MIT Technology Review’s EmTech Conference: What’s Next?
To help us with that, we have Peter Lee, president of Microsoft Research. Peter leads teams that incubate AI powered products and business models, any one of which may be tomorrow's next big thing. Before joining Microsoft in 2010, Peter was a research scientist at DARPA and before that professor at Carnegie Melon. He has a special interest in how AI can be applied to medicine. And last year,
Peter was named by Time magazine as one of the 100 most influential people in health and life sciences. Peter, welcome to MTech. Thank you. I think you've got a presentation to start us off. Sure. A
short one. Thanks for having me here. It's uh really a thrill and uh I'm I am a little bit nervous because I think what I was asked to talk about is what's next in the next one to three years in AI and I honestly the more I've uh fallen into the AI rabbit hole I more I feel I don't even know what's coming in the next one to three days and so uh it's a little bit intimidating. Um, and part of the reason I think uh that it's intimidating is we're in a state uh and the way I try to explain this to my wife is to imagine that you're in an alternate universe where electricity hasn't yet been uh delivered to people's homes. Uh you know people in their homes are using candles or fires to to light themselves uh you know light their environments. And now someone invents copper wire and discovers that it transmits electric current incredibly efficiently from point A to point B. And we just know in our guts that
this is going to transform life for the better for everyone. But we don't yet know why because no one has invented the light bulb. And now we have gigantic huge multi-t trillion dollar companies transforming themselves into copper wire companies stringing copper wire down every street into every home around the globe confident that those light bulbs are going to get invented. And not only light bulbs, but someone will figure out that you can wind these things around pieces of metal and make electric motors and maybe even someday transmit digital signals on these things. And that is basically the in a nutshell the situation that we're in today. And that's not to say that Microsoft isn't making a lot of money on AI. We are um
but the true value of AI I think we haven't yet imagined and it hasn't yet been invented. Uh and yet we are on a race to deploy have the deployment infrastructure to deliver AI to the entire world. So now uh what can I possibly say about what's next? And there is one area that at least for me uh has captured my imagination and has caused me uh in my leading role at Microsoft research to devote considerable amount of resource uh into investigating. uh and it is a
hot topic at MIT and at universities and at uh great companies around the world uh which you could just refer to as AI for science and that is the idea that the AI technologies the architectures around transformers diffusion models uh and so on might work just as well if not better for learning the languages of nature than they do in learning our language the languages of human beings. uh and uh as just one simple example uh this is a picture of a protein that is associated with uh the disease tuberculosis and if you look very carefully there's a little bit of a colored molecule uh that has been designed entirely by a system called TAMGEN uh which is built using three generative AI models one diffusion model and two uh transformer models uh that has uh designed on its own uh that potential drug molecule candidate that binds over 40 times more effectively in a key target in that tuberculosis protein. Um, incidentally, everything I'm going to talk about today, uh, I am going to talk about and try to ground it in the real world of one to three years. And so,
Tamjun is available to anyone. It's open source. Uh, and my colleague Asha Sharma, uh, delivers these things on the Azure foundry. But if you don't want to use something from Microsoft, there are dozens of laboratories around the world that are chasing after the same sorts of capabilities. And we even have Nobel prizes uh by uh some great scientists uh at Google and University of Washington and others. What is going on with these things? Well, you know, with generative AI, we dream of being able to write down in a prompt or some specification of what we want a molecule or material that we are designing to be. And then
generative AI should magically give us that thing. Uh and that is actually the backwards order of the way science has always been done computationally. Computational science traditionally has gone the other way where we uh you know uh take a simulator typically written in forran uh and we run that simulator to witness computationally uh what the molecu uh properties are of some proposed molecule some proposed physical system uh and if it's not right then we change the parameters and rerun our four trend codes. uh those forrren codes have been exceptionally important. They've driven a huge amount of computational science over the decades but uh of course they are uh incredibly slow and so for example on that tuberculosis drug candidate molecule it would have taken uh several million years of computation time using those for train codes uh to come up with the same type of molecule.
What is very interesting though is that these simulators are based on the precise physics of these molecules. And so these simulators actually generate perfectly labeled synthetic training data that allows us now with our new generative AI architectures to train what at least we in Microsoft have been calling emulators. They emulate these classical simulations, but because they're using AI, they are able to do so typically at an acceleration of 1,000 to 10 million fold. And in fact, for that TAMGen uh there was uh over a 10 millionfold speed up uh in the design of that molecule. And incidentally, this is one reason why we at Microsoft, but we're again we're far from alone. There are great labs around the world pursuing this. This is
another reason why we are so interested in quantum computation and we're so heavily invested there because uh the first and foremost quantum program that we've written that we are dying to be able to run is a molecular simulator uh and to be able to generate again training data much much more rapidly and much more cheaply uh than we're doing today to train these AI emulators. And so there are some early tantalizing uh examples and again every example I'm going to talk about here uh you can have it's open source it's on the Azure AI foundry or it's on hugging face whatever you want and again if you don't want to take it from our labs there are just amazing things going on uh here in the Boston academic and research ecosystem and as well as great companies around the world but biomolecular dynamics showing uh the confirmations of of molecules again with absolute chemical quantum um precision uh materials design. Yeah. Getting the precise sort of uh material and electrical properties in designed and and magnetic uh properties in design uh materials that's in a model called mattergen. Um, we've run millions of hours, uh, costing us quite a few million dollars, millions of hours of global climate simulations, atmospheric simulations, uh, at ex with the resolution turned way up down to 0.1 uh, degree of resolution
in order to generate training data to train a model called Aurora uh, that then shows tremendous foundational capabilities. And what I mean by that is it's trained on those climate simulations, but then we take a small amount of data from the Capernacus uh atmospheric modeling uh monitoring system uh that pertains specifically to nitrogen dioxide dispersion through the atmosphere. And we find that that fine-tuning of that foundation model gives us the most precise uh pollution dispersion models uh ever developed. And again trained purely on that synthetic but very uh physically precise training data. Uh and then um we even have our world uh action model just learning observing people playing video games and then uh from this uh this simulated data getting emulated is able to learn the world model. Where we're going with all of
this is to really drive at the heart uh of quantum physics. And you if you go back to Shortinger's equation, there was a Nobel Prize uh some time ago that really defined uh uh something called density functional theory or DFT. And there's a specific functional and there's a whole zoo of u machine learning and other approximation algorithms that are used uh to calculate uh various aspects of that DFT functional. uh what we're finding is precise simulation of these things at scale gives us training data that gives us a much more general and much faster emulator uh for DFT uh and that is something that has commercial uh impact uh we believe uh because about 60% of HPC workloads today around the world uh involve DFT solutions and so just to close um I think what I would try to claim here is that in the same way that We found that generative AI architectures are shockingly good at learning from the digital exhaust of human thought and expression. We're finding that the same is likely to be true for learning the languages of nature. Thank you.
Thank you. Please take a seat. Yeah. Thanks. That was great. Um, it's nice to see. Uh, I was nervous because what do you say? You asked me to talk, you know, one to three years from now. It's Yeah, I wouldn't answer that
question. I'm surprised you even came, but I'm glad you did. Um, yeah, I was going to say at the end of, you know, three days of of of talks and lots of conversations, you know, I'm going to do the foolish thing of trying to, you know, get you to help us tie tie a few things together. Um, it won't be a pretty bow, but we'll we'll do our best.
Um but there are already I'm seeing connections with you know you talking about these uh these simulations and this universal language of nature makes me think of um we had Scott Penathy from Google yesterday who's calling um you today's AI universal approximators it allows us to simulate anything we want and that's a very very powerful idea. Um I also loved the copper wire analogy. I mean, of course, you're not the first person to compare this to, you know, the invention of electricity, but that focus on the copper wire and every all the big companies wanted, you know, supplying copper wire, I think it's fascinating because, um, nobody, you know, back 100 plus years would have sort of set up a business putting copper wire in all our streets before we had the light bulb. So, why are we doing that now? What's why is everybody so convinced that putting the copper wire out uh is worth doing when we don't yet have the light bulb? Right. By the way, you know, uh
the copper wire thing is also, I think, interesting because um you know, what are regulators to make of this? Yeah. Because people are getting shocked uh all the time. Maybe some houses are burning down. There might be some crazy people putting their tongues on this thing. Yeah. You know,
and other people are saying, "Wait, it's it's it's just it's just a wire. It's just a wire." Yeah. or other people, you know, calling themselves copper wire companies, but what they're doing is selling uh uh clothes hangers made out of wire, you know. So, you know,
it so first off, you know, I think that um there are modernization opportunities at scale already. Um you these technologies are improving a lot of things. They're improving search, they're improving ad targeting, they're uh you know uh our uh enterprise co-pilots are selling extremely well. Uh actually the fastest
growing co-pilot project in Microsoft is uh an AI scribe for writing clinical notes. So things are happening but I think what I'm really trying to say is that 10 years from now we'll look at all these things as sort of trivial and foolish and maybe a way uh to which things exactly the things that we're trying to do with AI today they're limited by the limits of our imagination. So a way I try to explain this and this relates to quantum is um so you know the field effect transistor was invented by Lillenfeld in Germany in 1926. Um and then it took a couple decades for folks at Bell Labs to actually make one that worked. Um but here's what people don't remember. Then from 1945 or from 1946, it took until 1951 before the first commercial application of the transistor happened and that was the tone generator to make beeps and burps to switch telephones.
And then another four years before the second commercial application, which was the AM transistor radio. Notice there's no digital computer in there. and and so and if you go all the way back to 1826, Lillenfeld could not conceive of a digital computer. So I think the things we're
thinking about in AI today are the moral equivalent of tone generators and AM transistor radios. What will be the real things that will just transform our lives that we can't live without all day every day? I I think we're yet to see those things, right? Having said that, you know, at least speaking for Microsoft, we're doing very well and our shareholders rewarding us for selling a lot of AM transistor radios. So, it's it's a good business. It's a good business. Um, yeah, nobody quote you on calling those trivial things, right? Um, so I mean, another way of talking about this, I there's so much buzz, so much noise. How how do you go about you
finding the the signal? Um, I mean, and also I'm leaping between metaphors here. But when we spoke a few weeks ago, you had this another you had this other nice image of AI is like a a blanket that's just been dropped from the sky and it's it's settling now, right, into the sort of nooks and crevices, but once it's settled, there will be some obvious peaks. Um, you know, spikes where the applications are are really breakthroughs. I mean, is there any way we can sort of predict where those are going to be? I I think I said the same thing when we talked before. When you get the answer to that question, please tell me. Um
because in in a way I think the my most fundamental and important accountability to Microsoft is to answer that question and I honestly have no idea. You know I I there are fundamental values about the value of research that in my mind in my heart uh give us the best chance uh to do that uh to maintain a culture of uh independence to have an understanding of the need for balance between kind of bottomup independent ideiation and top down mission focused drive and and so on and on. Um, but these are all at the end of the day just my personal biases. Um, and you know, if I reflect, I I think I've gotten a lot of things right in my career, but I've missed a lot of incredible things also or just even dismissed them. Yeah. Um, you know, there was a lot of talk about deep learning and um, when I joined Microsoft in 2010, I was going through all of the existing projects that were funded in Microsoft research and there was this one project that was instigated by Jeff Hinton and his students who were spending the summer. Uh, it was actually the previous summer 2009. Um, and they
had proposed, hey, uh, you guys you do speech recognition. Why don't you use this layered cake of neural nets to do speech recognition? And by the time I arrived at Microsoft Research, uh Microsoft Research strangely had agreed to fund that project. I'm certain if I had joined Microsoft Research a year earlier, I never in a million years would have okayed that because I I knew enough about speech recognition to realize that that was completely ridiculous. Um, and of course, you know, the results were shocking. Andrew was hanging out at
Microsoft Research and so the next year he went and hung out with Jeff Dean and asked the question, would this work for computer vision? And of course, we're detecting kitten kitten videos all of a sudden. And so, you know, there have been so many times when I've just been wrong. But I think ultimately you just need to be able to admit both as a person and as an organization that you've missed some things, not be ashamed of that and and jump on it. Yeah. Yeah, I mean thinking about that trajectory you were just outlining, you know, we started the the the show a couple of days ago with, you know, discussion about sort of extra exponential growth and the advances and, you know, some people are happy to sort of project that exponential growth and and make predictions about where we're going. Um, and we've happily avoided the term AGI as much as possible in the last few days. I I I I personally dislike it.
Um, and I know it splits the industry. We also had um Ali Fahadi from the Allen Institute on stage yesterday and he he told me not so long ago joking I assume but if any of his if any of his PhD students use that term instant fail um and yet on the other hand you know there are there are the leaders of some of the biggest labs in the world you know I'd with a very straight face say you know we are making AGI um and you know to pick two names at random Sam Alman is on one hand making uh AGI happen and Satia Nadella you thinks it's a it's it's it's the nonsense distraction. I mean, how should we make sense of this wild disparity? Yeah, I think I side with Satcha. By the way, you know, um on the day that GP4 got revealed to the world in Microsoft research, um we published a paper on entitled sparks, yeah, of artificial intelligence. And Satcha never misses a chance to criticize me uh for that. and you know and I and I think uh he's got the right point and the way I like what the heck is AGI? Yeah. Um so here's the way I
think about it if I can just be an academic research geek for a minute. Um in my career I count let me ask this question. Uh who here is not carrying a mobile phone with them right now? I don't see any hands raised, which is a good thing because if if there were some unlucky person who maybe forgot it in their hotel room or at home, you'd be feeling pretty bad right now. Um naked,
vulnerable, uh all all of that stuff. And so in that mobile phone in uh throughout my research career have been six triumphs of academic computer science research that you literally can't bear to live without at any moment of the day. Um you know there is VSI design um emerging out of academia in in the late '7s. Um there
is uh softwaredefined mobile wireless network radio. Um we can just go through uh these six triumphs. Uh there's uh GPU uh uh ray tracing and uh 3D computer graphics. uh in each one of these things
academia sacrificed itself to make these things possible. They became incredibly hot topics in research funded by our government and then when they transitioned to industry to become the infrastructure of life all virtually all research in those areas in academia stopped. It matured. So my thought here is what counts for me is will generative AI become the seventh in my career for that to happen. Will generative AI power things that you cannot bear to be without for any moment of the day that you would literally feel incomplete and unsafe? And if you you could call that AGI but I don't think it's quite the same idea. And um and I would say the bet that at least Sachin Nadella has put Microsoft into making is that uh he's he's betting that it will happen and it will happen soon. Yeah. Which is a much
more concrete way of thinking about it. Um we have a few minutes left, but I'd love to go to questions in the room if there are some. Could we get the lights? Um I can see some hands, but who has the mic first? Okay. at the back. Thank you. Um, very nice. And I'll build on your academic uh
point. I'm a professor at a university and the question is, have you been dealing with universities in terms of rolling out AI? And more generally, how would AI help reimagine or reinvent academia? Yeah. Um, so I I would say uh so we have uh been doing that. We have several programs. I I can't say I'm particularly
impressed with the outcomes of what we've done so far and that's probably our fault, not academia's fault. Um, you know, we worked with several early uh almost two years ago, we worked with several universities to uh have private tenants with uh access to APIs and and models. and some universities like uh one that I would point out is the University of Michigan, you know, put in quite a bit of work to integrate things into student information systems and so on. Um um but it hasn't been the transformation that now two years later I was maybe hoping and expecting to see. I I I think things will happen. U but uh things along those
lines I think are important. Um uh one of the things I would like to see us do and industry do is to provide more access to compute infrastructure uh for open model development. Um I think that that uh could make a big difference. Um but um you know it has frankly just gone a little bit slower than I would have hoped.
Go ahead. Hi Kim. Yeah Kareem Mazeram at Novartis. Um so I wanted to come back to your premise about you know will AI how will AI give us insights into the natural world and specifically perhaps life sciences.
Yeah. Um so today we heard about Schroinger equation. Yesterday we heard about Na'vi Stokes. Um and you know as a
mathematician uh I you know I spent a good amount of time studying these. They give us significant insight into life and uh and we also had it took us quite a bit of time to develop the tools to be able to understand the systems that we're studying. Do you see these, you know, AI models as sort of more pattern recognizers? In other words, giving us information that's out there we haven't noticed or they're actually going to help us discover new aspects of our worlds. you know, for example, can they
lead us to a unified theory, for example? Yeah. Uh like another Schroinger, another Veniv Stokes. U so where do you see this going? Yeah, really great question. There's uh and we
probably don't have enough time to get into that. Maybe a couple of quick things. Um first off, I I think that these things will enable sort of abonio design of physical systems from the ground up. whether that will have true
practical value like will we ever uh abonio from the Schroinger's equation on up be able to design a drug that you would put in a human being I have no idea commercially today the way the pharmaceutical industry works you know those phase one designs are a dime a dozen so commercially it seems problematic so I would expect commercial drivers are going to not be based on ebonicio uh designs but instead be used to arbitrage of the multitude of people doing molecule design which are the ones I should bet on. Um and so that's a kind of a higher level kind of thing. So there's going to be a push and pull uh there. But on your more fundamental question, I would expect within five years we will start to see some physics, chemistry and biology discoveries that will be discovered autonomously, written up autonomously, submitted to major leading journals. And I think the
editors and chiefs of those journals will have a time trying to figure out what to do with those. Um, we I'm going to get myself in trouble and carry on taking questions even though we're out of time because these are great questions. So, one and two if they're quick, please. Uh, Dick Audi from Promex Santa Clara. How close are we to having machine consciousness and to having a personality live in what I'll call the cloud? Yeah, that's another one where I always say when you figure that out, please tell me. There's also supposed to be a
quick question but yeah um you know I think that one of the things we're finding uh in the healthc care space um this uh there was a beautiful study uh controlled uh clinical study done jointly by UC San Diego and Stanford on looking at email responses to patients um you know in the medical setting and you know it measured in a randomized study that the AI was just as accurate as human clinicians. But by a factor of nine to one, um the uh machine generated responses were deemed more empathetic, more kind, more personable, whatever you want to say. And so there's something strange, you know, going on uh there. You know, I think human beings are so frazzled today. Um, and so I don't I wouldn't call it consciousness, but the observation in that control study, there was a line in the report that said that human beings, the human doctors on seeing the draft written by the AI were often observed pausing for 15 seconds apparently reflecting on the life of that patient. So there's something about the relationship between man, machine, and other human beings that I think potentially could be made better by forcing us to remember and reflect on each other's lives in a different way. I hope that makes sense. What was
there one more question there? Oh, sorry. Just because I I did I did say go ahead. Uh so Rup from New York University. So you mentioned about the machine is able to now speak or interpret the language of nature and you showcased one of the examples of emulator but that is because we know the underlying physics. We know
how fundamentally data is generated. As a social scientist, we studied a lot of social behavior and we see this a lot more difficult to interpret and to make sense of know in this large uncertainty associated with human behavior, the policy, the society, how would you see that AI will be able to help? Thank you. Yeah, it's really a interesting question and even um with something like biology the amount of um diversity uh that uh is pos I mean there have only been what 106 billion human beings that have ever existed. So even if we had DNA samples from every single one in today's world of generative AI we might not have enough data to train a model um you know for certain types of problems. So I
think that uh the fundamental question and we saw this in the Aurora weather model um you know that foundation model what shocked me was uh it was not trained at all on any nitrogen dioxide dispersion data literally zero. And yet with a very small amount of fine-tuning uh on collected nitrogen dioxide data, it was able to surpass you know all of the specialized uh pollution dispersion models um out there. So I I think it's really unclear how this uh will play out. But it's not inconceivable to me that even in say trying to understand social systems that we may be able to build very capable foundation models on large amounts of simulated data. I don't know ultimately but it seems poss we we can't discount that possibility.
Thank you. Thank you. I there's nothing more reassuring to me to have you come on and say you don't know because makes makes me feel a lot better. All right. Thank you so much. [Applause]
2025-05-17 10:46
