YT YouTube2Text

YouTube2Text → Transcripts → Yann LeCun, AMI Labs | Evident AI Symposium 2026

Yann LeCun, AMI Labs | Evident AI Symposium 2026 — Transcript

by Evident Insights · 3,234 words · 554 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:04Good to have you even if it's remotely
0:07we're delighted to have you in the room.
0:08I just I don't know if you can see the
0:10room but
0:11about
0:12>> see part of it. Yes.
0:13>> Part of it. Yeah. In the room you will
0:15have um
0:17the leaders in the big enterprises in
0:20financial services on AI transformation
0:22journey. Everyone is working hard on
0:25deployment side of things. You'll have
0:28many engineers and scientists as well.
0:30So and I think everyone's very eager.
0:33We've spent uh most of the day up until
0:35now on everything from platform
0:37architecture to governance controls
0:40harnesses looking at um
0:43looking at sort of specific
0:45uh use cases
0:47uh in the demos. And and so it's really
0:50brilliant to have you here now. Only for
0:5320 minutes I know to sort of step back
0:55zoom out and talk about what's on the
0:57horizon. So
0:59just to to maybe to to start and kick
1:01off
1:02uh would love to hear where you are with
1:04AMI labs and GIP and
1:07um
1:07what sort of yeah where you are sort of
1:09today and of course would love to hear
1:12about the timeline for that and and many
1:14more questions but if we start there.
1:17>> Sure. So I'm actually calling you from
1:19the
1:20uh
1:21Anidance it's pronounced Anidance. Uh
1:24you know ami as in uh friend in French.
1:27Uh so Anidance is actually a French
1:29company uh headquartered in Paris. Uh we
1:31have two big offices in uh Paris New
1:34York
1:35and smaller offices in Montreal and
1:37Singapore. So we we're really a global
1:39company.
1:40Um uh we're we started operations in
1:45March. We uh raised
1:47uh about a billion dollars in a seed in
1:50the seed round.
1:51Um and which is uh the
1:54biggest seed round in history outside of
1:57the US I believe as far as I know.
2:00Um
2:01and uh uh we have uh
2:04around 65 employees at at the moment.
2:07And uh
2:08the the whole company is based around
2:11the idea of being of of building AI
2:13systems for the for the real world for
2:16for the the physical world. You know,
2:18people have been calling this physical
2:20AI, if you want.
2:22Uh and it's uh built around the idea of
2:25trading world models.
2:27And maybe we can go into what we mean by
2:29world model, which is not the science
2:31fiction way that other people mean by
2:32that.
2:33Um
2:34and the things are progressing really
2:37well and uh
2:39uh we we we are
2:41anticipating uh
2:43but you know, some
2:44some success in various industry uh over
2:46the next year or so.
2:48>> Over the next year.
2:49And if that's [clears throat]
2:50uh yeah. I
2:52Maybe if you could talk to us about like
2:54obviously in the in the news there's a
2:56big discussion around
2:58uh you know, reaching AGI. I don't know
3:00if we call it superintelligence now. But
3:03um
3:04who knows.
3:05Uh and
3:05and I know that you don't necessarily
3:08love that term, but again, to reach at
3:11the next level, if you then like of uh
3:14intelligence. Um can you talk to us a
3:16bit about how the you know,
3:19what Geeper is or
3:21uh and then and sort of how that is
3:23going to help to reach the next level of
3:25intelligence.
3:27>> Right. So, there is uh a mistake that
3:30people
3:32in AI and at the periphery of AI have
3:34kept making for the last 70 years.
3:37And it's the mistake that if a machine
3:39can do a complex
3:42uh task or a task
3:44identified by humans as intellectually
3:46complicated,
3:47uh
3:48then automatically this machine is
3:50smarter than humans or perhaps has
3:52reached human intelligence or perhaps
3:54superintelligence. And we know this is
3:56completely false because we've had
3:58systems that are capable of
4:00uh you know, solving equations, or
4:03computing integrals symbolically, or
4:05playing chess, or Go better than humans.
4:08So, we have computers that are
4:09superhuman in many different tasks, and
4:12the list keeps increasing thanks to
4:14recent progress in AI. But, this does
4:17not mean that we are anywhere close to
4:21having machines that have the type of
4:23intelligence we observe in humans
4:25and and and animals, for that matter.
4:27And, in fact, as it turns out, uh
4:30uh
4:31you know, it's much
4:32simpler for machines in general and for
4:36current AI approaches based on LNNs
4:39to manipulate uh sequences of discrete
4:42symbols, human language, code,
4:45mathematics,
4:46and things of that type,
4:48rather than handling the real world. The
4:51real world is much more complicated. Uh
4:54it's messy. It's not entirely
4:56predictable. It's continuous,
4:58high-dimensional, noisy. And, it turns
5:01out that the methods that have been so
5:02successful in the context of language,
5:04namely LNNs, actually do not work uh
5:08very well for for for the real world.
5:10And so, you can see this in the fact
5:12that we have systems that can, you know,
5:15pass the bar exam, and prove theorems,
5:17and and do all kinds of amazing feats.
5:19Uh but, we still don't have domestic
5:21robots. We don't even have level five
5:23self-driving cars.
5:24Uh we certainly don't have self-driving
5:26cars that can learn to drive in a few
5:28hours of practice like any teenager. Um
5:31so, there is a huge gap, and that gap
5:34has been known for a very long time.
5:35It's called the the Moravec paradox, the
5:37fact that uh you know, we uh
5:41we can handle
5:43uh things like, you know, logical uh
5:46reasoning and and mathematics and stuff
5:48like that with uh with computers
5:50relatively easily, but but not the real
5:52world. So, we're we're trying to kind of
5:54bridge that gap with uh with Amino Apps.
5:57Uh sort of, you know, coming up with uh
6:00new paradigms of AI, really, that that,
6:03you know, would be capable of uh
6:05handling the the messiness of the real
6:07world if you
6:09>> What what would be Yeah, what would be
6:11the proof points along the way that you
6:13are making progress on because it is a
6:15complicated
6:17um scientific endeavor to create a an
6:21understanding of the multi-dimensional
6:24world that we live in. I know you give
6:26examples of
6:28getting out of your house and taking a
6:29taxi and going to the airport and taking
6:31a plane. And I I love that example. I
6:34think that was Yeah, you you've you've
6:36used that a couple of times. How
6:38difficult it is to simulate to to
6:40actually build models that can fully
6:42understand that journey. So, what would
6:44be your proof points along the way that
6:47you're making progress?
6:49>> Well, the long-term proof point would be
6:52uh let's say uh
6:53uh
6:54humanoid robot that can, you know, do
6:56the tasks
6:57that a 10-year-old can accomplish,
7:00let's say,
7:01uh without
7:03much explicit training. When you ask a
7:0510-year-old,
7:07uh
7:08the first time you ask a 10-year-old,
7:10you know, uh
7:11uh empty the dishwasher, all right, or
7:14or or or clear the dinner table and then
7:17and fill the dishwasher.
7:19Uh
7:21a 10-year-old can probably do it the
7:22first time without any prior training.
7:24Maybe not perfectly, but
7:26uh
7:27but there's no
7:28requirement for, you know, thousands of
7:31examples by demonstration.
7:34So,
7:36what allows this? What what enables
7:38the ability of humans and many animals
7:40to uh
7:42essentially apprehend a new situation
7:44and come up with
7:45a plan, a sequence of actions to
7:47accomplish a task without ever having
7:49faced that that situation.
7:51Uh our our hypothesis for that is
7:55uh the the notion of world model.
7:58So, what is a world model? In in our
8:00opinion,
8:01a lot of people have different meanings
8:03for that, but
8:04what we mean by world model is that if
8:06you have an idea of the state of the
8:08world at time t
8:10and you imagine taking an action,
8:13your world model will predict the
8:14resulting state of the world at time t
8:16plus one after you uh have taken this
8:18action.
8:19And if you have such a world model, you
8:21can use it to plan a sequence of actions
8:23to accomplish a task. Okay, so a task
8:27would be some sort of cost function that
8:30uh
8:31uh
8:31indicates whether the task has been
8:33accomplished or not. If it's
8:34accomplished, the cost is zero. If it's
8:36not accomplished, it's a
8:37uh a higher cost.
8:39And by optimization, you can figure out
8:41a sequence of actions
8:43that will accomplish the task according
8:46to the predictions of your world model.
8:47Okay, this is very classical
8:49in the optimal control and motion
8:51planning and things like that.
8:53Uh except that what we
8:56the the type of world model we would
8:57use, our world models are not trained
8:59from from data as opposed to
9:02being a virtual equation that we turn by
9:04scientists or engineers.
9:05Uh
9:06so that that's the the the whole idea.
9:09And and there, the the question is
9:12uh
9:13you know, can that be done for systems
9:15that are too complicated to be used to a
9:18smaller of equations. And most of the
9:21most of the world is like that. You you
9:23You can of course simulate a lot of
9:24things in the world, which is what a lot
9:26of current robotics systems are using to
9:28to train themselves. But there's a lot
9:30of situations that are very difficult to
9:31simulate accurately, particularly
9:33manipulation and things of that type.
9:35But also all kinds of other uh other
9:37situations. So for that, you need a
9:40system to develop a phenomenal physical
9:43model of the the situation that it can
9:45learn from observation.
9:48Uh and uh
9:49and and then use that for for planning.
9:52Um the example you saw in would require
9:55what we call a hierarchical world model.
9:57We we had um
9:59just published a paper on this uh just
10:00last week where um
10:03that you
10:04it's relatively easy to predict with
10:06high accuracy what's going to happen in
10:08the world as long as the prediction is
10:10uh
10:10in the short term.
10:12But if you want to make long-term
10:13prediction,
10:15um then you cannot make accurate
10:18prediction in the real world because the
10:19real world has a lot of
10:21uncertainty, unpredictable things,
10:23unknowns. And so, indeed if I want to be
10:26if I'm in New York and I want to be in
10:27Paris tomorrow, uh I cannot predict
10:30exactly all the details about trajectory
10:33in terms of muscle control, but I know
10:35that at a high abstraction level, I need
10:37to go to the airport and catch a plane.
10:39Okay, now I have a sub-goal which is
10:41going to the airport. I'm in New York
10:42City, I need to go on the streets and
10:44hail a taxi.
10:45Okay, now I have a sub-sub-goal which is
10:47being on the street. I need to get out
10:49of the building and walk out the door,
10:51right? Etc. And then, you know, you can
10:53do this hierarchically
10:54and
10:55at the bottom come up with a very simple
10:57action which is, you know, standing up
10:59from your from from the chair,
11:01uh which you probably don't have to do
11:02much planning for. Uh so, ultimately we
11:05think that intelligent behavior
11:08requires
11:09uh the system to be able to anticipate
11:12the consequences of its actions so that
11:15it can plan a sequence of actions to
11:17arrive
11:18at a at an outcome.
11:19And the the
11:22happy use for us and use I think for the
11:24the state of the industry is that LLMs
11:26are basically completely incapable of
11:28doing this.
11:29>> Yeah.
11:30I I wanted to just bring it back into
11:33the into the question around the you
11:36know, once and I'm sure you will crack
11:38this. Um
11:40we're all, you know, rooting for you on
11:42this. But looking at the use of
11:45this kind of intelligence within
11:48regulated industries and especially the
11:50financial services,
11:52given what I think you're also saying is
11:54that they're going to be much more
11:55observability
11:57in these models if I'm correct.
12:00But if we just take the first question
12:01first, which is the application of this
12:05um
12:06is of course, you know, many and in the
12:08physical world in particular. But if
12:10you're if you're um
12:11you know,
12:12if you could comment on it for how it
12:15would be
12:16um, useful, which it which it will for
12:18the financial services, that would be
12:19great.
12:21>> All right, certainly. So, uh, so I have
12:23to explain uh, what the the basic
12:26you know, technical idea is.
12:29So, it's based
12:30on a architecture that I've called
12:32JETPA, J E T P A, and that means joint
12:35embedding predictive architecture.
12:37The characteristic of this approach is
12:40that it's not a generative model.
12:42So, you know how when you train an NLM
12:45or another generative model like a video
12:47generation system,
12:48uh, you you show it, uh,
12:51let's say a piece of text if it's an
12:52NLM, and you train the system to predict
12:54the next word in the text.
12:57And of course, you can never exactly
12:58predict which word follows a particular
13:00text, but what you can do
13:02is predict a kind of probability
13:05distribution over all possible words in
13:07your dictionary. Okay, so in a discrete
13:09world,
13:10handling the uncertainty in the
13:12prediction is relatively easy. It's
13:13just, you know, giving a a score to
13:16every uh, potential outcome. Now,
13:19think about
13:21if you want to
13:22transpose that kind of method
13:25of kind of
13:26learning about um,
13:28uh, a particular input modality by by
13:31training it to predict,
13:33um,
13:34generative prediction doesn't work. It
13:36doesn't work
13:38for video. If you If you would train a
13:40system to predict what's going to happen
13:41in the video, there's no way it can
13:43predict every detail in that video.
13:45Because most of the details in the world
13:48are completely unpredictable. You're all
13:49sitting in the room right now. If I
13:51point the camera to the left side of the
13:52room, and I slowly rotate the camera to
13:55look at the right side, I stop the
13:57camera there,
13:58and I ask the system predict
14:00the rest of the video,
14:01system will predict probably that the
14:03camera is going to keep panning.
14:06There is absolutely no way it can
14:07predict what every person on this on the
14:10right side of the
14:11room looks like.
14:13Because uh,
14:15the information is just not there. So,
14:17when you train
14:18a generative system, which is trying to
14:20predict every detail in the input
14:24at the pixel level in the case of video,
14:26you basically kill it.
14:28Um and so, you have to invent a a new
14:30set of technique for this. And by the
14:31way, this applies to financial data,
14:33right? If you
14:34uh
14:35much financial data is mostly noise with
14:37a little bit of information in it.
14:39If you do if you train the system to do
14:41sub time series prediction of financial
14:43data,
14:44um
14:45it has to be able to predict its own
14:47uncertainty.
14:48You cannot train it to predict a
14:50particular outcome because
14:52um
14:53because it's mostly noise.
14:55Uh so, so that this the same the same
14:58part and so, what what can you do?
15:00The the the technology that we've
15:02developed and we continue developing at
15:04NeurIPS is one in which the system
15:07learns an abstract representation of the
15:10input and it makes prediction in that
15:11abstract representation space. So, it
15:13does not have time to even all the
15:15details of the observation.
15:18It tries to predict whatever is
15:19predictable.
15:20But ignore all the details that are
15:22simply not not predictable.
15:25And I would think this is absolutely
15:26fundamental. And and we do this all the
15:28time in in in modeling. For example, let
15:30me take a simple example.
15:32Um
15:34if I want to uh let's say I'm a aircraft
15:37uh
15:38uh design engineer and I want to design
15:40uh
15:41an airfoil uh for an airplane. Um I can
15:45I can simulate the the the flow of the
15:47air around the the airplane using
15:49Navier-Stokes equations. But that's
15:51actually an approximation. It's a
15:54it's it's a it's an abstract
15:56representation of the underlying
15:57phenomenon, which is molecules of air
16:00bumping into each other and onto the
16:01airplane.
16:03But we never simulate this at that level
16:04because it's just too many details
16:06that we could not actually simulate
16:09accurately anyway. So, we invent an
16:11abstract representation that allows us
16:13to make predictions.
16:14Similarly, if I give you a box uh
16:17a closed box with air in it
16:20and uh I tell you I'm going to
16:22heat up this box by 10 degrees.
16:25You can predict exactly by how much the
16:26pressure will increase because you have
16:28the ideal gas law PV=nRT.
16:31But there's no way you can predict the
16:32trajectory of every molecule of air in
16:35that box. So so this idea that, you
16:38know, we need to invent abstract
16:40representation of reality to allow us to
16:42make prediction is really fundamental.
16:44Which is why you cannot reduce chemistry
16:47to physics or
16:49or you know atomic physics to particle
16:51physics or or or biochemistry to
16:54chemistry or biology to biochemistry,
16:56right? Every level in the abstraction
16:58ladder of how we understand the world
17:01has its own knowledge which ignores a
17:05lot of details about the level of
17:07description below
17:09below it.
17:10>> Yann, thank you so much. We've only got
17:121 minute left and I thought maybe just a
17:15quick fire question on geopolitics.
17:18How where do you know where China is and
17:20it's hard to ask a big question like
17:21that and say we've got 30 seconds left.
17:23Sorry about that. But
17:25where where is on on the on the deeper
17:28type work that you're doing, where where
17:31is China at on that and do you feel that
17:33there's the same race dynamics that it
17:36the sense of sort of who reaches what
17:39first right now is very intense. Is that
17:41is that the same in the in the deeper
17:44scientific type development?
17:47>> No, no. That's the short answer.
17:49So China
17:52and most of Asian countries are very
17:54interested in robotics and in physical
17:57AI because traditionally they they've
18:00been very strong in manufacturing. So
18:03that's true of China, of
18:05of Japan, of Korea where I was recently,
18:10Taiwan and a few other countries in that
18:13that part of the world. They're very
18:14strong in manufacturing which means
18:17they
18:18they're very strong in robotics. They're
18:19pushing this very very hard, which means
18:23because LLMs are basically hopeless for
18:25that kind of stuff, they they have a big
18:26interest in things like world models.
18:28So, a lot of the work that we see in in
18:33scientific publications
18:35in preprints, you know,
18:37on world models and on the joint
18:39embedding approach I just mentioned,
18:40many a lot of it comes from from Asia,
18:43China in particular.
18:45So, so they're catching up on this much
18:48faster than Silicon Valley, which seems
18:50completely hypnotized by LLMs and
18:53kind of too entrenched in LLMs to kind
18:55of look look around.
18:57>> Oh, yes, and it clearly has its
18:59limitations as you as you lay out. So,
19:02but I I know we're at time, Yann, and I
19:05cannot thank you enough for spending
19:07some time with us here today and giving
19:09us your view of the future and how to
19:12how to how to build it. So, a huge round
19:14of applause for Yann LeCun. Thank you so
19:16much.
19:17>> Thank you. [applause]
19:19>> [music]

About this transcript

This page contains the full transcript of Yann LeCun, AMI Labs | Evident AI Symposium 2026 by Evident Insights, generated from the public captions YouTube serves with the video. The transcript has 3,234 words across 554 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.