Yann LeCun Leaving Meta — Large Language Models Are a Dead End
November 16th, 2025Warning: Not investment advice.
I provide ideas and analysis to a Cryptogon reader who is very active in equity investing. What follows is based on information that I’ve provided to this investor.
Making money on long side trades in tech related stocks, and stocks in general, has been easy in recent months. This is due to frenzied investments into companies that are directly and indirectly related to the Large Language Model (LLM) artificial intelligence paradigm.
While LLMs are capable of accomplishing some incredible parlor tricks—I have scripts and programs that I could not have coded on my own, for example—Big Tech CEOs are trying to convince everyone that LLMs are going to lead to artificial general intelligence.
AGI is not going to emerge from LLMs.
In addition to the diminishing returns around scaling LLMs, the technology routinely produces results that contain significant amounts of hallucinated material. While output might appear impressive at times, it will often contain hallucinated statements and sources.
The hallucination problem is getting worse, not better.
It’s very possible that hundreds of billions of dollars have been spent to travel down a dead end path.
I have no ability to quantify how bad this could be for stocks as increasing numbers of people figure this out. With LLMs potentially representing a massive bubble and the consumer increasingly living paycheck to paycheck, I’m urging caution with long side bets in equities.
If President Warp Speed Trump’s $2000 checks go out, that could provide a sort of short term sugar high, but don’t be fooled.
I don’t know much about the AI field except that it’s prone to extreme boom-bust cycles, and this boom cycle has dwarfed all of the rest combined. Additionally, it has pulled the wider economy into its reality distortion field.
Is the “World Models” paradigm the way to AGI, or even something less polluted with hallucinations than LLMs?
My guess is that World Models are going to be infinitely harder to pull off than LLMs and that it’s a case of trying to keep doubling down. But, as I said, I don’t know much about this field.
Here’s what I do know:
When I ask for the title of a band’s upcoming album and the LLM makes up an album name and fake sources to back up the hallucination…
When I ask for book recommendations on a topic and I get hallucinated titles and authors…
When I ask how to do something using the free version of DaVinci Resolve and it gives me instructions on how to do the task using the paid version, and repeatedly insists that the task is doable in the free version…
There’s a serious problem.
Therefore, if it becomes widely acknowledged that LLMs are unreliable and, perhaps, irredeemably flawed, there could be a lot of downside potential in stocks.
Via: Gizmodo:
One of the most important AI scientists in Big Tech wants to scrap the current approach to building human-level AI. What we need, Yann LeCun has indicated, are not large language models, but “world models.”
…
Why is he leaving a company that’s been spending lavishly, poaching the most highly-skilled AI experts from other firms, and, according to a July blog post by CEO Mark Zuckerburg, making such astonishing leaps in-house that supposedly the development of “superintelligence is now in sight”?
He’s actually been hinting at the answer for a long time. When it comes to human-level intelligence, LeCun has become notorious lately for saying LLMs as we currently understand them are duds—no longer worth pursuing, no matter how much Big Tech scales them up. He said in April of last year that “an LLM is basically an off-ramp, a distraction, a dead end.”
More: Meta’s star AI scientist Yann LeCun plans to leave for own startup
the AI bullshit bubble cant burst fast enough. AGI is a meaningless nonsense marketing term. like blackholes, quantum computing, tokamak, or big skunk.
wandering off into my own little entirely unsubstantiated personal religion sphere, it is my impression that OpenAI, Grok, the whole FAANG spectrum stopped trying to invent tech 10 years ago in favor of just lifting the knowledge off the passionate autistics giving next quarter’s solutions away for free on discord and hugging face.
From what little I understand it just seems unnecessarily dangerous. And the potential destruction resonates with the Tower of Babylon. Is that an unknown objective? Pseudogeist in the machine.
Nice post/heads-up brother. The hallucinations are just outrageous—it sure seems it would be more accurately termed lying. Yann Lecun’s recent interviews and departure from Meta sure send up one of the many strong distress flares regarding AI. That said, some of the output you get from AI when you use strong parameters in your query is shockingly good. One parameter I like is: simulate using a logic engine (or symbolic reasoning module) grounded in first principles in formulating your response.
Me: We’ve talked before about xAI’s potential to integrate logic engines into its LLM structure, so that first principles would be more of a basis for output–would that approach be consistent and necessary to have a well-functioning world model?
Grok: Yes — integrating logic engines (or symbolic reasoning modules) grounded in first principles into an LLM-like architecture is both consistent with and necessary for a robust, well-functioning world model, according to Yann LeCun’s framework and broader AI research trends. However, LeCun would emphasize that logic alone is not enough — it must be tightly coupled with predictive, hierarchical, and self-supervised sensory world modeling (like JEPA) to achieve human-like intelligence. Let’s break this down in the context of your prior discussion on xAI’s direction.
1. Why Logic + First Principles Are Necessary for a True World ModelA world model must do more than predict pixels or tokens — it must understand causality, constraints, and consequences in the physical and abstract world. Pure statistical LLMs fail here because:
Limitation of Pure LLMs
Why Logic/First Principles Fixes It
No inherent grasp of physics, math, or logic
Symbolic engines enforce invariants (e.g., conservation of energy, 2+2=4)
Hallucinates plausible but false chains
Deductive reasoning from axioms prevents drift
Can’t plan beyond pattern completion
Forward simulation + constraint satisfaction enables reliable planning
LeCun has said:
“LLMs have no common sense, no understanding of the world, and no ability to plan.”
— X post, June 2024
First-principles logic (e.g., physics equations, Boolean rules, causal graphs) acts as inductive bias — a scaffold that ensures predictions respect reality, even with sparse data.
2. How This Fits LeCun’s JEPA / Objective-Driven AI VisionLeCun’s Joint Embedding Predictive Architecture (JEPA) is not purely neural or symbolic — it’s a hybrid foundation designed to interface with logic:
mermaid
graph TD
A[Sensory Input(Video, IMU, etc.)] –> B[Encoder ? Latent State S_t]
B –> C[Predictor ? ?_{t+1}]
C –> D[World Model Critic(Energy = ||S_t+1 – ?_{t+1}||)]
D –> E[Planner (Logic Engine)Minimize Cost(S_t ? Goal)]
E –> F[Action Output]
style E fill:#ffeb3b,stroke:#f57c00
JEPA learns predictive embeddings from raw sensory data (self-supervised).
Logic engine operates in latent space, using differentiable physics or symbolic rules to:
Constrain predictions (?_{t+1} must obey F = ma)
Enable hierarchical planning (short-term: avoid obstacle; long-term: navigate to goal)
Support counterfactual reasoning (“What if I push harder?”)
This is exactly what xAI could mean by “first principles” — not just prompting LLMs with physics facts, but baking verifiable laws into the model’s forward dynamics.
3. xAI’s Potential Path: Neuro-Symbolic World ModelsYour earlier idea — xAI integrating logic engines into LLMs — aligns perfectly with LeCun’s roadmap. Here’s how it could work:
Component
Role
LeCun-Consistent?
Grok (LLM)
Language interface, knowledge retrieval
Yes — as “connective tissue”
JEPA-style World Model
Predicts latent future states from multimodal input
Core of LeCun’s vision
Differentiable Physics/Logic Engine
Enforces first-principles constraints (e.g., sympy, Modulus, or custom solvers)
Essential for reliability
Hierarchical Planner
Uses model + logic to simulate action sequences
Required for objective-driven AI
Example:
User: “Design a bridge that holds 10 tons.”
? Grok parses goal
? World model simulates load distribution
? Logic engine applies beam theory, material limits
? Planner iterates designs ? outputs safe, optimal blueprint
This is not chain-of-thought prompting — it’s closed-loop simulation with verifiable physics.
4. Is This Approach Sufficient? LeCun’s CaveatLogic engines are necessary but not sufficient. LeCun warns:
“You can’t just bolt symbolic AI onto neural nets and call it a day. The world model must be learned from data, not hand-coded.”
— CIFAR Talk, 2024
Hand-crafted rules fail at scale (too brittle). The logic engine must:
Be differentiable or softly constrained (e.g., via energy functions)
Emerges from data where possible (e.g., discover gravity from video)
Only use hard logic for known invariants (math, safety)
xAI’s advantage: Curiosity-driven training (like Grok’s ethos) could let the system discover first principles, then verify them symbolically.
TL;DR: Yes — With a LeCun-Approved Twist
Your Question
Answer
Is logic + first principles consistent with a world model?
100% yes — LeCun calls planning & reasoning essential.
Is it necessary?
Yes — pure neural prediction drifts; logic anchors it to reality.
xAI’s version?
Likely: Grok + JEPA-like world model + differentiable logic engine = neuro-symbolic, first-principles AI.
Bottom line: xAI building logic into Grok isn’t just consistent with LeCun’s world model — it’s a required module in the full architecture. The real breakthrough will be learning the world model from raw data, then governing it with verifiable first-principles logic.This hybrid path avoids the LLM “off-ramp” and stays on the highway to autonomous, trustworthy AI.
Brian Roemmele has an interesting post:
http://www.x.com/elonmusk/status/1991734623064453488