中文

The Evolution of Artificial Intelligence: Statistics, Economics, and Rule Inference · Thomas Sargent

2026-06-21 · A faithful, transcript-grounded reading by PodLens

Original episode:https://youtu.be/B2m-T3QQiqQ?si=KQd2LE89eI5RbuMl　·　Timestamps are clickable — they seek the player in place

structural modelsrule inferencecognitive limitationsscientific revolutioninformation theoryeconomic thinking

What This Episode Is About

In a lecture at Peking University's HSBC Business School, Thomas Sargent dissects the nature and history of artificial intelligence from the distinctive vantage point of an economist and statistician [02:53]. He defines human intelligence as three core activities — pattern recognition, generalization, and decision-making [05:48] — and argues that AI is fundamentally a machine humans invented to simulate these activities. Borrowing from Steven Pinker's cognitive science, Sargent reveals a profound paradox: evolution hardwired humans with cognitive deficits in statistics, biology, economics, and physics — precisely the fields that form the technical pillars underlying AI's algorithms [11:39]. He reconstructs in detail how Ptolemy, Copernicus, Kepler, Galileo, and Darwin drove the scientific revolution through what amounts to "artificial data compression," analyzes how AlphaGo and genetic algorithms fuse economic pricing mechanisms, game theory, and Monte Carlo simulation from physics, and explores the gap in contemporary large language models between "describing patterns" and "inferring the rules of the game."

Timeline Theme Map

Core Viewpoints

  1. The core of human intelligence is a causal chain made up of "pattern recognition," "generalization," and "decision-making." Pattern recognition compresses and classifies past sensory data; generalization is a belief that assumes past patterns extend into an unknown future; decision-making is choosing among options based on that generalization.
  2. The natural intuitions humans evolved are "disabled" across nearly every core modern scientific domain. As a product of hunter-gatherer evolution, human hardwired intuition is bad at assessing extremely low-probability events (statistical disability), resists interest and middlemen (economic disability), and cannot grasp subatomic particles or non-absolute spacetime (physical disability).
  3. The scientific revolution was, in essence, "artificial intelligence" practiced in an age without computers. Ptolemy used combinations of circles and a least-squares-like method to compress vast celestial position data into 250 parameters; Galileo generated data with inclined planes and performed curve fitting; Kepler did data mining — all were, at their core, data compression and rule generalization.
  4. The selection of scientific theories is often decided by aesthetic intuition rather than goodness of fit to data. When Copernicus proposed heliocentrism, his 36-parameter model's actual predictive accuracy for planetary orbits was far lower than Ptolemy's elaborate 250-parameter geocentric model — but scientists chose to believe it was closer to the truth because of its simplicity and elegance.
  5. Structural econometrics and Feynman's exploration of physics are logically perfectly isomorphic. Feynman described physics as "inferring the entire rule set of chess by watching a single game" [56:15]; structural economics does the same thing — observing only scattered prices and transaction volumes (the moving of pieces) and reverse-engineering the underlying rules of the game (Von Neumann's game definition: agents, choices, payoffs, and information sets).
  6. Large language models are extremely good at "describing patterns" but still have a fundamental deficiency at "inferring rules." AI systems like ChatGPT or Claude can produce excellent descriptive output through pattern recognition over vast amounts of text, but they still cannot, like a physicist or econometrician, reverse-engineer the underlying causal game structure (structural models) from a phenomenon.
  7. The core value of scientific discovery lies in "surprise that cannot be predicted." Per Turing's and Shannon's information theory, information is the "surprise content" within data. Because every genuine scientific discovery is a disruption (a surprise) of the existing trajectory, it follows logically that we can never predict what future scientific discoveries will be.

Plain English Retelling

In an era when everyone is either thrilled or terrified by large models, Nobel laureate economist Thomas Sargent offers an unusually calm and deep perspective: don't be fooled by the hype — artificial intelligence is nothing new. It's actually been running for over a hundred years, and at its core, it's a clever fusion of four classic disciplines: statistics, economics, biology, and physics [05:21], [17:30].

There's a darkly funny paradox here: over the long course of hunter-gatherer evolution, in order to hunt animals and avoid danger in the forest, our brains got hardwired with a very practical set of intuitions. But the moment that intuition runs into advanced science in modern society, it turns out to be riddled with "cognitive disabilities" [10:53]. For instance, we're not naturally equipped to assess a one-in-ten-thousand chance of disaster (statistical disability); we instinctively despise middlemen who buy low and sell high, feeling that only growing crops with our own hands counts as production (economic disability); we can't comprehend why time slows down, or how a subatomic particle can exist in two places at once (physical disability) [13:02], [14:52], [15:49]. And the most ironic part: the four foundations holding up today's entire edifice of AI algorithms happen to be exactly statistics, economics, physics, and biology [17:59]! So, borrowing Steven Pinker's words, Sargent says the essence of modern education isn't to teach you knowledge — it's to invent a "corrective technology" that fights against humanity's natural intuitive deficits [16:56].

If you zoom out to the scientific revolution centuries ago, you'll find that scientists were actually the pioneers of "machine learning" in an age without computers. Take Ptolemy, two thousand years ago: to describe five erratically wandering planets in the night sky, he turned centuries of observational records into a giant spreadsheet, then layered circles upon circles into a mathematical model with 250 parameters, fitting the data using a primitive version of least squares [20:41], [24:12]. Although he was wrong (he put Earth at the center), this data-compression model was so effective that it dominated humanity for 1,500 years. Later, Copernicus came along and said the geocentric model was too ugly, building a heliocentric one with just 36 parameters. Even though Copernicus's predictions were actually less accurate than Ptolemy's at the time, scientists simply preferred the "beauty of fewer parameters," and so heliocentrism won [27:59]. This is exactly the same logic behind today's pursuit of "regularization" and "parameter compression" in large models.

Today's AlphaGo, and today's large language models, are both, underneath, borrowing ancient tools from these same four disciplines. When AlphaGo calculates each move, it uses economics's notion of "price and value" to assess the position; it uses Monte Carlo simulation, invented by physicists in the 1940s to build nuclear weapons, to predict win probability; it uses the "exploration vs. exploitation" balance from evolutionary biology to decide whether to try a new move [41:12], [43:00].

But Sargent points to the biggest bottleneck facing AI today. Feynman once said that doing physics is like watching two people play a game of chess whose rules you don't know — your job is to infer the underlying rules just by observing how the pieces move [56:15]. Economists doing structural econometrics do exactly the same thing, trying to find the underlying rules of the game hidden behind chaotic prices and quantities [57:59]. Today's ChatGPTs are nothing short of perfect at "pattern recognition" and "describing how the pieces move," but when it comes to truly inferring, like a scientist would, the entire invisible set of underlying rules (a structural model), they're still completely lost [01:00:52].

As for whether AI will ultimately replace humans entirely, Sargent offers an answer that is remarkably warm and personal: AI may be capable of the most complex calculations, but it will never, in his entire life, create someone like his wife Carolyn — someone whose presence makes him smile every day, who can point out his mistakes respectfully in conversation, and who gives him genuine warmth and support [01:38:47].

Recommended Segments for Close Listening

Resonances with past episodes

A faithful reconstruction and plain-language retelling of the episode, generated by PodLens.

This is one source-grounded reading, not a replacement for the original. Every point is anchored to its source, so you can check it yourself — and corrections are welcome.