Yann LeCun, who is hailed as the Godfather of AI, recently published a paper that changes the way AI models are designed and developed. Unlike the current generation of AI models that rely on what they have read endlessly for hours after training, LeCun’s paper says you can make AI models have common sense, i.e., understand the world like a human being.
The AI Godfather walked away from Meta to prove that today’s generative AI boom is heading toward a dead end. His new startup, AMI Labs, is building an architecture designed to learn the laws of physics before it ever learns language.
Yann LeCun, who is also a Turing Award winner, a co-pioneer of deep learning, and Meta’s long-standing Chief AI Scientist, decided he had seen enough of the generative AI hype cycle. He knew a secret that the industry’s most vocal tech leaders, CEOs and executives were trying to hide – today’s smartest AI systems are blind to reality, i.e., they don’t understand the world.
By March 2026, however, the tech universe saw what LeCun could do outside of Meta’s influence. His newly created startup, Advanced Machine Intelligence Labs (AMI Labs), announced a major $1.03 billion seed funding round, pushing the valuation to $3.5 billion before it had even shipped a single line of commercial code.
His new mission? To abandon the ‘text-obsessed, brute-force’ scaling laws of Silicon Valley and solve the deepest and tricky puzzle in computer science, i.e., giving AI true common sense.
AI and the limits of language
To understand why LeCun is risking his legacy on a completely different AI architecture, one must understand his total dissatisfaction with AI systems like ChatGPT or Claude.
To the average user, an LLM feels magical. It writes essays, debugs code, and mimics human empathy. But to LeCun, these systems are merely operating on “System 1” thinking, i.e., highly sophisticated digital reflexes that predict the next word in a sequence without an underlying comprehension of what those words actually mean.
“AI systems still lack the general common sense of a cat,” LeCun has frequently pointed out.
Consider a feline. A cat can navigate a cluttered room, judge the trajectory of a jumping mouse, grasp gravity, and adapt to a changing physical environment instantaneously, all powered by a biological brain that consumes a fraction of the energy of a household lightbulb. It achieves this without ever reading a single word of text.
Contrast that with today’s frontier LLMs, which eat more text than a human could read in ten lifetimes, and yet still hallucinate facts, fail at basic logic puzzles, and lack any concept of physical reality.
This gap becomes painfully obvious when AI attempts to step out of digital environments and into the physical world. A human teenager can learn the basics of driving a car in about 20 hours of practice because they already possess an internal model of physics – they know intuitively that if they steer into a brick wall, the car will smash. An AI agent trained via text or pure digital reinforcement learning, however, requires millions of miles of simulated driving data (crashing thousands of times over) because it has no understanding of cause and effect. It lacks a “world model” — the internal mental simulator we call common sense.
Brute force won’t be enough for AI
For the last several years, the prevailing dogma in AI has been ‘scaling laws’, i.e., the belief that if you add more parameters, feed it more text, and burn more compute, true intelligence will eventually emerge. LeCun fiercely disagrees, considering auto-regressive language models as a dead-end in engineering.
Because LLMs predict text one token at a time, their errors accumulate exponentially, thus making them fundamentally weak. They are optimised for probabilistic likelihood rather than factual grounding or physical constraints.
Furthermore, the internet’s supply of high-quality, human-generated text is finite, and tech companies are already running up against a data ceiling.
Then there’s the resource cost. The massive, energy-guzzling data centers required to train and run these trillion-parameter models are economically and environmentally unsustainable. AMI Labs’ strategy completely flips the script. Instead of training massive generalist models on text scraped from the web, LeCun’s team is focused on teaching machines how the physical world moves, shifts, and reacts by training them primarily on continuous, high-dimensional visual data, like video.
The JEPA blueprint: When AI thinks before speaking
At the heart of LeCun’s AI common-sense revolution is an architecture known as the Joint-Embedding Predictive Architecture (JEPA). Developed alongside collaborators during his final years at Meta and now being used at scale within AMI Labs, JEPA completely rejects the traditional generative paradigm.
Generative AI models try to predict and reconstruct every single pixel in an image or every token in a sentence. If an AI tries to simulate a video of a car driving down a tree-lined street, a generative model wastes immense computational power trying to predict the exact motion of every single leaf blowing in the wind — an unpredictable and ultimately irrelevant detail.
JEPA, by contrast, is non-generative. It ignores the superficial “noise” and focuses strictly on the “signal,” predicting abstract, high-level representations in a latent feature space. Instead of reconstructing the exact pixels of the leaves, it understands the macro-dynamics – the car is moving forward, the road is solid, and objects persist even when they momentarily go out of frame.
This abstract representation allows the AI to form a true, functioning world model. When shown a video where something physically impossible occurs, such as a thrown ball suddenly morphing into a cube or vanishing into thin air, AMI Labs’ early prototypes immediately flag it as impossible.
They aren’t just matching language patterns; they are executing “System 2” thinking, i.e., deliberate reasoning, memory, and hierarchical planning.
It is essentially an architecture designed to think before it speaks.
AMI Labs has five years to make a true world model
AMI Labs has a five-year runway before it is expected to deliver a commercially saleable product. Instead of building consumer chatbots or quick software wrappers, they are partnering with data-intensive industries like advanced manufacturing, biomedicine, and robotics – environments where physical reality reigns supreme and where a single hallucination can mean catastrophic failure.
Their first official collaborator is Nabla, a healthcare AI startup, aiming to ground machine intelligence in the highly precise, high-stakes world of clinical workflows.
Furthermore, AMI Labs is running a dramatically leaner ship. While traditional foundational models require clusters of thousands of GPUs, JEPA-based world models run on a fraction of that hardware. Some of LeCun’s recent breakthroughs, like the LeWorldModel paper, have demonstrated advanced physical reasoning and planning capabilities utilising only 15 million parameters, trained in just a few hours on a single GPU.
In essence, it is a victory of elegant mathematics over brute force.
AI won’t end the world: LeCun
LeCun’s mission is as much philosophical as it is technical. In an era dominated by apocalyptic warnings from tech executives claiming that artificial general intelligence (AGI) could trigger human extinction, LeCun has emerged as a voice of sanity.
“Don’t listen to CEOs,” LeCun recently advised students. “They have a vested interest in propping up the power of the products they sell.” He has openly called existential risk predictions “extremely destructive and wrong,” noting the severe psychological toll alarmist rhetoric takes on young researchers.
To LeCun, the fear that AI will break its bounds and shackles to enslave humanity is a sci-fi fantasy. By building objective-driven, modular architectures constrained by hard-coded safety guardrails that are grounded in the laws of physics, intelligence can be engineered to be controllable.
LeCun envisions a future where AI doesn’t replace humanity, but amplifies it.
Conclusion
The stakes for LeCun and his AMI Labs could not be higher. Critics wonder whether enterprise clients and investors will eventually lose patience with a research lab that rejects quick monetisation in favour of long-term scientific breakthroughs. Others question whether world models can truly catch up to and outperform the staggering, multi-billion-dollar inertia of the current LLM paradigm.
But if Yann LeCun is right, the current AI race is chasing a mirage. The AI Godfather wants to prove that real intelligence doesn’t start with language – it starts with common sense, just like humans.