Waiting on the World to Change
When Meta's Chief AI Scientist of 12 years, Yann LeCun, quit the company, it was because he disagreed about the future of LLMs, placing him at odds with nearly everyone. What if he was right?
Yann LeCun made headlines last year when he left Meta, but he didn’t just leave the company; he threw cold water on the entire AI sector on the way out and at a time when excitement and enthusiasm were reaching a fever pitch. He’s back in the news for criticizing Elon Musk and xAI as a “failure,” in part because most of the founding team has left the company. The man is unafraid of controversy, but he saves his most savage barbs for LLMs (Large Language Models), declaring the technology a dead end.
To be specific, when LeCun was criticizing the technology, he was criticizing its ability to achieve AGI (Artificial General Intelligence); think Jarvis from Iron Man. When people talk about “the singularity” and are fearful of AI’s ability to become sentient, they’re generally talking about AGI, even when they don’t mention it specifically.
LeCun left Meta, made this bombastic declaration, and then… nothing; but what if he was right? Yann LeCun has been right about AI before, and he’s generally considered to be one of the “Godfathers of AI,” so when he declares LLMs a “dead end,” you would think people would listen, unless they didn’t like what he was saying.
If LeCun is right, then that means that billions of dollars have been invested in the wrong technology, that is, if you’re trying to create Artificial General Intelligence. That would pose a threat to OpenAI, Anthropic, Google, Meta, and any number of companies who have pushed their chips in on this specific branch of Artificial Intelligence.
LeCun even said it himself in this excerpt from Brown University, where he recently spoke on AGI: “Here’s another controversial statement: There’s literally hundreds of billions invested in an industry that basically is counting on the fact that LLMs [are] going to reach human-level intelligence,” he said. “It’s complete BS.”
It’s true, there’s a lot of money riding on Large Language Models and Gen AI leading to an AI revolution similar to the Industrial Revolution. LeCun’s assertion comes at an inconvenient time for those heavily invested in the technology, but the question remains, is AGI a requirement for success, or can these technologies succeed on their own merits to deliver the first phase of this revolution?
A Whole New World
LeCun’s World Model concept is actually quite brilliant, which probably isn’t surprising. To understand why he feels we need to take this next step into the World Model, let’s first take a look at LLMs and why he feels they’re insufficient to tackle the tasks we’re asking of them, or at least what we expect of AI.
“We have systems that can manipulate language, and they fool us into thinking they are smart because they manipulate language. But in fact, they are completely helpless when it comes to the physical world.”
He’s not wrong. LLMs have been derogatorily referred to as autocomplete on steroids, or some variation of that notion. The characterization is apt, and it gets to the point of what LLMs try to do, which is to say, they use machine learning to find themes and return outputs/answers.
The reason why all the foundational LLMs indiscriminately scraped nearly everything from the internet to train on is so when you input “What are some themes in Steve Jobs’ biography?” the LLM first has the terms “biography,” “Steve Jobs,” and any number of other phrases. It can then make its connections with the text it’s sucked in to give some kind of answer, though that answer may be a total hallucination.
The issue with language is that it’s already abstracted from the original thing. It’s the Magritte’s Ceci n’est pas une pipe (“This is not a pipe”) painting with every input. This has a knock-on effect that means the LLM doesn’t actually know anything; the tokens (words and phrases LLMs use to process requests) are abstracted from the things they represent. So Steve Jobs isn’t Steve Jobs to an LLM; it’s just the words “Steve Jobs,” and those words appear frequently together. They also appear in the context of certain phrases and descriptions. If you brought an LLM to life somehow and asked it to point out a picture of Steve Jobs, it wouldn’t be able to do that, not without some kind of external tool, in which case the tool is doing the work and the LLM is just expressing it.
Which brings me to my next point: today we compensate for LLMs’ shortcomings by using scaffolding, a secondary layer which can be comprised of machine-learning functions with specific actions or traditional functions executed in standard programming languages. The end product is something greater than the sum of its parts; it can give a really smooth experience, if a deceptive one.
That’s not good enough for LeCun because, let’s be honest, it’s a hack. What he’s looking for is a single approach that can take inputs and contextualize them without scaffolding, or perhaps the scaffolding is more specific to giving the model context about the world around it.
In LeCun’s world model, the model compensates for these shortcomings. Using my previous example, the model would presumably know that Steve Jobs was the founder of Apple, that he was adopted by an Armenian family, that he founded Next and co-founded Pixar, and that he sadly passed away from pancreatic cancer in 2011. Additionally, the model would presumably know that Tim Cook is the current Apple CEO and that the new CEO John Ternus starts on September 1st, 2026.
This is all stuff LLMs don’t have context for yet, not without tooling and scaffolding, so you can see why LeCun is dubious of them as a viable path to AGI. So while LeCun’s model sounds magnificent and like the correct approach, it also seems like a purist approach to AI versus a pragmatic approach to AI. That distinction is important because LeCun is a scientist and researcher first.
My question is, does that cloud his perspective on LLMs and their practical applications as he pursues the perfect solution for Artificial General Intelligence? More importantly, just how far off is this solution? We have LLMs today; the world model hasn’t even been demoed yet, so while the world model sounds intriguing, there is no “world” model until it can meet the rigors of the real world itself.
Is it AGI or Bust?
LeCun’s whole premise for leaving Meta and taking his position on LLMs is that they aren’t the path to Artificial General Intelligence (AGI), but is that really a requirement for AI to succeed in the short term, or are his goals different from the sector at large?
LeCun isn’t the only one talking about AGI. Sam Altman has used AGI as his North Star for OpenAI and frequently uses it to intimidate the world at large for reasons unknown.
Just a quick recap of Altman’s history discussing AGI:
2023: Altman discusses the possibilities of AGI in a blog.
2024: Altman predicted AGI would do most marketing work by 2029.
2025: Altman said AGI is generally pretty well understood, which is different from existing.
“We are now confident we know how to build AGI as we have traditionally understood it. We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies. We continue to believe that iteratively putting great tools in the hands of people leads to great, broadly distributed outcomes.”
Obviously, this didn’t happen in real terms. We had companies adopt AI on some level and deploy agents via MCPs, but they were far from self-sufficient, singular beings.
Early 2026: Altman shifts the goalposts on what AGI is (10:00), about what the definition of AGI is.
Today: Altman declares that college students will graduate into a world with AGI in existence.
So clearly, these two tech luminaries have different opinions on the timeline of when AGI will descend upon the unsuspecting public and whether it will happen at all with the current tech stack.
We’re still only scratching the surface of agentic AI, and really, agentic AI can more be defined as AI working on behalf of a user, not necessarily in a sentient way. That’s a really important distinction because AI working in a domino effect of functions and LLM processes is very much doable today; sentience, not so much.
So is the former enough, or is the Valley going to be hard-lined about what qualifies as agentic AI? Moreover, if it works, do users care? The bar today is to make AI functions that are easier, faster, and better than the previous generation.
As someone building a product, I can tell you that the bar is nosebleed high. The Overton window of what’s acceptable has shifted from rudimentary prototypes to a refined, high-functioning product. There is no more passing along crudely taped-together MVPs to users and clients to get feedback. These products are competing in the real world; you’re not trying to beat the idea of something or something offline that is slow and clunky, you’re trying to beat something slick, fast, simple, and polished most of the time.
I’m a pragmatist, so if you ask me, no, I don’t think that AGI is a requirement for this wave of AI to succeed and to bring about meaningful change. This wave of AI doesn’t represent sentience; it’s more like streaming video. Mind you, streaming video unlocked an enormous marketplace, one that upended all of broadcast TV and continues to grow in unexpected ways. I make this comparison because LLMs change how data is processed, similar to how streaming video changed how footage was transmitted. Both go through a unique process that allows a transformation and ultimately increases the speed, but also the nature, of the original thing. Streaming video turned footage into bits, and LLMs turn words into tokens.
I feel that analogy can give us a sense of what we’re in for in terms of disruption until we land upon a new method of contextualizing our world for machines. That’s really what LLMs are trying to do: use this universal tool, language, to provide context.
This is ultimately why I believe Yann LeCun to be correct about LLMs. I’m very much eager to see what he or others come up with in their pursuit of transforming our world into something comprehensible for machines, but for those of us building products today, there is an opportunity right now to build on something that, while imperfect, has the potential to unlock amazing things, maybe not sentient things, but amazing nonetheless.