knew about all those neural nets and the base foundations as well as the trust that this technology was going to work. Uh, we could actually work on it and then eventually made it work. So, execution was a bit different. And of course, situation was very different from nineties. Yes. It is an argument that there's more to unlearn, right?

And that there's almost like suspended belief, right? That you have to— Mm-hmm. —you have to shake away by not overfitting to what's already been known or discovered in the literature. And also, this is not to mention the time lag between when you read something and when it was actually done, which could be many years in some disciplines too. Um, absolutely. Andy, I think maybe it's a good time now to transition into, uh, to KC's work.

Yeah. Well, I, we've been touching on some of it a little bit. We've been with you at the beginning of the large language model explosion. You've continued a lot of that work, but you also have this interesting work in biology and drug discovery. You are a co-founder of a company called Prescient. You've done a lot of work in diffusion models of biology and things like that. Could you talk about how you got interested in biology and drug development and what led you to that?

So, it was about 2017 or so, I noticed something weird about working on this kind of neural net for language processing or machine translation. If there is a problem in machine translation, let's say I want to build a very low resource machine translation system, and there was a very easy way to make it work better. Either I'm going to collect more data or mix in some of the data that are relevant but are that are not directly about this problem. So, I'm going to increase the size of the data.

The problem was solved better, and if I couldn't actually get further, then what I had to do was to make the model larger. I would make the model larger and then make the model solve more than one problem. So, we made a multilingual machine translation system, and then every time we made a model larger and then make it solve more problem, everything got better.

And then even after, you know, sometimes we see that, okay, it's still not working as well as we want it to, uh, at the level of what our desire than what we do. We're going to just let it train longer. Literally, we had to have a bigger patience and then it would work better. And then that's when I realized that the, this is probably not the thing that we should do at universities anymore. This is effectively the product development.

If there is a clear, predictable improvement based on the predictable amount of the investment, either be it computation or the data, then what that means is that the, you can now make a prediction on the financial aspects or the returns, all of this product building and then deployment. And then when that happens, all you need to do is to build a company or sell it to a company and then start making an actual product out of this technology.

And the universities are lousy places to do any kind of product development, as almost everyone knows. So, I started to actually explore a bit, what are the other problems that I want to solve? What are the other problems that require the breakthrough that may be brought on by the similar set of the technologies that we've been working on. And then I, I even worked on some, a bit of a particle physics, text processing for political science, and on and on and on, and eventually I dabbled on the

protein modeling. Because I was having a beer with Richard Bonneau, Rich Bonneau, who is the co-founder of the Prescient Design and who is now the VP, uh, at Genentech. And then he started to tell me about the problems that he's interested in the molecular biology. And as I listened to his description, I started to see this kind of parallel. So, at the end of the day, what we care about in natural language processing is

the meaning. Can this system extract the meaning out of the text and then try to produce a text that's going to convey this extracted out meaning. That's the everything about machine learning from natural language processing, and then text is going to be nothing but at the surface level, just a sequence of strings. Now, the interesting thing is that in biology, in molecular biology as well, if I just abstract out every possible detail that are really important. Yes, absolutely.

But that is not really critical for our discussion is that we have a string of base pairs, and then what we really care about is this very high-level physiological phenomena that we see. And then you can think of this physiological phenomena as a meaning that is encoded by this string of the discreet symbols. Then it's exactly same as a natural language processing, or at least the structure behind these problems are same.

So we started to actually work on it together with Vlad who is yet another, so third, let's say co-founder of Prescient Design, who was back then the senior scientist at Flatiron Institute, where Rich Bonneau was leading the Center for Computational Biology. And then we really literally started to add in all our knowledge that we gained from the natural language processing machine, translational research into this kind of protein modeling.

We start with the protein function prediction and eventually protein design as well. So, that's how we, actually, I got into this kind of AI for biology or drug discovery. So, I would just like to say anytime you'd like to have a beer with me, I'd be open to that 'cause it seems like some of these amazing stories have always been, uh, instigated by having a beer with KC. So, I hope we get to do that at some point. Absolutely, yes. Um, so I'd like to follow up on that.

First, let me plug your blog. You have an amazing blog where you have lots of really insightful technical things, but also personal things. I think Raj will touch on one of those in a bit. You did a nice job at summarizing the technical motivation for what led you from machine translation to protein engineering. Actually, I'd been working in computer vision and there was lots of overlaps and that was reminiscent of my own journey.

But you have a blog post, or you have a tweet that was drug discovery maybe in the Cold War era. Um, could you walk me through exactly what you meant by that? Right. It turned out that there was actually a mis, uh, received or the misinterpreted in all possible ways that I did not even imagine. But what I meant was trying to make a parallel between the early years of the machine translation research with the drug discovery.

So, if you think about the origin of machine translation, it actually comes from the beginning, uh, it started at the beginning of Cold War. So, of course, the American intelligence wanted to read all those texts, as well as the materials that are produced by Soviets. Of course, they're all produced in Russian as quickly as possible, but relying on the human translator takes too much time. There's too much delay, and then there's a chance that there's gonna be a lot of error.

And obviously Soviets wanted to do the exactly same thing. So, what was the first thing that they did was to hire a lot of linguists as well as the polyglots and then trying to extract out the rules behind machine translation.

And then what they thought was that in order to solve machine translation, we are going to extract out the rules for each of the languages separately, and then we're going to figure out how the rules from one language translates to the rules of the other language, and then eventually we're going to reconstruct the same sentence or the sentence that conveys the same meaning in the target language following this rule books.

Now, thinking about from this angle, it's very natural that they hire tons of linguists and they hire tons of polyglots and then try to figure out these rules because that's how we think we actually approach the problem of machine translation or the just translation, in our case, I guess. I read the Korean text trying to get the meaning out of it, and then based on the meaning, I'm going to write the same text in English. But then obviously that one did not work at all.

We actually improved our understanding of individual languages dramatically during that time because there are a lot of funding for computational linguistics as well as the people who can speak multiple languages. And then they produced a lot of data, but those data wasn't that used as, you know, that kind of as it goes against what we think now. But back then, data was a side effect. They were trying to come up with this kind of rule books.

And then if you think about it and then try to make a knowledge here to the drug discovery, what is the first thing that any biotech or the pharmacist do is to hire biologists. Hire very good biologists and hire very good medical professionals.

And, in particular, when they hire biologists, the reason why they hire biologists because they think they or we all believe that knowing precise, uh, having a precise knowledge about our biological knowledge, about how things work at the lowest level possible is something that we really need in order to design a drug. And then that sounds exactly like what happened with the machine

translation during the Cold War era. At the very end of the Cold War, so in 1980s, late eighties, IBM in Yorktown Height decided that the, well, what if we actually hire statisticians and physicists and mathematicians, and then try to approach machine translation directly, not by understanding language, but by looking at it from a coding. In fact, English sentence is nothing but a encrypted version of the original sentence that was written in another language.

Then what we do is we're doing a decryption. If you approach it from this angle, there's nothing language or linguistic about it. It's all mathematics. And then that was the beginning of the statistical machine translation.

And I feel like for the drug discovery as well, what we need is that to, is to acknowledge that a lot of what we are doing, what we have done so far has been about understanding biology, but understanding biology may not be the reason how we are going to change or the is not going to the factor that's going to change the drug discovery. In fact, drug discovery is a drug discovery. We need just approach it much more directly. That's what I wanted to convey in the blog post.

But of course, everyone read the Cold War and then, you know, everyone started having all those crazy thoughts about the idea, what it means for anything to be in a Cold War era. So, it was an interesting experience I had. There are some geopolitical sensitivities around drug discovery right now that you may have accidentally backed into, um, uh, with that. True. So, I understand the reaction there. Mm-hmm.

I think that's like one of the most fascinating and nuanced ways to think about, like, obviously the knee jerk reaction would be, like, of course you need domain experts to be able to do things. But fundamentally, you may be using the wrong set of abstractions to solve a problem if you don't bring the sort of statistics and computer science and physics perspective.

Maybe on the other side of that, though, I have thought that pure statistical or machine learning methods have not advanced drug discovery maybe as much as we would. So, machine translation, huge progress. Modeling language in general, huge progress. Language models of DNA and language models of protein, and even, like, generative models of protein structure and things like that have pushed out at the margins.

But we have not seen everything inverted in the same way that machine translation, for example, has been inverted. What are your thoughts on that? No, you're absolutely correct. Well, you know, in some sense we made too much progress in language processing to the point that the progress in any other domain seems to be tiny. That gives us a bit of a distorted, uh, perception of what's happening.

But the major difference I see, and the major factor behind this tremendous improvements in language processing has been the availability as well as the collection of the data with the right purpose. So, in fact, many of the technologies or the algorithms that are used to train these language models already existed for some of them many decades.

But what has changed during the past few years is of course the availability of the compute as well as the, uh, let's say financial resources, but also thereby collecting the data that are relevant to the problems that we want to solve.

So, in fact, I'm pretty sure I'm just speculating here, the amount of data OpenAI has collected by pushing out ChatGPT-2 users, however noisy, that is is probably a way larger than any of the academic scale corpus that we have ever used in the entire history of the natural language processing research. You think that OpenAI has a bigger internal data set than like the pile or the common crawl, or?

No. So, that's actually the point is that the, but the pile open, uh, the common crawl and so on, those are actually the data that was collected without purpose. They were collected with a different purposes than the purpose of the data purpose what we have in our mind. Now, what OpenAI has by deploying these systems and have these systems interact with the millions, if not hundreds of millions of users, and then getting the feedback from them.

Now, all those data is a purpose-driven data that makes a huge difference, and that's probably the reason why many of these so-called Frontier Labs that put out the amazing models are the ones who tend to have a large number of the clients and customers as well. How can Claude be so good at coding? Because they actually have pushed out their models.

Anthropic has pushed out their models in the applications of the coding, and then they collect all those feedback and the data that allows them to bootstrap their model to become better and better. Can I propose something that maybe is contrary to your thesis? Just 'cause— Please. —I'm curious how the reason Claude has gotten really good at coding is because everyone at Anthropic is a coder and so they have tighter iteration cycles where they can internally do something and then push it out.

Whereas that's why maybe they are not as good at science or drug discovery because they don't have the internal, like, quick feedback mechanisms that they have for areas of their own area of competence. You're absolutely correct, although Google employees are going to feel a bit sad about that comment because somehow the Geminis, Gemini models, I use them very, like, obsessively, but are not good at coding. But as far as I can tell, Google also has a lot of software engineers.

Maybe they're good at product management. I'm just kidding. Sorry. I, no, that makes sense, so and then yeah, but of course the difference there difference to let a difference in drug discovery or the biotech and pharma is that we are very careful. And then for good reasons. We cannot be at the I don't know, willy-nilly and would just inject these people with that. And that now a lot of the people in Silicon Valley these days are going crazy with this kind of Chinese imported as peptides.

I don't know why they're doing it. My guess is that if they remove the GLP-1 that's going to be, it's going to be just a water. But still some people are doing that. But generally, there are some good reasons why we are being very careful, but then we are actually being careful to the point that we are not, we are being too conservative in our way of operation as well. So, what I think is that, yes, these models are great.

We are building all these models based on the various, say, molecular data as well as the scientific articles and whatnot. But what we really need is the data that is collected with these models in the loop. So that we collect the data with the purpose that aligns very well with the technology we have and then the problem we want to solve. And then that is the part that is actually largely missing these days.

Amazing. So, I just wanna transition to yet another one of your I think very wide range of pursuits and topics that you've tackled over the last decade or two. This was a paper that I read a few years ago. I think it came out maybe back in 2023, if I'm not mistaken, in Nature. And the title of the paper was Health System-Scale Language Models are All-Purpose Prediction Engines. So, you did a few very interesting things in the paper.

One of them was that I think you coined the term NYUtron for the name of the model, and you really got people talking. I can remember this. Everyone was talking about this.

We were reading this up at one of the Harvard AI research seminars or something we were discussing in the paper, and everyone was talking about the value of fine tuning, using electronic health record data, taking the sort of base foundation models from OpenAI from others, and then asking the question, which I think a lot of us were interested in, but you were very fast at doing and set up around interesting tasks, was essentially what is the value of the sort

of local data to localize or to— Mm-hmm. —you know, further specialize the general purpose model. Where I want to go is to jump ahead, like, what do you think now, but maybe before, I'm gonna resist my urge to do that, and just ask you to maybe set the stage for us for that paper back in 2023. What was your motivation in actually going down that, that path? And maybe just set up the paper and tell us some of the top line findings as you see them.

Yeah, so I mean not a lot of people know, but I've been actually working with the people at NYU Langone health system ever since 2015. When I joined NYU, one of the first things that I did when I joined NYU to go to the NYU radiology and then start knocking on the doors and then trying to see if any one of those radiologists would be open to collaboration.

Because it was so clear that we had amazing convolutional neuro net-based classifiers back then that it was shame, I thought, that they were not using this technology for a large scale screening of the various types of the modalities. And yes, I went to radiology, and then some of them, such as Linda Moy actually responded very positively. And then we started a whole collaboration on using convolutional neural net for the breast cancer screening and diagnosis.

And then that went on for many years. And then during those years, what I realized is that at least I was thinking about AI for health care from a completely wrong angle. What I thought, and then what I think a lot of the computer science still thinks, is that the AI is going to come up with the amazing way to make a diagnosis on extremely difficult to diagnose, the cases, or the symptoms, or the diseases, so that we can find those rare cases.

And then AI is going to tell us exactly what to do in terms of treatment, in order to make the patient healthier, that particular patient. And then that's what I thought I was doing and that's what I thought we were going to do. But as I spend more and more time together talking with the clinicians at NYU Langone, as well as just generally looking into what is happening within the hospitals, it became very clear that in fact, health care is a constraint optimization problem.

And then what we really need is the predictor that's going to help us optimize it better. Because one thing I realized is that no one opts out of health care. In fact, in modern society, everyone is born inside health care. Even if they are born at home, there are always the midwives who are all clinicians as well under the health care nowadays. And then everyone dies within health care. And then everyone, including myself, becomes a patient at one point or another during their lifetime.

And then we all go to hospitals, clinics, or whatnot. So, we have to somehow solve this huge problem of optimization 'cause everyone is part of it. And in other words, everyone needs to pay for everyone else's, let's say, health care bill. Now, how do we actually ensure that when the client population is the entire population, we can provide the best care to each and every one of them given this constraint.

So, that's why I started to think a lot about the prediction, not in terms of let's solve the problem that we don't know how to solve, but more like, can we solve these problems just a little better so that overall optimization problem can be solved better? And then that's that's when Eric Figueroa joined the NYU Langone as well. So, we were interviewing Eric and then Langone wanted me to actually talk with Eric as well. We talked, we hit it up immediately. I told Eric, please join NYU.

And he did, thankfully. And then we started to actually look at

all our problems within the health care system. Beyond what our individual, let's say, specialty were. And then that's how we decided to think about it. And the one thing that we realized is that, well, what is the most, the richest possible source of information about any patient within the hospital? It is the clinical notes.

Because every clinician is trained and incentivized to write the note as carefully as possible so that the future clinicians, whoever they are, who are going to read these clinical notes, will be able to pick up on what happened to the patient as quickly as possible. So, it's almost like a highly enriched feature set of the patient. And of course I was working on language model quite a bit. Eric is always on top of everything that is happening in the field of AI.

So we, and then we admitted a Lavender who is a PhD student and the first author of the uh, NYUtron. And then she was very motivated and also mission driven to the point that she actually did everything from beginning to the end. Training infrastructure, she set it up. Training codes, she set it up. Crawling all the clinical notes from Epic, she actually set it up. Cleaning up the data, she did it. And then she even trained it and tested it.

So, with the three of us, and then, of course, a lot of our collaborators, we were able to create it. And then there was a point at which we had to decide which task we were going to look into. Eric told us that, oh, let's try the readmission prediction. And then I gotta say, my initial reaction was that, oh, that sounds very boring. I mean, is this really something that we need AI to do?

But of course, whenever the patient comes back in within 30 days for the same cause for which they were admitted before, obviously they're taking up the, uh, let's say very precious, let's say bed in the hospital. Also, taking away the time from other patients. So we want to minimize that if you could, uh, minimize it.

And then predicting whether the patient is likely to come back is going to help us develop further, let's say intervention measures to minimize the kind of, let's say, readmission. And then, you know, I got sold and thereby we actually, uh, worked on it and then wrote a paper. Yes. That's very, very helpful. And so, maybe can you tell us about the very, sort of quick version of what the main finding was?

So you were able to take a large amount of EHR notes from NYU and then do what with what model and how does it compare to the, the other models? Yeah, absolutely. So, the paper was published in 2023, but as you can imagine, uh, the whole work was done by the end of 2021 and the beginning of 2022. So, just take us back in time. This was, this is now ChatGPT came out? Yes. During this time, during the time you're working? No, actually it was a pre-ChatGPT. Okay. Wow. So is, so yeah.

The paper, well, the paper was published after ChatGPT came out. Right. Yeah. So the right after ChatGPT came out, but all the work was done before ChatGPT. Right. So, it is a very interesting era. Exactly. That's what I'm getting at. That's fascinating. You guys started the project before the release of ChatGPT. But then Nature— Yes. —the Nature paper came out after ChatGPT came out. Exactly, exactly. And then, yeah, there. So, what we did was to reuse the idea from the BERT.

So, BERT is a mass language model, and unlike the language models that are prevalent these days, that is going from left to right, trying to model one word at a time, given the entire context. This mass language models look at the entire text altogether and then try to figure out some of the words or the phrases that are missing in the input.

Uh, what it learns from the process is how to figure out the meaning of the missing tokens by looking at the neighbors and also, inferring the meaning that is induced by the competition of these words. And then it was shown to be extremely effective and many language processing problems in 2017 and 18. So, this is a model that was proposed by the team at Google back then. So, we started from this backbone.

We didn't use the pre-trained model, but we started from the fully, let's say, randomly initialized version. And then we collected all the clinical notes we could from the NYU hospital. And then we started just training it as if clinical notes were just a usual natural text and then we continued to train it. And then once the model was trained, fully pre-trained, we actually fine-tuned it by adding a classifier on top and then training to solve many of these operational tasks.

And then we focus on the many of these operational tasks because those are the tasks that we believe and we continue to believe are really important if we wanted to really make the health care efficient. Thereby, we can actually provide a benefit to increasingly more people. And then it worked really well.

So, and then what we mean by worked well is that for each and every one of those tasks you can, and we can still build amazing predictor, bespoke predictor by looking at the EHR very, very carefully, figuring out what are the features that we think are going to be really important, and then mix and mesh those features using a highly complicated, let's say classifiers such as the grad boosted trees or the neural network, and then

train the hell out of it so that we know that this one is going to work well. But that's not that great, because if you think about the operational or the clinical test that we want to solve eventually, there will be hundreds of them. We are not going to be able to build hundreds of the bespoke classifiers and then maintain them over time. What we wanted was just one standardized way that we can actually use to solve all these problems.

And then indeed, the NYUtron actually showed us that it is possible. Just taking me back to Andy's comment about second place and pretty much everything, but in, in this instance, you wanted it to be the, the best, but have the sort of generality to, to function on many of these tasks and— Mm-hmm. —the baselines in that paper were the sort of standard, you know, gradient-boosted models, right?

Yes. And in general, what you found is that there's a pretty nice, AUC lift for most of your, uh, your prediction tasks. Mm-hmm. Um, so, so now let's fast forward a few years. So, we're in 2026. It's, you know, uh, two, three years later after the paper came out. First, I'm curious, like, is NYUtron still an active research or, or clinical activity? Is it being used that way?

Um, or, um, whether it is, whether it's not, uh, have the sort of intellectual lessons you've learned around kind of fine tuning these types of models with the HR data? Have they shifted at all? 'Cause everything is an agent, uh, agent-based workflow era that, that we're in or that everyone seems to, to want to run towards right now. Yeah. So, uh, we've been actually—. uh, are we still fine tuning? Are we still fine tuning? You're absolutely correct. We still do need to fine tune it turned out.

So, Lavender, uh, of course, uh, has been continuing on this line of the work. She just put out the pre-print a couple of weeks ago where we actually present, we call it LANGE1. So, LANGE1, but a Langone kind of thing. I like it. Yeah. Auto regressive model this time, left, right generation and then much larger than before. So, the NYUtron had about a hundred million or so parameters.

But now LANGE1, we trained that actually for variety of models, ranging from the a hundred million parameter models to the 7 billion parameter models, trained only about 10 to 20 times more clinical node collected from the NYU Langone. And then what we do is that because it's now the order regressive model, we train on all these five or six tasks that we have selected simultaneously. So, it does become a conversational multitask model and it works pretty well.

And in particular when the number of the label data is small, this model works beautifully. And then compared to the existing open source, uh, off the shelf models, even when we fine tune the LANGE1 tends to work better.

So, this line of the work is going well, and because we are in a health care where we have to be extremely careful, I think this, uh, way of doing a fine tuning and careful modal selection, this loop is going to continue to happen regardless of what kind of, let's say, research paradigm we're going to. Can, can you share what the base model is for that? So, base model, we took, uh, the network architecture from Llama family, but we actually trained everything from scratch.

again. We actually tried to from scratch, continue pre-training from the already pre-trained model as well as just fine tune the pre-trained model. We compare all of them and it turned out that at every stage, if we mix in more data that is clinically relevant and then also fine tuning on the ally relevant tasks, they all work better.

And then it turned out many of the off the shelf, let's say, are commercial language models cannot really solve these clinical problems well enough because these are, clinical tests are not the things that people actually talk about. It just happens in the way of operation within hospitals. So, they tend to make a really weird, uh, errors here and there. That makes it really difficult for us to even compare them to many of those, especially, uh, crafted models.

Although I'm very, uh, you know, like, I'm very bullish about these kind of general-purpose models, but I feel like there are areas where we need to really go deep and sharp and then health care is probably one of them. Amazing. Alright, uh, one more quick, quick topic for us and then I think we're gonna jump into the lightning round. So, I think Andy was talking a couple moments ago about your blog, uh, which is, uh, really great.

So, it's actually great to chat with you because I've been a big fan of some of the posts over the years. Um, and one that, um, we wanted to talk about just, you know, maybe just, uh, quickly here before we jump into the next segment is this post that you had on, I think you called it MyChartExplorer: A Vibe Coding Journey Towards Accessing My Own Medical Records. Um, so can you tell us about that? What you were trying to do and, uh, maybe what you, what you learned?

Yeah, so, uh, by the way, mychartexplorer.com still online. So, you go there, sign up on the waiting list, and then I'm going to actually send you the, uh, access and the API keys so you can try it out. So, one thing that I find both fascinating and frustrating is the fact that the medical records are all of the patients, but patients don't carry their own medical records.

And of course, it's for mainly the historical reasons, is that, you know, historically people were not good at keeping records. You know, records are all written down on papers. If you lose them, they're gone forever. So, what the physicians did was to actually keep the records on behalf of the patients at their offices, and then if necessary, when requested, they're going to make a copies to the patients.

So, because of this kind of historical reason, all the medical records, wherever in the world you go to, are maintained by the hospitals. But one thing we know, and or at least I have learned, is that the hospitals are good at providing care. Hospitals are not good at software engineering, nor the data management. It's not their job.

And then thereby in these days where everyone has smartphones, so smartphone accessibility is in fact greater than the health care accessibility throughout the world. And then we have access to all those cloud storages. It's weird that even still now, all our medical records are spread through across the different EHR systems as well as the different clinical systems.

It's worse if you're an expat like myself. I was born and raised in Korea, lived six months in Melbourne as an exchange student, lived in Finland for five, six years, Canada for another two years, and then now I'm here. You know, like, and then you at the, of course, every single country I've been to, I've been to some of the clinics for various reasons, and then I don't have access to any of the medical records they have of me.

So, I got a bit frustrated and then I thought, okay, so I'm gonna just download everything. And then download is actually legally mandated, at least in US. So you can download your medical records from any of the EHR vendors, all of the, uh, used by your hospitals or the clinics. You can download them. But downloaded files are not at all easy to parse. It's just a massive XML files you get. No one can read them. And then the XML file schemas are so massive.

There is no way anyone that is not trained as a data scientist nor clinicians can actually read any one of them. So, I actually downloaded all of them. I couldn't do anything. And I was sitting down and I was like, well, you know, everyone tells me that the vibe coding really works. Coding agents works well, let's try it out.

So, I spent about, uh, two, three days working with the Gemini, you know, like the GPT, and the Claude, all of them, all together, going back and forth, building the full lab, following the instruction, coming out of these LLMs, and I could actually make it work. And then that's when it hit me. It said, yes, this is a time for us to actually take control of our own medical records. And then that's really going to change.

That's going to be one way to change the health care because a lot of the inefficiencies as well as in fact, the lack of capabilities come from the fact that the, all this, the data about ourselves are fragmented and are not collected well enough. So, that's the reason why I actually did it. It's still there. You can try it out. Let me know. I'll give you the access. Awesome. KC. I love the empowerment that vibe coding tools are gonna give patients going forward.

And obviously someone with your level of expertise will be showing what the Frontier can do. I think Anthropic announced that they had this huge vibe, coding prize and like a cardiologist was like the third place, um, person here. So, I think that it's gonna democratize access to health care data in a big way. So, I think we're gonna transition now to the lightning round.

Um, these are gonna be a bunch of rapid-fire questions on various topics, and you can give short answers to them, but try and refrain from giving overly long answers. Okay. Are you ready? Alright, let's try. Sounds good. Some of this was researched with help of Claude, so any hallucinations are Claude's fault, not my own. Uh, but we're gonna take a shot at some deeper cuts here. So, I believe that both of your parents are language experts. Yes. Have you ever shown them an LLM?

And if so, what do they think of it? So, my mom is a big fan. So, my wife, uh, doesn't speak Korean at all. So, my mom speaks only Korean, but somehow they can talk to each other on WhatsApp group these days. And it's all thanks to LLMs. And I'm so grateful. But also at the same time, I feel a bit threatened because my job as the someone who actually controls the information between my parents and my wife has, has essentially disappeared.

Can, can, do you ever go your wife and say, I invented the technology that allows you to speak to your mother-in-law? Do you ever play that card? I actually do. Perhaps a bit too much. It's a bold, bold strategy there, KC. Alright. Awesome. Next, uh, next lightning round question. You've worked in academia, big tech, and pharma. If you were starting your career now in 2026, where would you go? University, academia, still.

Can I maybe, uh, exercise my prerogative and just ask maybe very briefly, why? Uh, academia is where I can actually learn and to do things that are beyond all three of these options that you just mentioned. Perhaps by going into academia next time I'll be able to go into some completely different field. Yes. Cool. So, this question will be a self-consistency test with the previous question. So, have we entered the singularity yet?

No. So, that's a self-consistent answer I think with the previous answer, because if you had said yes to that, I think that. That's true. Cool. Follow up. Do you think that we will within the next five to 10 years? No, because I don't think the singularity, uh, exist in terms of the intellectual activities. We'll follow, I think we'll follow up on that at the, at the big picture. Okay. Um, so put a pin in that for now.

Mm-hmm. KC, will AI and medicine be driven more by computer scientists or by clinicians? By clinicians. Because they'll become computer scientists themselves. Do you think we're seeing the early, the early parts of that with this, like, vibe coding cardiologists and things like that? Absolutely. Absolutely. Nice, nice. It is funny how much that has flipped my own opinion. Uhhuh, um, 'cause I used to say computer scientists, but now the programming is the easy part. True. It's the taste.

The taste, the taste, right? Mm-hmm. The taste and the real problems. Yeah. Yeah. Cool. Okay, so the next question, uh, you received a prize called the Samsung Ho-Am Prize, and you donated this entire prize to scholarships named after your parents, one for a woman in CS, and one for classic Korean literature. What drove that decision? So, for the first one.

I feel like we need to ensure that whatever comes out of this kind of AI research is going to benefit everyone and provide opportunities to everyone. But when we see who are actually entering these fields and then benefiting from these fields or the outcome from these fields, we do see that it's not uniformly across the different demographics, different sects, and then different, let's say societies. So, and then one particular aspect that is very clear in Korea is this gender imbalance.

So, I wanted to really, uh, make sure that the, this time around, so my mother's generation, of course know they did not really benefit anything from the economy, let's say a booming in South Korea and so on, if you're a woman. But this time around, I want to make sure that it is not going into the case. So, that was for my mom's, uh, the scholarship named after my mom.

Scholarship named after my dad, one thing that I find really frustrating is how much, let's say many of the governments all over, all around the world, including US, are putting increasingly more money on AI research when there are so many companies that are already making tons of money out of AI. And then those are the people who need to invest in AI research.

What are the things that we need to actually invest in are the important things such as the humanities and all those literatures that are not directly tied to this economic activities, but still enriches us. And then I feel like that's our obligation as AI researchers and you know, AI, AI researchers and developers to in fact support this kind of activities that are not, let's say, driven by economic value only. Yeah. Amazing.

Alright, the last lightning round question, KC. What's one book everyone in AI should read that has nothing to do with AI? It ties in pretty well to your previous answer, too. Yeah. So, uh, in fact, uh, one, one, I mean, I actually don't read that many books. I used to read a lot of books. Uh, but one book that left huge impact on me was the One Hundred Years of Solitude. Uh, I read it when I was in high school.

Just the fact that this, someone could actually wrote a whole book in that manner and then it was a very thick book that I had to read overnight, nonstop, from beginning to the end. That just left so much impression. Not the content itself, the fact that the, some writing can make people stay up and then all the way could not just stop. I think that that was really powerful. And then being able to write like that. I don't know how he did it.

Of course you, but I think that that's what we can actually see. And then you enjoy the beauty of it. Yes. An AI researcher embracing magical realism is fantastic. So true. I love it. Alright, I think you, uh, you did very, very well. That was, that was a great, great lightning round and thanks for, uh, the short, but I think very, very memorable answers. Thank you. Cool, KC. So, maybe just like one set of questions, uh, to close us out.

We touched on it a little bit in the lightning round, but I think you've been very vocal about pushing back against AI doomerism and AI hype. I think you've said that existential risk discourse is sucking the air out of the room. So, you have been a key player in this last decade of AI and you've trained under Yoshua Bengio. You've been at the epicenter of this. So, what should we actually be worried about in, in AI?

So, in fact, there are so many things that we need to worry about actually today, not for, let's say a hundred years from now or the, uh, you know, couple of the centuries from now. And then I cannot, I'm, I'm not going to tell you exactly what those are, but I can, I can say one thing that we are not doing is that whenever we have this discourse about the danger or safety of this AI-driven, let's say developments, we are always missing one group of people.

And then that group of people are the ones who are being subjected to these AI driven systems, or are, uh, who cannot opt out of using these AI based systems essentially, unlike in let's say, uh, NIH panels or the FDA panels and whatnot, we actually don't have the, essentially the patient advocates. There's a user advocate or the users who are actually involved in this whole discourse around, let's say, AI's you at these danger or the safe or whatever.

So, in that sense, I feel like what we really need now is to ensure that there is a representation of the users and the people who are subject to this AI technologies. When we have this kind of discussion, it's not easy. It's very difficult. And then new idea, we all are very frustrated whenever we have to actually ensure that they will, in this panel, we have to have a patient advocate and so on, right?

'Cause they'll bring up the things that we don't really think about or bring up the things that we almost feel like are actually hindering the progress. But there is a reason why that has to happen because we are now touching upon everyone, billions of the people.

So, the reason why I said that, you know, I did this kind of existential risk, uh, discussion, is sucking the air out of the room is this is a room that is filled up with the actual people who are being subjected to this AI technologies every single day, and now we're sucking the air out of, out of this room so that, that these people actually suffer on the spot. So, that's what I meant. Yeah. I think that's a super important point.

Maybe if I could push in on that a little more to understand exactly the kinds of things that you think we should have in place. So, for drug safety, um, there's post-marketing surveillance and pharmacovigilance that has to happen so that when a drug is released on the market, adverse events and safety incidents are reported. Is there something like that that you think we need for AI? Is it something more upstream of that?

So, like, patients being involved in the development process directly of an AI model, or where should we, where should we inject some oxygen back into the conversation? Right. So, we want to have both of them. So, the first one has to be done at the legislation level. So, we need to have this kind of regulatory framework that compels these companies to report back to the, let's say, federal authorities.

Or there is a public forum based on the user safety experience or the users reporting directly. And then this is actually very similar to what already happens in many other areas and perhaps is possible that demand of the existing regulatory framework. Such as the one for the privacy, right? If there is a privacy violation, they are actually compelled to report everything and release it and then trying to figure out how to compensate the users as well.

So, we probably have the kind of framework, but we can make it sharper for AI-related products. Now, second part of course, I don't think it's a good idea in general to force how each and every company operates and how the products are developed.

However, I hope that the many of these large companies feel that they do actually have a responsibility toward the society that enable them to grow this big and being able to, to develop these products so that they're going to voluntarily, in fact, involve more of the users and the potential users and the potential people who may be harmed by these technologies from the early on. I want them to do it voluntarily.

I mean, I know it's a difficult thing to happen, but hopefully there is a sense of the responsibility that is still left in this society. Awesome. I think that's a great place, um, to stop it. So, KC, thanks so much for coming on. It was great to hear— Well, thank you very much. —both about, you know, being close to the, um, center of the current AI revolution for the past decade. Again, I was joking with Raj on the side.

I feel like a poor man's KC 'cause I've worked in health care, protein design and, and neural nets. And, uh, you set a super high bar. So, really impressive body of work and thanks for coming on to talk about it with us. Thanks for the invitation. It was great. And then Andy, what do you mean? Uh, I look up to you. You know, don't forget this. Okay. Thank you. Uh, KC, uh, it was an honor meeting Andy Beam squared. So, this was just, uh, completely fantastic.

Yeah. For Andy Beam. Greater than one. Just to be clear. For Andy Beam. Greater than one. So good. Oh man. You guys are both great. Alright, thanks, KC. That was fantastic. That was great. Thank you very much. Thank you. Sure. This copyrighted podcast from the Massachusetts Medical Society may not be reproduced, distributed, or used for commercial purposes without prior written permission of the Massachusetts Medical Society.

For information on reusing any NEJM Group podcasts, please visit the permissions and licensing page at the NEJM website.