The $1B Scientific Split: Why Artificial General Intelligence Has Two Competing Paths
Almost every mainstream conversation about artificial general intelligence assumes there's one race, with one strategy: make the language models bigger, feed them more data, and eventually general intelligence falls out the other end.
Yann LeCun — a Turing Award winner, one of the three researchers credited with founding modern deep learning, and Meta's chief AI scientist for twelve years — left specifically to bet against that entire premise. By early 2026, his new company had raised over a billion dollars to build something deliberately different.
That's the part of the artificial general intelligence story almost no explainer covers: it's not one race. It's at least two genuinely competing scientific bets about what intelligence actually requires — and the disagreement is coming from inside the field's own most credentialed researchers, not from outside skeptics.
Two legitimate scientific bets, not one obvious path: scale language models further, or build systems that understand the physical world the way LeCun argues intelligence actually requires.
What Artificial General Intelligence Actually Means — and Why the Path There Is Contested
Artificial general intelligence refers to an AI system capable of understanding, learning, and applying intelligence across a broad range of tasks at or above human level — as opposed to "narrow AI," which excels at one specific domain (chess, image recognition, language generation) without that capability transferring elsewhere.
Where the real disagreement lives isn't in that broad definition — it's in what kind of system could actually achieve it. The dominant industry bet, behind OpenAI, Anthropic, and most of Google DeepMind's public roadmap, is that scaling large language models — more parameters, more training data, more compute — will eventually produce general intelligence as an emergent property.
Yann LeCun represents the most credentialed public counter-bet: that language models, no matter how large, are fundamentally limited because they're trained to predict text rather than to understand and simulate the physical world. His term of choice isn't even "AGI" — he prefers "Advanced Machine Intelligence" (AMI), deliberately avoiding language he considers overly dramatic in favor of framing focused on capability and collaboration.
2026 Not One Race Competing Scientific BetsThe Numbers Behind the Industry's Biggest Scientific Disagreement
Two Competing Definitions of What "General" Intelligence Actually Requires
📝 The Scaling Bet
General intelligence emerges from making language models larger, trained on more data with more compute. The dominant approach behind OpenAI, Anthropic, and most of Google DeepMind's public roadmap. Evidence: genuine, measurable capability jumps across reasoning, coding, and multi-step tasks in recent years.
🌍 The World-Model Bet
General intelligence requires an internal model of physical reality — cause and effect, spatial relationships, object permanence — learned from sensory data like video, not just text. LeCun's AMI Labs and Meta FAIR's V-JEPA-2 research represent this camp. Evidence: even top language models still fail at basic physical reasoning tasks a toddler handles intuitively.
Neither camp claims the other's approach produces nothing useful. The actual disagreement is narrower and more interesting: whether the scaling approach can reach general intelligence on its own, or whether it's a genuinely different kind of system that's needed to close the remaining gap.
Five Facts About the AGI Debate Most Coverage Skips
🧭 What's Actually Happening Inside This Disagreement
- LeCun's Most Quoted Line Is Deliberately Provocative — and Specific: LeCun has repeatedly argued that today's most advanced language models remain, in his words, "dumber than a cat" at basic physical and causal reasoning — even while dramatically outperforming humans at language-based tasks. The specific point isn't that LLMs are weak generally; it's that raw task performance on language benchmarks doesn't necessarily track progress toward the kind of grounded, physical-world understanding he believes general intelligence actually requires.
- He Left Behind Real Research, Not Just an Opinion — V-JEPA-2: During his final years at Meta's Fundamental AI Research division, LeCun's team produced V-JEPA-2, a model trained on video specifically to learn how physical objects and environments behave, rather than to generate or caption content. It's a concrete example of the "world model" research direction rather than an abstract talking point, and his new venture is expected to continue that research line independently.
- Meta Is Partnering With AMI Labs — But Explicitly Not Investing: LeCun has stated publicly that Meta will not be a financial investor in AMI Labs, describing the relationship instead as a commercial partnership. It's a notable structural detail: the company whose resources funded over a decade of this research isn't the one bankrolling the independent venture built to pursue it further, even as it maintains a working relationship for future technology access.
- LeCun Isn't the New Company's CEO — Someone Else Is Running It: Alexandre LeBrun, founder of the French health tech startup Nabla and a former Facebook AI research engineer, serves as CEO of AMI Labs. LeCun's role centers on the underlying research direction rather than day-to-day company operations — a detail easy to miss in coverage that treats the venture as simply "LeCun's startup."
- He's Not Alone — Other Senior Researchers Are Making a Similar Bet: Fei-Fei Li, the Stanford researcher widely credited with building ImageNet and catalyzing the deep learning revolution, has separately argued that "spatial intelligence" — understanding and reasoning about three-dimensional physical space — represents the next necessary frontier beyond language models, which she describes as eloquent but lacking direct experience of the world. It's a meaningfully similar thesis from a different, independently credentialed researcher, suggesting the world-model camp is a genuine scientific position rather than one dissenting voice.
The Honest Case for Each Side
✅ The Case for the Scaling Approach
- Demonstrated, repeated capability jumps across successive model generations
- Massive existing commercial infrastructure and investment already built around it
- Language and reasoning performance improvements are measurable and reproducible
- Test-time reasoning and search techniques have extended scaling's returns further than expected
- Currently backed by the largest frontier labs and the most total research investment
⚠️ The Case for the World-Model Approach
- Addresses a specific, demonstrated weakness: physical and causal reasoning gaps in LLMs
- Backed by multiple independently credentialed researchers, not a single dissenting voice
- Grounds intelligence in the same sensory learning process humans and animals actually use
- Real published research (V-JEPA-2) exists behind the argument, not just critique
- Represents genuine scientific diversification rather than betting everything on one approach
Both camps agree on one thing: pure scale alone eventually shows diminishing returns on certain classes of problems. Where they split is what to add next — more of the same architecture at greater scale, or a structurally different approach to how the system learns about reality.
4 Ways to Follow This Debate Without Getting Lost in Hype
🧭 Tip #1: Notice Which Definition of "Progress" a Claim Is Actually Using
When you see a claim about AGI progress, check whether it's measuring language and reasoning benchmark performance (the scaling camp's preferred evidence) or physical/causal reasoning tasks (the world-model camp's preferred evidence). Progress on one doesn't automatically mean progress on the other — they're genuinely different capability categories.
🧭 Tip #2: Track V-JEPA-Style Research as a Distinct Progress Signal
Separate from standard LLM benchmark leaderboards, world-model research (V-JEPA-2 and its successors, plus similar work from Google DeepMind and World Labs) has its own evaluation criteria centered on physical prediction accuracy. Following this research line gives you visibility into a genuinely different axis of AI progress that pure language-model coverage won't show you.
🧭 Tip #3: Don't Treat Funding Size as Proof of Correctness
AMI Labs' billion-dollar raise is a meaningful signal of investor conviction and researcher credibility, not proof that the world-model approach will succeed where scaling falls short. Both research directions remain genuinely unresolved bets — funding size measures belief and resources, not outcomes.
🧭 Tip #4: Watch for Convergence, Not Just Competition
Several frontier labs are already incorporating elements of both approaches — reasoning-focused scaling alongside multimodal and video-trained components. The scaling-versus-world-models framing is useful for understanding the current disagreement, but the eventual path to general intelligence may end up combining elements of both rather than vindicating one side outright.
✅ Artificial General Intelligence — Quick Reference
- ✅ AGI = broad, human-level or above capability across tasks — not one narrow specialty
- ✅ The dominant industry bet is scaling large language models — more parameters, data, and compute
- ✅ Yann LeCun (2018 Turing Award, 12 years leading Meta's FAIR) left in Nov 2025 to bet against that approach
- ✅ AMI Labs raised $1.03B by March 2026 at a $3.5B valuation — the largest European startup seed round on record
- ✅ LeCun prefers "Advanced Machine Intelligence" (AMI) over "AGI" as deliberate, less dramatic terminology
- ✅ V-JEPA-2, built at Meta FAIR, is real published world-model research — not just critique
- ✅ Meta partners with AMI Labs commercially but is explicitly not an investor
- ✅ Alexandre LeBrun, not LeCun, is AMI Labs' CEO
- ✅ Fei-Fei Li's "spatial intelligence" thesis independently echoes similar concerns — this isn't one lone dissenting voice
🖥️ The Hardware Needed to Build and Test Local World Models
Testing video-trained "world models" offline requires serious VRAM. A 16GB GPU (like the RTX 5060 Ti) is the most cost-effective entry point for prototyping smaller spatial models locally. For complex causal physics and 70B-class architectures, a flagship 32GB card (like the RTX 5090) provides the raw processing power to run everything under your own roof.
Check 16GB AI Starter GPUs on Amazon → Check 32GB Flagship GPUs on Amazon →⚙️ Whichever Path Wins, Local Silicon Is Already Shifting
Whether AGI scales via the cloud or localized world models, the computing bottleneck is shifting directly to your device. Read our complete guide to understand how dedicated Neural Processing Units (NPUs) are powering these complex, offline AI workloads.
Read the Complete NPU Guide →Frequently Asked Questions — Artificial General Intelligence
What is artificial general intelligence in simple terms?
Artificial general intelligence (AGI) refers to an AI system capable of understanding, learning, and performing at or above human level across a broad range of intellectual tasks — as opposed to "narrow AI," which excels at one specific domain without that skill transferring elsewhere. The concept itself is broadly agreed upon; what's genuinely contested is what kind of system could actually achieve it, with different credentialed researchers backing meaningfully different technical approaches.
Why did Yann LeCun leave Meta, and what is he building instead?
Yann LeCun, a 2018 Turing Award winner and Meta's chief AI scientist for 12 years, left the company in November 2025 to found a startup (AMI Labs) focused on "world models" — AI systems trained on video and sensory data to build an internal understanding of physical cause and effect, rather than being trained primarily to predict text. LeCun has long argued that large language models, no matter how large, are fundamentally limited for reaching general intelligence because they lack grounded understanding of the physical world. By March 2026, his venture had raised over $1 billion at a $3.5 billion valuation — the largest seed round ever raised by a European startup.
What are "world models" and how do they differ from large language models?
A world model is an AI system designed to develop an internal understanding of its environment — physical objects, spatial relationships, cause and effect — so it can simulate and predict outcomes, typically learned from video or sensory data rather than text. This differs from large language models, which are trained primarily to predict the next word in text sequences. Proponents of world models, including Yann LeCun and separately Fei-Fei Li, argue this grounded, physical understanding is necessary for genuine general intelligence in ways that text prediction alone cannot provide, even at massive scale. Meta's V-JEPA-2, developed under LeCun's research division, is a concrete published example of this approach.
Is scaling large language models enough to reach AGI, or is a different approach needed?
This is genuinely unresolved and represents the field's most significant current scientific disagreement. The dominant approach behind most major frontier labs, including OpenAI and Anthropic, bets that continued scaling — more parameters, data, and compute, often combined with test-time reasoning techniques — will eventually produce general intelligence. A separate camp, represented most visibly by Yann LeCun's new venture and echoed by researchers like Fei-Fei Li, argues that language-based training alone cannot produce genuine physical and causal reasoning regardless of scale, and that a structurally different, world-model-based approach is necessary. Both camps have credible published research and significant funding behind them, and the disagreement remains actively contested rather than settled.
Does Meta still support Yann LeCun's AGI research after his departure?
Meta maintains a commercial partnership with LeCun's new venture, AMI Labs, that gives it access to resulting technology, but LeCun has publicly stated that Meta is not a financial investor in the company. AMI Labs is led day-to-day by CEO Alexandre LeBrun, a former Facebook AI research engineer and founder of the French health tech startup Nabla, while LeCun's role centers on the underlying research direction. This structure — a former employer as commercial partner but not investor, with a different executive running daily operations — is a nuance often missed in coverage that frames the venture simply as "LeCun's startup."
