What this covers

This is a commentary video from TheAIGRID built around clips of Yann LeCun, Meta's chief AI scientist, speaking at Nvidia's GTC 2025. The narrator opens with LeCun's remark that he is no longer interested in LLMs, then plays his explanation of what he finds more promising. LeCun names four problems that matter more to him: getting machines to understand the physical world, giving them persistent memory, enabling reasoning, and enabling planning. The narrator frames the whole thing as a prediction that future general systems will be hybrids rather than pure language models, a claim the video presents as contested.

The first technical turn concerns tokens. LeCun points out that a token comes from a finite set of roughly 100,000, while the physical world is continuous and high-dimensional, so token-level prediction fits natural phenomena poorly. He then introduces the world model, a simplified picture of physical reality that humans and animals build in the first months of life. He leans on the psychologists' split between System 1, automatic and reactive behavior, and System 2, deliberate planning that anticipates consequences. Repeated practice compiles System 2 into System 1 policies, and System 2 is what lets people handle a new task zero-shot. His examples are mentally rotating a cube and a cat planning a jump onto furniture, both offered as reasoning that happens without language.

The later portion covers his proposed alternative, Joint Embedding Predictive Architecture (JEPA)). These systems learn abstract representations and can discard unpredictable detail, where generative models waste effort reconstructing pixels such as a specific face in a video continuation. The most contested argument is a data comparison. A four-year-old has taken in about 10^14 bytes through vision, which LeCun equates to text that would take 400,000 years to read. From this he concludes that text alone cannot reach AGI.

Sharpest takeaway

Yann LeCun argues that large language models alone cannot achieve AGI because they lack world models, persistent memory, and reasoning capabilities that operate in abstract representational space rather than token space, and proposes Joint Embedding Predictive Architectures (JEPA) as a fundamentally different architectural approach.

  • LLMs predict discrete tokens but the physical world is continuous and high-dimensional, making pixel-level or token-level prediction inefficient and ineffective
  • Human reasoning operates in abstract mental space independent of language, requiring System 2 planning capabilities that current transformer architectures cannot provide
  • A 4-year-old acquires as much visual data (10^14 bytes) in four years as text data requiring 400,000 years to read, proving vision-based world models are necessary for AGI

The claims · ranked73 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.77

Yann LeCun stated at Nvidia GTC 2025 that he is not interested in large language models anymore, contrasting them with more promising developments in AI.

factualhigh valuecontestednovelty 3/4durability 3/4· Unknown Speaker (video narrator/analyst)

“Um, I'm not so interested in LM anymore. That was one of the statements that Yanlukan made at the Nvidia GTC 2025.”

0.76

Tokens are discrete representations from a finite set (typically ~100,000 tokens in LLMs), but the physical world is continuous and high-dimensional, making token-based prediction unsuitable for modeling natural continuous phenomena.

factualhigh valueestablishednovelty 2/4durability 4/4· Yann LeCun

“Tokens are discreet. Okay. So when we talk about token generally we talk about uh a a a finite set of possibilities. is in a typical LLM the number of possible tokens is on the order of 100 thousand or something like that right”

0.75

System 2 reasoning enables zero-shot task accomplishment: humans can perform novel tasks they've never encountered before by using their world models and planning abilities, without requiring training or prior examples.

factualhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“this idea that you start by uh you know using your world model and you're able to accomplish a task even a task that you've never encountered before. Zero shot right? You don't have to be trained to solve that task. you can just so you can just accomplish that task without learning anything just on the basis of your understanding of of the world and your planning abilities.”

0.75

Humans take for granted substantial implicit knowledge about physics and physical affordances when making predictions, which token-level models must attempt to explicitly predict and learn, making token-level prediction inefficient.

causalhigh valueestablishednovelty 2/4durability 3/4· Unknown Speaker (video narrator)

“if you're just predicting the next token, there are a lot of things that are just implicit from your understanding of the physical world and you being there and all of this reasoning that goes on in your brain that you really just take for granted.”

0.74

A world model is a simplified representation of the physical world that humans acquire in the first few months of life, allowing manipulation of thoughts and prediction of physical interactions, which is much more difficult to model than language.

definitionhigh valueestablishednovelty 1/4durability 4/4· Yann LeCun

“World model is we all have world models in our in our mind. This is what allows us to um kind of you know manipulate thoughts essentially... we have models of the physical world that we acquire in the first few months of life and that's what allows us to deal with the real world and it's much more difficult to deal with the real world than to deal with language”

0.74

World models are mental representations that allow humans and animals to manipulate thoughts and predict the consequences of actions in the physical world, acquired in the first few months of life, and are fundamentally different from language-based models.

definitionhigh valueestablishednovelty 1/4durability 4/4· Yann LeCun

“World model is we all have world models in our in our mind. This is what allows us to um kind of you know manipulate thoughts essentially. So you know we have a model of the of the current world. You you know that if I if I push on this on this bottle here from the top it's probably going to flip but if I push on it at the bottom it's it's going to slide. Um and you know if I press on it too hard it's might pop. So we have models of the physical world that we acquire in the first few months of life”

0.74

Psychologists distinguish between System 1 and System 2 thinking: System 1 is automatic, reactive task execution; System 2 is deliberate planning and reasoning that recruits prefrontal cortex and world models to anticipate consequences before acting.

definitionhigh valueestablishednovelty 1/4durability 4/4· Yann LeCun

“psychologists talk about system one and system two. System one is tasks that you can accomplish without really sort of thinking about them. they they've become you become used to them and you you can accomplish them without thinking too much about them... But if you are a a if you drive for the first time or the first few hours you are you are at the wheel you have to really focus on what you're doing right and you're planning all kinds of catastrophe scenarios and stuff like that... So that's system two. You're recruiting your entire prephoto cortex to your your world model your internal world model to uh figure out the you know what's what's going to happen and then plan action so that good things happen.”

0.74

Imagining spatial transformations (e.g., mentally rotating a cube 90 degrees) is something humans can do without language, demonstrating that abstract mental reasoning operates independently of language.

factualhigh valueestablishednovelty 1/4durability 4/4· Yann LeCun

“if I if I tell you you know imagine a cube floating in front of you and now rotate that cube by 90 degrees around a vertical axis. Okay, you can do this mentally. It has nothing to do with language.”

0.71

JEPA (Joint Embedding Predictive Architecture) models learn abstract representations and are capable of manipulating those representations, reasoning, and producing sequences of actions to arrive at particular goals.

definitionhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“it's called we call it Japa uh or JPA world models and we've you know um my my colleagues and I have kind of put out a bunch of uh papers on this kind of you know first steps towards towards this over the last few years. So JPA means John predictive architecture. This is those world models that learn abstract representations and are capable of sort of manipulating those representation uh and and and perhaps reason and produce sequences of actions to you know arrive at a particular goal.”

0.71

When training systems to predict at the pixel level, the system wastes computational resources trying to generate low-level visual details that are fundamentally unpredictable and irrelevant to understanding high-level physical principles and affordances.

causalhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“if you train a system to predict at a pixel level, it spends all of its resources trying to come up with details that it just cannot invent. And so that's just a complete waste of resources.”

0.71

Joint embedding architectures that learn abstract representations of images or videos work better than generative pixel-reconstruction approaches because they can discard irrelevant information rather than wasting resources reconstructing unpredictable pixel details.

causalhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“every technique that works works by reconstructing an image from a corrupted or transformed version of it basically have failed not completely failed. They kind of work but they don't work as well as alternative architectures that we call joint embedding which essentially don't attempt to reconstruct at the pixel level. they try to learn a representation an abstract representation of the image or the video or the natural uh signal that is is being trained on so that you can make prediction in that abstract representation space.”

0.69

Repeated task performance eventually compiles learned behaviors into a policy—a reactive system that allows automatic accomplishment without planning (System 1).

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“But if you accomplish that task multiple times, then eventually it gets compiled into what's called a policy, right? So a sort of reactive system that allows you to just just accomplish that task without without planning.”

0.69

Humans can accomplish novel tasks zero-shot using only their understanding of the world and planning abilities, without prior training on those specific tasks.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“this idea that you start by uh you know using your world model and you're able to accomplish a task even a task that you've never encountered before. Zero shot right? You don't have to be trained to solve that task. you can just so you can just accomplish that task without learning anything just on the basis of your understanding of of the world and your planning abilities.”

0.69

When predicting video continuation, systems cannot invent unpredictable details like specific faces; attempting pixel-level prediction wastes resources trying to generate details that cannot be inferred from initial frames.

causalhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“if I take a video of this room and I kind of pan a camera and I stop here and I ask the system to predict you know what's the continuation of that video. It's probably going to predict it's a room and there's people sitting blah blah blah. it can there's no way it can predict what every single one of you looks like, right? That's completely unpredictable from the initial segment of the video.”

0.69

Cats perform complex planning and trajectory calculations (e.g., jumping on furniture) that involve abstract mental operations far more sophisticated than those achievable through language or token prediction.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“a cat could do this... cats do things that are much more complex than this when they plan like you know some trajectories to jump on a piece of furniture right they they do things that are much more complex than that and um that is not related to language it's certainly not done in so you know token space”

0.69

Cats and other animals perform complex planning and physical reasoning (e.g., trajectory planning to jump on furniture) that is more sophisticated than discrete linguistic reasoning, and this reasoning is not dependent on language or token space.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“a cat could do this. uh we can't specify the problem to a cat obviously through language but you know cats do things that are much more complex than this when they plan like you know some trajectories to jump on a piece of furniture right they they do things that are much more complex than that and um that is not related to language it's certainly not done in so you know token space which would be kind of actions it's done in sort of abstract mental space”

0.69

Repeated successful accomplishment of a task gradually compiles System 2 reasoning into System 1 reactive policies, allowing automatic execution without planning.

causalhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“But if you accomplish that task multiple times, then eventually it gets compiled into what's called a policy, right? So a sort of reactive system that allows you to just just accomplish that task without without planning.”

0.68

The challenge for the next few years is to develop new architectures that enable reasoning in abstract mental space rather than token space, fundamentally different from current LLM approaches.

normativehigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“that is not related to language it's certainly not done in so you know token space which would be kind of actions it's done in sort of abstract mental space so that's uh that's kind of the challenge of of the next few years um which is to figure out new architectures that allow this type of of things”

0.68

Yann LeCun identifies four main areas of focus for AI beyond LLMs: (1) getting machines to understand the physical world, (2) giving them persistent memory, (3) enabling them to reason, and (4) enabling them to plan.

factualhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“I tell you one thing which may surprise a few of you. Um, I'm not so interested in LLMs anymore... I think there are more interesting questions in uh four four things how you get machines to understand the physical world and Jensen talked about this this morning in this keynote how do you get get them to have persistent memory which not too many people talk about and then the last two are how do you get them to reason and plan”

0.68

The type of architectures needed for systems that can deal with the real physical world are fundamentally different from current transformer-based LLM architectures.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“the the type of architectures that I think we need for systems that really can deal with the real world is completely different from the ones that we deal with at the moment right predict tokens”

0.68

Reasoning in humans occurs in abstract mental states that have nothing to do with language or tokens; it is fundamentally different from current LLM-based approaches to reasoning.

factualhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“when we reason when we think we do this in some sort of abstract mental mental uh state that has nothing to do with language. You don't like kicking the tokens out. You want to be reasoning in your um latent space and not abstract space”

0.68

System 1 thinking in humans refers to automatic, reactive tasks accomplished without conscious thought (e.g., experienced driver driving without focus), while System 2 refers to deliberative thinking requiring sustained attention and planning.

definitionhigh valueestablishednovelty 0/4durability 4/4· Yann LeCun

“System one is tasks that you can accomplish without really sort of thinking about them. they they've become you become used to them and you you can accomplish them without thinking too much about them. So if you are an experienced driver you can drive even without driving assistance... But if you are a a if you drive for the first time... you have to really focus on what you're doing right and you're planning all kinds of catastrophe scenarios”

0.68

Future general AI systems will likely be hybrids combining multiple capabilities rather than being based solely on LLMs.

forecasthigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (video narrator/analyst)

“I think that you know in the future the systems that really are general AI probably will be a hybrid of some sort. They'll be a mixture of all of those capabilities and we've actually seen, you know, AI companies move towards omnimodels.”

0.68

A 4-year-old human child, awake for 16,000 hours, receives approximately 2 megabytes of visual data per second through the optic nerve, totaling about 10^14 bytes of visual input—equivalent to the text training data of current LLMs.

factualhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“a 4-year-old has been awake a total of 16,000 hours. And we have about 2 megabytes going to our visual cortex through our optic nerve. Um, every second, 2 megabytes per second roughly. Multiply this by 16,000 hours times 3600, it's about 10^ the 14 bytes.”

0.68

In 4 years of life through vision alone, humans see as much data as would take 400,000 years to read as text, indicating that training language models on text alone cannot achieve AGI.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“In four years through vision, you see as much data as text that would take you 400,000 years to read. I mean, that tells you we're never going to get to AGI, whatever you mean by this, u by just training from text. It's just not happening.”

0.68

A 4-year-old child has been awake approximately 16,000 hours and receives roughly 2 megabytes of data per second through the optic nerve to the visual cortex, totaling approximately 10^14 bytes of visual data over four years—equivalent to the amount of text data required for LLM training.

factualhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“the psychologists tell us that a 4-year-old has been awake a total of 16,000 hours. And we have about 2 megabytes going to our visual cortex through our optic nerve. Um, every second, 2 megabytes per second roughly. Multiply this by 16,000 hours times 3600, it's about 10^ the 14 bytes.”

0.68

VJepa version 1, trained on short 16-frame videos with masked prediction at the representation level, is able to detect when physical laws are violated in videos by measuring prediction error, indicating it has learned physically plausible models.

factualhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“The first version of this is trained on very short videos, just 16 frames, and it's trained to uh basically predict the representation of full video from a version of of a partially masked one. And that system apparently is able to tell you whether a particular video is physically possible or not, at least in restricted cases. And it gives you a binary output. This is feasible. This is not or maybe well no it's simpler than this. You you measure the prediction error that the system produces.”

0.68

Yann LeCun stated at Nvidia GTC 2025 that he is not so interested in LLMs anymore, focusing instead on four key areas: getting machines to understand the physical world, enabling persistent memory, developing reasoning capabilities, and planning abilities.

factualhigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (analyzing Yann LeCun's statement)

“I'm not so interested in LLMs anymore... I tell you one thing which may surprise a few of you. Um, I'm not so interested in LM anymore. That was one of the statements that Yanlukan made at the Nvidia GTC 2025.”

0.68

VJepa (Vision Joint Embedding Predictive Architecture) is an alternative approach that learns by predicting masked or missing parts of videos in abstract representation space rather than at the pixel level, allowing efficient concept learning from few examples.

definitionhigh valueestablishednovelty 2/4durability 3/4· Meta/Yann LeCun

“VJEP is pre-trained on video data allowing it to efficiently learn concepts about the physical world similar to how a baby learns by observing its parents. It's able to learn new concepts and solve new tasks using only a few examples without full fine-tuning. VJEPA is a non-generative model that learns by predicting missing or masked parts of a video in an abstract representation space.”

0.68

Vision-based training data (2 megabytes per second through the optic nerve) is fundamentally different in character from text data, requiring different computational approaches despite similar total data quantities.

causalhigh valueestablishednovelty 2/4durability 3/4· Yann LeCun

“we have about 2 megabytes going to our visual cortex through our optic nerve. Um, every second, 2 megabytes per second roughly... In four years through vision, you see as much data as text that would take you 400,000 years to read.”

0.68

Achieving AGI is impossible through text-only training because the quantity and richness of visual data humans acquire through direct environmental interaction is fundamentally different from and likely more important than text data.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“In four years through vision, you see as much data as text that would take you 400,000 years to read. I mean, that tells you we're never going to get to AGI, whatever you mean by this, u by just training from text. It's just not happening.”

0.66

LLMs are currently being improved only at the margins by industry product teams through additional data, more compute, and synthetic data generation, representing incremental rather than transformative progress.

factualhigh valuecontestednovelty 2/4durability 2/4· Yann LeCun

“you know they're kind of the last thing they are in the hands of you know industry product people kind of you know improving at the margin uh trying to get you know more data more compute generating synthetic data”

0.66

Tokens are discrete (typically on the order of 100,000 possibilities in LLMs) and therefore cannot adequately represent continuous physical phenomena in the real world.

causalhigh valueestablishednovelty 1/4durability 4/4· Yann LeCun

“Tokens are discreet. Okay. So when we talk about token generally we talk about uh a a a finite set of possibilities... in a typical LLM the number of possible tokens is on the order of 100 thousand or something like that right”

0.65

Pixel-level video prediction wastes computational resources by forcing systems to invent unpredictable details rather than learning robust abstract representations of the world.

causalhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“if you train a system to predict at a pixel level, it spends all of its resources trying to come up with details that it just cannot invent. And so that's just a complete waste of resources.”

0.65

The problem of predicting high-dimensional continuous data at the pixel level is unsolved and has proven resistant to multiple architectural approaches across multiple research efforts.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“we don't know how to do this with u with deal with what know natural data that is highdimensional and continuous and every attempt at trying to get system to understand the world or build mental models of the world by being trained to predict videos at the pixel level basically have failed”

0.64

V-JEPA trained on natural videos can detect physically implausible events in synthetic test videos (objects disappearing, changing shape, violating physics) by measuring elevated prediction error, indicating the model learns physics without explicit instruction.

factualhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“when something really strange happens in the video like an object disappears or changes shape or or you know something like that or sponty appears or doesn't obey physics so it's learn physically realistic just by observing videos... you train it on natural videos and then you test it on synthetic video where something really weird happens right?”

0.64

Current AI systems can do System 1 (automatic reaction) and are inching toward System 2 (reasoning), but ultimately need a fundamentally different architecture for genuine System 2 capabilities.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“can do system one and are trying to inching their way towards system two but ultimately I think we need a different architecture for system two”

0.64

Current approaches attempting to augment LLMs with reasoning capabilities are not the proper solution because reasoning should operate in abstract latent space rather than in token space.

normativehigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“You know, Jensen is absolutely right that uh you get ultimately more power in a system that that can sort of you know reason. I I disagree with the fact that the proper way to do reasoning is the way you know current NLMs that that have are augmented by by reasoning abilities. You're saying it works but it's not the right way. It's not the right way. I think uh you know when when we reason when we think we do this in some sort of abstract mental mental uh state that has nothing to do with language. You don't like kicking the tokens out. You want to be reasoning in your um latent space and not abstract space, right?”

0.64

Transformers designed to predict the next token in a sequence are not suitable architectures for learning representations of and reasoning about the physical world because of their fundamental architectural constraints.

causalhigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (video narrator)

“you know basically what he's stating here is that you know using a transformer to basically predict the physical world just doesn't work because of the architecture.”

0.64

Dealing with the real physical world is much more difficult than dealing with language, therefore the architectures needed for systems that can handle the real world must be completely different from current LLM architectures based on next-token prediction.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“it's much more difficult to deal with the real world than to deal with language and so the the type of architectures that I think we need for systems that really can deal with the real world is completely different from the ones that we deal with at the moment right predict tokens”

0.64

Current AI systems have System 1 (reactive) capabilities and are only beginning to develop System 2 (planning and reasoning), but achieving true System 2 reasoning requires a fundamentally different architecture from current LLM-based approaches.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“So the first thing this reasoning is system two the sort of automatic subconscious reactive policy that's system one can do system one and are trying to inching their way towards system two but ultimately I think we need a different architecture for system two”

0.64

AI companies are already moving toward omnimodels that integrate multiple modalities and architectural approaches, rather than pure text-based LLMs, as exemplified by Google's recent work.

factualhigh valueestablishednovelty 1/4durability 2/4· Unknown Speaker (video narrator)

“we've actually seen, you know, AI companies move towards omnimodels. We've seen Google actually doing that recently.”

0.64

VJepa version 2 is expected to be released soon and shows the most promising results of any model developed so far for learning abstract world representations.

forecasthigh valueestablishednovelty 1/4durability 2/4· Unknown Speaker (video narrator)

“And this is where Yanakan talks about his famous VJ architecture. And apparently, they're actually coming out with version two. very soon and this one is probably having the most promising results out of any model so far.”

0.63

Chinese research on the Sora video model demonstrated that these video prediction architectures do not actually predict the physical world but rather mimic the world based on learned patterns.

factualhigh valuecontestednovelty 2/4durability 2/4· Unknown Speaker (video narrator/analyst)

“I've done a video on where, you know, China did a lot of research on Sora and basically said that, you know, these kind of architectures don't really predict the physical world. In fact... they essentially proved that, you know, these video models aren't really predicting the physical world. It's more sort of mimicking the world based on the architecture.”

0.63

Current LLMs operate at the margin through improvements like more data, more compute, and synthetic data generation, but these incremental advances do not address the fundamental limitations of the LLM paradigm.

causalhigh valuecontestednovelty 2/4durability 2/4· Yann LeCun

“they're kind of the last thing they are in the hands of you know industry product people kind of you know improving at the margin uh trying to get you know more data more compute generating synthetic data”

0.63

Future systems that are truly general AI will likely be hybrids combining multiple architectural components and capabilities rather than pure LLMs.

forecasthigh valuecontestednovelty 2/4durability 2/4· Unknown Speaker (video narrator)

“I think that you know in the future the systems that really are general AI probably will be a hybrid of some sort. They'll be a mixture of all of those capabilities and we've actually seen, you know, AI companies move towards omnimodels.”

0.63

System 2 thinking recruits the entire prefrontal cortex and internal world model to figure out what will happen and plan actions to achieve desired outcomes in novel situations.

factualhigh valueestablishednovelty 0/4durability 3/4· Yann LeCun

“you are at the wheel you have to really focus on what you're doing right and you're planning all kinds of catastrophe scenarios... you are recruiting your entire prephoto cortex to your your world model your internal world model to uh figure out the you know what's what's going to happen and then plan action so that good things happen.”

0.63

Yann LeCun is one of the godfathers of AI research with decades of expertise, making his statements about AI limitations and future directions based on deep knowledge rather than speculation.

factualhigh valueestablishednovelty 0/4durability 3/4· Unknown Speaker (video narrator/analyst)

“he is actually one of the godfathers of AI research, and he's actually been in the AI space for quite some time. This isn't someone's first rodeo when they're making statements like this. It's based on years and years of knowledge and expertise in that field.”

0.63

Meta has publicly released VJepa to allow other researchers to build upon the work and further develop joint embedding predictive architectures.

factualhigh valueestablishednovelty 0/4durability 3/4· Meta (organization)

“To allow our fellow researchers to build upon this work, we're publicly releasing VJeppa.”

0.61

Every attempt at training systems to understand the world or build mental models by predicting videos at the pixel level has failed to produce effective world models.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“every attempt at trying to get system to understand the world or build mental models of the world by being trained to predict videos at the pixel level basically have failed”

0.61

Reading 10^14 bytes of text would take any human over 400,000 years to accomplish, representing the totality of all text available on the internet.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“That would take any of us over 400,000 years to read through that because that's kind of the totality of all the text available on the internet.”

0.61

VJepa can learn new concepts and solve new tasks using only a few examples without full fine-tuning, similar to human few-shot learning capabilities.

factualhigh valueestablishednovelty 1/4durability 3/4· Meta (organization)

“It's able to learn new concepts and solve new tasks using only a few examples without full fine-tuning.”

0.61

After 20 years of work on self-supervised learning by predicting videos, every attempt to train systems using pixel-level video prediction has failed; prediction only works at the representation level, not at the pixel level.

factualhigh valueestablishednovelty 1/4durability 3/4· Yann LeCun

“every attempt that we've tried and I've been working on this for 20 years uh of training a system using self-s supervised learning by predicting video doesn't work. It only works if you do it at a representation level.”

0.61

VJepa is trained on video data and enables learning of concepts about the physical world similar to how babies learn by observing their environment.

factualhigh valueestablishednovelty 1/4durability 3/4· Meta (organization)

“VJEP is pre-trained on video data allowing it to efficiently learn concepts about the physical world similar to how a baby learns by observing its parents.”

0.61

Joint embedding architectures that learn abstract representations of images and videos are more effective than generative approaches that attempt to reconstruct images at the pixel level.

causalhigh valuecontestednovelty 2/4durability 3/4· Yann LeCun

“every technique that works works by reconstructing an image from a corrupted or transformed version of it basically have failed not completely failed... They kind of work but they don't work as well as alternative architectures that we call joint embedding which essentially don't attempt to reconstruct at the pixel level. they try to learn a representation an abstract representation”

0.61

Transformers are not suitable architectures for predicting the physical world because they are fundamentally designed for token prediction rather than abstract representation learning.

causalhigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (video narrator/analyst)

“basically what he's stating here is that you know using a transformer to basically predict the physical world just doesn't work because of the architecture.”

0.61

A research study by China showed that video generation models like Sora do not actually predict the physical world; instead, they mimic patterns in training data based on architectural constraints, without developing genuine physical understanding.

factualhigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (video narrator/content creator)

“There's been research papers. I've done a video on where, you know, China did a lot of research on Sora and basically said that, you know, these kind of architectures don't really predict the physical world. In fact, they I can't remember exactly what they did, but in the paper, they essentially proved that, you know, these video models aren't really predicting the physical world. It's more sort of mimicking the world based on the architecture.”

0.57

Google and other AI companies have recently moved toward omnimodels that combine multiple modalities and capabilities rather than pursuing single-modality (text-only) approaches.

factualhigh valuecontestednovelty 1/4durability 2/4· Unknown Speaker (video narrator/analyst)

“we've actually seen, you know, AI companies move towards omnimodels. We've seen Google actually doing that recently.”

0.56

Modern LLMs are typically trained on approximately 30 trillion tokens, which represents roughly the totality of all text available on the internet, equivalent to about 10^14 bytes of data.

factualhigh valueestablishednovelty 1/4durability 2/4· Yann LeCun

“current LLMs are trained typically with something like on the order of 30 trillion tokens, right? Token typically is about three bytes. So that's 0.9 10 to the 13 bytes. Let's say 10 to the 14 bytes. Um that would take any of us over 400,000 years to read through that because that's kind of the totality of all the text available on the internet.”

0.55

Yann LeCun's critique and position on the future of AI beyond LLMs is credible because he is one of the founding figures ('godfathers') of modern AI research with decades of experience in the field.

normativehigh valueestablishednovelty 0/4durability 3/4· Unknown Speaker (video narrator)

“he is actually one of the godfathers of AI research, and he's actually been in the AI space for quite some time. This isn't someone's first rodeo when they're making statements like this. It's based on years and years of knowledge and expertise in that field.”

0.50

Current large language models are trained on approximately 30 trillion tokens, with each token representing roughly 3 bytes, totaling about 10^14 bytes of text data.

factualhigh valueestablishednovelty 0/4durability 2/4· Yann LeCun

“current LLMs are trained typically with something like on the order of 30 trillion tokens, right? Token typically is about three bytes. So that's 0.9 10 to the 13 bytes. Let's say 10 to the 14 bytes.”

0.50

JEPA (Joint Predictive Architecture) world models learn abstract representations and are capable of manipulating those representations to reason and produce sequences of actions to arrive at particular goals.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Yann LeCun

“it's called we call it Japa uh or JPA world models... JPA means John predictive architecture. This is those world models that learn abstract representations and are capable of sort of manipulating those representation uh and and and perhaps reason and produce sequences of actions to you know arrive at a particular goal.”

0.45

Yann LeCun and colleagues have developed the V-JEPA (Video Joint Embedding Predictive Architecture) with a second version coming soon that achieves the most promising results of any model so far for learning physical world models.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unknown Speaker (video narrator/analyst)

“this is where Yanakan talks about his famous VJ architecture. And apparently, they're actually coming out with version two. very soon and this one is probably having the most promising results out of any model so far.”

0.45

V-JEPA is trained on video data to learn concepts about the physical world similarly to how a baby learns by observing, and it can learn new concepts and solve new tasks using only a few examples without full fine-tuning.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unknown Speaker (Meta/video description)

“VJEP is pre-trained on video data allowing it to efficiently learn concepts about the physical world similar to how a baby learns by observing its parents. It's able to learn new concepts and solve new tasks using only a few examples without full fine-tuning.”

0.44

LeCun has worked on video prediction using self-supervised learning for 20 years and found that it only works when training at the representation level, not at the pixel level.

factualhigh valuespeaker onlynovelty 0/4durability 3/4· Yann LeCun

“every attempt that we've tried and I've been working on this for 20 years uh of training a system using self-s supervised learning by predicting video doesn't work. It only works if you do it at a representation level.”

0.43

V-JEPA is a non-generative model that learns by predicting masked or missing parts of videos in abstract representation space, allowing it to discard irrelevant information and achieve more efficient training than generative pixel-reconstruction approaches.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Unknown Speaker (Meta/video description)

“VJEPA is a non-generative model that learns by predicting missing or masked parts of a video in an abstract representation space. Unlike generative approaches that try and fill in every missing pixel, VJA has the flexibility to discard irrelevant information, which leads to more efficient training.”

0.42

V-JEPA version 1 is trained on short video clips (16 frames) to predict representations of full videos from partially masked versions, and can determine whether a video is physically possible by measuring prediction error on held-out test videos.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unknown Speaker (video narrator/analyst)

“The first version of this is trained on very short videos, just 16 frames, and it's trained to uh basically predict the representation of full video from a version of of a partially masked one... that system apparently is able to tell you whether a particular video is physically possible or not... You measure the prediction error that the system produces.”

0.40

Meta released V-JEPA publicly to allow other researchers to build upon the joint embedding predictive architecture work for AI systems that can understand the world, plan, reason, and accomplish complex tasks.

factualhigh valuespeaker onlynovelty 0/4durability 2/4· Unknown Speaker (Meta/video description)

“To allow our fellow researchers to build upon this work, we're publicly releasing VJeppa. We believe this work is another important step in the journey towards AI that's able to understand the world, plan, reason, predict and accomplish complex tasks.”

0.29

LeCun wrote a comprehensive paper about JEPA and how it might work approximately three years ago.

factualestablishednovelty 0/4durability 3/4· Yann LeCun

“I I think that's the that's the future. I wrote a long paper about this that explains uh how this might work about about three years ago.”

0.27

The clip of LeCun's statement about not being interested in LLMs has performed well on social media ('Twitter') because LLMs are receiving the most hype in the current AI space.

factualestablishednovelty 0/4durability 1/4· Unknown Speaker (video narrator)

“And I think this clip has done the numbers on Twitter because honestly, we do know that right now in the AI space, LLMs definitely are receiving most of the hype when it comes to AI.”

0.24

The narrator has created a previous video analyzing Chinese research on Sora that provides deeper understanding of why video generation models fail to predict physics.

factualestablishednovelty 0/4durability 2/4· Unknown Speaker (video narrator)

“There's been research papers. I've done a video on where, you know, China did a lot of research on Sora and basically said that, you know, these kind of architectures don't really predict the physical world.”

0.22

LeCun wrote a comprehensive paper about three years ago explaining how JEPA world models might work theoretically.

factualspeaker onlynovelty 0/4durability 2/4· Yann LeCun

“I wrote a long paper about this that explains uh how this might work about about three years ago.”

0.19

LeCun's statement that he is not interested in LLMs has generated significant attention on Twitter because it contradicts the prevailing focus in the AI community on LLM development.

factualspeaker onlynovelty 0/4durability 1/4· Unknown Speaker (video narrator/analyst)

“this clip has done the numbers on Twitter because honestly, we do know that right now in the AI space, LLMs definitely are receiving most of the hype when it comes to AI.”