What this covers

This is a video essay from the AI Copium channel. The creator walks through a recent interview with Yann LeCun and adds commentary, including a closing section that asks whether LeCun is right. The background is his departure from AI research leadership at Meta to start AMI (Advanced Machine Intelligence)), a company built around world models. The creator offers his earlier bet on neural networks, made against widespread skepticism, as a reason to take his current contrarian position seriously. The argument centers on a split between language and reality. Large Language Models (LLMs)) work where language is the substrate of reasoning, such as coding and math. The physical world is high-dimensional, continuous, and noisy. LeCun says intelligent behavior needs two things: predicting the consequences of one's actions, and planning through search and optimization. He says autoregressive token prediction provides neither.

The middle of the video turns to robotics and driving. LeCun calls vision-language-action models a failure because they are unreliable and need too much training data. He also dismisses synthetic-data efforts at Nvidia and Google, pointing to teenagers who learn to drive in about 20 hours. He adds that humans predict outcomes at an abstract level, not at the pixel level. Other threads include his claim that benchmark gains partly come from contamination, once a publicized failure case enters the training set. He also explains why industry herd behavior favors scaling over data efficiency, and what zero-shot agentic performance would have to look like to change his mind.

The most contentious section is on safety. LeCun says LLMs are intrinsically unsafe because the gap between training and test error never closes. He contrasts them with objective-driven systems that enforce constraints through a cost function, while conceding that those systems can still err. The creator pushes back with evidence that coding agents, reasoning models, and robotics demos keep improving. LeCun's forecast is that the shift away from scaling will be obvious by early 2027.

Sharpest takeaway

Yann LeCun argues that current LLM-based AI systems cannot achieve human-level or animal-like intelligence regardless of scale, and that a paradigm shift toward world models capable of planning and predicting action consequences is necessary and will become obvious by early 2027.

  • LLMs excel only at language-based reasoning but lack ability to predict consequences of actions or plan through search, both essential for true agency
  • Current approaches like VLAs are failing in real-world robotics despite massive data, while 17-year-olds learn to drive in 20 hours, proving data efficiency through world models is the missing ingredient
  • LLMs are intrinsically unsafe due to training-test gaps that cannot be closed, whereas objective-driven world model systems can enforce safety constraints by construction

The claims · ranked27 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.

0.71

Current AI companies have little business incentive to pursue more data-efficient approaches because they can continue scaling existing systems and seeing improvements, and competitors cannot afford to fall behind, creating industry-wide herd behavior where everyone works on the same scaling approach.

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

“There's no incentive for the other companies to do anything different either because they're all chasing the same like they can't afford to kind of fall behind the others, right? So, they all work on the same thing. Yeah. And and there's a bit of this sort of you know, kind of herd behavior uh and and you know, in in mostly in Silicon Valley where everybody is digging the same trench.”

0.68

Companies like Nvidia and Google are pursuing synthetic data generation approaches to improve robotics, but this is the wrong approach when humans can learn to drive in 20 hours without synthetic data.

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

“I'm sort of asking again the question, you know, why can a 17-year-old learn to drive in 20 hours? You don't need millions of hours of demonstration. And you don't need synthetic data. Uh you don't need any of that. So, you know, I I want to see something that can learn as fast as that. If we crack that, then we don't need, you know, generated data, right?”

0.68

Humans predict consequences of actions at an abstract level of representation rather than at pixel level, understanding physical principles and abstractions rather than precise pixel-perfect outcomes, which is how world models should operate.

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

“There's no way we can predict this at a pixel level. And so, our mental model of the world predicts, but at an abstract level of representation.”

0.66

Truly intelligent systems can solve new tasks zero-shot after being exposed to their environment, something humans are completely capable of and many animals also demonstrate.

normativehigh valueestablishednovelty 1/4durability 4/4· Unidentified Speaker — Yann LeCun Says LLMs Have 2 Years Left… [85M0cTnNKCI]

“the core idea here is that truly intelligent systems shouldn't just predict the next token or action. That's not what humans do. They should be able to internally model reality itself, predict the consequences of their actions abstractly, and then plan accordingly. And according to Jan, one of the biggest pieces of evidence that current AI systems are still missing this ability is the fact that even after millions of hours of driving data, we still don't have truly generalized self-driving cars. But the hope with uh world models is that the system can solve new tasks zero shot, which humans are completely capable of doing, right? And many animals as well.”

0.66

Cost functions in world model systems might be inaccurate or world models might make incorrect predictions, so objective-driven systems can still make mistakes.

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

“there's many things that can go wrong there. In in particular, uh the cost function might be inaccurate. It could be that the cost function you think is actually measuring to what extent the task has been accomplished, but perhaps it's not accurate, okay? Uh you the world model might be inaccurate. So, the prediction that the system makes is actually not the right one.”

0.64

LLMs are intrinsically unsafe because they cannot close the gap between training error and test error—there is always a new prompt where they will do stupid things, unlike objective-driven world model systems which can enforce safety constraints by construction through cost functions.

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

“LLMs are intrinsically unsafe. Uh I don't think there is any way to fix that in the current um paradigm. Um and what I've been proposing is the architecture I've been talking about is objective-driven AI. So, basically, you give an objective to uh an AI system, which is accomplish this task. Now, how does the system knows it will accomplish this task? It has a world model, and it predicts uh you know, the outcome of a sequence of actions it imagines taking...the system cannot violate those things. By construction, it will have to satisfy those conditions. Not the case for an LLM. The LLM can always escape. There's a gap between your training error and test error. There's always going to be a prompt where the system is going to do really stupid things.”

0.64

With world models, systems could solve problems with a small amount of training data or no training data at all, plus perhaps some RL-style fine-tuning.

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

“So, uh so that's really the the hope, like, you know, solving a lot more problems with uh either a small amount of training data or or no training data at all. And just a little bit of maybe, you know, RL style uh fine-tuning.”

0.64

Yann LeCun recently left his position leading AI research at Meta to start a new company called AMI (Advanced Machine Intelligence) specifically focused on building world model-based systems as the next major AI paradigm.

factualhigh valueestablishednovelty 1/4durability 2/4· Video Creator/Analyst

“He spent over a decade leading AI research at Meta and recently left to start his own company called AMI or Ami, entirely focused on building what he believes could become the next major AI paradigm beyond LLMs.”

0.64

Most conversations about AI focus solely on scaling current systems larger, and this is where most investment money is being directed, yet Yann argues this approach fundamentally won't achieve AGI or human-level intelligence.

factualhigh valueestablishednovelty 1/4durability 2/4· Video Creator/Analyst

“Most conversations around AI right now are solely focused on scaling current systems bigger and bigger. That's also where all of the money is being pumped into. And yet, Yann is saying that that isn't going to work.”

0.63

Large Language Models are not a path towards human-level, human-like, or even animal-like intelligence, though they are useful for specific applications like coding and mathematical problem-solving where language is the substrate of reasoning.

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

“There's nothing wrong with LLMs in the sense of LLMs, you know, are the basis for a lot of very useful AI products that all of us use. Including me. They're great, okay, for what they do. They're just not a path towards human-level or human-like intelligence, or even animal-like intelligence.”

0.62

Reinforcement learning could theoretically continue to improve with scale, but this continued scaling masks a deeper need for more data-efficient approaches.

factualhigh valuecontestednovelty 1/4durability 3/4· Unidentified Speaker — Yann LeCun Says LLMs Have 2 Years Left… [85M0cTnNKCI]

“if you continued scaling RL the way we're scaling, you get more, you know, you continue getting very impressive results. But, I I he felt, "God, there's just got to be some like way more efficient way to do this."”

0.61

Reality and the physical world are fundamentally more complex than language—being high-dimensional, continuous, noisy, and messy—making it much harder to train systems to understand the real world compared to language manipulation.

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

“What about the real world? What about like understanding the physical world? Turns out reality is way more complicated than language. Because it's high dimensional, it's continuous, it's noisy, it's messy. And uh training a system to understand the real world is much, much harder.”

0.61

Yann LeCun's previous contrarian bet on neural networks in the face of widespread skepticism validates the credibility of his current contrarian position on world models versus scaling.

factualhigh valueestablishednovelty 1/4durability 3/4· Video Creator/Analyst

“I mean, and you've obviously had this uh you know, this experience before of of kind of making this really contrarian bet on neural nets and and being certainly proven abundantly right in the in in the history books.”

0.61

Intelligent behavior requires two essential characteristics: the ability to predict consequences of your actions, and the ability to plan through optimization and search to find sequences of actions that accomplish goals, neither of which LLMs can do since they operate through autoregressive token prediction rather than search.

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

“If you have this ability, then you can plan a sequence of actions to accomplish a task to, you know, satisfy a goal. And you do this by planning, reasoning, uh, by a process of search and optimization. You don't do this by predicting one action after the other autoregressively, like a VAE we do. Uh, you do this by searching for a sequence of actions that will accomplish the task you set you set for yourself. So, the blueprint for this is completely different from what, you know, LLMs can do at the moment. Uh, LLMs do not have the ability to predict the consequences of their actions, and they do not have any planning abilities. Because inference is by predicting the next token, right? It's not by search.”

0.60

LLMs appear to improve on benchmarks through training set contamination rather than genuine capability improvement—when people see Yann discuss an LLM limitation, they test it, the example enters training data, and the next version appears to have 'learned' the skill when it was actually fine-tuned on that specific question.

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

“There are issues and sensors, right? Where where I said like, you know, an LLM can do this and then 6 months later it was keep people doing it. And it's simply because, you know, as soon as people watch the podcast of me saying LLMs can do this, they of course type it into ChatGPT. So, now it becomes part of the training set. And now, of course, you know, the next version has that uh you know, that that thing in the fine-tuning set and of course it can answer the question, but it's not because it's it became smart all of a sudden. It's just because it was it specifically trained with that question.”

0.57

Visual Language Action (VLA) models—systems using LLM technology to produce robot control actions from vision and language input—are now seen as a failure because they are not reliable enough and require too much training data.

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

“I'm not going to talk about VLA because VLA is clearly now being seen as not going anywhere. Like it's really not working. Uh so VLA is, you know, visual language action models, right? So basically use the LLM technology to train a system to produce actions for like controlling a robot or something like this, right? So you have vision in, language in, action out. Maybe language out, too. Uh, and that's pretty much now seen as a failure. Uh, not being reliable enough, requiring too much training data, you know, things like that.”

0.56

LLMs lack common sense reasoning, demonstrated by the fact that they consistently fail simple reasoning tasks like determining whether a 17-year-old should walk 100 yards to a car wash when they need to wash their car at home.

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

“Well, I mean, also they don't have common sense, right? So, I mean, there's the the joke that was circulating like a month ago of you know, I need to wash my car and, you know, the car wash is 100 100 yards from my house, should I walk? I tried it again like maybe 2 weeks ago. Uh they all say yes, you should walk, except Gemini.”

0.56

While some current AI systems like self-driving cars in certain environments are pretty capable, they are not generalized AI drivers—they cannot be deployed into any random car on any road in any condition the way humans can.

factualhigh valueestablishednovelty 1/4durability 2/4· Video Creator/Analyst

“Sure, we have some pretty capable self-driving systems right now in certain environments. I'll be the first to tell you that, but we still don't have what you'd really call a generalized AI driver. It's not like we have a system that we can just throw into any random car on any road in any condition and trust that it'll drive relatively well. Humans can do that.”

0.55

The blueprint and architecture for planning through search and optimization is completely different from what LLMs can do, making this a fundamental architectural divide.

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

“So, the blueprint for this is completely different from what, you know, LLMs can do at the moment. Uh, LLMs do not have the ability to predict the consequences of their actions, and they do not have any planning abilities. Because inference is by predicting the next token, right? It's not by search.”

0.51

Coding agents are getting better almost every month, reasoning models are improving, and robotics demos are becoming more frequent and impressive, contradicting the claim that current approaches are fundamentally limited.

factualhigh valueestablishednovelty 1/4durability 1/4· Video Creator/Analyst

“Current AI systems are still improving at an insane pace. Coding agents are getting better almost every month. Reasoning models are getting better. Robotics demos are becoming more impressive and frequent.”

0.47

A world model is fundamentally defined as a system that allows an agentic system to anticipate and predict the consequences of its own actions, which LeCun considers essential and unavoidable for any truly intelligent agentic system.

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

“A world model at a regional level is something that allows an agentic system to anticipate the consequences of its own actions. Okay, predict the consequences of its own actions. From my point of view, I cannot imagine how you can even think of building an agentic system without that system having the ability to predict the consequences of its actions. I mean, that's pretty essential, right?”

0.45

There is a race between developing data-efficient capabilities through world model approaches versus continued scaling of current data-hungry architectures, with the key question being whether systems can learn as efficiently as 17-year-old humans without needing millions of training demonstrations.

forecasthigh valuespeaker onlynovelty 1/4durability 3/4· Yann LeCun

“I want to see something that can learn as fast as that. If we crack that, then we don't need, you know, generated data, right? I mean, we might need to train the system in simulation, but not with the same amount of uh uh you know, of time or or trials as as current systems require. It's really a a question of data efficiency.”

0.45

The shift in the AI industry toward recognizing that current approaches don't work for real-world data and that a paradigm change is necessary is happening now and will become completely obvious to people by early 2027.

forecasthigh valuespeaker onlynovelty 2/4durability 1/4· Yann LeCun

“And a lot of people are realizing that VL suck. And you know, LLMs don't work for real-world data. Industry has realized this already. Certainly on the on the on the user side. And I think because of the importance of the robotics industry, uh you know, a lot of people are kind of trying to figure out like how how do we how do we get there? How do we get How do you make those robots uh useful? So so I think it's I think the realization that you need a change of paradigm is is happening as we speak. It will become completely obvious to people by early 2027, I think.”

0.43

What would convince Yann that LLMs are a path to AGI is if they could solve zero-shot agentic tasks they haven't been trained on without having scripts for them, which requires predicting consequences and planning.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Yann LeCun

“What would an LLM like need to do to convince you uh otherwise? So, I mean, like a zero-shot agentic system, right? You have an agentic system, give it a new problem. It's not been trained to solve that that particular problem. Doesn't have a script for it. Uh is it going to be able to uh accomplish this task that it's never been trained to solve? And unless this system has the ability of predicting the consequences of its actions and then use using use that for play for planning. It's not going to be able to do it.”

0.34

When humans act, they have the ability to anticipate consequences; the example of politicians making decisions without thinking through consequences demonstrates how incompetent such behavior is.

factualestablishednovelty 0/4durability 4/4· Yann LeCun

“And when we, uh, take an action without thinking about the consequences, we're taking a big risk, and very often, you know, other people think we're we're an idiot. Uh, we're plenty of examples on the international political scene at the moment, but people who have complete, you know, no ability to predict the consequences of their actions.”

0.17

AMI stands for Advanced Machine Intelligence and has the motto 'AI for the real world', focused on applying AI techniques to physical world understanding rather than language manipulation.

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

“I mean, really stands for advanced machine intelligence. And the the the kind of subtitle, the motto, if you want, is uh AI for the real world. So, basically, a lot of you know, AI techniques that people know about today are good for language manipulation, either human language or computer code or mathematics or or legalese which barely qualifies as human language...what about the real world? What about like understanding the physical world?”

0.17

LeCun hopes to have solutions for world model-based systems in place by early 2027, but acknowledges this is aspirational and outcomes are uncertain.

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

“Now, that doesn't mean we'll have a solution by then. We hope we will, but you know, we'll see.”