YT YouTube2Text

YouTube2Text → Transcripts → Yann LeCun on What Comes After LLMs

Yann LeCun on What Comes After LLMs — Transcript

by Unsupervised Learning: With Jacob Effron · 15,072 words · 2,582 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00You're one of the godfathers of AI.
0:01What's your kind of view of the path of
0:03progress here? Five years complete world
0:04domination. The best way to get
0:06breakthrough research is you hire the
0:08best people and you get the [ __ ] out of
0:09the way.
0:10>> Pardon my French.
0:11>> You shared the Turing Award with two
0:12others. When did your views start
0:13diverging? In 2023. How do you know it
0:15was time to leave Meta? It sounds like
0:17you were thinking through some of these
0:17things over a period of time.
0:19>> There is a big misconception about my
0:21role, my relation to AI and how AI was
0:24run at Meta. What's like one thing
0:25you've changed your mind on in the last
0:26year? I mean, the whole idea of uh Yann
0:29LeCun is one of the godfathers of AI.
0:31He's an absolute legend in the field, uh
0:33someone I've admired for a long time.
0:35And so it was such a treat to get him on
0:37on Unsupervised Learning.
0:39Uh he's been a noted skeptic of of LLMs
0:41in many ways, and so we dug into what
0:43LLMs can do, what they can't do, uh some
0:45of the limitations he sees, and why he
0:47ultimately decided to pursue a different
0:49architecture. Uh and we also talked
0:51about his time at Meta,
0:52um you know, the things he's proud of in
0:53in setting up FAIR, how the last few
0:55years proceeded, and what ultimately led
0:57him to uh spin out and start his own
0:59company, uh me. Um I think it's just
1:01fascinating to get Yann's thoughts on
1:03everything happening in the AI ecosystem
1:05today, this tension between basic
1:07research and then pushing LLMs forward,
1:09and how that's happening in in a bunch
1:11of organizations today, as well as his
1:12thoughts on just where the the whole
1:14space is headed. Uh he's just an
1:16absolute giant in the field, and when I
1:18started this podcast, I hoped we'd get
1:19guests like him, so it is just such a
1:21treat. I think folks will really enjoy
1:23hearing the conversation we had. Without
1:24further ado, here's Yann.
1:29Yann, this is such a pleasure. You're
1:30one of the godfathers of AI. I feel like
1:32when I started doing this podcast years
1:34ago, I was really hoping we might one
1:36day get someone like you on. You know, I
1:38don't like that term because I live in
1:39New Jersey, when you're a godfather in
1:40New Jersey, it [laughter] doesn't mean
1:42the same thing.
1:43Very fair, very fair. You know,
1:45obviously, you know, your bet on on
1:46neural nets when everyone doubted them
1:47is legendary, and I feel like today
1:49you're making uh a similar bet in many
1:51ways against LLMs and the kind of
1:52predominant generative architectures
1:54that that so many believe in.
1:55Uh you've recently started a new company
1:57uh behind this theme. And so, you know,
1:59our goal today in the conversation is to
2:01leave our listeners with a lot for more
2:03information about AMI, what you're doing
2:04there, some of your work at Tapestry,
2:06um you know, why you think the rest of
2:08the field is is is pointed in the wrong
2:09direction around some of these
2:10generative models, and then also just
2:12get your reflections on the way the
2:13field's unfolded, your time at Meta, and
2:15and all that. So, you know, modest goals
2:16for uh for for for a single podcast
2:18episode. I feel it'd be great to start
2:20with AMI um because the company feels
2:22like the clearest statement of your
2:23technical thesis going forward. And so,
2:25you recently launched the company that's
2:26focused on world models uh and scaling
2:28the Jeff bar architecture, which you
2:29obviously pioneered uh over at Meta. And
2:32so, I'm wondering if you could talk a
2:32little bit about the origins of that
2:34architecture and the extent to which you
2:35drew inspiration from the human brain
2:37and the way that works. So, first of
2:38all, I want to say there's nothing wrong
2:40with
2:41LLMs
2:42in the sense of
2:44LLMs, you know, are the basis for a lot
2:46of uh
2:47very useful AI products that all of us
2:49use,
2:50including me. Uh
2:52They're great, okay, for what they do.
2:55They're just not a path towards
2:57human-level or human-like intelligence
3:00or even animal-like intelligence.
3:02Uh so, that's my claim, okay? I'm not
3:04saying LLMs are useless, right? I'm I'm
3:06just saying
3:07they're not a path towards I mean, you
3:09helped build some of the first major
3:10open-source ones.
3:11Right. Right.
3:12>> [laughter]
3:12>> Right, absolutely. So, what is uh AMI?
3:15So, AMI really stands for advanced
3:17machine intelligence. And the the the
3:21kind of subtitle, the motto, if you
3:22want, is uh AI for the real world.
3:26So, basically, a lot of
3:28you know, AI techniques that people know
3:30about today are good for language
3:32manipulation, either human language or
3:35computer code or mathematics or or
3:38legalese,
3:40which barely qualifies as human
3:42language.
3:43>> [laughter]
3:43>> Unfortunately, a lot of human language
3:44is for it. Right. Right, sadly.
3:47You know, language is very special in a
3:49way, and it's uh particularly
3:52well suited for the type of uh you know,
3:55architectures that
3:56have been so successful uh recently, the
3:59the you know, large language models,
4:01GPT-style architectures. But what about
4:04the real world? What about like
4:05understanding
4:07the physical world? Turns out reality is
4:09way more complicated than language.
4:12Uh because it's high-dimensional, it's
4:14continuous, it's noisy, it's messy.
4:18And uh training a system to understand
4:20the real world is much, much harder. So
4:21that's really what we're after. That's
4:23what I've been after for most of my
4:25career and really kind of,
4:27you know, working on in an accelerated
4:29fashion over the last uh 5, 6 years or
4:31so and making significant progress over
4:33the last 2 years.
4:35And so, it made sense to really do a
4:37startup around it and sort of go to into
4:40high gear, you know, in pushing that.
4:42And it became clear, you know, by the
4:44end of last year that
4:45Meta was really not the right place for
4:47that.
4:48So, which is why I left and started I
4:51mean, Labs. I think it's an interesting
4:53like, you know, trend that we're seeing
4:54across the board, right? Where it feels
4:56like um there you're there's there's
4:58many folks spinning out of, you know,
4:59either some of the large companies or
5:01research labs, you know, that have a
5:03particular direction of research they're
5:04excited about and
5:05you'd have some interesting vantage
5:06point of this from your time at Fair of
5:07this
5:08uh almost tension that exists between,
5:10you know, go pursue as many different
5:11research directions as possible in these
5:13companies versus hey, something's really
5:15working. This is the thing that we're
5:16going to sell for the next 6, 12 months.
5:18Like, go focus on that. You know, I'm
5:20curious your your thoughts on that and
5:21and what you kind of seen in the
5:22industry at large. Well, it's a strange
5:25uh
5:26trade-off. There's really two modes of
5:28operating, right? There's a lot of
5:29exploratory research, a lot of research
5:31directions, right? And sometimes
5:33something
5:34kind of seems to work and you you need
5:36to push it further. And it's not
5:38research anymore. I mean,
5:40the people working on it are still
5:41researchers or they're called
5:43researchers at least in the press, but
5:45uh but really it's becoming more
5:46engineering and pushing for for
5:48products, right? So,
5:51that happened a number of times at Meta
5:54because of things that were started at
5:56FAIR, such as seeing happened in, you
5:59know, early 2023, essentially. Uh, when,
6:03you know, Llama, which was developed at
6:05FAIR, Llama 1, um, was very promising.
6:09And, uh,
6:10Meta created a a whole organization,
6:12GenAI, to turn it into something real
6:15and a series of products. Uh, and
6:17produce, you know, Llama 2, Llama 3,
6:19Llama 4, which was a bit of a
6:21disappointment. Uh, and because, you
6:23know, Mark Zuckerberg was disappointed
6:25by it, he kind of rebooted the entire
6:27organization, reorganized it, and hired
6:30new people, etc. But, what also
6:32happened,
6:34uh, over the last year is that
6:36uh,
6:37basically the company, Meta, realized
6:39that,
6:40um, they'd fallen behind a little bit,
6:42and so that kind of refocused the the
6:45strategy on trying to catch up with the
6:47industry. And the sad side effect of it
6:51is that a lot of the exploratory
6:53research
6:54was basically not
6:56given high priority anymore. I mean, it
6:58didn't concern the stuff I was working
7:00on, all the Jeppa and world models,
7:02uh, cuz, you know, Mark himself and and
7:05do buzzwords, the CTO, and a bunch of
7:07other people in the company were really
7:08interested in that project and really
7:09believed in the long-term impact. But,
7:12the rest of the company was just, you
7:13know, totally entirely focused on LLMs,
7:16and made it clear to me that Meta was
7:18really not the the right place to push
7:20on that project anymore. And then we
7:22started to have good results, and so it
7:24was clear that, you know, we had to kind
7:26of make that transition between research
7:29and actually kind of uh, developing the
7:32technology, scaling it up, and building
7:33products out of it. And we realized also
7:35that most of the
7:37applications were probably
7:40for things that Meta was not
7:41particularly interested in. A lot of
7:44applications of the kind of stuff that
7:46we've been working on is in the
7:48industry, like manufacturing industry
7:50and stuff like that. Obviously, you're
7:52you're kind of pursuing world models and
7:54and and in that broader world. And I
7:55think there's other people that have
7:56come at the world model pace from a more
7:58like generative approach. And so I think
8:00you've got folks, you know, you've got
8:01the Google folks and Genie in the video
8:03models. You've got folks, you know,
8:04building VLAs on the robotic side.
8:05You've got Feifei and and kind of like
8:08the 3D spatial models. As you think
8:10about kind of the the body of of of of
8:12evidence that got you excited about the
8:14Japa models and how you kind of compare
8:15them to what the generative folks have
8:17done, you know, where do you think we
8:19are today in in terms of like comparing
8:20these architectures and approaches?
8:22Okay, so world model is quickly becoming
8:24a buzzword right now, right? Certainly
8:27in research, but also in industry to
8:28some extent. And uh and then there are
8:31two factions, if you want. I'm not going
8:33to talk about VLA because VLA
8:35is clearly now being seen as not going
8:38anywhere.
8:40Like it's really not working. Uh so VLA
8:42is, you know,
8:43vision language action models, right? So
8:45basically use the LLM technology to
8:48train a system to produce actions for
8:50like controlling a robot or something
8:52like this, right? So you have vision in,
8:54language in, action out. Maybe language
8:57out, too.
8:58Um and that's pretty much now seen as a
9:02failure.
9:03>> [laughter]
9:04>> Uh not being reliable enough, requiring
9:06too much training data, you know, things
9:07like that.
9:08Okay, then there is world models. Okay,
9:10so what is a world model? Uh a world
9:12model at a regional level is something
9:15that
9:16allows an agentic system to anticipate
9:19the consequences of its own actions.
9:22Okay, predict the consequences of its
9:24own actions. From my point of view, I
9:26cannot imagine how you can even think of
9:29building an agentic system without that
9:31system having the ability to predict the
9:33consequences of its actions.
9:35I I that's pretty essential, right? When
9:37we
9:39act in the world, we have this ability.
9:42And when we
9:43uh take an action without thinking about
9:45the consequences,
9:47we're taking a big risk. And very often,
9:49you know, other people think we're we're
9:51an idiot.
9:53Uh we have plenty of examples on the
9:55international political scene at the
9:57moment of people who have complete, you
9:59know,
10:00no ability to predict the
10:01consequences of their actions. So,
10:03that's the one model. That's what it is,
10:05right? Ability to predict the
10:06consequences of your own actions. If you
10:08If you have this ability, then you can
10:10plan
10:12a sequence of actions to
10:14accomplish a task, to you know, satisfy
10:17a goal. And you do this by
10:20planning, reasoning,
10:22uh by a process of search and
10:24optimization. You don't do this by
10:27predicting one action after the other
10:28autoregressively, like a real AI we do.
10:31Uh you do this by searching for a
10:33sequence of actions that will accomplish
10:35the task you set you set for yourself.
10:37So, the blueprint for this is completely
10:40different from what, you know, LLMs uh
10:42can do at the moment.
10:44Uh LLMs do not have the ability to
10:45predict the consequences of their
10:47actions, and they do not have any
10:48planning abilities. Because
10:50inference is by
10:52predicting the next token, right? It's
10:54not by search. Okay, so right there,
10:56you have the two characteristics that I
10:58think are essential for intelligent
11:00behavior.
11:02Ability to predict consequences of your
11:04actions. And second, uh ability to plan
11:07by optimization, by search. Um find a
11:10good sequence of actions that will
11:11produce the correct outcome. And then
11:13there is a third characteristic, which
11:15is
11:16uh how do you pre- how do you predict
11:18the consequences of your actions?
11:20Okay, so, you know, if uh if I have a
11:25water bottle in front of me. I realize
11:26some people would just listen to this
11:28and not have the picture. So, I have an
11:30open, uncapped water bottle in front of
11:33me. If I push at the bottom, it's going
11:35to slide on the table. If I push
11:38near the top, it's probably going to
11:39flip. We can't predict exactly
11:42how the
11:44the bottle will will fall in which
11:45direction.
11:47Uh we can't exactly predict how it's
11:48going to slide, you know, how the water
11:50will spill, you know, whether the table
11:52is tilted in one way and the water will
11:55uh
11:55you know, kind of flow in one direction
11:57or another.
11:58There's no way we can predict this at
12:00the pixel level.
12:01So, our mental model of the world
12:03predicts that at an abstract level of
12:06representation. So, as you were working
12:07on this architecture, was a lot of it
12:08inspired by the human brain? I mean,
12:10obviously, like the you know, the way
12:11you're articulating things is exactly
12:12how how we do things. Right, or at least
12:14by, you know, cognitive science, right?
12:16Whether you can sort of translate this
12:17into a neural architecture and things
12:19like this, that's there's a big gap
12:21there. Um okay, so that that, you know,
12:24certainly uh cognitive science was a bit
12:26of a motivation or or, you know, what uh
12:29psychological system two, which is this
12:31idea of the way you behave in sort of
12:34deliberate reflective behavior is that
12:37you do imagine, predict the consequences
12:39of your actions, and you plan
12:41uh accordingly. Contrary to system one,
12:44where you just act, you know, reactively
12:47and instinctively. So, yeah, there is an
12:49inspiration, but also there is a lot of
12:52empirical evidence
12:53that you don't want to generate pixels.
12:56Okay, I've been I've been really
12:58interested in that problem of
13:01learning models of the world by
13:03prediction for a very long time. And
13:05then had an epiphany about 5 years ago,
13:08realizing that
13:10all of the architectures that have have
13:12been successful to learn
13:14representations of images and videos
13:17are non-generative architectures. And
13:20all the generative ones basically have
13:21been failures, right? So,
13:25VAE, right? Variational autoencoders, or
13:27auto encoders more generally,
13:30uh is kind of a natural
13:32way to think about like learning
13:34abstract representations of inputs,
13:36right? So, you put a an image at the
13:38input of a
13:39of a neural net, and then you train it
13:40to just reproduce the input on its
13:43output.
13:44Uh now, with a big neural net. Now, if
13:47you just do it this way, your neural net
13:48will not do anything interesting. It
13:50will just learn the identity function.
13:51Yeah. Completely uninteresting. It
13:53doesn't work. Now, if you train a VAE to
13:55learn representations of images, you get
13:57something, but it's really not that
13:58great. Same with sparse auto encoders.
14:00Then, you have another set of
14:01techniques,
14:03uh and it's kind of derivative of
14:05something called denoising auto encoder,
14:07uh masked auto encoder is a version of
14:10this. BERT is a version of this for NLP.
14:12So, you take the image, you corrupt it
14:14in some way, and then you train this big
14:15neural net to recover the original uh
14:19the original image. There's a huge
14:20project that at FAIR on this called MAE,
14:23masked auto encoder. It was very
14:25disappointing.
14:27A lot of computation, and not not really
14:30great satisfying result. Simultaneously,
14:33uh some of the same people working on
14:35MAE, and and some other people
14:37in Paris and in New York were working on
14:39other techniques using
14:41non-generative architecture, joint
14:43embedding architecture. So, take an
14:45image, corrupt it in some way, and then
14:47run the two images through encoders, and
14:49then try to predict the representation
14:51of the original image from the
14:53representation of the corrupted one.
14:55Uh that's JEPA. Yeah. Okay. So, JEPA
14:58means joint embedding predictive
14:59architecture, right? So, you have one
15:01encoder that makes an observation,
15:03another encoder that makes a different
15:04observation. You try to predict the
15:06representation of the first one from the
15:08second one with a predictor. And those
15:11techniques turned out to work much
15:12better for representing images and
15:15video. So, things like DINO,
15:18uh DINO V1, V2, V3, um project that is
15:21still going on at at fair in Paris.
15:25Projects like I Jepa
15:27and then V Jepa and then before that
15:28there were like Sim Siam and Moco and a
15:31bunch of different techniques mostly
15:33from Meta. There was a bunch of others
15:34from other groups.
15:36Um,
15:37but
15:38that turned out to be a much better way
15:39of learning representations of images
15:42than
15:44predicting pixels. Yeah. And so
15:46it just clicked in my in my mind but you
15:48know, not just mine.
15:51That this was the way to go and
15:52predicting pixels was kind of a a losing
15:54proposition. You know, it feels like
15:56there's all these robotics demos that
15:57are released
15:59you know, from from some of the model
16:00companies that are feel increasingly
16:02impressive and maybe you know, seem to
16:04resemble things like planning and
16:05reasoning when you know, they maybe
16:07haven't seen a a room or or a specific
16:09and you know, a version of a task before
16:11and are still able to execute that task.
16:13You know, what would you say to our
16:14listeners I guess that that observe that
16:16stuff and feel like it feels like we're
16:17trending toward some real progress with
16:19some of the general approaches. Well,
16:21there is real progress and some of those
16:22demos are really impressive. Um,
16:25but
16:26>> [laughter]
16:27>> they are trained with enormous amounts
16:29of data collected either from
16:32teleoperation
16:34or from just you know, human action with
16:36things you hold in your hand that look
16:38like grippers.
16:39Grippers that you know, and you and you
16:41collect the data for that. Or just you
16:44know, tracking hands and fingers of of a
16:47person.
16:48And then translating this into kind of
16:50commands for for a robot. And so those
16:52things are trained with imitation
16:54learning mostly, right? And a little bit
16:56with you know, reinforcement learning to
16:58fine tune in mostly in simulation. So
17:01the issue with this is that you need a
17:03lot of data to train the systems
17:06to
17:07to imitation.
17:09And it
17:11it becomes expensive and it's a little
17:12brittle
17:14in the sense that you know, you need to
17:16collect lots of data for every task you
17:18want the robot to uh uh to solve.
17:21Whereas, if the system had a world model
17:23that allowed it to predict the
17:26you know,
17:27the outcome of an action, it would just
17:29plan an action to solve a new task
17:32without actually having to be trained
17:34to accomplish this task. So, the degree
17:37of generalization you would get with a
17:39world model-based system is much, much
17:41larger
17:43uh
17:43you know, kind of wider spectrum of of
17:46tasks
17:47with less training data that would be
17:49required than a a system trained with
17:51imitation learning and
17:53and you know, fine-tuning
17:54>> No doubt those approaches require more
17:56data. And I guess this question of
17:57generalization really is is the big
17:58question, right? Of you know, and I
17:59think you know, some folks have have uh
18:02have shown some results around, you
18:03know, uh getting better at task A helps
18:05with task B, but that obviously feels
18:06like there's still the big unanswered
18:08question uh you know, around those
18:09architectures. I mean, you get this uh
18:12you know, synergy between tasks. So, the
18:13more tasks that you train the system to
18:15solve, the more tasks it's being
18:17it's going to be able to acquire with
18:18with small amount of data, regardless of
18:21what what technique you use. But, but
18:22the hope with uh
18:24world models is that the system can
18:26solve new tasks at zero shot, which
18:28humans are completely capable of doing,
18:30right? And many animals as well.
18:32So, uh so, that's really the the hope.
18:35Like, you know, solving a lot more
18:36problems with uh
18:40either a small amount of training data
18:42or or no training data at all.
18:45And just a little bit of maybe, you
18:47know, RL style uh fine-tuning. Yeah.
18:49Like, you know, how
18:51how is it that a 17-year-old can learn
18:53to drive in
18:54like, a dozen hours or maybe 20 hours?
18:57Uh we have millions of hours of training
18:59data of
19:00you know, people driving cars. We still
19:02don't have level five self-driving cars,
19:04right? So, imitation learning obviously
19:06does not work even for just the task of
19:08autonomous driving. Yeah, I guess it'll
19:10be a race between the ability to develop
19:12some of those capabilities, which may
19:13take time and lots of data versus this
19:15kind of architecture. I feel like
19:16there's this dream of using video models
19:18to just generate like tons of synthetic
19:20data for for, you know, simulation and,
19:22you know, even if it's not perfect,
19:23these video models from a physics
19:24perspective, it's like helpful enough
19:26to, you know, improve
19:28robotics and in the underlying physical
19:30world. What have you made of some of
19:31those approaches? Obviously, I think
19:32Nvidia's been focused there. Google
19:33seems to be going down that road.
19:35>> I'm sort of asking you again the
19:36question,
19:37you know,
19:38why can 17-year-old launch a driving 20
19:41hours? You don't need millions of hours
19:43of demonstration. And you don't need
19:45synthetic data. Uh you don't need any of
19:47that. So, you know, I I want a system
19:49that can learn as fast as that. If we
19:51crack that, then we don't need, you
19:53know, generated data, right? I mean, we
19:55might need to train the system in
19:57simulation, but not with the same amount
19:59of uh
20:01uh
20:02you know, of time or or trials as as
20:04current systems require. It's really a
20:07question of data efficiency.
20:08>> You know, I was interviewing Jerry
20:09Tworek on the podcast. He was at OpenAI
20:11and spun out to start his own lab, and
20:13you could sense a similar tension where
20:14I think he actually might even agree
20:15that, you know, if you continued scaling
20:17RL the way we're scaling, you get more,
20:19you know, you continue getting very
20:20impressive results. But, I think he
20:22felt, "God, there's just got to be some
20:23like way more efficient way to do this."
20:25And it's interesting. It's an
20:26interesting tension because you could
20:27imagine if you're OpenAI and you know
20:29something is going to continue like you
20:30could continue scaling it and it will
20:31keep getting better. There's not a ton
20:33of incentive necessarily from a business
20:35perspective to do something more
20:36data-efficient.
20:37>> Right. And there's there's no incentive
20:39for the other companies to do anything
20:41different either because they're all
20:43chasing the same like they can't afford
20:45to kind of fall behind the others,
20:47right? So, they all work on the same
20:49thing. Yeah. And And there's a bit of
20:51this sort of, you know,
20:53kind of
20:55herd behavior
20:57uh
20:58and and, you know, in in mostly in
21:00Silicon Valley where everybody is
21:02digging the same trench. Yeah. Uh and
21:05you know, so I pur-
21:07purposely
21:09set up the headquarters of Amy Labs
21:12in Paris. Yeah.
21:13>> [laughter]
21:14>> Uh
21:16the American office being in New York,
21:18not Silicon Valley.
21:19>> [laughter]
21:19>> It's really interesting cuz I think it
21:20it it it points to a tension that, you
21:22know, it it exists in the broader
21:23ecosystem today where
21:25uh you could imagine the other side
21:26being sure, maybe there are more
21:28data-efficient methods out there, but
21:29like almost who cares because we can
21:31keep scaling what we have to to better
21:33and better results. And then obviously I
21:35think from both, you know,
21:37new things you can accomplish from these
21:39models as well as just the joy of being
21:41a researcher and finding these new
21:42things. I get why there's such an
21:43attraction to to to these other
21:45architectures as well.
21:46>> And it's a bet.
21:47But, you know, we're pretty confident
21:49because, you know, we we have results
21:51already, actually.
21:52>> And as you think about like the the kind
21:54of um the initial spaces you're most
21:56excited about for the Amy technology,
21:58like what gets you know, where do you
21:59think you know, the the technology goes
22:01and and what are you most excited about?
22:03Well, I mean, you know, AI for the real
22:04world. Um
22:07like, you know, can
22:08where is your domestic robot? Where is
22:10your level five self-driving car? Yeah.
22:12Where is uh and that's you know
22:14>> When am I going to get a domestic robot?
22:15I'm excited about this.
22:17Well, so this is several years down the
22:19line. Okay? Despite the fact that there
22:21is like
22:22huge number of companies building
22:24robots, none of those companies
22:26actually has any idea how to make them
22:28smart enough to be useful, right? Or
22:29trusted around with a baby in the house
22:31or something or
22:31>> Certainly not that. Uh but but even for
22:34like, you know, relatively narrow
22:35manufacturing task, right? You know, I
22:37mean
22:38uh none of them really knows
22:40how how to do this reliably other than
22:42you know, for by imitation learning for
22:44a small number of tasks.
22:46Uh so, how how do we make those things
22:48useful? So, that's kind of a
22:50relatively long-term objective.
22:53Shorter term, there is a huge amount of
22:55applications in industry
22:57where you need to have
22:59a a system, an intelligent system that
23:01has the ability of
23:04you know, predicting what's going to
23:05happen if I change this
23:07control variable on this complex system,
23:10be it uh
23:12a jet engine, a chemical plant, a power
23:15plant, a some manufacturing line,
23:19a patient, a human cell, right? Those
23:22are systems
23:23that are sufficiently complex that you
23:25can't
23:26model their behavior with a small number
23:28of equations. Right? So, the traditional
23:30way of modeling
23:32does not work. And what you need to do
23:34is train a neural net, deep learning
23:37system,
23:38uh to to
23:40um you know, model the dynamics of that
23:42system
23:43from data.
23:44And what you get at the end is a a
23:46phenomenological model of of that
23:49uh process, of that
23:51uh system.
23:52Um and if it's action condition, then
23:54you get
23:55basically a a world model of that system
23:58that allows you to control it optimally
24:00for whatever purpose you have. And I
24:02think the
24:04number of applications of this in
24:06industry is mind-boggling. Where do you
24:08think we'll be with uh you know, general
24:10models over the next couple years? Are
24:11there like, you know, milestones you'd
24:13point to or like, what what's your kind
24:15of view of the path of progress here?
24:16Okay, couple of years is a little short.
24:18Like, 5 years, complete world
24:19domination, essentially. [laughter]
24:21Okay. So, somewhere between on the path
24:23to world domination in 5 years. I mean,
24:25this is kind of a joke, obviously, but
24:27uh this is a quote from Linus Torvalds,
24:29right? You know, when people ask him,
24:30"What's your goal with Linux?" He said,
24:32"Total world domination." [laughter]
24:34Um he actually managed to do that.
24:36>> Yeah, very fair. To first approximation,
24:38every computer in the world runs Linux,
24:40right? So, um so, that's kind of a joke.
24:42But but in the end, I think this is the
24:44blueprint for intelligent systems of the
24:46future.
24:48There still be a a small place for LLMs,
24:51you know, for
24:52like a language interface, basically.
24:55But uh
24:56but what we're designing are are systems
24:58that are capable of thinking. They They
25:00may not be capable of talking or
25:02listening initially,
25:04but they'll do the thinking.
25:06And then you can add the talking and
25:08listening
25:10uh on top of that. I'm sure you and the
25:12team are are are eagerly working to kind
25:14of, you know, get the early proof points
25:15of this. And obviously, you've already
25:17had some in the work you've done. How do
25:18you think about like the interim steps
25:19of what you'll be able to show on that
25:21path to to 5-year world domination?
25:23Well, so I think uh
25:25you know, within a year or so, um we'll
25:28have
25:29I think a a general methodology
25:32to train
25:33hierarchical world models
25:35on, you know, a a very wide variety of
25:38modalities.
25:40We know we can do a good job on video
25:42uh with some techniques that we're not
25:44completely happy with because they have
25:46some shortcomings, but
25:48um
25:49and we have
25:50sort of small-scale demonstration of a
25:53methodology that we think is
25:55really what we want.
25:57So, we need to scale that one up
25:59and get it to the same level of
26:00performance as the
26:03the other techniques that are not as uh
26:06satis- satisfying, if you want, on on
26:08things like video, but also on other
26:10types of data sets that we would get
26:12from industry partners. Okay, so we'll
26:14have
26:15demonstrations that we can train world
26:17models, perhaps action-conditioned world
26:18models that allow us to plan
26:20for uh a number of different use cases.
26:23Some of them will be robotics, some of
26:25them will be industrial process control
26:27of various types, maybe some of them in
26:29health um health care as well cuz we
26:32have partners in that
26:33>> Yeah. in that domain. And
26:36that should be within a year or two, 18
26:38months. Um
26:40And then we'll push the this methodology
26:43and those models into
26:45uh those use cases with partners, some
26:48of which are investors already, you
26:49know, in our company, and gain
26:51experience on how to kind of
26:54essentially build a somewhat universal
26:56world model if you want. I mean, you've
26:57obviously had this uh you know, this
26:59experience before of of kind of making
27:01this really contrarian bet on neural
27:02nets and and being certainly uh proven
27:05abundantly right uh in the in in the
27:06history books. I guess as you think
27:08about this bet which I think, you know,
27:09if you talk to the majority of people uh
27:11maybe at at at at the cutting edge of
27:13various parts of AI maybe would would
27:14say is contrarian today. In what time
27:16frame do you think it will become
27:17apparent like, you know, that this was
27:19right?
27:20I think it'll happen faster than
27:24expected perhaps because I mean, you can
27:26see that world model is already becoming
27:28a buzzword, right?
27:30At least at the research level.
27:32Uh
27:34and it's starting to kind of permeate
27:35into the industry. Yeah. And a lot of
27:37people are realizing like VLs suck
27:40and, you know, LLMs don't work for real
27:42world data. Industry has realized this
27:45already. Certainly on the on the
27:48on the user side.
27:50And I think because of the importance of
27:52the robotics industry
27:54um you know, a lot of people are kind of
27:55trying to figure out like how how do we
27:57how do we get there? How do we get how
27:58you make those robots
28:00uh
28:01useful. So so I think it's
28:03I think the realization that you need a
28:05change of paradigm is is happening as we
28:07speak and will become completely obvious
28:09to people by
28:11early 2027, I think. Yeah. Now, that
28:14doesn't mean we'll have a solution by
28:15then. We hope we will, but you know,
28:17we'll see. I guess, you know, switching
28:18gears to the LM side you mentioned some
28:20of this work you're doing with uh with
28:21Tapestry which I think would be really
28:22interesting for our listeners. And so
28:24maybe to speak to that a little bit.
28:25Okay, so this is kind of a little bit
28:27orthogonal to uh to ML Labs. Yeah, as if
28:30that wasn't enough to keep you busy.
28:32>> [laughter]
28:32>> Well, it's a it's a kind of an idea I've
28:34I've been uh forming over the last uh
28:36three years or so
28:38is the fact that uh
28:41people increasingly use AI assistants
28:43for various things, right? I mean, uh
28:45you see a decrease in the use of general
28:48traditional search engines and you just
28:50ask a question to your favorite AI
28:52assistant.
28:54Um and you know, if the plan that Meta
28:58and others are are
29:00developing of, you know, having smart
29:01devices like smart glasses and stuff
29:03like that,
29:04uh
29:05you know, is realized,
29:07basically you'd just be talking to your
29:09AI assistant, you know, by voice with,
29:11you know, to your smart glasses or maybe
29:13some other smart device.
29:15And so, all of your information diet
29:18will be mediated by AI assistants.
29:22And
29:23if you are someone, you know, somewhere
29:24in the world, let's say outside the US
29:26or China, and you have an AI assistant,
29:28and that AI assistant was built in
29:29California or
29:32you know, Beijing
29:33or Shanghai or Shenzhen, uh
29:37it's not good for you. Like you may
29:39speak a language that those systems
29:41really haven't been trained to handle
29:43particularly well.
29:44Uh you may have a culture that is not
29:47particularly well understood by people
29:48in Silicon Valley and China.
29:50Not well represented by the training
29:52data that is publicly available on the
29:54internet.
29:56Um
29:57you may have a value system that is
29:58absolutely not represented by
30:01uh you know, people building those
30:02models.
30:03And certainly you'll almost certainly
30:05have political opinions that are
30:07absolutely not represented by the
30:09handful of AI assistant you you might be
30:12able to get from the
30:14you know, um West Coast tech companies
30:16or from Chinese companies.
30:19So, what is the solution to this? Like
30:21how do you serve,
30:23uh you know, a farmer in India,
30:25uh or um even a philosopher in France,
30:30uh or Germany? And
30:33what you need is
30:35a platform
30:37which basically is uh an open free
30:41foundation model, LLM style,
30:44that is fine-tunable
30:46by anyone
30:48to cater to the interest of people
30:51speaking a particular language, having a
30:53particular culture,
30:55having particular
30:57value systems, political biases,
30:59uh
31:00creeds, whatever it is.
31:03And so, what you need is a wide
31:05diversity of AI assistants. There's a
31:08lot of countries around the world uh
31:11that are neither the US nor China, who
31:14absolutely want some level of
31:16sovereignty for AI, not just for their
31:18industry, but also for their citizen.
31:20They don't want their citizen to get
31:22brainwashed by
31:24a Chinese model or a Canadian model,
31:26actually.
31:27Uh And so,
31:29they want sovereignty. How do you get
31:31that? So, the way you get
31:34a platform like the an open platform
31:36like this to get to the frontier is you
31:38just train it on more and higher quality
31:40data than the than the proprietary
31:43systems.
31:44If you talk to
31:46people in India, in France, in Vietnam,
31:49in
31:51Morocco, in Switzerland,
31:54in Korea, Japan,
31:56uh
31:57Kazakhstan,
31:59everyone wants
32:02basically sovereignty.
32:04And you tell them like, you guys have
32:06been training your model, you know,
32:08locally. You don't have to share your
32:09data. So, that's the crucial aspect of
32:11Tapestry.
32:12You would have international contributor
32:15contributors to Tapestry
32:17contributing to training a a global
32:19model that would basically constitute a
32:23repository of all the world's knowledge
32:25and culture, if you want. But the
32:26contributors would contribute uh
32:30data and uh computing resources, but
32:33they would preserve the control on their
32:35data. They would not have to share their
32:37data with the other
32:38uh contributors.
32:40What they would contribute is
32:42parameter vectors. Interesting. So, it
32:44would be kind of a kind of federated
32:46learning style thing where uh you have a
32:48bunch of data centers. Uh
32:51you know, they they get the parameter
32:53vector from the
32:55the the global consensus of a model.
32:58Think of it as an average of all the
33:01all the parameter vectors of all the
33:02contributors, right?
33:04So, all the contributors uh periodically
33:06tell
33:08everyone else through maybe a central
33:10server, here is my parameter vector,
33:13what is yours? Okay.
33:15Uh and so, you exchange parameter
33:16vectors like this. And a local worker
33:19basically, whatever it updates its
33:20parameter vector, it tries to also
33:24makes make it as close as possible to
33:26the global consensus vector. So, as the
33:30training of this thing kind of
33:31progresses, all those parameter vectors
33:34converge towards
33:36like a a consensus model, essentially,
33:38which is kind of a repository of all
33:41human knowledge. Now, you have an open
33:43an open model
33:45that is as good as if it had been
33:47trained on all the data in the world.
33:50And now, you can fine-tune it for your
33:52own purpose, your own
33:55political, cultural, and linguistic
33:56biases.
33:57Whatever you want or centers of
33:58interest.
33:59And I think there is a natural force for
34:01this to happen
34:03uh because, you know, most countries
34:05that are not the US nor China want
34:09sovereignty, but also because
34:11uh
34:12AI is fast becoming a platform. And
34:15there is a natural tendency for
34:16platforms to become open.
34:18That's what happened with Linux,
34:20right? And that's what happened with the
34:23software infrastructure of the internet
34:25or the wireless network. It's all open
34:27source.
34:28Um
34:29it was proprietary initially, but that
34:32was all
34:33wiped out. It's a really clever way to
34:34get around you know it would seem that
34:36this trend of you know decreasing open
34:38source and and obviously I think there's
34:41been many fears that is like the closed
34:42source models get better they'll be held
34:44back and they'll be used to train the
34:45next generation and you know they'll
34:46they'll kind of be this almost like a
34:48scape scenario for for closed source
34:49models where they get you know so much
34:51better than than their open source
34:52counterparts. So remember what you know
34:54who the big players of the internet
34:56infrastructure were
34:58in 1990 six?
35:01Sun Microsystems HP
35:04Dell
35:05and a few others.
35:07Um so Sun Microsystems was selling you
35:09Solaris with their you know proprietary
35:11hardware.
35:12HP with HP-UX
35:14uh they were claiming you know Unix is
35:16so much more reliable than Windows
35:17you're not going to run a web server on
35:18Windows. Dell was doing this you know
35:21with Windows NT but like who is running
35:23Windows NT now [laughter] as a web
35:25server?
35:26All of this was totally wiped out by
35:28Linux like the entire internet runs on
35:30Linux.
35:31Um even Azure right? Even Microsoft
35:34>> [laughter]
35:34>> it runs Linux. So
35:37uh
35:39basically OpenAI and Anthropic
35:41etc. of today are the
35:44Sun Microsystem and HP-UX
35:47of yesterday.
35:49Yeah I mean I guess it it implicit in
35:51that is is obviously um you know I think
35:53your you know your view of like the
35:55limitations of of what like you know the
35:57these models can only get so good and so
36:00it'll be possible over time for for the
36:01open source folks to to catch up.
36:03They've already run out of data right? I
36:04mean the the
36:06the open openly available publicly
36:08available data text data
36:11uh
36:12is already all used.
36:13I mean there's not more of it right? So
36:15what what those companies are doing is
36:17licensing uh
36:19commercial copyrighted data
36:22or training on
36:24synthetic data. I guess I'm curious cuz
36:25obviously there's been some impressive
36:27results uh in the last few years that
36:29they that they have been able to drive,
36:30you know, post these large-scale
36:31pre-trainings. Um you know, IMO gold, uh
36:34you know, the meter
36:35task horizon benchmark keeps going up.
36:37Okay, that's Okay, that's very
36:39interesting. Now, think about those two
36:40domains, right? Mathematics and code.
36:43Those are two domains where the language
36:45itself is the substrate of reasoning.
36:49It's not the only substrate of
36:50reasoning, but a lot of
36:52when you do mathematics, right? The the
36:55formal way on a piece of paper, not the
36:57intuitive stuff, but the
36:58you manipulate language, right? And LLMs
37:01are really good at this. So,
37:03um
37:03you know, proving theorems and stuff
37:05like that, that's that's what LLMs are
37:07really good at.
37:08They're not so good at the sort of
37:11you know, coming up with uh good
37:12concepts and definitions and things like
37:14that. It it's more like, here is a
37:15problem, solve it. They're problem
37:16solvers. Mathematics is not just problem
37:19solving, right? Most of it
37:21uh is actually a creative act that those
37:23things don't do.
37:25Um
37:27and same for code. So, LLMs are good
37:29programmers. They're not software
37:31architects.
37:33They're not computer scientists, right?
37:36Uh
37:37but they can program for us.
37:39So, they they they're not in a in a
37:41state where they can just, you know,
37:43replace humans uh entirely. It changes
37:46the world of humans. So, humans now,
37:48you know, kind of go one level up in the
37:50abstraction hierarchy and
37:53our world is to decide what to build.
37:54But like building it, you know, you can
37:56you can get help from LLMs. But okay,
37:59that's the the important point is that
38:01uh LLMs are particularly successful at
38:04domains where the language itself is the
38:07substrate of reasoning.
38:08Uh not for anything else. Yeah. What
38:11would an LLM like need to do to convince
38:12you uh otherwise? So, I mean, like a
38:15zero-shot agentic system, right?
38:18You have an agentic system.
38:20Give it a new problem. It's not been
38:21trained to solve that that problem.
38:23Doesn't have a script for it.
38:25Uh
38:26is it going to be able to uh accomplish
38:28this task? That it's never been trained
38:30to solve.
38:32And unless the system has the ability of
38:34predicting the consequences of its
38:36actions and then use using use that for
38:38play for planning
38:41it's not going to be able to do it. And
38:42you're not going to do this with an LLM.
38:44You're going to do this perhaps with a
38:46significantly
38:47augmented LLM that is capable of
38:50you know, search and planning blah blah
38:52blah. And currently
38:54you know, LLMs that do math and code
38:55actually do this.
38:56>> Yeah. Right? Cuz they search for you
38:58know, sequences of tokens that actually
39:00accomplish a particular task and you
39:03know, they can run the code or verify
39:05that the proof is correct or whatever.
39:08Um so, you have like a way of checking
39:10whether something that's produced is is
39:11is correct. Um but that's not a very
39:14efficient way of of doing planning.
39:16And it only works in domains where this
39:19type of search can be performed in token
39:21space.
39:22What I'm talking about with Jeppa is you
39:24don't do this in token space. You do
39:25this in you know, abstract thoughts
39:28space. And I'm sure some people
39:30listening might think, well, you know,
39:31hey, if if even if it's inefficient and
39:32it works uh and it works at you know, at
39:35things that are done in token space,
39:36that's still a large part of the uh of
39:38the economy that
39:39>> I mean, if it works, it's fine. I mean,
39:41there's again, there's nothing wrong
39:42with you know, using an LLM for what
39:44they're good at.
39:45Uh it's just not a path towards human
39:48level AI. You're missing you know, an
39:50>> like a huge
39:51uh domain.
39:52>> You seem like you you know, hey, it's
39:53going to tap out before it can become a
39:55software architect, whereas I'm sure
39:56>> going to tap out. It's it's just going
39:58to have like a a limited you know,
40:00ability to be deployed for like an it's
40:02going to become like increasingly
40:03difficult to kind of deploy it for an
40:05increasingly large number
40:07you know, of of use cases because you're
40:09going to have to collect tons of
40:10training data for each of those use
40:12cases. And
40:14there's a basically you're not going to
40:16be able to make those systems completely
40:18reliable, you know, without
40:19hallucinations or or dangerous stuff or
40:22uh etc.
40:24Unless those systems have the ability to
40:26predict the consequences of their
40:27actions, which means they're going to
40:28have to have explicit world models.
40:30Yeah, so I guess it's a bet against, you
40:31know, the uh 100% accuracy and then also
40:34the generalization uh across different
40:36tasks.
40:36>> Right. I guess, you know, one thing that
40:38that's so interesting about the way that
40:39the field has developed is obviously you
40:41uh share the the Turing Award with two
40:43others and I feel like they seem much
40:44more convinced of like maybe the the
40:46power or potential threats or safety
40:47risks of LLMs over time. Um
40:50I'm wondering like when did your view
40:51start diverging?
40:53Uh in 2023.
40:56And what like drove that in your mind? I
40:58didn't change my mind.
40:59They changed their mind, okay?
41:00[laughter]
41:01And I just about the same time, and it
41:03was basically GPT-4.
41:05I mean, Jeff basically had
41:07was not connected to any of that. He was
41:09never really interested in
41:11LLMs and discovered uh
41:14GPT-4 or, you know, 2023 when it came
41:16out.
41:17And basically he had an epiphany and
41:18said, "Oh my god, those systems, you
41:20know, are really close to human-level
41:22intelligence and they have
41:24possibly they have subjective
41:25experience.
41:26Uh
41:28and he he did a a quick calculation
41:29saying like, "Okay, the human cortex has
41:32about 16 billion neurons. If you want to
41:35um
41:37do something like backprop, okay? The
41:39brain doesn't do
41:41backprop directly.
41:42But if it does something like backprop,
41:44like some sort of, you know, gradient
41:45estimation for some sort of objective
41:47function,
41:48you would probably need like a a network
41:50of a few neurons to kind of reproduce
41:52the functionality of a virtual neuron in
41:54a in a neural net."
41:56So he said like, "Let's assume, you
41:58know, maybe you need you need
42:00a circuit of 10
42:02actual neurons
42:03to reproduce what a a backprop neuron
42:05does.
42:07Then all of a sudden your your cortex is
42:09only 1.6 billion neurons.
42:12Oh my god, GPT-4 is really close to
42:13this. Okay, so maybe it's as smart you
42:15know, it's going to get as smart as
42:16humans. I do not believe in this claim
42:18at all. This is kind of you know, Jeff's
42:22uh
42:23way of
42:25saying
42:26Okay, basically
42:28I can retire. I can declare victory.
42:31You know,
42:32I searched for the learning algorithm of
42:34the cortex all my career.
42:36Uh
42:37maybe I didn't discover what it really
42:39was, but backprop seems to be like a
42:41good substitute for it.
42:42It works really well.
42:44And so
42:45maybe that's all we need. So, I can
42:47retire.
42:49Uh
42:50and [laughter]
42:51and go around the world and give talks
42:52about you know, the potential
42:54uh
42:56promises and dangers of uh of AI.
43:00Uh that's basically what you know, I
43:02think what his uh
43:03uh intellectual kind of trajectory has
43:06been.
43:07Uh
43:08he's much less
43:10vocal about the potential dangers now
43:12than he was
43:14uh
43:15a year or two ago.
43:16He kind of realized there's probably a
43:17way to design
43:19truly intelligent systems. So, first of
43:20all, he probably you know, he realized
43:22that
43:23current LLMs are not that smart first of
43:25all. And and second
43:27uh
43:28that there's probably a need for a few
43:30breakthroughs like conceptual
43:31breakthroughs before we get to
43:33human-like intelligence.
43:35And third, that the the blueprint of
43:37those systems will be quite different
43:39from LLMs and we have probably have a
43:42way of
43:43you know, making them controllable and
43:45things like that. Yeah. I've been saying
43:47this for years, but
43:49Okay, he's sort of discovered this
43:51recently. Yeah. Same kind of there's a
43:53similar thing with Yoshua. I think what
43:55they are both worried about is
43:58the ability of
43:59society
44:01and the political system to make sure
44:04that the benefits of of AI will be
44:06maximized.
44:07And AI would not, you know, just
44:10profit you know, make a few rich people
44:13even richer.
44:14And uh you know,
44:17accentuate inequalities and and you
44:20know, cause major catastrophes because
44:23of bad usage. Okay, this is not like the
44:25the doomer scenario of AI taking over
44:27the world. It's more bad use users. What
44:31seems possible with the LLMs of today?
44:32Which is a danger, but you know, I don't
44:35I don't think it's as
44:37apocalyptic as you know, what some
44:39people have claimed it is.
44:41Certainly not as apocalyptic as what
44:43even Anthropic has claimed.
44:45And it's trying to kind of lobby
44:46governments into you know, scaring
44:48governments into kind of regulating AI
44:51because because of that.
44:52I don't I don't I don't subscribe to
44:55this at all. They seem to genuinely
44:56believe it. I think they genuinely
44:58believe it, but also I think there is,
45:01you know, some kind of commercial good
45:03commercial reasons for them to believe
45:04that. And to kind of
45:06uh
45:08you know, brainwash
45:09some people and governments into
45:11thinking their systems are are
45:12dangerous. And it sounds like you know,
45:14with these other architectures, do you
45:15think they're cuz obviously it doesn't
45:16you know, as maybe
45:18bearish as you are on LLMs being the end
45:20state of everything, you know, you have
45:22some pretty ambitious timelines too for
45:23for these new architectures. And so it
45:25doesn't seem like you think we're
45:26particularly far away from from from
45:28some very compelling capabilities. How
45:30do you think about I guess the the
45:31safety around, you know, if it ends up
45:33if these breakthroughs end up coming
45:34from new architectures and whether that
45:35should make us rest easier or not. I'm
45:38going to say something that's again
45:40might be controversial. Uh
45:42and certainly my some of my colleagues
45:44at Meta didn't like me saying this, but
45:46I think LLMs are interestingly unsafe.
45:49I don't think they can be made
45:51reliable and safe. Okay.
45:53They cannot be made reliable because you
45:55can't stop them from hallucinating.
45:57Uh and if they're agentic, you cannot
45:59guarantee they're not going to like take
46:01an action that, you know, they didn't
46:03predict the outcome of and that
46:05>> I mean, does it surprise you they can do
46:06these like 15-hour coding tasks given
46:08the concerns around reliability?
46:09>> Well, but coding is something where you
46:10can actually verify that, you know, the
46:13the the code that you generate uh you
46:15know, satisfy your specification.
46:17Um
46:19but
46:20but not everything is coding. And and
46:23there are examples of, you know,
46:25uh
46:26coding agents like wiping up your
46:28your hard drive,
46:30like
46:31uh
46:32or or doing stupid things, right? That
46:33makes you lose a lot of money or data or
46:36whatever.
46:37So, I think I think uh you know, LLMs in
46:40their current forms
46:41uh are are intrinsically unsafe
46:44because they cannot predict the
46:45consequences of their actions and
46:46because the way the task that they
46:49accomplish is determined
46:51uh is
46:53is subject to their training. You know,
46:56you you give them a prompt
46:59and then they will accomplish a task
47:01that correspond to that prompt only to
47:03the ex-
47:04to the extent that their training
47:06has conditioned them to actually do the
47:08right task corresponding to this prompt.
47:11But there is no, like, you know,
47:12hardwired constraint that will force
47:15them to accomplish this task
47:17and then, you know, predict that the
47:19task will be accomplished properly.
47:20Yeah, I mean, I think people were saying
47:21in the early days, right? They would
47:22you'd ask them a question and they'd
47:23keep asking the they'd keep asking the
47:24question, right?
47:25>> Right. Right. For example. [laughter]
47:27Uh
47:28or I mean, also they don't have common
47:29sense. Right. So, I mean, there's the
47:31the joke that was circulating like a
47:33month ago of
47:35you know, I need to wash my car and you
47:37know, the the car wash is a 100 yards
47:39from my house, should I walk?
47:43I tried it again like maybe 2 weeks ago.
47:45Uh they all say, "Yes, you should walk."
47:47Except Gemini.
47:49Gemini says
47:50>> they're training on your video of of
47:51having done having given that speech
47:53before? The The was not my video because
47:54[laughter] I I come up with this
47:55example.
47:56>> Whoever came up with it. Yeah, right.
47:57Whoever came up with it. But they are
47:58issuing sentences, right? Where where I
48:00said like, you know, an LLM can do this
48:02and then 6 months later it was like
48:03people are doing it and it's simply
48:04because, you know, as soon as people
48:07watch the podcast of me saying LLMs can
48:10do this, they of course type it into
48:12ChatGPT. So now it becomes part of the
48:14training set. And now of course, you
48:16know, the next version has that uh you
48:19know, that that thing in the fine-tuning
48:21set and of course it can answer the
48:22question but it's not because it's it
48:23becomes smart all of a sudden. It's just
48:25because it was explicitly trained with
48:27that question. So LLMs are intrinsically
48:29unsafe. Uh I don't think there is any
48:32way to fix that in the current
48:34um paradigm.
48:36Um and what I've been proposing is
48:38the architecture I've been talking about
48:40is objective-driven AI. So basically,
48:43you give an objective to an AI system,
48:46which is accomplish this task. Now,
48:48how does the system
48:50knows
48:51it will accomplish this task? It has a
48:53world model and it predicts
48:55uh
48:57you know, the outcome
48:59of a sequence of actions it imagines
49:01taking.
49:02Uh and if this uh outcome
49:05satisfies
49:07a
49:09cost function that you know, describes
49:11to what extent the task has been
49:12accomplished or not accomplished,
49:14then that system, if
49:16if the way that system works is by
49:19optimizing by optimization, finding a
49:22sequence of actions that accomplishes
49:24this task, minimizes this cost according
49:26to its world model,
49:28it can do nothing else. Yeah. Okay. And
49:32of course there's many things that can
49:33go wrong there. In in particular,
49:36uh the cost function might be
49:38inaccurate. It could be that the cost
49:40function you think is actually measuring
49:42to what extent the task has been
49:44accomplished but perhaps
49:46it's not accurate. Okay?
49:48Uh you the world model might be
49:49inaccurate. So the prediction that the
49:51system makes is actually not the right
49:52one. So, its prediction of what was
49:54going to happen as a consequence of its
49:56action wasn't right. Okay, so the system
49:58can still make mistakes, but but it can
50:01predict the consequences of its actions
50:02to some extent, which is I think
50:05indispensable for any agentic system.
50:07Now, you can add to that system is not
50:09just
50:10a cost function that guarantees a task
50:12has been accomplished, but you can also
50:14add a bunch of other
50:17objective functions, other other cost
50:18functions, or even constraints
50:21that are safety constraints.
50:23And say, "Okay, you know, don't hurt
50:24anybody on the way, right?" And you
50:26cannot specify this at a at an abstract
50:28level, but you can have, you know,
50:30low-level objective functions that
50:33put together will guarantee that the
50:34system will not be dangerous. Uh and the
50:37system cannot violate those things by
50:39construction. It will have to satisfy
50:41those conditions.
50:42Not the case for an LLM. The LLM can
50:44always escape. There's a gap between
50:47your training error and test error.
50:49There's always going to be a prompt
50:50where the system is going to do really
50:51stupid things. To talk to one specific
50:53space around LLMs, like, you know, I I
50:55think you're obviously really excited
50:56about LLMs in healthcare, and I think
50:57you know, people have been using LLMs in
50:59healthcare for for all sorts of things.
51:00And so, I'm curious how you think about
51:02like the set of things where LLMs are
51:04just not going to work in healthcare,
51:05and you need like a
51:07model that understands the world better.
51:09So, uh I mean, designing a course of
51:11treatment for a chronic disease, for
51:13example,
51:14or even a non-chronic disease,
51:17uh for a particular patient,
51:19uh which may not completely fit into,
51:22you know, templates that you've observed
51:23before.
51:24But if you have a good mental model of
51:26the
51:27dynamics of the physiology of the
51:29patient, then you might design a course
51:31of treatment that will actually bring
51:33the the patient to a good state. Yeah.
51:35Uh when I'm saying and when I'm saying a
51:36patient, it can be
51:38a cell. Okay, how do you tell
51:41a a
51:43stem cell
51:44to turn into a
51:46uh pancreas beta cell that produces
51:48insulin.
51:49Okay, you have a patient with type 1
51:51diabetes.
51:53Um you know, they have
51:55you know, their immune system basically,
51:57you know, kind of
51:59eats up their own beta cells, right?
52:01It's autoimmune.
52:02Um how do you keep making beta cells?
52:04You know, can you send a message? Do you
52:06have a a model of a of a human cell that
52:09will allow you to figure out what
52:11sequence of message you need to send to
52:14uh a stem cell so that it turns into a a
52:16beta cell. The less LLM-pilled camp and
52:18the LLM-pilled camp talk past each
52:19other, but it's like I think it's
52:21actually very possible that both
52:23what LLMs can do, which is maybe scaling
52:25what a top doctor, the treatment you get
52:27at like the top doctor or at the top
52:29place, scaling that around the world,
52:30like unbelievable potential impact of
52:32that, right? If you're able to do that.
52:34And then, you know, I think what you're
52:35talking about, which is certainly still
52:36on the come for for a lot of these
52:37things, is okay, well, even better than
52:40the top doctor. Like, how do you how do
52:41you go do that?
52:42>> more than just a top doctor, right?
52:43Because I mean, what the LLM can do well
52:45is
52:47you know, it can it can sort of
52:48regurgitate like knowledge that you can
52:50read in books, mostly.
52:52Um but if medicine was
52:55only kind of
52:57about accumulating
52:59uh declarative language that declarative
53:02knowledge that exist in books,
53:05you can be a doctor by just reading
53:06books. And you can be a doctor by
53:08reading books. You have to do, you know,
53:09residency and, you know, actually kind
53:12of listen to the heart and like press on
53:14the belly and things like that to, you
53:16know, diagnose a disease or whatever it
53:18is. Yeah, it's interesting. I'll I'll be
53:20very curious to see whether LLMs
53:21themselves can provide like, you know,
53:23top-quality health care
53:25globally. We'll have to we'll have to
53:25check back in on that one. It seems it
53:27seems like they're pretty pretty close.
53:28You know, I definitely also want to hit
53:29on your your time at Meta cuz you spent
53:30over a decade building like one of the
53:32most respected research labs in the
53:33world. You know, obviously, you recently
53:35left. As you reflect back on on the time
53:37there, what do you think you got like
53:38most right and most wrong in your time
53:40running FAIR? So, the thing we got right
53:42is
53:44uh you know, building a a a top research
53:47lab
53:48that really sort of innovated, produced
53:51a lot of the sort of basic
53:53methods and science and tools like
53:54PyTorch
53:56um that are useful to the entire
53:57industry, right?
53:59>> [snorts]
53:59>> Uh I mean, the entire industry is built
54:01on PyTorch basically, except for a few
54:03people at Google.
54:04>> [laughter]
54:05>> And I think a a culture of
54:07uh
54:08you know, openness and and
54:11and kind of you know, scientific process
54:14which I think is is necessary for
54:17breakthrough innovation.
54:19Yeah. Um because you know, there there
54:21is a lot of there's a whole chain of
54:23innovation, right? You have blue sky
54:26research
54:27uh new concepts, a lot of that takes
54:29place in universities, some of that
54:31takes place in advanced research labs in
54:33industry
54:35which can be counted on the fingers of
54:36one hand.
54:38Uh
54:38you know, Google is a good one, uh
54:40you know, fair was a good one.
54:42Hopefully, it will still be, I'm not
54:44sure.
54:44Um
54:46and you know, a few others. Then you
54:48have okay, this is a good idea, but
54:50let's push it forward and see see if it
54:52can be
54:53uh made useful, but still at the
54:56research level. In a in a sense of
54:59we're not going to fool ourselves. We're
55:00not going to try to just you know, find
55:04a solution that just works for this
55:05problem. We we're going to see if
55:08this technique that we imagine or we
55:10picked up from other people in the
55:12community
55:13can actually be pushed and and be be
55:16made uh practical, not as a product, but
55:19like we can show that it beats some
55:21record on you know, some uh
55:24task or benchmark. And then
55:26the next stage is for the the company
55:29that hosts the research lab to say,
55:30"Okay, now we're going to push the
55:31button
55:32devote a you know, big engineering
55:34effort to that uh to that vision and
55:37push it forward.
55:39That is where a lot of projects fail.
55:42That's That's where a lot of companies
55:44kind of fail to pick up. Meta was
55:46actually pretty good at this, okay? But
55:49far from perfect.
55:51It was not like, you know, textbook
55:52example of how you do it wrong, like,
55:54you know, Xerox PARC like totally
55:56missing out on,
55:57>> Yeah. you know, you GUI interface and,
55:59you know, mouse and windowing systems,
56:02right? Meta was, you know, kind of
56:04missed a few steps, essentially. And it
56:07And it it's partly partly just
56:09organizational. It's partly because
56:12uh
56:13you need a an organization that is
56:16pretty close to research, but not
56:18completely a product organization, to
56:20take the relay
56:21of, you know, pushing the technology a
56:23little further.
56:25Not making product with a 3-month
56:27deadline, but like, you know, pushing
56:28things.
56:30And
56:31we had that at one point. Yeah. At at at
56:34Facebook and Meta.
56:35Uh and then we lost it.
56:38And
56:39FAIR was basically isolated within the
56:41company, had lots of ideas that nobody
56:44picked up on. And then in 2023, the
56:46GenAI organization was created by
56:48basically taking about 60 or 70
56:52scientists and engineers from FAIR,
56:54right? Initially, and then it built up.
56:56Um but then it was under so much
56:59short-term pressure that basically that
57:01organization, GenAI, didn't have time to
57:03talk to FAIR.
57:05And so, instead of
57:07being at the forefront and innovating in
57:10LLM,
57:11uh GenAI basically had to focus on
57:14short-term things and become very
57:15conservative.
57:17And so, there was a gap, basically, in
57:18periods of mismatch between research and
57:22uh and the Is that kind of what happened
57:23with Llama 4? Yeah.
57:25Well, even with, you know, Llama 3,
57:27starting with Llama 3.
57:29So, Llama 1 was
57:31a small project within fair. 2022 early
57:332023 GenAI was
57:36created. The Lama people were basically
57:39moved to GenAI.
57:41They started working on Lama 2. And then
57:43a bunch of them realized
57:45like I could do a startup.
57:47So that was the genesis of Mistral.
57:50>> Yeah. Okay. Two of the
57:52authors of
57:55Lama 1 basically created Mistral with
57:57another guy
57:58from Google. And
58:01and you know a few people kind of left
58:03and sort of did other things.
58:05This is not a kind of a happy time at uh
58:08at Meta for various reasons. And so
58:10there were you know a bunch of people
58:12kind of left. And then the the the GenAI
58:15organization was kind of took over
58:17uh
58:18Lama
58:202 to some extent and Lama 3 and 4 was
58:23under so much short-term pressure that
58:25they became very conservative.
58:27And you know it's a combination of what
58:30is apparently of of the groups but but
58:32like pressure from the leadership and
58:36I mean there's many ways things can go
58:37wrong and you can't blame anyone in
58:39particular but
58:40um but yeah that's kind of what
58:43happened.
58:43>> I mean it feels like a lot of these
58:44organizations obviously are under
58:46short-term pressure right now because
58:47there's just a incredible race going on.
58:49And so I'm curious like obviously this
58:51this you know fair setup you had and
58:52kind of there's similar one you know at
58:54Google for for many years and certainly
58:56many researchers running around Open AI
58:57and Anthropic trying many different
58:58things. Do you think like
59:01that is still possible going forward or
59:03like is the only you know is one of the
59:04only paths to leave and and do your own
59:06company or or you know are there still
59:08places within the industry that you
59:10think have this like original ethos of
59:12fair even amidst the race that is race
59:14dynamics that are happening? I think
59:15there are a few places within Google
59:17research and DeepMind that where where
59:19people actually do research.
59:21Um
59:22but increasingly the industry has become
59:24more kind of closed right? I mean Google
59:26has certainly climbed up and, you know,
59:29Meta and Fair even is kind of going a
59:31bit in the same direction. There are
59:32restrictions on publication now, like
59:35more restrictions.
59:36Uh and so, it's still less appealing for
59:39people who really want to kind of do
59:41breakthrough research and, you know,
59:43they they don't get as much resources.
59:45If they do something that is
59:47relevant in medium term, they are told
59:49not to talk about it. And and so, it's
59:51it's not, you know, it's not a good
59:53atmosphere, I think, for for
59:55breakthrough. It's not conducive. You
59:57you, you know,
59:58I mean, basically, the get the best way
1:00:00to get breakthrough research
1:00:03of the type that, you know, you
1:00:05we were getting it in the early days of
1:00:06Fair and
1:00:08uh you know, at Bell Labs in the good
1:00:09days and Xerox PARC is you hire the best
1:00:12people and those are people who have a
1:00:13good nose to know what to work on,
1:00:16what projects to kind of attack.
1:00:18You give them the means to succeed and
1:00:21you get the [ __ ] out of the way. All
1:00:23right, pardon my French.
1:00:24>> [laughter]
1:00:27>> Yeah, I mean, I'm curious like what you,
1:00:28you know, what impact it then ends up
1:00:29having on the broader research
1:00:31community. So, obviously, one of the
1:00:31legacies of Fair is you trained, you
1:00:33know, uh so many researchers, right? And
1:00:35and like they're all throughout the
1:00:36ecosystem. Um and it feels like now the
1:00:38maybe equivalent to those people that
1:00:39came in younger in their careers at
1:00:41Fair, you know, they're joining these
1:00:42these labs uh with maybe shorter term
1:00:45priorities and focus. And I guess I'm
1:00:46wondering like, you know, uh in this
1:00:48current ecosystem where it feels like a
1:00:49lot of younger people getting into the
1:00:51field are thrust much more into these
1:00:53like short-term dynamics. Does that
1:00:54change anything about the way the the
1:00:56ecosystem evolves? Well, I mean, the
1:00:58people who tend to want to work with me
1:01:00are generally people who
1:01:03uh
1:01:04you know, sufficiently crazy to do it,
1:01:06first of all.
1:01:07>> Very very fair. And uh or or, you know,
1:01:09kind of subscribe to the the whole idea
1:01:11that
1:01:12uh in academia and during your PhD, you
1:01:15should work on the next generation of
1:01:18of AI system. You shouldn't work on the
1:01:19current generation. Yeah.
1:01:21>> Like if you work on an in in academia
1:01:23now, it's incredibly boring. At least to
1:01:24me it's boring. It's basically kind of
1:01:26studying how how and why LLMs work and
1:01:30explaining why they work or what their
1:01:32limitations are. It's like descriptive
1:01:34science. It's It's really not, you know,
1:01:36kind of creative
1:01:37very creative. Like, I I don't find that
1:01:40particularly interesting. It's useful.
1:01:41Yeah. Uh
1:01:43And, you know, if you really want to
1:01:45kind of show how to do new things with
1:01:47LLMs, like
1:01:49you're not going to have the GPUs you
1:01:50need for that. So, like, forget that.
1:01:52Like, don't work on LLM if if you're
1:01:54doing a PhD. Like, there's no point. You
1:01:56cannot contribute. How do you know it
1:01:57was time to leave Meta? It sounds like
1:01:58it was, you know, uh you know, you were
1:02:00thinking through some of these things
1:02:01over a period of time. You know, was
1:02:02there a moment that it crystallized or
1:02:04Well, it was a combination of things,
1:02:05right? Uh So, first of all, you have to
1:02:07understand uh a lot of people have like
1:02:10completely wrong idea about what my role
1:02:12at
1:02:13uh Facebook and Meta was. So, I joined
1:02:15in late 2013. Really kind of started
1:02:19early 2014. The first 4 and 1/2 years, I
1:02:22was director of FAIR. So, I built the
1:02:24FAIR organization,
1:02:25uh set up the culture, hired the key
1:02:27people,
1:02:28and and sort of managed it.
1:02:30Um
1:02:32And
1:02:33after 4 and 1/2 years, I stepped down
1:02:35from that uh role
1:02:38for a number of reasons. I then I became
1:02:40chief AI scientist. Okay, so
1:02:42uh the the
1:02:44the reason is uh
1:02:46you know, I was
1:02:49basically getting close to
1:02:52um
1:02:53turning 60.
1:02:54>> [laughter]
1:02:56>> First of all, 58. And
1:02:58uh
1:03:00I just don't want to do management.
1:03:02Okay. I mean, I was willing to do it for
1:03:04a while to get the the organization
1:03:06started, but I'm just not good at it.
1:03:08It's not the thing I'm I'm more like a,
1:03:10you know,
1:03:11scientific or technical
1:03:13visionary and engineer and scientist.
1:03:16So,
1:03:17uh
1:03:18other people are much better at
1:03:19management than I am.
1:03:20>> [laughter]
1:03:20>> So, I basically stepped down uh
1:03:22you know, two other people uh Joelle
1:03:24Pineau and uh Antoine Bordes basically
1:03:28took over uh
1:03:29the directorship of of FAIR. And I
1:03:32became chief AI scientist. So, um
1:03:35I was reporting to the CTO.
1:03:38And uh
1:03:39and
1:03:40you know, had roles of
1:03:43uh
1:03:45you basically we're starting a research
1:03:46project that I thought
1:03:48was necessary because the ambition of
1:03:50FAIR was always to build intelligent
1:03:52systems.
1:03:53Right? And I thought
1:03:55you know, I put my own research in in
1:03:57parentheses while I was running FAIR. I
1:03:59just didn't didn't have the time.
1:04:01And I thought it was important to
1:04:03basically kind of
1:04:05design the architecture of
1:04:08of like human-level,
1:04:10you know,
1:04:11human-like AI systems.
1:04:14Uh and
1:04:16you know, I had come up with the
1:04:19concept that this was going to be based
1:04:21on self-supervised learning on on, you
1:04:23know, prediction from
1:04:25sensory signals like video, things like
1:04:27that. I mean, this is these are old
1:04:28ideas.
1:04:29And uh and world models. I actually gave
1:04:32a keynote at NeurIPS in 2016
1:04:35where I I said like this is the way AI
1:04:37research should go like world models
1:04:38predict, you know, consequences of your
1:04:40actions and plan. And I said like, you
1:04:42know, RL is not the thing that will take
1:04:45us there cuz it's too inefficient.
1:04:47Supervised learning has shown its
1:04:49limits. And so, the future is
1:04:50self-supervised learning and world
1:04:52models.
1:04:53So, how do we do self-supervised
1:04:54learning and world models? And and I
1:04:56started a few projects on this with like
1:04:58a few avenues that didn't pan out.
1:05:00Uh some projects on video prediction and
1:05:03stuff like that. And uh and then came up
1:05:05with this uh concept that you could
1:05:07train self-supervised learning from
1:05:09video. Um but you have to train the
1:05:11system to make prediction in your
1:05:12representation space. So, that's the
1:05:14idea of JEPA. Yeah. And if you have
1:05:15JEPA, you can turn it into a world model
1:05:17by making it action conditioned, and
1:05:19then you can use it for planning. So, I
1:05:21had this idea around 2020, and in 2022,
1:05:23I wrote a long vision paper. So, I said,
1:05:25"I'm just going to write a paper with my
1:05:27entire vision, okay?
1:05:29Spill all my secrets like I don't care.
1:05:31Uh, but maybe that will rally a bunch of
1:05:33people to to that vision."
1:05:35And boy, did it work.
1:05:37>> [laughter]
1:05:38>> Because not only did I
1:05:41rally, you know, a bunch of students who
1:05:43kind of came working with me at NYU or
1:05:45in Paris because they wanted to work on
1:05:47this,
1:05:48but also a whole team at at at FAIR who
1:05:50said like,
1:05:51"This sounds great. That's what we want
1:05:52to work on." And then Joel Pino
1:05:55uh, said, "Well, maybe this should be
1:05:56like a major mission of uh,
1:05:59of of FAIR." Uh, we called it
1:06:01advanced machine intelligence. Yeah.
1:06:03That was the internal name of the
1:06:04project.
1:06:05>> Interesting. Okay.
1:06:06>> And they let you leave with it.
1:06:08And now it's the name of the company.
1:06:09Um, and you know, Mark Zuckerberg, you
1:06:12know, kind of
1:06:14kind of read that paper and knew what it
1:06:16was about and subscribed to the project.
1:06:18And Andrew Bosworth, the CTO, also.
1:06:20And uh, Mike Schroepfer, the uh,
1:06:23previous CTO,
1:06:24uh, Chris Cox, who was my my direct
1:06:27manager, chief product officer, also
1:06:28loved the idea. So, like, you know,
1:06:30there was a lot of support in the
1:06:31leadership uh, about this project that
1:06:34we internally called AMI.
1:06:35Uh,
1:06:37and uh, and you know, and and
1:06:40and it started
1:06:42really kind of working uh, for for
1:06:45video.
1:06:47But then,
1:06:48you know, company kind of refocused all
1:06:50of its effort on LLM.
1:06:52Despite support from Mark and Andrew,
1:06:56uh,
1:06:57Bos, we call him Bos. Um,
1:07:00you know, the all the layers below,
1:07:02like,
1:07:03didn't see the point, I think. And so,
1:07:06politically, it sort of became a little
1:07:08difficult. Uh the applications as I as I
1:07:11said of
1:07:12Japan world model are there are
1:07:14applications in like, you know, wearable
1:07:16agents and stuff like that, but and
1:07:18robotics, but but Meta chose to get rid
1:07:21of its entire robotics AI group um that
1:07:26was led by Jitendra Malik who's now at
1:07:28Amazon.
1:07:29And so,
1:07:31you know, clearly it wasn't the right
1:07:33environment anymore. Most of the
1:07:34applications were in industry that Meta
1:07:36had no interest in. Uh
1:07:39FAIR was increasingly getting pressure
1:07:42to kind of basically help MSL with uh
1:07:46LLMs.
1:07:47Um so,
1:07:49yeah, you know, it it make clear it make
1:07:51clear. And and that, you know,
1:07:54sort of ramming uh worked really well
1:07:57with investors, too, because
1:07:59when I had to raise money for Emmy,
1:08:02everybody knew my story. And you Anybody
1:08:04knew, you know, many investors
1:08:07um
1:08:08you know, staff at various VCs that read
1:08:10my paper and or had listened to my talks
1:08:13and had bought my story. They were
1:08:14realizing, you know, LLMs had
1:08:15limitations and, you know, were kind of
1:08:20interested by the idea of like building
1:08:22the next generation AI systems.
1:08:24>> I guess was was like the Scale
1:08:25acquisition like part of this catalyst
1:08:26of of like the pure LLM focus
1:08:28internally? Yeah, definitely. I mean,
1:08:30there's probably some, you know, other
1:08:31reasons to it. I think, you know, maybe
1:08:34um
1:08:35uh [snorts] I don't have any sort of
1:08:36inside information to comment on this,
1:08:38but uh
1:08:39it's possible that Mark sees in Alex
1:08:41kind of a potential successor to
1:08:43himself, like a younger version of
1:08:44himself. Yeah, I feel like that like uh
1:08:47a lot of the popular narrative or or,
1:08:49you know, in the media has been like,
1:08:50oh, like, you know, when Alex comes in,
1:08:52it then gets harder to run like a
1:08:53research organization. You know, I don't
1:08:54know if that the extent you felt that or
1:08:56>> Well, okay, so here is a big
1:08:58misconception uh about my role, my
1:09:01relation to Alex, and how AI was run at
1:09:04Meta.
1:09:05I had
1:09:06zero
1:09:08technical contribution to Llama, like
1:09:10none whatsoever. My one contribution to
1:09:12Llama
1:09:14was to argue for open-sourcing Llama 2
1:09:16because there was a big internal debate
1:09:17whether we should open-source. Like the
1:09:20legal department was against it.
1:09:22The policy department was
1:09:24kind of against it. Uh
1:09:27the comms department was for it. All the
1:09:29engineering side was for it. Like Boz
1:09:32was for it.
1:09:33Uh so there were like enormous internal
1:09:35discussions at a very high level, you
1:09:36know, 40 people from Mark Zuckerberg
1:09:39down every week for 2 hours
1:09:41>> [laughter]
1:09:41>> for months. So So really it was, you
1:09:45know, kind of a a big debate internally,
1:09:47and I really really you know, pushed um
1:09:50argued for for the fact that uh
1:09:53you know, and and Boz also was was very
1:09:55vocal about it that
1:09:57um
1:09:58the uh
1:09:59you know,
1:10:00uh safety risks were basically
1:10:02overblown.
1:10:03Uh the opportunities to create an
1:10:05industry were
1:10:07extremely strong. Um
1:10:10and that we were going to jump-start the
1:10:11AI industry by open-sourcing Llama 2,
1:10:13and in fact, that's exactly what
1:10:14happened. So but I had zero contribution
1:10:18to to Llama
1:10:19positive or negative. Like I I didn't do
1:10:21anything to stop it or slow it down or
1:10:22anything. There was a lot of people
1:10:24working on LLMs within FAIR, and it was
1:10:26fine.
1:10:27Uh
1:10:28and I never said anything against it.
1:10:30Okay.
1:10:31Um
1:10:32Other than saying this is not a path to
1:10:34a human-level intelligence, but it's
1:10:35fine. Uh
1:10:37it's useful.
1:10:38>> [laughter]
1:10:39>> Uh
1:10:40You know, same thing for speech
1:10:41recognition or translation, right?
1:10:43Uh so
1:10:45uh
1:10:46and particularly since uh 2018 when I
1:10:48stepped down from being director of
1:10:50FAIR,
1:10:51uh I didn't have any direct influence on
1:10:54what people were working on other than
1:10:57you know, basically you publishing my my
1:10:59vision and then rallying people
1:11:02uh around
1:11:03uh around my project, but
1:11:06you know, they they were working with me
1:11:07because they wanted not because I was
1:11:08their boss. I wasn't telling them to
1:11:10work with me.
1:11:12Um
1:11:13and so
1:11:14um
1:11:16so I had no positive or negative
1:11:17influence on LLM.
1:11:19>> [laughter]
1:11:20>> Within
1:11:21within Meta.
1:11:22Uh
1:11:23uh and uh I had some influence on the
1:11:25strategy, but it was more like the
1:11:27long-term and and like how how you
1:11:29maintain a research lab and things like
1:11:30this. And in the last uh year, you know,
1:11:33I mean, starting maybe early '24
1:11:36uh and certainly in '25, the the the way
1:11:40FAIR was kind of
1:11:42the direction in which it was moved and
1:11:44managed basically did not correspond to
1:11:46what I thought was necessary to preserve
1:11:49um
1:11:51you know, innovation, research, and
1:11:52breakthrough and preserve the good
1:11:54people. Like a lot of good people have
1:11:56left already. Yeah. And I guess a lot
1:11:58of, you know, it probably was harder to
1:11:59get people to work on the stuff you were
1:12:01working on internally and and I'm sure
1:12:02there's pressure for your you yourself
1:12:03to work on a lot of the LLM stuff. Yeah.
1:12:06Yeah. No, but a lot of other people also
1:12:08have left, right? No, it's it's it's
1:12:10fascinating. I mean, one thing I'm
1:12:11struck by throughout our whole
1:12:11conversation is I feel like you're
1:12:13you've like had a remarkably consistent
1:12:15point of view like, you know, on the
1:12:17in the space like FAIR for a long time
1:12:19and you can go back to your, you know,
1:12:21to a bunch of the earlier talks you
1:12:22referenced.
1:12:23You know, obviously it is a fast-moving
1:12:24space and and a ton of interesting
1:12:26things have happened in the last year.
1:12:28What's like one thing you've changed
1:12:29your mind on in the last year? I mean,
1:12:30the whole idea of uh
1:12:32what we used to call unsupervised
1:12:33learning that we now call
1:12:34self-supervised learning.
1:12:36Uh you know, until about
1:12:382003, the whole idea of
1:12:41unsupervised pre-training
1:12:43where you get a good representation for
1:12:46the input data and then you either
1:12:48fine-tune the the model with a little
1:12:50bit of supervised labeled data. And it
1:12:53sort of give us, you know, some evidence
1:12:55that this whole technique could work. I
1:12:56tried to apply this to video because
1:12:59ultimately what I wanted to do is
1:13:01train a system to understand how the
1:13:02world works by just
1:13:04watching the world go by, right? I mean,
1:13:06that's the basic idea.
1:13:08Uh and sort of started to argue for this
1:13:10in the sort of
1:13:12you know, early 2010s.
1:13:14Um
1:13:15did some some work on
1:13:17simple video prediction. We didn't have
1:13:19GPUs. Okay. Um
1:13:22and uh
1:13:24and then sort of doing this more
1:13:25seriously about after the creation of
1:13:27fair
1:13:28um by doing pixel level video prediction
1:13:32realizing that wasn't working. Uh but
1:13:34then arguing for self-supervised
1:13:35learning. Okay, this whole idea of like
1:13:37training a system generically not to
1:13:39solve a task but to basically just
1:13:41predict and then using the
1:13:42representation that is learned this way
1:13:44as input to a downstream task that you
1:13:47can train supervised or reinforcement or
1:13:49whatever.
1:13:50Uh so this that was a bit of the topic
1:13:52of my
1:13:53second half of my keynote at
1:13:55at NIPS in 2016. It was still called
1:13:57NIPS at the time.
1:13:58>> Yeah, of course. in 2016. And then I I
1:14:01kept kind of, you know, kind of pushing
1:14:02for this idea and tried to kind of
1:14:04discover some methods to to get that to
1:14:06work. And
1:14:08what surprised me is that that became
1:14:10incredibly successful but not for video,
1:14:12for language.
1:14:13LLMs basically are
1:14:16a a
1:14:18blindingly successful example of
1:14:21self-supervised learning.
1:14:23>> No, that that they are. Well, I feel
1:14:25like that's a that's almost like the
1:14:25perfect note to end on but I want to
1:14:27make sure to leave the last word to you.
1:14:29Um I feel like there's I mean, all our
1:14:30listeners are are very familiar with you
1:14:32but I want to at least give you the mic
1:14:33to point them to anything that you think
1:14:35they should they should check out with
1:14:36some of the new stuff you're doing or I
1:14:37don't know, any of your your work you
1:14:39want to point to.
1:14:40The mic is yours. Okay, let me tell you
1:14:43um
1:14:44one thing, an LLM
1:14:46works because when you have a sequence
1:14:48of discrete symbols, making predictions
1:14:51is easy.
1:14:52There's only a finite number of possible
1:14:54symbols in your language.
1:14:57100,000 possible tokens or something
1:14:58like that, right? And you can
1:15:00have your neural net produce a a
1:15:02probability distribution over all
1:15:04possible uh tokens, and then you can
1:15:07sample from that distribution, shift the
1:15:09token into the input, and then produce
1:15:10the next token, and you can do auto
1:15:12aggressive prediction. Okay, so that's a
1:15:14special case. If you have the real
1:15:15world, you can't use a generative model.
1:15:17So now you have to train a system that
1:15:19learns a representation and makes
1:15:20prediction in the representation space.
1:15:22There's a big issue with this, which I
1:15:24didn't think until about 5 years ago
1:15:26that was
1:15:28easily solvable,
1:15:30even though I invented one taking to
1:15:31solve it,
1:15:32yeah, you know, decades before that.
1:15:35Uh and it's a problem that
1:15:37um
1:15:39if you take two inputs, let's say the
1:15:41initial segment of a video and the
1:15:42continuation of that video, or you take
1:15:45one image and a corrupted version of it,
1:15:47you run them both through an encoder,
1:15:49and you train a predictor to predict the
1:15:51representation of one from the
1:15:52representation of the other.
1:15:54There's a very simple solution
1:15:56where the system basically predicts a
1:15:58constant representation, and now the
1:15:59prediction problem becomes trivial.
1:16:01That's called a collapse.
1:16:03Representation collapse.
1:16:05So the big question of self-supervised
1:16:07learning for Jappa, for the joint
1:16:08embedding architecture, is how do you
1:16:09prevent collapse? Yeah. The solution
1:16:11that uh I came up with many years ago,
1:16:151993, is uh contrastive learning.
1:16:18So basically you have
1:16:20examples of things that should be
1:16:22predictable from one another, and then
1:16:23an example of things that should not be
1:16:25predictable from one another.
1:16:26Uh it turns out this method works, but
1:16:29uh
1:16:30it doesn't scale with dimension. It
1:16:32doesn't scale very well.
1:16:34Um there's another technique that was
1:16:35actually invented by uh
1:16:38Jeff Hinton and Subbiah Ecker in the
1:16:40late '90s, late '80s, I'm sorry. Uh
1:16:44where you have those two networks and
1:16:45you try to maximize the mutual
1:16:46information between them.
1:16:48Uh
1:16:49Jürgen Schmidhuber is mad at me because
1:16:50he also came up with a version of this
1:16:52[laughter]
1:16:53in 1992
1:16:55and he says that's JEPA. It's not JEPA.
1:16:57It's just another way of preventing
1:16:58collapse of a joint embedding
1:17:00architecture. Okay.
1:17:01Uh
1:17:03>> [snorts]
1:17:03>> which is
1:17:04fine, but it's not
1:17:06you know, it's a particular way of doing
1:17:08it which I I don't think it's
1:17:09particularly good. Um so
1:17:13um
1:17:15Okay. So, now you have the JEPA
1:17:16architecture. You have to come up with a
1:17:17good way of preventing collapse.
1:17:19And there is a couple ways. So, as
1:17:22already said, contrastive methods I
1:17:24think is not a good uh a good approach.
1:17:27Uh
1:17:27there's another set of methods that are
1:17:29kind of
1:17:30called distillation methods.
1:17:33And they do prevent collapse. We we
1:17:35don't know why.
1:17:36So, a good example of that is uh DINO or
1:17:39DINo. Yep. Um that's a joint embedding
1:17:42method using the distillation method.
1:17:44Basically, one of the encoders trains
1:17:45the other one is like used as a
1:17:48teacher for the other encoder.
1:17:50Uh
1:17:51and the encoder that is being trained,
1:17:53you do backprop to it. The one that is
1:17:55not being trained, you don't do
1:17:56backprop, but you share the weight with
1:17:58the other one with some exponential
1:17:59moving average.
1:18:00It's a collection recipe. There was a a
1:18:02paper from from DeepMind about it called
1:18:04BYOL, Bootstrap Your Own Latent, which
1:18:06uses this trick. That trick is derived
1:18:08from some intuition from reinforcement
1:18:10learning. And somehow it prevents
1:18:12collapse, but we don't know why. Okay.
1:18:14There's a few theoretical papers on it
1:18:16that explain why it
1:18:19possibly might work in some simple
1:18:20cases, but it's not satisfactory. Uh
1:18:25the function the cost function you think
1:18:26you're minimizing, you're not actually
1:18:28minimizing and so you can't monitor.
1:18:30It actually goes up when you train. It
1:18:32makes sense. So, we don't like this
1:18:34method, but it works.
1:18:36And some of the models we've trained,
1:18:38large scale video representation
1:18:40learning system, VJPA, VJPA2, VJPA2.1,
1:18:44they train using this method.
1:18:46Uh I Jepa also.
1:18:48But we're moving away from this and now
1:18:49we have uh
1:18:51a few papers that came out recently on
1:18:54a a specific regularizer to prevent this
1:18:57collapse, which basically tries to
1:18:58maximize the information content coming
1:19:00out of the encoder. So, it's in the same
1:19:02family as the Becker and Hinton from '89
1:19:06and the Schmidhuber 1992 and a bunch of
1:19:09others since then. And to some extent
1:19:11also contrastive techniques also it's
1:19:12not although it's not simple
1:19:14contrastive.
1:19:15Um
1:19:17And then the question is how do you
1:19:18measure information content? How do you
1:19:19maximize
1:19:21the information content coming out of a
1:19:23neural net?
1:19:24And the problem is if you want to
1:19:25maximize the quantity, you
1:19:27either need to be able to measure it or
1:19:29you need to have a lower bound on it.
1:19:31Yeah.
1:19:32Uh information content, we only have
1:19:33upper bounds.
1:19:35We cannot measure it. We can only come
1:19:37up with upper bounds. And so, we take an
1:19:39upper bound and we cross our fingers.
1:19:41Okay. And it kind of works. So, the
1:19:43latest one is called SigReg.
1:19:46That means sketch as isotropic Gaussian
1:19:50regularization.
1:19:51We had a previous one called
1:19:54VCReg or VICReg, variance invariance
1:19:56covariance regularization.
1:19:59Um
1:20:00And the SigReg stuff is really cool.
1:20:02Um so, this is some work by
1:20:04uh
1:20:05Randall Balestriero who's uh was a
1:20:07postdoc with me. He's
1:20:08he's a
1:20:09assistant professor at Brown
1:20:11uh now. And uh it basically consists in
1:20:14forcing the distribution of variables
1:20:17coming out of the encoder to be
1:20:19uh joint Gaussian essentially. So,
1:20:21maximizing information if you want. It's
1:20:24just a very different way of doing it
1:20:25than, you know, what what Jürgen
1:20:27Schmidhuber
1:20:28>> [laughter]
1:20:29>> and Subbarao and and Jeff Hinton were
1:20:31doing.
1:20:32Um
1:20:33And so uh uh
1:20:35This this is super promising in my
1:20:36opinion and we have you know variations
1:20:38of it, you know, when that we can
1:20:39produce sparse representations.
1:20:42Another one that can produce uh
1:20:44uh anisotropic representations but not
1:20:46necessarily Gaussians. And we have uh
1:20:48uh a paper with Randall
1:20:51uh and student at at Mila uh Luca Mice
1:20:55that
1:20:57where we train a world model with this.
1:20:58It's still small scale.
1:21:00But we think it's super promising. So if
1:21:03you want to
1:21:05read one paper
1:21:06read that paper. It's Le World Model l e
1:21:09world model. Awesome. I'll definitely
1:21:10link to it, too. Yeah. I'm I'm not
1:21:12responsible for the name. Randall
1:21:13[laughter] picked up the name.
1:21:15Amazing. Well, Jan, seriously, thank you
1:21:16so much. It is such a privilege to get
1:21:18to spend the last bit of time with you
1:21:21and really appreciate you coming on the
1:21:23podcast. Thanks for having me, though.
1:21:24It's fun. I'm Jacob Effron and this has
1:21:26been unsupervised learning. A podcast
1:21:28where I get to talk to the smartest
1:21:29people on AI and ask them tons of
1:21:32questions about what's happening with
1:21:33models and what it means for businesses
1:21:35in the world. As I hope is clear, I have
1:21:36a ton of fun doing this. It's a nights
1:21:38and weekends project in addition to my
1:21:40day job as an investor at Red Point, but
1:21:42our ability to get these incredible
1:21:43guests on really comes from folks like
1:21:46you subscribing to the podcast, sharing
1:21:47it with friends. It's really what
1:21:49ultimately makes this whole thing work.
1:21:50And so please consider doing that and
1:21:52thank you so much for your support and
1:21:53listening. We'll see you next episode.

About this transcript

This page contains the full transcript of Yann LeCun on What Comes After LLMs by Unsupervised Learning: With Jacob Effron, generated from the public captions YouTube serves with the video. The transcript has 15,072 words across 2,582 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.