Learning Abstractions: A Conversation with Yann LeCun

James Manyika. Yann LeCun is widely considered one of the godfathers of the modern era of artificial intelligence. His pioneering research with Geoffrey Hinton and Yoshua Bengio on deep learning won them the 2018 Turing Award, considered the Nobel Prize for computer science, and in 2025, the three were awarded the Queen Elizabeth Prize for Engineering. Yann has long focused on foundational and scientific advances in AI, from biologically inspired convolutional neural networks and graph transformer networks to energy-based models for machine learning and world models. In this dialogue, we focus on foundational advances in AI, what he sees as progress to date, the limitations of current mainstream approaches, and what's needed to further advance AI and benefit science, including the role of scientists and the importance of open science.Manyika. The last decade in artificial intelligence has been extraordinary. How would you characterize the progress? What has surprised you?Yann LeCun. I'm actually going to go back fifteen years to when Geoffrey Hinton, Yoshua Bengio, and I were trying to rekindle interest in neural networks, which we rebranded as “deep learning.” And we were struck by how fast it was adopted: within eighteen months of the first papers showing that deep learning could improve speech recognition, basically every mobile speech recognition system was using a neural network.1 This was my first exposure to how quickly technology is widely adopted when it actually brings something to the table.There was a similar phenomenon in 2013 when convolutional neural networks became the rage for image recognition. Within months they were widely deployed, and people got excited about the prospect of driving assistance and autonomous driving. Within a few years, we had automatic emergency braking systems in cars, all using convolutional neural networks. These systems have become mandatory in Europe, leading to a 40 percent reduction in frontal collisions.2 This technology was saving lives, and it made me proud.The year 2015 was more about natural language processing (NLP). Yoshua's lab made people pay attention to attention mechanisms, which direct deep-learning models to focus their attention on the most relevant data.3 Then “Attention Is All You Need,” which introduced the transformer architecture, surprised quite a few by building large networks as stacks of associative memory modules based on so-called self-attention that compares inputs to each other.4 I had been advocating for self-supervised learning (SSL) for many years, but I did not expect that its success would be primarily in NLP - though we had hints that “shallow” SSL worked for text embedding, with Word2vec from Google and fastText at Meta.5 Then there was BERT from Google, which used transformers and attention instead of recurrent neural networks, which a lot of us suspected would not go very far.6I wasn't particularly interested in NLP myself, but it was certainly a very useful thing to do. What I found surprising was that a simple form of temporal prediction works so well for text. I had been working on temporal prediction for video for at least ten years before that, hoping that if you train a system to predict what's going to happen in a video, it will understand how the world works: that the world is three dimensional and that objects move independently, for example. Some objects are inanimate and obey laws that are relatively easy to understand and make predictions about. Animated objects are much more difficult to predict. The hope and the whole idea of SSL is that by training a system through prediction, it can understand the underlying structure of what it is trained on. I was focusing on video because I believe that if we can do video, we can do everything. As it turns out, SSL from video is much harder than from text.Manyika. As you think about where we are with the development of AI's capabilities, are there scientific disciplines you're most excited for AI to help make breakthrough advances?LeCun. Certainly. Materials science, chemistry, biochemistry, and proteins are major topics. In medicine, can we understand and cure chronic diseases or cancers? Predicting protein structure is major progress, but the next step is designing proteins with particular properties.7 Can we build predictive mesoscopic AI models for complex systems like organelles, cells, or pathways in cells that are too complex for a single human to understand? The same applies to new materials and chemical compounds for improving energy storage (like batteries) or for catalysis (such as separating hydrogen from oxygen to store electrical energy as methane).Importantly, the models that work well for proteins are really specialized; they're not generic models. There is a hope that in the future we'll come up with models that are more generic, that can make wider sets of predictions. This includes efficient prediction in large systems controlled by partial differential equations (PDEs) that are too complex for us to solve accurately and quickly. In fluid dynamics, for instance, sometimes you only need to predict lift and drag or turbulent behaviors. These collective behaviors can be captured by training a neural net (essentially 3D convolutional networks that mirror PDEs) to make predictions at a coarse scale. This allows simulating much larger or turbulent systems to predict general properties rather than requiring a detailed simulation.Shirley Ho at the Flatiron Institute demonstrated a striking example of this.8 You can put the equations of general relativity into a simulator and simulate the first instants of the universe. Then you take that simulation and change the values of some of the parameters, like mass density or the properties of various fields. Then you run the simulation again to see if the result looks like our universe, if you get galaxies like those we see. That could validate or invalidate some theories about the nature of dark matter, for example. You can't do that simulation at the scale of the universe if you reduce it to the equations that we know. But if you run the simulation in a small cube, and you train a neural net to make a prediction of what goes on within that cube, but without requiring a detailed simulation of everything, then you can scale that up and perhaps run that simulation at a scale of the universe.That is model prediction validation. The next step is design. If you have a predictive model, you can by to the a for a chemical a energy of hydrogen from This is the of a world model with very as in or and What AI need to help advance and advance can machine learning to train predictive systems that can be models. not models that but they can make predictions that are and much than a think this is a and is what I all of my AI systems that can of the phenomenon and make predictions in that This a lot of about the And a because language and models are trained to predict every of the In not too much of a You predict which a of but you can a easy because a of But when you train the model to predict future in a video, you can't a useful You have to make predictions in not at the a lot of people in the last few years the take from and them into and that into a that from the one that a and this to a model to predict the next few it work that The is that you predict what's going to happen in a video at the There are so many that are not in the how to a all video because a people have with for in what we do as scientists is to a of the that all the we and we make predictions in that not a about progress in AI, been about the limitations you see in What are those particularly in the of AI to There are three The first is that you predict that in the world or in from There is are we only partial about the to train a system to make prediction in do particularly You have to the system the within which to make predictions. This is really what is all could that us in this in of instead would be because we would need to the of the universe within a that of a Then need to run some large to make predictions. do we do is we that us to make and then or organelles, cells, and You have this whole What a of is the of at which you work to is that you the system to be to the of a model or its or what many people a idea of the of the world at and you a world model will predict the of the world at from that You can think of it as a next you have a world model, you can it for The is by this predictive I can a If I put the world in a particular and take a of what is the Then I can do the and in the world what actually if you have a world model, you can do and You can a and to a of that will to that do this all the as using our model of the world to a of in a new that This is what you train the system for all it but a really have the prediction and I'm of by in a which are do through by the of for of and then a neural net that them and the In those you or more systems in which you have and them with a or of to then a form of You need neural net the most that a or of In like and you have a neural net that every and you only the that the You need a to and a to them so that you can the and the of The idea has been the though training are Google has a lot with and the is to You a before you if a of was learning is all about how you to going If my at a move not I need to that to the it was or what learning the of a similar but the of and or learning from human If the are or you can run and You can the system to But more you need a human to one was this one was not Then the system to a to the and a for the brings us to the lot of are using with to the The is that as you the universe of it of and you for a long of and not in training Then you get The system that looks not to a because it wasn't trained for a particular of the been a predictive us more about it and for training a with self-supervised you the system a or of and train it to the with all of its If you this to is and like can't predict all the The is to that the you predict and predictions in that of is about a me take a example. back people were quite at the of and objects in the a complex to a of to how some to be going in the go But if you change of the and you that is a the and the the it that some sometimes go because you're going than them on The is example is the which the of a as a of the into each This prediction is in a that about the underlying of the like the and of all the can't so we we about the we or This whole what is useful to make predictions and the we at the of I would at the of the idea of training a neural net to a of the that not all the and training a in that You make the system a that as much as about the but allows you to make it not all the you like the of in the example. to a with a The make predictions in make predictions in The is in which the the to a prediction You used to work on energy-based models and as to think about Is that The energy-based model is the only to what goes on You can think of as energy that is the of it if the is or small trained on this But if is that it actually in the the You energy that energy the that you train it on but values The energy a that where are and where there are this by the of and their energy like NLP systems and for are You take it in some to a the of and then you train to this of into the is you the instead of training the to the you train it only to take you to the You from a then you you a neural I you to this so that when they you this you're it You do this for all the in the so that you get a neural net to a that you back to the you could do this by training energy and then the of this energy which will take you back to the as What you're is trying to and build from which is in the of But in some of science, there be or limitations with to that AI's to advance in those The hope is you be to train a generic and with of from as long as the not the of you have a generic model that is and if you have for a particular you You not have to train but the put you in a If you're training on from or at of and at very how do you with the I think we would need models. And when you go up a you some about the at the You have a but that allows for predictions of more complex the same we to human If we to make predictions of complex we need to and This is how you would train those make detailed predictions. are trained to make predictions for more complex more to make those predictions. which to for each is a of the been is what you machine and general Can you and I that there is in my that, at some in the will be as as in all including there is a of I the machine intelligence to the of intelligence that we in and because are really AI systems come to that, at least in their of the world and to there is AI, which includes systems that have all the of perhaps in some What people really by is AI, and there is a the that human intelligence is But human intelligence is not general at very specialized; what found was useful for our to we have general intelligence. you be to a that you at at a lot of that computer has the last or can be by than used to be that you could have a as a by up we have that do this much than we used to be in like that the you the intelligence of a is to it to can do this with turns that for but not that I to intelligence using the artificial intelligence because it make about the nature of human you that as long as we that to human is really a intelligence as to a general really a of all There can be a general for the same that in the that there are only a small of that learning system can to some more like the that of you come up with will that is for a of that of as at the of a small of of a system are considered because of going to go from what we to because there are many more of our of what is only a small of are I think there is a similar for intelligence. There be a system that can solve every in the has to be the of the general that I the idea of to more intelligence all the up to to limitations of their Can you that, particularly with to AI for The of that people in AI have for is that you for a to a by You need machine with one to if a is - Is this a to my Is it the - and to the for the But single for every some for the of AI to this was a In the and up with this that they very the that if I can a in of a and I have a of a is a then I can solve What they is that the is most of the In the which you the through a of through each only and back at the of you the of the That the whole of computer science, of when people that most that are are actually idea that you to a new by is and do There is in a that of and which are are using a very form of this of the first this not be in or but instead in That is the and is relevant to AI in science.Manyika. But to do in some of for general intelligence or for science, we not what that or or our that that because AI all that there be a to the of for I think it that at because what is you make a prediction about something that happen as a of or there are If prediction was model is you need to it and the there is thing that are and that I think future AI systems will be or be the of a If there are for which prediction of the is you take the the and model if prediction was That is a form of about like the of that every But at ten If you put some on with a of they will the on the and them do the and that applies to everything. If you them a a and it they are because is that the that they they where the of a world model There are the and the In the you with the of the world at The world is a Can I predict the of the world at The the you to run prediction, the more difficult it is to make detailed predictions. What you can do is make predictions at a of where many are not Then there are models the of the world at and a can I predict the of the world at really what we do in have a system in a we do and we at the If our is the prediction is If what our model from our model at the of the of a of that the relevant of the and the next a The you are is that with the to or the you get to building a of intelligence. one could that intelligence you very about the underlying nature and of the Is that You can have predictive models that make predictions about how the world without is example. not you can't actually the system you're But you can come up with models that are of You can predict the of The with which you can do this on the that you the of became a of a lot we all the the including The of can we come up with a of that allows us to make and predictions on system to be to do the advances in AI, and particularly if we go human and AI systems come to understand more than we what are the for of science, how we do science, and the nature of scientific I think that is not a new we differential with did the fluid simulator understand than we can make a prediction and using based on equations that up next step AI is training a system to make predictions from without the step of the to AI allows us to to first build a model of that can then be This is because many in are collective complex of in a particular and the for this is not The of particularly complex be from the equations of too example is the at which you of to form a a collective phenomenon that is difficult to There are various properties of materials of this that be to a small of equations from which you can this collective How intelligence from in a of how a complex like intelligence can from a large of relatively simple in but a a are of how from the This is what has scientists for a long the from the to the This is where like for a new of science, which is and but is a model that us to the and In the do you AI and by or will it be a for scientists to for I think going to be a for a long There are some in which you can the thing run for a long and come up with something and it will be to the or the model is There could be some in some but I think the of scientists with AI systems for a long will be the same as a and a that the of the is for the to be is something to in this you work on this and then come back when you get something This is the the research of a and the research will as a of research if AI systems are than us at or scientific to work with people are than will be their will them what to work because we to help as a or what AI systems is a role we'll In that could we be by the this of they're or AI that some of to the that we can I think that a or not more than is a to what a can come up with a idea the about. You put them on the but they a think about the of science, so when scientists in for work on a there is a of from their but not a if a research there is too much on the scientists working in you are only going to get If you you have to people a very long and basically get of their a of the I at and to at AI the them the to them in a and get of the that that to progress AI and science, they are we the AI them and get of the if you think all the and all we need to do is scale up and the thing a lot of people in AI are that we have the and we need to scale it I believe I believe that we need a few before we get to something like or need to do research that is not as as the current in the to be but there is some and some on the work you can do in that you need to to pay for and which from a particular that or some was been a very of open and How in and in AI, The very of is You have you got some and you The you need to it is because you need people in the to be to the to the result and see if to do The progress of on scientists and on you do some research and you about as if you about If you it a some years the be to build a of But how do you that idea is and that they you to You need to You need to is how it works and I it on this but I'm a in a in a of the and I can't this on a without is the will have the to it to something and then at some it some from and up or a new a long and people that can take This was the for deep AI on had become and would months The of the “Attention Is All You had something like before it was at I that if advance is not not because you can this in research where up the and but one or is and but it can you from The whole of is to you from really what the scientific is all of open has progress in AI in like and have progress, and the that their you can and the work of That is the of most the are But some in and make open based on and as well as How do you the of a from research to are more useful the of the As you move the work and to there is a at which if you to you to But you if you can't really through a to the have a a if you them from their This is to have AI research the of research If scientists are to they will to And the of research be years the If you scientists to be they can't be by because the of that can be if is a could be years the have to something a at a and are of And in are a for to a of is difficult to the of a of on the research can be a to If a from lab at on or on and you pay attention to it thing if in a few or if that attention to got ten on and is used by the whole about the and the from research to or you could the for research for or rather than for of scientific But you pay a The more you the more you And if you can do research and solve a in then that was not that difficult to You a on to solve but if a it will take the scientific to take interest and of or is of that a scientific not a of up or a scientific and has a on which is I think some of the that are that come up with or they within a few years, with a small of ten people working in That can't one of the most scientific and of our You solve it with a of people working in that only in As you to the what are the in that you hope AI will have to There are three major in the universe made all And how the much all of into one of those three I think the of AI for is