You’ve probably noticed that every AI company is racing to build the same thing: a bigger brain. More parameters. More GPUs. More data. The logic seems airtight — make the model enormous enough, and intelligence will simply… emerge.
But what if the entire industry is solving the wrong problem?
What if intelligence isn’t a destination you arrive at by brute force, but a skill you develop by learning how to learn?
That’s the bet Liang Wenfeng is making. And if he’s right, the billions flowing into GPU farms and robot fleets might be backing the wrong horse.
Most of the AI world treats intelligence as a static state to be reached. Liang treats it as a dynamic process to be cultivated. That distinction changes everything.
Let me explain why.
The AGI race has split into two camps, and you’ve likely encountered both. The first camp believes in the Scaling Law — the idea that if you make neural networks big enough, feed them enough data, and throw enough compute at them, capabilities the system was never explicitly trained for will simply appear. This is the “sudden enlightenment” school. OpenAI’s former chief scientist Ilya Sutskever championed this view. Sam Altman operates within it. The entire Nvidia empire is built on its back.
The second camp says language models are a parlor trick. Real intelligence requires understanding the physical world — predicting what happens next, learning from consequences, grounding abstract knowledge in reality. This is Yann LeCun’s world model thesis. It’s why billions are pouring into autonomous driving, humanoid robots, and embodied AI. Call this the “gradual cultivation” school.
Both camps sound reasonable. Both have brilliant defenders. Both have raised staggering amounts of capital.
And both might be missing the point.
Because here’s what Liang Wenfeng said at a four-hour investor meeting that recently leaked: his roadmap to AGI runs through Chain of Thought, then Agents, then continuous learning, then a gradual singularity of AI self-iteration — and only after all of that, embodied intelligence.
Notice what’s missing? Scale for its own sake. Physical world interaction as a prerequisite. Liang explicitly said he doesn’t think world models or 3D generation have much to do with the upper limit of intelligence right now. He’s not building robots. He’s not chasing video generation. He’s not even particularly interested in the physical world — yet.
He’s interested in something more fundamental: can an AI system learn the way a living being does? Can it absorb experience and reshape itself accordingly?
The question isn’t how much an AI knows. It’s whether an AI can grow. Everything else is just storage.
Think about what current large language models actually are. They’re frozen. Once training ends, the model doesn’t evolve. It can answer questions and chain reasoning, but it can’t wake up tomorrow and be fundamentally different from what it was yesterday. It’s a photograph of intelligence, not intelligence itself.
Liang sees this clearly. In his view, the path forward isn’t a bigger photograph — it’s a system that can keep developing. That’s why DeepSeek’s roadmap prioritizes continuous learning above almost everything else. As he put it: once continuous learning is solved, AGI might become easy.
This is a radically different bet. And it has a distinctly ascetic quality to it.
While every other AI lab is scrambling for users, scenes, and commercial entry points, DeepSeek is turning inward — exploring the architecture of learning itself. It’s the difference between a monk in a cave and a merchant in a marketplace. The merchant accumulates. The monk transforms.
While the market chases milestones, Liang chases the mechanism that produces milestones. That’s not a strategy difference — it’s a philosophical one.
But let’s be honest about the challenge here. It’s enormous.
Learning doesn’t happen in a vacuum. ChatGPT got better not just because of algorithmic improvements, but because hundreds of millions of users interacted with it, generating feedback that shaped subsequent versions. AlphaGo didn’t become superhuman through architecture alone — it played millions of games against itself and the environment. Waymo’s self-driving works because of billions of miles of real road data.
If Liang’s AI isn’t deeply embedded in the physical world, if it doesn’t have massive user interaction generating continuous feedback, then where exactly does it learn? Is continuous learning without rich environmental interaction just a beautiful theory?
This is the paradox at the heart of his approach. And it mirrors an ancient tension. A monk can spend years in meditation and achieve profound self-knowledge — but can you truly understand the world without entering it?
Training gives a model capability. But only the world gives it meaning. The question is whether you need meaning before capability, or capability before meaning.
Now here’s where it gets interesting for anyone with money in this game.
For the past few years, the capital narrative has been straightforward: bet on compute. Nvidia is the new oil company. Data centers are the new infrastructure. Whoever buys the most GPUs wins. This narrative works because it’s legible — investors understand infrastructure plays, and the Scaling Law provides a clean story about why more compute equals more intelligence.
The world model camp offers a different but equally legible narrative: bet on physical interaction. Robots, autonomous vehicles, embodied AI — these generate proprietary real-world data that becomes the moat. The asset isn’t the model; it’s the experience stream.
But Liang’s route scrambles both narratives. If the key to AGI isn’t scale and isn’t physical interaction — if it’s the ability to learn continuously and self-iterate — then what exactly do you invest in?
You stop betting on the people selling shovels and start betting on the players who are actually mining something new. Not more parameters. Not more robots. But the algorithmic breakthrough that makes recursive self-improvement real.
The market has been betting on infrastructure. Liang is betting on insight. Infrastructure scales linearly. Insight scales exponentially. But insight is also unpredictable, untrackable, and deeply uncomfortable for quarterly-minded capital.
And that’s the real tension here. Capital can tolerate losses. What capital cannot tolerate is losses without a clear thesis. The Scaling Law story gives investors a metric: watch the parameter count, watch the compute budget, watch the benchmark scores. The world model story gives them a metric too: watch the fleet size, watch the data pipeline, watch the deployment scenarios.
Liang’s story gives them… patience. That’s it. Wait for the algorithmic breakthrough. Wait for continuous learning to work. Wait for the gradual singularity that isn’t really a singularity but a slow climb.
How many investors are built for that?
This is why DeepSeek behaves differently from every other AI lab. They’re not chasing the leaderboard for its own sake. They’re not shipping consumer products to generate feedback loops. They’re not building a robot army. They’re sitting in a room, working on the hardest, least glamorous problem in AI: how to make a system that teaches itself.
It’s lonely work. It doesn’t generate viral demos. It doesn’t produce neat quarterly updates. And it might be the most important work happening in AI right now.
Or it might be a beautiful dead end.
The Scaling Law says intelligence comes from size. The world model camp says it comes from experience. Liang says it comes from the capacity to grow. Only one of these can be fundamentally right — and the winner won’t be decided by who raises the most money, but by who correctly identifies what intelligence actually is.
A thousand years ago, Siddhartha Gautama pursued wisdom through extreme asceticism. He starved himself, meditated relentlessly, pushed his body to its limits. After years of this, he realized that extreme self-denial wasn’t enlightenment itself — it was just another trap.
The lesson wasn’t that asceticism was wrong. It was that no single path — not accumulation, not denial, not even disciplined practice — contains the whole answer.
The AI industry is living this parable right now. The Scaling Law camp accumulates. The world model camp engages. Liang cultivates. Each believes they’ve found the way.
But the history of intelligence — both human and artificial — suggests the answer is rarely where you expect it. And the person who wins this race probably won’t be the one with the most GPUs, the most robots, or even the best algorithms.
It’ll be the one who correctly asked: what does it mean to learn?
FAQ
Q: Isn't Liang's approach just theory without evidence that continuous learning actually works?
A: Fair criticism. DeepSeek hasn't yet demonstrated a system that truly learns continuously in the way Liang describes. But neither has the Scaling Law camp produced AGI despite spending tens of billions. At some point, you have to ask which unproven bet has better theoretical grounding — and Liang's focus on recursive self-improvement has deep roots in both AI research and cognitive science.
Q: What does this mean for AI investors right now?
A: If intelligence is about learning capability rather than knowledge storage, then the valuable asset isn't compute infrastructure or robot fleets — it's the team most likely to crack the continuous learning problem. That shifts investment from infrastructure plays to algorithmic bets, which are harder to evaluate but potentially exponentially more valuable.
Q: Isn't this just DeepSeek spinning a narrative to justify not having Nvidia's compute budget?
A: That's the cynical read, and it's worth considering. But Liang explicitly rejected world models and embodied AI — directions that don't require massive compute either. If this were just a resource constraint story, he'd be pivoting to cheaper domains, not pursuing what might be the hardest unsolved problem in AI. Sometimes the contrarian with limited resources is the one most forced to think clearly.