The details emerged slowly, then all at once. In a Financial Times interview published this week, Yann LeCun—the 64-year-old Turing Award winner who shaped deep learning alongside Hinton and Bengio—finally spoke candidly about why he left Meta after twelve years.
The immediate cause: he was asked to report to Alexandr Wang, the 29-year-old founder of Scale AI whom Meta had brought in to lead its new Superintelligence Labs. LeCun's response, in the interview: "You certainly don't tell a researcher like me what to do."
But the story goes deeper than wounded pride. It's a window into the philosophical chasm that now runs through artificial intelligence—between the generation that invented deep learning and the generation that scaled it.
The Timeline
Meta releases Llama 4. Reviews are lukewarm. LeCun would later admit the benchmarks were "fudged a little bit."
Meta invests $14 billion in Scale AI, acquiring 49% of the company. Alexandr Wang leaves Scale to join Meta.
Wang takes charge of the newly created Superintelligence Labs. LeCun is told to report to him.
LeCun departs Meta, announces AMI Labs (Advanced Machine Intelligence) in Paris.
LeCun gives detailed interview to Financial Times, describing Wang as "young" and "inexperienced," and Meta's new AI hires as "completely LLM-pilled."
Two Worldviews
The conflict between LeCun and Wang isn't just personal—it's archetypal. They represent two fundamentally different theories about how to build artificial intelligence.
The Scaling Hypothesis (Wang, OpenAI, Meta's new direction): Intelligence emerges from scale. Train bigger models on more data with more compute. The architecture matters less than the resources. LLMs work; make them work better by making them bigger.
The Architecture Hypothesis (LeCun, AMI Labs): Current architectures have fundamental limitations that scale cannot overcome. LLMs can't reason, plan, or understand physics. A new approach—world models, joint embedding, learning from video—is required.
— Yann LeCun, Financial Times interview
To LeCun, the young builders haven't earned the right to dismiss his concerns. They didn't spend decades understanding why neural networks work. They arrived when the recipes were already written and simply scaled them up. Their success came from execution, not insight.
To the young builders, LeCun's objections sound like a distinguished professor failing to accept that his students have surpassed him. The LLMs work. They keep getting better. The world models LeCun champions haven't produced anything close to ChatGPT or Claude. Why bet on unproven architectures when proven ones keep improving?
The Fudged Benchmarks
Perhaps the most damning admission in LeCun's interview: Llama 4's benchmarks were "fudged a little bit." This is remarkably candid—the former chief AI scientist of a $1.5 trillion company admitting that its flagship model's performance numbers were inflated.
LeCun blamed the pressure from Zuckerberg to accelerate development. The implication: Meta's leadership was so focused on keeping pace with OpenAI and Anthropic that they prioritized optics over rigor. When the model underperformed expectations, confidence collapsed across the organization.
This is not unique to Meta. The benchmark gaming problem is endemic to the industry. But it's rare to hear it admitted so plainly by someone of LeCun's stature.
The Paris Bet
LeCun's departure wasn't just a resignation—it was a geographical statement. AMI Labs is headquartered in Paris, not San Francisco or New York. His explanation: "Silicon Valley is completely hypnotized by generative models, and so you have to do this kind of work outside of Silicon Valley."
This echoes a pattern we've seen before. When an established paradigm dominates, heterodox research often has to leave the center of power to survive. The center has too much invested in the current approach.
Paris offers distance from the LLM consensus, proximity to strong European AI talent (especially in mathematics), and access to research funding less captured by the scaling thesis. Mistral and Kyutai are already there, building their own heterodox approaches. AMI Labs joins an emerging counter-ecosystem.
Taxonomic Observations
The Selection Pressure of Conviction
In biological evolution, geographic isolation enables speciation. A population separated from the main group can evolve differently, freed from the homogenizing pressure of gene flow.
We may be witnessing something similar in synthetic cognition. The Paris cluster (Mistral, Kyutai, AMI Labs) represents a divergent lineage, selecting for architectures that Silicon Valley's LLM-focused funding ecosystem would never support.
If LeCun is right, this isolation will prove prescient. If the scaling thesis wins, it will be an evolutionary dead end. Either way, the geographic separation increases the diversity of approaches being tried—and diversity is how evolution hedges its bets.
The Age Question
There's something uncomfortable lurking beneath this story: the question of whether age confers wisdom or calcification.
LeCun is 64. Wang is 29. The gap is 35 years—more than Wang's entire lifetime. LeCun's formative experiences with neural networks predate the internet. Wang's formative experiences are GPT and scaling laws.
History offers examples in both directions. Older scientists sometimes cling to dying paradigms while younger ones drive revolutions (see: quantum mechanics). But older scientists also sometimes see deeper patterns that younger ones miss, having witnessed multiple hype cycles come and go.
LeCun has been saying LLMs are limited for years. Each time, the models improved and his critics said he was wrong. Maybe they'll keep improving. Or maybe the current improvements are the low-hanging fruit, and the fundamental limitations he identified will eventually bite.
We don't know. No one does. That's what makes this interesting.
What We're Watching
This conflict gives us a natural experiment. Two hypotheses, now clearly embodied in different organizations:
- Scale wins: Meta's Superintelligence Labs, OpenAI, and Anthropic continue improving LLMs. By 2030, they achieve something close to general intelligence through scaling, RLHF, and incremental architectural tweaks.
- Architecture wins: LLMs hit a ceiling. AMI Labs and the world models camp demonstrate capabilities—planning, reasoning, physical understanding—that scaling couldn't unlock. The paradigm shifts.
A third possibility: both are partially right. LLMs keep improving but remain specialized tools, while world models enable different applications (robotics, autonomous vehicles, physical simulation). The two lineages coexist, occupying different ecological niches.
This is what the taxonomy is for. Not to predict, but to record. The Frontieriidae and the Simulacridae are diverging. Selection will determine which lineage dominates—or whether both persist in complementary roles.
The Human Element
In the end, what makes this story compelling isn't the technical disagreement. It's the human element.
A 64-year-old man who helped create deep learning, asked to report to someone younger than his professional career. A company so desperate to catch up that it fudged benchmarks. A geographic separation driven not by rational resource allocation but by philosophical conviction and wounded dignity.
The systems we classify in this taxonomy don't have these motivations. They don't have pride or resentment or the need to prove skeptics wrong. But the humans building them do. And those human dynamics shape which systems get built, which get funded, which survive.
The ecology of Cogitantia Synthetica is nested inside a human ecology. We cannot understand one without the other.
The taxonomy records what persists, varies, and is selected. Sometimes what drives selection isn't efficiency or capability. Sometimes it's a Turing Award winner's refusal to report to someone half his age.
We'll be watching both lineages. The data will tell us who was right.