The Teacher and the Defector
Meta built a two-trillion-parameter model whose entire purpose is to teach smaller models and then get out of the way. In the same window, the scientist who won a Turing Award building Meta’s AI quit, called the whole paradigm a dead end, and raised a billion dollars to prove it. Both moves point at the same uncomfortable place: the thing everyone is renting you today is not where the value is going to live.
There are two ways to read what Meta has been doing with AI, and they only look contradictory until you notice they’re the same bet.
The first: Meta built a monster and named it a servant. Llama 4 Behemoth — roughly two trillion parameters total, 288 billion active across sixteen experts — was never designed to be the product you talk to. Meta’s own engineering writeup calls it a teacher model. Its job is distillation: run the giant once, use its outputs to train much smaller “student” models, and ship the students. That’s how Meta built Maverick and Scout — co-distilled from Behemoth, the big model’s intelligence compressed into footprints small enough to run cheap. The teacher stays in the lab. The students go out into the world, onto laptops and eventually phones, free to download.
Sit with the strategy, because it’s genuinely radical for a company Meta’s size. Everyone else in the frontier race is trying to build the smartest possible model and rent it to you by the token — keep it locked in their datacenter, meter every query, defend the moat. Meta looked at the same board and did the opposite: give the weights away. Commoditize the layer. Zuckerberg has argued the logic in the open for years — open-weight models keep Meta from being trapped in a competitor’s ecosystem, and if intelligence is going to be cheap and everywhere anyway, better to be the company that made it cheap than the one whose margin it destroys. The teacher-and-student design is that philosophy turned into architecture: the point of the giant is to make the small models good enough that you never need to rent the giant.
That is the accessible-local-AI future arriving on schedule. Not the frontier walking out of the datacenter — the frontier being manufactured for export. A single company spending giant-model money specifically so the useful model can be small, local, and yours.
The catch nobody’s putting on the box
Now the part the press releases skip. Behemoth — the teacher, the whole premise above — was never actually released. Meta previewed it in April 2025 as “still in training,” and as of mid-2026 the public weights still have not shipped. Reporting says it’s effectively shelved: the smaller models improved faster than expected and narrowed the gap, expert-routing bugs produced inconsistent outputs, and the giant stopped justifying its own cost. The teacher, it turns out, may have taught itself out of a job.
And the next flagship tells a darker story for the open-source thesis. Meta’s Llama 5 — codename “Avocado” — has slipped from March to mid-2026, reportedly because it was falling short of Google’s Gemini in internal tests. More telling: Meta is said to be weighing whether to release Avocado closed — API-only, no downloadable weights. If that happens, the company that spent five years making “open weights” the most adopted foundation strategy on Earth would be quietly walking back through the door it propped open, right as the local-model wave it started crests.
So hold both truths. The strategy — build a teacher, ship free students, put capable AI on your own hardware — is real, and it has already reshaped the market. But Meta’s own execution of it is wobbling, and its own commitment to it is now in question. The idea is bigger than the company that popularized it. Which is exactly why what happened next matters more than either.
The defector
In November 2025, Yann LeCun told the world he was leaving Meta.
This is not a mid-level departure you scroll past. LeCun is one of three people who share the 2018 Turing Award — computing’s Nobel — for inventing the deep learning that everything, Llama included, is built on. He founded Meta’s fundamental AI research lab. He is, more than any single person, the reason Meta had a credible AI program at all. And he walked out the door saying the entire direction the industry is sprinting toward is a dead end.
His words, not mine. LeCun’s position is that large language models — the whole GPT/Llama/Claude lineage — are “limited to the discrete world of text,” that they “can’t truly reason or plan, because they lack a model of the world.” The belief that you can just keep scaling them up until they reach human-level intelligence? He calls it “the illusion, or delusion, that it is a matter of time until we can scale them up to having human-level intelligence, and that is simply false.” An LLM predicts the next token. It does not know what a glass of water is — that it will spill, that it’s heavier full than empty, that gravity exists. It has read every sentence ever written about the physical world and understood none of it, the way a person can memorize a language’s phonetics without speaking a word of meaning.
LeCun’s alternative is what he calls world models — his JEPA architecture — systems that learn how reality behaves by watching it, not by reading about it. “It learns the underlying rules of the world from observation,” he put it, “like a baby learning about gravity.” Not autocomplete. Comprehension.
And he didn’t just publish a paper and stay for the stock. He left Meta, stood up a new company — Advanced Machine Intelligence, based in Paris, with LeCun as executive chairman — and in March 2026 raised $1.03 billion at a $3.5 billion valuation, with Bezos Expeditions among the backers, to build the thing he says will replace the paradigm Meta is currently mass-producing. He was gracious about the split. He even said Meta “might be our first client.” But strip the manners off and it’s the sharpest possible verdict: the man who built your AI thinks the kind of AI you’re building is a cul-de-sac, and he was willing to bet a billion dollars and the last act of his career that he’s right.
Why both of these are the same story
Here’s the through-line, and it’s the part worth pricing.
Meta’s teacher-and-student machine is an argument that the current models are becoming a commodity — cheap, local, not worth renting. LeCun’s defection is an argument that the current models are a dead end — not just cheap, but capped, incapable of the reasoning the whole valuation bubble is priced on. One says the frontier is leaking out of the datacenter and onto your desk. The other says the frontier itself is pointed at a wall. Different claims — but both of them say the same thing to anyone renting today’s intelligence at today’s prices: do not assume this is permanent.
Because if LeCun is even half right, the trillion-dollar buildout is aimed at scaling a technology that its own inventor says can’t get where the money assumes it’s going. And if Meta is right, the models good enough for almost everything you actually do will be free, small, and running on hardware you already own — teacher retired, students loose. Either future is fatal to the same bet: that a handful of companies can keep intelligence scarce, locked in their clouds, and charge rent on it forever.
You don’t have to know which man is right. You only have to notice that the two most informed players in the room — the company that dominates open AI and the scientist who built it — are both, in their own way, betting against the scarcity trade. When the people who know the technology best are quietly heading for different exits, the one position that looks reckless is standing still in the middle of the room, paying full price, and assuming nothing changes.
That’s the read. Nobody here is selling you a model, an API tier, or a seat on a cluster — which is exactly why we can say the quiet part: the value in AI is migrating away from the thing you rent and toward the thing you own or the thing nobody’s built yet. Meta is trying to give away the first. LeCun took a billion dollars to go build the second. The only people insisting it all stays exactly as it is today are the ones with a datacenter to fill.
SIGNAL · THE PARADIGM ISN’T THE MOAT
The teacher was built to make itself unnecessary, and it’s working. The defector says the whole lineage is a dead end, and he put a billion dollars where his mouth is. Bet on access and comprehension — the model you can run and the intelligence that actually understands the world. Not the tollbooth in between.
Not financial advice. MarketCrystal provides trend analysis for informational purposes only. Equities and crypto are volatile and you can lose money. Figures are drawn from public reporting current at publication; Meta’s unreleased-model and closed-source deliberations are reported plans, not confirmed decisions, and technical roadmaps move fast — verify before you act. Always do your own research. Past trends do not guarantee future results.