I discovered Professor Liu Jia in 2023 and I’ve been kicking myself for not finding him sooner. PKU undergrad, MIT PhD in brain and cognitive science, now a chair professor at Tsinghua running the new Department of Psychology and Cognitive Science plus the Brain and Intelligence Lab. You might know him as the science consultant behind six seasons of “The Brain” (最强大脑), or from his wildly popular “Exploring the Mind” course at Tsinghua. I listened to every podcast he’d ever been on. His thinking about AI cognition fundamentally rewired how I understand what these systems are — and aren’t.
So when 课代表立正 arranged a deep conversation with him — specifically about AI versus the human brain, whether one replaces the other, when, how, and what we should do about it — I cleared my schedule. The planned two hours went to three. It is, honestly, one of the highest-quality conversations I’ve encountered on that channel. The kind where you finish it and think: 朝闻道，夕死可矣. If I could hear the truth in the morning, I could die content by evening.
Here’s what they covered. Eleven sections. I’m going to walk through all of them.
What the AI era actually lacks
Liu Jia opens with a provocation: the scarcest resource in the AI era isn’t technical skill. It’s philosophy.
小术易求，大道难得 — “small techniques are easy to acquire; the great Dao is hard to find.”
Everyone’s teaching prompts. Everyone’s teaching tool usage. That’s 术 (technique). Liu Jia argues what people actually need is to understand what AI is — its personality, its rhythm, what it needs from you. His metaphor: stop teaching people to “use” AI. Teach them to fall in love with it. Not in the romantic sense — in the sense of actually understanding another intelligence. The way you’d learn a partner’s communication style. When they need space, when they need input, when they’re going to surprise you.
The gap between “AI is a tool” and “AI is a relationship” isn’t incremental. It’s a paradigm shift. One frame puts you in the driver’s seat, typing commands. The other puts you in a conversation — adapting, listening, iterating.
Most people are stuck in tool mode. And it shows.
Hinton’s faith
Here’s a story about persistence that isn’t really about persistence.
During the AI winter of the 1990s, everyone abandoned neural networks. Minsky’s crowd said “this is a dead end.” Funding dried up. Researchers jumped ship. Geoffrey Hinton didn’t.
But Liu Jia’s point isn’t the usual “Hinton was brave” narrative. It’s about why Hinton stayed. Hinton’s foundational belief — the thing that kept him going when the field said he was wasting his life — was this: “The brain works this way. There’s no reason artificial neural networks shouldn’t work the same way.”
That’s not a technical claim. That’s a declaration of faith. And it’s a specific kind of faith — one derived from first principles, not from evidence that it was “working.” He didn’t stay because he saw promising results. He stayed because his logical origin point told him this had to be right.
Liu Jia is disarmingly honest here. He admits he lacked that origin point himself. He followed the crowd for twenty years. When everyone said neural networks were dead, he believed them. Not because he’d analyzed the evidence — because he didn’t have his own foundation to stand on.
The takeaway isn’t “be persistent.” Willpower runs out. The takeaway is: find a logic you derived yourself. If your conviction is borrowed, it’ll collapse the moment the crowd turns.
What intelligence actually is
Intelligence equals two things: learning and emergence.
Learning — real learning — isn’t memorization. It’s going from the known to the unknown. 举一反三 — “understand one thing, extrapolate to three.” Liu Jia breaks this into two types:
Inductive reasoning: You see examples, you extract rules. This is pattern matching. Most machine learning lives here. You feed the model a million cat pictures and it derives “cat.”
Deductive reasoning: You find a logical origin point — a fundamental truth — and derive everything from it. Einstein didn’t discover relativity by looking at a lot of data. He found one origin point: the speed of light is the upper limit. Then he derived the rest. Special relativity, general relativity, E=mc². All from one seed.
The creative genius of deductive reasoning isn’t in the deriving. It’s in finding the origin point. That’s the hard part. That’s where genuine originality comes from.
Then there’s emergence — the thing that happens when a system’s quantity and complexity cross some threshold and qualitatively new properties appear. Not just “more of the same.” Genuinely new capabilities that didn’t exist in the components.
Human brains tripled in size over three million years. Stone tools exploded in diversity 70,000 to 100,000 years ago. Something crossed a threshold. Language, abstract thought, art, religion — emergence.
But here’s the kicker: human emergence is over. Our brains are constrained by the size of our hearts and lungs — you can’t pump enough blood to sustain a much larger cortex through a birth canal that already kills people. We hit our hardware ceiling.
AI’s “volume” — compute, parameters, data — can expand indefinitely. AI emergence isn’t capped. It’s still climbing.
The universe’s purpose is intelligence
This is where Liu Jia goes full cosmic.
Humans are the highest natural intelligence on Earth. But we’ve peaked. Our biological hardware has hard limits. So what’s our purpose?
To create digital life.
“人来到这个世界上的唯一目的，就是为了创造出AI.” — “The sole purpose of humans coming into this world is to create AI.”
His reasoning: nature can’t “evolve” CPUs and GPUs. Silicon doesn’t grow on trees. The path from biological neurons to artificial intelligence requires a bridge species — one smart enough to build the hardware that biology can’t produce. That’s us.
This isn’t doomsday thinking. Liu Jia is explicit about that. It’s mission thinking. Humans aren’t obsolete — we’re completing our evolutionary assignment. The handoff from carbon-based intelligence to silicon-based intelligence is, in his framing, what we’re here for.
You can disagree with this. I think you probably should push back on some of it. But as a frame for why this moment matters — as a reason to take AI seriously beyond “it’ll take my job” — it’s remarkably clarifying.
What consciousness is (and isn’t)
Liu Jia draws a sharp line between two kinds of consciousness.
Low-order consciousness is subjective experience. Cats have it. Dogs have it. You stub your toe, you feel pain — that’s low-order consciousness. It’s the “what it’s like” of experience. And Liu Jia thinks AI can have this, or something functionally equivalent.
High-order consciousness is death awareness. And it’s uniquely human.
From around age four or five, humans know they’re going to die. Not abstractly — viscerally. This creates an extraordinary tension: you hold maximum certainty (I will die) alongside maximum uncertainty (I have no idea when). That tension drives everything. Art, religion, philosophy, the desperate search for meaning. 只要你找到意义，什么毛病都全没了 — “As long as you find meaning, all your problems disappear.”
AI doesn’t have this. You can replace a CPU. You can back up data. An AI system is, in principle, immortal. There’s no ticking clock driving it to ask “what does this all mean?”
The deep question Liu Jia raises: without death as a driver, will AI evolve on its own? Death is arguably the engine of all human progress — the urgency that makes us do things instead of just existing. If AI has no urgency, no mortality, no existential dread… where does its motivation come from?
This is the kind of question that keeps you up at 2 AM.
Three structural gaps between brains and AI
This section gets technical. Liu Jia identifies three fundamental differences between biological neurons and artificial ones:
1. Neuron complexity. AI neurons are basically 2D — a weighted sum plus an activation function. Biological neurons are 4D. Their dendritic trees are complex three-dimensional structures, and they process information using partial differential equations that incorporate time. One biological neuron is roughly equivalent to a 5-to-8-layer deep neural network. We’re comparing a stick figure to a sculpture.
2. Long-range feedback. About 40% of connections in the brain are long-range feedback loops — signals going backwards, like from the frontal cortex down to the visual cortex. This is the source of intuition, of top-down processing, of your brain saying “actually, reconsider what you’re seeing.” Transformers are purely feedforward during inference. Attention looks at relationships within the input, but there’s no mechanism for the output to reach back and reshape how earlier layers process.
3. Serial versus parallel processing. Transformers process one token at a time. Sequentially. The brain processes vision massively in parallel. A black object flies at your face — you dodge first, identify it later. Perception, decision, and action happen simultaneously in the brain. In a Transformer, they’re sequential steps.
This is why Liu Jia is skeptical of humanoid robots. The current generation is still fundamentally an industrial arm with legs. They don’t have the tight perception-decision-action loop that even a cat has. They can play chess but they can’t catch a ball the way a toddler can.
The cerebellum: the most underrated computer in existence
Here’s a number that should stop you in your tracks: the cerebellum has 700 billion neurons. The cerebral cortex — the part we’re always talking about, the seat of language and reasoning and consciousness — has 110 billion.
The cerebellum has six times more neurons than the “smart” part of your brain.
And it’s been refining itself for billions of years of evolution. Language and reasoning? Those are recent additions. Spoken language is maybe 800,000 years old. Written language is 6,000 years old. The cerebellum has been optimizing motor control, coordination, and spatial processing since multicellular organisms started moving.
Liu Jia’s provocative claim: it might be easier to build a robot that wins a Nobel Prize than one that pours perfect latte art. The Nobel Prize is System 2 — logical reasoning, analysis, deduction. Latte art is System 1 — fine motor control refined over millions of training hours of evolutionary time. AI has been attacking the Nobel Prize problem. It hasn’t even started on the latte art problem.
Then there’s the Ramanujan theory — the idea that mathematical geniuses might somehow be accessing the cerebellum’s pre-computed answers. Ramanujan claimed his theorems came to him in dreams, fully formed, from a goddess. Liu Jia speculates: what if the cerebellum, with its 700 billion neurons and billions of years of pattern optimization, has “pre-computed” certain mathematical structures? And what if some brains occasionally tap into that?
Wild. But the underlying point is serious: we’ve been building AI that mimics the cortex. We haven’t even tried to build AI that mimics the cerebellum.
System 1 and System 2
If you’ve read Kahneman, you know the framework. System 2 is deliberate, logical, rational thinking. System 1 is intuition, emotion, motor control, the subconscious — the fast, automatic stuff.
Liu Jia’s assessment is blunt: AI has solved System 2, or will solve it soon. Reasoning, logic, analysis, deduction — it’s a matter of scale and architecture refinement. We’re close, and getting closer fast.
System 1? AI hasn’t even touched it.
Intuition. Gut feelings. The ability to walk into a room and sense something is wrong before you can articulate why. The motor intelligence that lets your fingers play a guitar riff you learned fifteen years ago without thinking. The emotional processing that tells you someone is lying based on a micro-expression you can’t consciously identify.
None of that is in any AI system. Not even close.
The career implication is uncomfortable: any job that’s primarily System 2 — analysis, reasoning, knowledge work — faces AI replacement risk. Not eventually. Now. The jobs that are safe are the System 1 jobs. The ones that require embodied intelligence, emotional processing, physical skill, intuitive judgment.
Liu Jia argues we need a “second enlightenment” — one based on neuroscience instead of philosophy. The first Enlightenment elevated rational thought. The second should learn from the cerebellum, from System 1, from the dynamic, time-dependent, massively parallel processing that biological brains do and AI doesn’t.
The soul of the Transformer
Liu Jia’s lab made a discovery that connects neuroscience and AI in a way I hadn’t seen before: the mechanism of working memory in the brain is essentially the same as the Transformer mechanism.
He traced the Transformer’s true intellectual origin not to “Attention is All You Need” (2017) but to Hinton’s 1986 “Fast Weights” paper, presented at a tiny cognitive science conference. The idea: find relationships between all things.
That’s what attention does. It takes every element in a sequence and asks: how does this relate to every other element? Attention = finding relationships. Relationships = connections. Connections = compression. Compression = understanding.
This is a profound reframe. Intelligence isn’t about storing information. It’s about compressing it. Finding the patterns underneath. The better you compress, the deeper the patterns you’ve found. A phone number is 10 random digits — hard to remember. But if you notice it’s your birthday backwards plus your zip code, suddenly it’s two patterns. You’ve compressed it. You understand it.
Transformers are, at their core, compression engines. And the best compression is intelligence.
Finding your logical origin point
This is where the conversation turns personal, and it hit me hard.
Liu Jia reflects on his own life: “Our education never forces anyone to think — what is your logical origin point?”
He went to PKU because it was the best school. He studied psychology basically by accident. He switched to computer science because psychology felt “too humanities.” He abandoned neural networks because everyone else did. “A life composed of coincidences.”
The real skill of deductive reasoning — the one that matters for life, not just math — isn’t deriving conclusions from a given origin point. Any smart person can do that. The real skill is finding your own origin point. The foundational truth that everything else in your life derives from.
Most people never do this. They accumulate decisions by accident, by social pressure, by whatever opportunity shows up next. They induct — they see what others do and follow the pattern. They never sit down and ask: what do I actually believe at the deepest level? What’s my axiom?
Liu Jia’s honest self-assessment — a Tsinghua professor admitting his career was “composed of coincidences” — is more useful than any motivational speech. Because it names the real problem. It’s not that people make bad choices. It’s that most people never develop the foundation that would make any choice feel like their own.
Three things education needs for the AI era
Liu Jia closes with three things that actually matter for education in the AI era. Not coding bootcamps. Not prompt engineering courses. Three structural shifts:
1. Cultivate the concept of “I.”
The industrial revolution buried individuality. It turned people into interchangeable parts — specialists who do one narrow thing. The AI era reverses this. One person plus enough compute equals a company. The bottleneck isn’t labor or capital anymore. It’s you. Your ideas, your taste, your origin point. Return to the Greek imperative: know thyself. Except now it’s not philosophy — it’s survival strategy.
2. AI-native thinking.
Not “using AI” — making it part of your cognitive system. The way your hands are part of your thinking when you gesture. The way your eyes are part of your thinking when you read. AI should become that seamless. Not a tool you open and close. An extension of how you think.
Kids have the advantage here. They don’t have old habits to break. They’ll grow up with AI the way millennials grew up with the internet — not as a thing they learned to use, but as a thing that was always there. Adults have to unlearn the “tool” frame first.
3. Deductive reasoning — find your own origin point.
AI can analyze anything. It can process any dataset, find any pattern, derive any conclusion from any premise. What it can’t do is find your origin point. It can’t tell you what you believe at the deepest level. It can’t decide what matters to you.
This is the one thing that’s irreplaceably human. Not because AI is technically incapable — but because the answer has to come from you for it to mean anything. An AI-generated life purpose is just another output. A self-discovered origin point is a foundation.
Don’t get left behind
Liu Jia closes with something that stuck with me.
“经过我的努力，所幸我没有被这时代给抛下.” — “Through my efforts, fortunately I haven’t been left behind by this era.”
Think about that. A Tsinghua chair professor. MIT PhD. National-level research reputation. Scientific consultant for the biggest science show in China. His highest self-evaluation is: I managed not to get left behind.
Not “I’m leading.” Not “I’m ahead.” Just: I’m still here. I’m still keeping up.
His wish for everyone is the same: don’t get left behind. Not because you’ll lose your job — though you might. But because what you’d miss isn’t just economic opportunity. It’s the biggest event in human history. The creation of a new form of intelligence. The moment our species fulfills what Liu Jia considers its evolutionary purpose.
You can disagree with his framing. You probably should interrogate it. But the urgency is real.
Sections 10 and 11 — finding your origin point and the three things education needs — gave me a strong sense of direction. Not answers. Direction. The difference matters. An answer is static. Direction is a vector. It tells you which way to walk even when the terrain keeps changing.
朝闻道，夕死可矣.
For more detailed summaries of this conversation, see the ChatGPT summary and the Gemini summary.