A Collectivist, Economic View of AI

About this talk

Speaker: Michael I. Jordan (UC Berkeley / Inria Paris) — identified from the talk; see Speakers
Event: Invited talk at Berkeley (“home terrain”); a Berkeley dean of CDSS is in the room. Exact event not confirmed
Date: 2024-11-13 (Berkeley EECS colloquium “A Collectivist Vision for AI”; recording confirmed as this talk, uploaded 2024-11-18)
Length: ~1:05:00 (transcript ends 1:04:45)
Channel: UC Berkeley EECS
Watch: Michael I. Jordan: A Collectivist Vision for AI

How to read this page. It has two kinds of content, always marked:

🎙 In the talk. What Jordan (or the audience) says, with [MM:SS] timestamps. Quote marks mean exact words from the captions. Only these parts are his views.
📚 Added context. Background from the traditions he draws on, plus later facts and objections. It explains why he thinks what he thinks and who disagrees. It is not his claim unless a timestamp says so. Facts marked “from memory; not checked” come from my own knowledge and were not looked up.

Suggested route. Summary → Map → Key terms → Part 2: The intellectual map (read once, slowly) → Deep dives in order → “Putting it together”.

Speakers

Label	Who	Role / confidence
Introducer	Not named in the transcript. Lists prizes and thanks “Mike”	Host at Berkeley. The first words are cut off
Jordan	Michael I. (Mike) Jordan	Main speaker. High confidence. The intro lists the World Laureates Association Prize (inaugural), the John von Neumann prize, and “most influential computer scientist in the world”. He calls Rumelhart “my advisor”, says he “just arrived at Berkeley around 1998”, and an audience member asks why he “moved to France”. He answers: “That’s absolutely one of the reasons I’m in France”
Moderator	Not named	Runs Q&A
Audience Q1	Not named	Asks how strong the economic assumptions are
Audience Q2	Not named	Asks why the UnitedMasters market needs three sides
Audience Q3	Not named	Says French officials are École Polytechnique graduates; raises operations research history
Audience Q4	Not named	Asks about four-way or five-way markets
Named in the room	”John” (quant-finance leader), “Eli”, the CDSS Dean, student Meena (job market), Alireza Fallah	Mentioned as asides only

Caption warnings (auto-captions; names and terms are often wrong). Corrections used on this page:

“David Romer Hult prize” = David E. Rumelhart Prize; “Alan noble award” = probably Allen Newell Award; “AMS of Granada prize” = not identified (uncertain); “John Von Neumann” = IEEE John von Neumann Medal (from memory).

“Sam Mormons” at [1:01:53] = Sam Altman (the words “the West Coast people are all crazy” are about the same group).

“quantum disorder” / “quantum fluctuation(s)” [38:12]–[41:47] = almost certainly intrinsic (structural) disorder: loops and strands of a protein with no fixed fold. Nothing quantum is involved. The captioner or speaker mixed the word up. Treat as a caption error; unverified which.

“principle” for principal at [32:09] and [47:00].

“Dirk Bergman” = Dirk Bergemann (with Ben Brooks and Stephen Morris, “The Limits of Price Discrimination”, 2015, from memory).

“Stackelberg” is correct. “Marchenko-Pastur” is correct (a law from random matrix theory).

“Steven” = probably Stephen Bates (MIT); “Michael” = probably Michael Sklar (Stanford); “Jake” = probably Jake Soloff (Chicago). All three are guesses from the paper “Principal-Agent Hypothesis Testing” (from memory; not checked).

“Nika”, “Nivasini”, “Meena”, “Jacob Steinhardt”, “Alireza” = probably Nika Haghtalab, Nivasini Ananthakrishnan, Meena Jagadeesan, Jacob Steinhardt, Alireza Fallah.

“Random Forest … gradient-based” at [05:47]: Jordan says Breiman’s trees are “gradient-based”. Random forests are not gradient-based in the usual sense. See the slip list below.

“[AUDIO OUT]” at [58:12] hides a few words of the answer to Q1.

Factual slips and loose claims by the speaker (checked from memory; not looked up unless linked):

At [04:01] he says Rumelhart “reinvented the backpropagation algorithm and published a paper on it in 1983”. The famous paper (Rumelhart, Hinton, Williams) is 1986. Earlier derivations exist (Linnainmaa 1970; Werbos 1974). From memory.

At [04:32] he says Rumelhart “would have been aghast” at a “Nobel Prize”. The Physics Nobel 2024 went to John Hopfield and Geoffrey Hinton. The Turing Award for backpropagation-based deep learning (2018) went to Hinton, LeCun and Bengio. This is the basis for dating the talk after October 2024. From memory.

At [05:47] “gradient-based” about random forests is loose. Forests are built by greedy splitting and averaging. The shared idea is “adjust predictions from data at scale”, not gradient descent.

At [16:21] “That’s a Nash equilibrium, meaning highest maximal social welfare”. A Nash equilibrium is not in general welfare-maximal (the prisoner’s dilemma is the counter-example). In the duck problem the equilibrium happens to match the food. He is describing the ideal free distribution of ecology (Fretwell and Lowe 1970; Harper’s 1982 duck experiment, from memory).

At [06:20] “30 years ago” for Amazon in the late 1990s. This is fine as a rough date for a 2024 talk.

At [05:47]–[06:20] the claim that the cloud “emerged from Random Forest run on supply chain data at Amazon” is his version of history. The usual account of AWS (2006) is that it grew from Amazon’s internal infrastructure work and the need to sell spare capacity. Whether a random-forest supply chain system was the cause is not something I could confirm. Treat as his claim.

At [35:40] “a recent paper showing how badly it’s hurting small firms in Europe”. Several studies find GDPR reduced the number and funding of small firms or apps. I did not identify the paper he means. Mark as not verified.

At [38:12] “only about 100,000 sequences whose structure is known” vs “hundreds of millions” from AlphaFold. The PDB held roughly 200,000 entries in 2024 (from memory), and the AlphaFold database holds about 200 million predictions. Order of magnitude holds.

At [1:02:31] the audience member says it “all came out of operations research and linear programming in the 1950s”. That is an audience claim, not Jordan’s. He agrees in part.

At [51:43] he calls the Bergemann–Brooks–Morris result something that “may get a Nobel Prize itself”. An opinion, not a fact.

Summary

Central question. Jordan asks “why are we doing this?” about AI. He thinks the public and policy dialogue is “horrific”, full of science-fiction promises: “solve hunger”, “solve climate change”, “make the economy grow exponentially fast”. He wants a “more sober” answer to “what are we really doing”.

Main thesis. The right picture is not one superhuman mind that replaces people. It is a federated architecture of computers and humans, a collective. To build such systems well you need three ways of thinking, not one:

Computer science (algorithms and logical recipes),
Statistics (evidence from partial observation, uncertainty),
Economics (incentives, prices, contracts, markets).

Most of AI is the first two blended (that is “machine learning”). What is missing is the third. He lists three gaps in today’s AI talk: collectives, real uncertainty, incentives.

How he argues it.

History [02:49]–[07:30]: three visions from the 1950s (Wiener, McCarthy, Engelbart). Machine learning won in practice, through supply chains, then transactions, then language.
Uncertainty [11:28]–[16:55]: ChatGPT is “very agreeable”, not calibrated. Humans handle uncertainty by relying on other people. The duck experiment shows a collective explanation for what looks like irrational behaviour.
Markets [18:08]–[21:37]: markets give equilibrium, less uncertainty, and stability. The “tower on the hill” super-entity is a “fever dream”.
Worked cases [21:37]–[56:26]: a three-sided music market (UnitedMasters), three-layer data markets with differential privacy, price discrimination with privacy, barriers to entry for LLM firms, and, in the most technical part, statistical contract theory for clinical trials, plus prediction-powered inference for AI-generated data.

Close. [55:47]–[57:40]: a new engineering discipline is emerging, about “all of us being connected in new ways”, like chemical engineering after chemistry and electrical engineering after Maxwell. It needs about “20 years”. He says: “I am an engineer.”

Stance. Sober, mathematical, pro-technology, anti-hype. He is not against markets (“they’ve been around for longer than humans have been around”) and not against regulation; he wants regulation designed with economists, at the level of the equilibrium, not the mechanism. He is hostile to the “superintelligence” story and to the US discourse (“completely broken”), and critical of free-with-ads internet business models (“very distortionary”).

Map — talking points

#	Point	Time
1	Introduction and prizes	[00:00]
2	Why this talk: the AI dialogue is lost; “why are we doing this?”	[01:07]
3	Three 1950s visions: Wiener (cybernetics), McCarthy (symbols), Engelbart (aiding humans)	[02:49]
4	What really happened: machine learning. Rumelhart and Breiman	[03:56]
5	Three eras: supply chain (Amazon, cloud), transactions (recommendation), language (LLMs)	[04:32]–[07:30]
6	Reframe: new federated architectures of computers and humans	[07:30]
7	Three ways of thinking (CS, stats, economics) and the diagram; econometrics, algorithmic game theory, the missing middle; CDSS	[08:01]–[10:17]
8	What the AI dialogue lacks: collectives, uncertainty, incentives	[10:17]
9	ChatGPT confidence is “very agreeable”; calibration plot	[11:28]–[12:34]
10	Coping with uncertainty through collectives; the duck and probability matching; Nash equilibrium	[12:34]–[16:55]
11	Networks need incentives: why would anyone connect, why not lie?	[16:55]–[17:29]
12	The Silicon Valley “thing on the hill”; replace it with an embedded collective	[17:29]–[19:19]
13	Markets defended: equilibrium, less uncertainty, stability (tomatoes, pizzeria). Regulate the equilibrium	[19:19]–[21:37]
14	Prices and incentives replace bare predictions	[21:37]
15	UnitedMasters: a three-sided music market, NBA website, unraveling vs Spotify	[21:37]–[24:41]
16	Three-layer data markets (users, platforms, data buyers). Differential privacy as a platform choice	[24:41]–[27:32]
17	Theory of incentives = inverse game theory. Nash → auctions. Stackelberg → contract theory	[27:32]–[29:50]
18	Contract theory via airline pricing; menus of contracts; principal and agent; statistical contract theory	[29:50]–[33:15]
19	The model: linear regression layers, utilities, the trade-off parameter beta, equilibrium theorems, GDPR comparison	[33:15]–[36:56]
20	Prediction-powered inference: AlphaFold as biased “ground truth”, odds ratio, corrected intervals	[37:30]–[43:27]
21	Clinical trials as a contract problem: FDA as principal, pharma as agent, type I/II error, e-values	[43:27]–[49:52]
22	Price discrimination plus differential privacy: the Bergemann triangle shifts	[50:27]–[52:52]
23	Barriers to entry for LLM firms: reputation damage, scaling laws, random matrix theory, French government	[52:52]–[55:47]
24	Close: a new engineering discipline, like chemical or electrical engineering. “I am an engineer”	[55:47]–[57:40]
25	Q&A 1: how strong are the economic assumptions? Data inform the curves	[57:40]–[59:23]
26	Q&A 2: why three sides? Unraveling and Akerlof’s lemons; no ads; “I would pay for YouTube”	[59:23]–[1:01:20]
27	Q&A 3: why France? Polytechnique-trained officials	[1:01:20]–[1:02:31]
28	Q&A 4: operations research and linear programming vs backpropagation; stochastic gradients; who will dominate AI?	[1:02:31]–[1:04:12]
29	Q&A 5: four-way, five-way markets; brokers as “new statistical agents”	[1:04:12]

Part 1 — Before you read

Key terms

Collective. Here, a group of people (and machines) whose data, knowledge and choices jointly produce the intelligence of a system. Jordan: “the data that drives all of this stuff is coming from collectives” [10:52]. Not the same as “collectivism” in politics.
Federated. Many separate computers and people working together without one central controller. He first applies it to Amazon’s computers [05:47] and then to “federated architectures that involve computers and humans” [07:30]. Watch out: in machine learning, “federated learning” is a specific technique (training on local devices). He uses the word loosely.
Mechanism design / theory of incentives. The “inverse of game theory” [28:44]. You pick the outcome you want, then design the rules (the game) so that self-interested players produce it.
Game theory vs mechanism design. Game theory predicts what happens given the rules. Mechanism design chooses the rules.
Nash equilibrium. A state where no player gains by changing strategy alone. Players move simultaneously. Auction design uses it.
Stackelberg equilibrium. A sequential game: a leader moves first, a follower reacts. Jordan sees it as the right model when agents “have different knowledge and they’re not symmetric” [29:17].
Contract theory. The mechanism-design side of Stackelberg games. Handles asymmetric information between a principal (who designs the contract) and an agent (who knows something private). Airline price menus are his example.
Principal / agent. Words that carry a lot. A principal sets terms; an agent picks among them using private knowledge. Note that a person can be both: “the platform is a principal … or an agent” [33:15].
Asymmetry of information. One side knows something the other does not (willingness to pay; whether a drug works).
Differential privacy. A method that adds calibrated noise to data so that no single person’s record can be confidently inferred. Here it is a dial that a platform sets, with a price attached, not a legal mandate.
Unraveling. A market collapsing because the good participants leave, leaving only the bad. Akerlof’s “lemons”. Jordan: Spotify’s model needs ads or subscriptions to survive; the three-sided model does not unravel [1:00:00].
Equilibrium vs optimum. An optimum asks “optimal for whom?“. An equilibrium is where “everybody gets a piece” [19:19]. Warning: in everyday English “equilibrium” sounds like “fair”. It is not; it just means stable.
Probability matching. Choosing options in proportion to their reward odds (left 2/3, right 1/3), instead of always choosing the better option. Looks irrational; has a collective explanation.
Calibration. Whether a stated 70% confidence is right 70% of the time. His ChatGPT plot shows it is not [12:02].
Prediction-powered inference (PPI). A statistical method that combines a large pile of model predictions with a small set of real labelled data to get confidence intervals that stay valid. Not a prediction method; an inference method.
Type I / type II error. False positive (approve a useless drug) and false negative (reject a good one). Power is 1 minus type II.
E-value. A non-negative random quantity whose expected value under the null is at most 1; large values are evidence against the null. A non-negative supermartingale is the sequential version. In the talk, contracts are incentive-aligned “if and only if” the payoff functions are e-values [49:13].
Scaling laws. Empirical rules for how model error falls as data and compute grow. Jordan uses them as inputs to an economic model of market entry [54:04].
Price discrimination. Charging different customers different prices. Government worries about it; economists often find it improves total welfare.
Broker. An intermediary that turns many suppliers and many users into a market. He predicts a new kind: “new statistical agents” [1:04:12].

People named (a reading list)

Person	Why they come up	Start with
Norbert Wiener	Cybernetics; control and feedback	Cybernetics (1948); The Human Use of Human Beings (1950)
John McCarthy	Coined “artificial intelligence” (1955–56)	The Dartmouth proposal (1955)
Douglas Engelbart	Computers should augment human intelligence	”Augmenting Human Intellect” (1962)
David Rumelhart	Jordan’s PhD advisor. Backpropagation	Rumelhart, Hinton, Williams (1986)
Leo Breiman	Random forests; the “two cultures” paper	”Statistical Modeling: The Two Cultures” (2001)
Steve Stoute	UnitedMasters CEO; “legendary hip hop producer”	UnitedMasters’ own materials
Alireza Fallah	Co-author on data markets and privacy-price-discrimination	Fallah, Jordan et al. (from memory)
Dirk Bergemann (with Brooks, Morris)	The price-discrimination triangle	”The Limits of Price Discrimination” (AER 2015)
Meena Jagadeesan	Market entry and scaling laws	Her papers with Jordan and Steinhardt
Jack Kiefer	Early stochastic gradient paper (Kiefer–Wolfowitz 1952)	The original paper
Cynthia Dwork	Differential privacy	Dwork, McSherry, Nissim, Smith (2006)
George Akerlof	”Market for lemons” (unraveling)	“The Market for ‘Lemons’” (1970)
Heinrich von Stackelberg	Sequential leader–follower games	Marktform und Gleichgewicht (1934)

Jordan’s own frame

Jordan is a professor at UC Berkeley (EECS and Statistics), long-time leader in machine learning and Bayesian statistics, and in recent years also affiliated with Inria in Paris. He has argued for years against the label “AI” for what is mostly engineering. Two earlier essays set up this talk: “Artificial Intelligence — The Revolution Hasn’t Happened Yet” (Harvard Data Science Review, 2019) and “Dr. AI, or: How I Learned to Stop Worrying and Love Economics” (HDSR, 2019). The first calls for an engineering discipline of “intelligence infrastructure”; the second brings in economics. Both from memory; not checked. This talk is a progress report on that programme with his group.

The current debate (2025–26)

Jordan has put the thesis in writing. “A Collectivist, Economic Perspective on AI” (arXiv 2507.06268, first submitted 8 July 2025, v3 dated 15 December 2025). As summarised by the search results, it argues that “AI” mainly evokes the 1950s goal of “thinking machines”, that an LLM is also a “collectivist” artifact (you interact implicitly with the many humans whose data went in), and that a market is a closer metaphor for emerging AI systems than a search engine, chatbot or personal secretary (arXiv v2, abs v3). I did not read the paper in full; the summary comes from search snippets.
The talk travels. The same idea appears as “A Collectivist Vision for AI” at the Berkeley EECS colloquium on 13 Nov 2024 (EECS page), as a public lecture “A Collectivist, Economic Perspective on AI” at LSE (LSE Player), at Oxford on 29 January 2026 per the search results (OxTalks), and as a Machine Learning Street Talk episode in May 2026 (Pulse summary). The Berkeley abstract says agents join learning systems only when it serves their interests.
The web bargain is being re-offered. In the paper summary, Jordan says the early-internet bargain (data free for the taking, services free in return) is now “implicitly” offered again by LLM developers. That ties to live fights over training data, copyright and creator payment. In this talk his answer is the UnitedMasters model [22:15]–[24:41].
Opposing views. Critics of market framing argue that markets for data cannot price harms to people who are not at the table, that consent in data markets is weak, and that concentrated power is the real issue. I know these as general positions (for example, work on “data as labour”, the Data Dignity idea of Lanier and Weyl, and antitrust critics of platform power); I did not search for a direct reply to Jordan. From memory; not checked.
The “normal technology” view. Narayanan and Kapoor’s 2025 essay “AI as Normal Technology” is close to Jordan on the hype and the ‘superintelligence’ story, though not on economics. From memory; not checked.

Part 2 — The intellectual map

The talk is 57 minutes of argument for an audience that knows its Berkeley statistics and CS. It leans on eight families of ideas, only some of which it names. They are laid out below, one at a time, then shown meeting.

 1950s visions        Statistics           Economics           Markets in nature
 Wiener  McCarthy     (inference,          (equilibrium,       (ducks, ideal free
 Engelbart            uncertainty)         incentives)          distribution)
     \                     |                    |                    /
      \                    |                    |                   /
       +---- Machine learning (CS + stats) ----+                  /
                           |                                     /
          [what the AI dialogue misses: collectives,            /
           uncertainty, incentives] ---------------------------+
                           |
        Mechanism design --> Contract theory --> Statistical contracts (clinical trials)
                           |
      Privacy (differential privacy) + Data markets + Price discrimination
                           |
          Prediction-powered inference (bias of AI-made data)
                           |
          Scaling laws + random matrix theory --> market entry
                           |
          NEW ENGINEERING DISCIPLINE (chemical / electrical engineering analogy)

2.1 Three visions of intelligence — control, symbols, augmentation

Problem. What is a computer for? Is it a mind, a controller, or a tool for minds? Chain of thinkers.

Norbert Wiener (1894–1964). Cybernetics (1948): feedback and control in animals and machines. Intelligence as a property of a loop, not of a thing. Jordan: his “vision was kind of control theory esque at large scale” [02:49].
John McCarthy (1927–2011). Coined “artificial intelligence” for the 1956 Dartmouth workshop. The programme was symbolic: logic and representation, “a robot thinking with logical symbols in its head” in Jordan’s words [02:49]. Dominant until the 1980s.
Douglas Engelbart (1925–2013). “Augmenting Human Intellect” (1962); the 1968 “Mother of All Demos” (mouse, hypertext, shared screens). Computers amplify human intelligence. Close to J.C.R. Licklider, “Man-Computer Symbiosis” (1960).
Later: Michael Jordan himself names this lineage “intelligence augmentation” (IA) vs AI in his 2019 essay (from memory). How it feeds the talk. Jordan sides with Wiener and Engelbart: the search engine is “not intelligent, but it aids all of our intelligence” [03:23]. His “collective” is the Engelbart picture made precise, with economics added to say why people would take part. Watch out. The 1950s people did not form a school. Jordan says “They all hated each other, I think” [03:23].

2.2 Machine learning as engineering, and the two cultures

Problem. Why did prediction at scale, not reasoning, win? Chain.

Stochastic approximation: Robbins and Monro (1951), Kiefer and Wolfowitz (1952). Jordan cites Kiefer at Berkeley at [1:03:07].
Linear and nonlinear programming and operations research (Dantzig, Kantorovich, 1940s). The audience member in Q&A stresses this lineage.
Backpropagation: Linnainmaa (1970), Werbos (1974), Rumelhart–Hinton–Williams (1986). All from memory.
Leo Breiman (1928–2005). Classification and Regression Trees (1984), bagging (1996), Random Forests (2001), and “Statistical Modeling: The Two Cultures” (2001), which argued statisticians should care about predictive accuracy as much as about models.
Cloud computing: distributed systems, MapReduce (Dean and Ghemawat 2004), AWS (2006).
Transformers and next-token prediction (Vaswani et al. 2017; GPT series). Language as just another data set. How it feeds the talk. The “three eras” story [06:20]–[07:30]: “Same algorithms, same cloud. Everything was the same, just the data changed.” The point is that nothing in this lineage is about intelligence. It is about large-scale prediction in industry. That is why he says “this is all kind of engineering” [07:30] and why he later wants engineering done properly. Watch out. “Gradient-based” and “stochastic” carry the argument at [1:02:31]. Jordan’s claim that the 1950s operations researchers “weren’t being statisticians” means they optimised exact objectives, whereas the winning idea is to optimise sums over huge data with noise.

2.3 Statistics: evidence, partial observation, honest uncertainty

Problem. How do you learn about a world you only partly observe, and say how sure you are? Chain.

Frequentist school: Fisher (significance, 1920s), Neyman and Pearson (type I and II error, power, 1933), confidence intervals (Neyman 1937).
Bayesian school: Bayes, Laplace, de Finetti, Savage. Probability as belief; the “Bayesian optimal” decision in the duck example.
Calibration and proper scoring rules: Brier (1950), Dawid’s calibration (1982).
Causal inference: randomised experiments (Fisher), potential outcomes (Neyman, Rubin), clinical trials as the gold standard.
Sequential analysis and e-values: Wald (1945) sequential tests, Ville’s inequality (1939), recent work by Vovk, Shafer, Grünwald, Ramdas, Wang. From memory.
Robust and semi-parametric inference: using machine learning predictions inside valid inference. This is where prediction-powered inference sits (Angelopoulos, Bates, Fannjiang, Jordan, Zrnic, Science 2023, from memory). How it feeds the talk. Three places: (a) the claim that ChatGPT has no real uncertainty [11:28]; (b) AlphaFold-as-data gives “very sure of yourself and … very far wrong” [40:32]; (c) the clinical-trial contract is built from Neyman–Pearson error control [44:39]–[49:52]. Watch out. “Uncertainty” has three meanings in play. (1) Sampling error (error bars). (2) Model uncertainty (am I using the right model?). (3) Real-world unpredictability (“I don’t know what’s going to happen tomorrow”). Jordan’s phrase “I don’t mean just a little bit of error bars” [10:52] points at (2) and (3).

2.4 Economics I: markets, prices, equilibrium

Problem. How can millions of people with private knowledge coordinate without a planner? Chain.

Adam Smith (1776): the invisible hand.
Walras, Arrow and Debreu (1870s; 1954): general equilibrium exists under conditions. The welfare theorems: competitive equilibria are efficient.
Friedrich Hayek, “The Use of Knowledge in Society” (1945): the price system is a way to pass on dispersed knowledge. Jordan’s line “Prices are ways to decentralize computation” [21:37] is Hayek in computer-science words. He does not name Hayek; this is my link.
The calculation debate (Mises, Lange, Hayek, 1920s–40s): can a planner replace the market? The Silicon Valley “thing on the hill” is the planner’s dream in AI form (my reading, not his words).
Market failures: externalities (Pigou), public goods (Samuelson), information (Akerlof, Stiglitz, Spence; Nobel 2001).
Regulate the equilibrium, not the mechanism. Jordan’s rule of thumb [21:06]; the idea that policy should set outcomes (levels, prices, caps) rather than micro-manage the rules. He says he is “not an expert here”. How it feeds the talk. The tomato market [19:54] is a Hayek story told with food: thousands of drivers raise prices and move goods and the city gets fed. The pizzeria example [20:34] is about stability enabling investment: reliable inputs let new layers be built on top. Watch out. “Market” means very different things: a physical bazaar, a stock exchange, a designed matching platform, and “the market” as ideology. The history department hated his talk because they heard the last meaning [19:19]. He means a designed or evolved institution with prices and incentives.

2.5 Economics II: game theory, mechanism design and contract theory

Problem. If people act in their own interest, how do you design rules so that good outcomes follow? Chain.

Von Neumann and Morgenstern (1944): game theory. John Nash (1950): equilibrium for simultaneous games.
Heinrich von Stackelberg (1934): leader–follower (sequential) games.
Leonid Hurwicz (1960 onward), Eric Maskin, Roger Myerson: mechanism design, the “revelation principle”. Nobel 2007.
Vickrey (1961), Clarke, Groves: auctions and truthful mechanisms. Auction theory in practice: spectrum auctions (Milgrom and Wilson, Nobel 2020).
Contract theory: Mirrlees (1971), Mussa and Rosen (1978) on screening; Holmström; Oliver Hart and Bengt Holmström, Nobel 2016. Laffont and Martimort’s book The Theory of Incentives (2002) is the standard text. Jordan says he learned this “by reading a book about it” [29:50], likely that one or a similar text (not stated).
Screening: the airline offers a menu so that customers reveal their type by choosing. The “fences” in airline fares (Saturday-night stay, refundability) are the real-world version.
Algorithmic game theory: Nisan, Roughgarden, Tardos, Papadimitriou (1990s–2000s): computational aspects of auctions and equilibria. Jordan places it at “economics meets computer science” [09:40]. How it feeds the talk. This is the core machinery. Nash gives “auctions” [29:17]; Stackelberg gives “contract theory” [29:50]. The step Jordan adds is statistical contract theory: contracts whose terms are written in terms of data and tested by statistics [47:32]. Watch out. “The inverse of game theory” is a slogan. Technically, mechanism design is a reverse problem (choose the game for the outcome), with constraints: players must opt in (participation), and must find it best to follow the intended behaviour (incentive compatibility). “Incentive aligned” in the clinical-trial part [49:13] is his short form.

2.6 Evolution, ecology and collective behaviour

Problem. Why do animals (and people) behave in ways that individual rationality calls irrational? Chain.

Ideal free distribution: Fretwell and Lowe (1970). Animals spread across patches in proportion to resources. Experiments with ducks and bread (Harper 1982, from memory).
Probability matching in psychology: Herrnstein’s matching law (1961); Estes (1950s). Often read as an error, as Jordan says [15:08].
Evolutionary game theory: Maynard Smith and Price (1973), evolutionarily stable strategies. A Nash equilibrium of a population game, with no one deciding anything.
Collective intelligence and the “wisdom of crowds”: Galton’s ox (1907), Condorcet’s jury theorem (1785), and more recently Surowiecki (2004). Science of “collective intelligence” (Malone, Woolley).
Social learning: humans evolved to ask others (Henrich, The Secret of Our Success, 2015, from memory). How it feeds the talk. The duck is the central metaphor for ‘collective rationality’ [13:53]–[16:21]. Note the move: the duck’s apparent error disappears once you add the other ducks. His lesson: “We evolved in worlds where we’re collective, and we should be living in worlds where collective thinking is actually favorable” [16:21]. Watch out. It is an explanation, not a proof. The same behaviour can also come from other causes (hedging against model error; exploration). Jordan rules those out only informally [15:08].

2.7 Privacy, data markets and price discrimination

Problem. Personal data are valuable and exposing. Who decides how much privacy there is, and who pays? Chain.

Differential privacy: Dwork, McSherry, Nissim, Smith (2006). A formal, tunable guarantee; the “epsilon” is the dial. Used by the US Census 2020 and big platforms.
Economics of privacy: Acquisti, Taylor, Wagman survey (2016, from memory). Privacy has costs and benefits on both sides.
Data markets: Ghorbani and Zou on data valuation; Shapley values of data (2019, from memory). Also the “data as labour” proposals (Posner and Weyl, Lanier).
GDPR (EU, in force 2018): a fixed legal regime. Studies debate its effects on competition and small firms.
Price discrimination: Pigou’s three degrees (1920); Bergemann, Brooks and Morris (2015): over all ways of splitting a market into segments, the achievable pairs (consumer surplus, producer surplus) form a triangle. One corner gives consumers everything the seller can’t take; another gives the producer the maximum (perfect discrimination) with zero consumer surplus.
Information design: Kamenica and Gentzkow (2011): how an informed party should reveal information. How it feeds the talk. Two projects: (a) the three-layer data market [24:41]–[36:56], where each platform sets its own noise level and the equilibrium depends on a parameter beta; (b) the shifted triangle with privacy [50:27]–[52:52], where “consumer utility is bounded away from zero”. Watch out. “Privacy” here is a continuous good that is bought and sold, not a right. That is a normative choice. A rights-based view (GDPR’s own logic) treats privacy as not for sale.

2.8 Bias of AI-generated data and prediction-powered inference

Problem. If your “data” come from a model, your statistics inherit its blind spots, and confidence goes up while truth goes away. Chain.

Missing data and surrogate endpoints: Rubin on missing data; semi-supervised inference; “doubly robust” estimators (Robins, 1990s); Chernozhukov et al. on “double machine learning” (2018). All from memory.
Model collapse and synthetic data: Shumailov et al. (2023–24): training on model output degrades models. A related worry.
AlphaFold: Jumper et al. (2021); Nobel Chemistry 2024 to Baker, Hassabis and Jumper (from memory). A model trained on the Protein Data Bank.
Prediction-powered inference: Angelopoulos et al. (2023). Use the model on a large unlabelled set, and a small labelled set to estimate the model’s bias; combine into an interval that is valid whatever the model does. How it feeds the talk. The odds-ratio example [37:30]–[43:27]: the AlphaFold-only interval is tight and wrong; the labelled-only interval is honest and wide (it “covers one, sadly”); the PPI interval is honest and narrower. Jordan frames it as an information asymmetry [43:27]: the tower on the hill knows a lot; the scientist knows a little about her particular problem. The point of the method is to put them together. Watch out. It does not say AlphaFold is bad: “I’m not going to assert that AlphaFold is a bad idea at all” [42:19]. It says the new science is done where the model was not trained, which is where it is most likely wrong.

2.9 Engineering disciplines and “how fields are born”

Problem. What does it take to turn a body of science into dependable infrastructure? Chain.

Chemical engineering: George E. Davis’s Handbook of Chemical Engineering (1901) and the “unit operations” idea (Arthur D. Little, 1915). Chemistry first; then a generation of work to build working factories.
Electrical engineering: Maxwell (1860s), then Edison, Tesla, Westinghouse and the grid.
Software engineering and systems engineering: the 1968 NATO conference coined “software engineering”.
Science and technology studies on infrastructure: Susan Leigh Star, Paul Edwards: infrastructure is invisible until it breaks. From memory. How it feeds the talk. The close [55:47]–[57:05]: we have proto-principles (gradient ascent, hypothesis testing) and “need about 20 years” to build systems that “for good reasons, lead to human happiness”. His 2019 essay makes the same analogy (from memory). Watch out. The analogy is a prediction, not evidence. Chemical and electrical engineering produced dangers too (pollution, accidents) that regulation later handled. Whether AI will follow a similar path is open.

Part 3 — Deep dives

1. Why the AI dialogue is “lost” — history and the reframe

🎙 In the talk.

He opens with a complaint: “the dialogue about this overall field that we’re now calling AI is so horrific” [01:40].
The people with money give science-fiction reasons: “We’re going to solve climate change” [02:16]. He asks what a student would think: “what the heck is this?”
The three 1950s visions: Wiener (“intelligence arising out of control mechanisms”), McCarthy (“the thought in a computer idea”), Engelbart (“they should aid human intelligence”) [02:49]–[03:23].
“So I do tend to think these guys got it right. And sadly, this guy’s terminology won out.” [03:56] (He means Engelbart’s and Wiener’s picture was right; McCarthy’s word won.)
His own history of what happened: “machine learning is really just gradient descent algorithms. Make small adjustments to large bags of parameters.” [03:56]
Three eras of the same recipe [06:20]–[07:30]: supply chain, transactions, human language. “Same algorithms, same cloud. Everything was the same, just the data changed.” [06:53]
The reframe: “We’re not building superhumans where we take out a human and put in a computer. We’re building new federated architectures that involve computers and humans to solve real world problems.” [07:30]

📚 Deeper theory.

The AI-as-replacement frame is the McCarthy line (2.1). Jordan’s claim about his history is that the academic visions mostly did not cause the engineering: “Those academic perspectives just informed everything and we all went marching off” [03:56]. That is a claim about causation in the history of science; historians of computing would add that funding (DARPA, Cold War), military control research and industry all shaped the field. From memory; not checked.
The word “AI” was chosen partly for funding reasons in 1955–56 (McCarthy later said he wanted to avoid being tied to Wiener’s cybernetics). This is a commonly told story. From memory; not checked.
The “same recipe, new data” thesis is a version of the claim that data, not algorithms, drive the third wave. Compare Halevy, Norvig and Pereira, “The Unreasonable Effectiveness of Data” (2009).
Calling the output a system of people and machines connects to sociotechnical systems research (Trist and Emery, 1950s; later STS). He does not cite it.
The Amazon cloud story [05:47] is a nice causal chain but is contested (see slip 6).

Where others push back.

Researchers inside the large labs argue that scale produced real capabilities that were not just “the same algorithms with new data”: emergent abilities, in-context learning, reasoning-like behaviour. Jordan acknowledges “it is surprising” [06:53] but treats it as prediction.
Symbolic-AI defenders say the logic tradition did succeed (databases, theorem provers, verification), just not under the name.
Historians point out that “AI” as a public frame now does political work (investment, regulation, national strategy). Calling it “just engineering” does not remove that.

What they leave out.

Labour: who produces the training data and the labels, and under what terms. (He turns to this through markets, but not through labour rights.)
Compute and energy politics. He mentions “vast energy consumption” in passing [18:08] only.
Military and state uses.

Check yourself.

Which of the three 1950s visions does Jordan prefer, and which one’s word survived?
What exactly changed between his “three eras”, and what stayed the same?
Why does he say “new federated architectures” instead of “superhumans”? What follows for regulation?

2. Three ways of thinking, and what is missing

🎙 In the talk.

“There are three ways of thinking that are contributing to all this.” [08:01] Computer science is “algorithms and formulae, logical recipes”; statistics is “thinking about what’s in the world and how can I actually get evidence about that, even though I only have partial observations” [08:34].
“If you blend the two, that’s what machine learning really is” [08:34]; “it’s really mostly statistics, frankly”.
“What’s that’s missing, though, almost entirely is incentives.” [08:34] He adds that AI “talks a little bit about fairness because things are broken, but it doesn’t have a good solution for solving it.”
The pairings: economics + statistics = econometrics (“It’s analyzing the economy. It’s not trying to change the economy”); economics + CS = algorithmic game theory (“almost no statistics in this field”) [09:07]–[09:40].
“there’s something in the middle … that’s where most real world, large scale industry is” [09:40].
On Berkeley’s CDSS: “Very hard because there’s a resistance from all of them. They think they are ready to solve all the problems, but they can’t.” [10:17]
What the AI dialogue lacks: “There’s not very much talk about collectives”, “Lack of focus on uncertainty”, “lack of focus on incentives” [10:17]–[11:28].

📚 Deeper theory.

Jordan’s Venn diagram is a claim about disciplinary styles, which he says are “hundreds of years old ways of thinking” [08:01]. Compare C.P. Snow’s two cultures (1959) and Breiman’s two cultures (2001). Jordan wants three, and a blend.
Econometrics (Frisch, Haavelmo, Tinbergen, 1930s; Nobel 1969 onward; Heckman, Angrist, Imbens in the causal-inference wave) is retrospective analysis. The “Lucas critique” (1976): a model fitted on old policy fails under new policy. This is a reason economics got wary of prediction without mechanism. From memory.
Algorithmic game theory (Roughgarden and others) is strong on worst-case guarantees but mostly assumes known value distributions. Jordan’s complaint (“no learning from let’s do the auction again and again”) is partly answered by the “learning in auctions” and “online mechanism design” literatures (from memory). He may be generalising.
Fairness and incentive blindness: the machine-learning fairness literature (Dwork et al. 2012, Hardt et al. 2016) is mostly about equal treatment under a fixed predictor; it generally does not model strategic response (Hardt et al. 2016, “Strategic classification”, is a partial exception). From memory.
Economists have also been moving into this space: Acemoglu, Athey, Bergemann, Bonatti, Rajan and others on data markets and AI. His claim of a gap is about the AI discourse, not about all of academia.
The “middle” he describes (“where most real world, large scale industry is”) is where operations research, marketplaces design and ad auctions already sit (Google ads, ride-hailing, Amazon). His pitch is to bring that into public-interest settings.

Where others push back.

Social scientists would add sociology, law, political science and ethics as ways of thinking that CS, stats and economics also lack. Three disciplines is already a narrow circle.
Behavioural economists say rational-incentive models miss norms and fairness preferences. (He addresses this in Q&A, see deep dive 9.)
Some in the ML community say “incentives” is already in reinforcement learning (reward) and in RLHF.

What they leave out. Power, law and politics. Incentives assume people can choose to walk away; many cannot (workers, citizens).

Check yourself.

Why does Jordan say econometrics “doesn’t have much mechanism in it”?
What does he mean by “almost no statistics” in algorithmic game theory?
Name one thing each of the three fields can’t do alone, in his account.

3. ChatGPT, agreeableness and real uncertainty

🎙 In the talk.

“It’s not doing reasoning and/or uncertainty in any way that I think we’d recognize.” [12:02]
“What it’s doing is being very agreeable.” [12:02] A doctor who changes the diagnosis whenever you push back would be useless.
Evidence: a plot of ChatGPT’s stated confidence. It “often a saying a zero or a one, or sometimes it says, I can’t make up my mind at all. 50/50.” Humans, “you want to be on this line. It’s not even close” [12:02]–[12:34].
The mechanism he proposes: “in the past, some human was asked, how sure are you about that, and the human said something sensible. So it’s kind of copying that forward. That’s OK. That’s prediction, but that’s not reasoning or uncertainty.” [12:34]
He wants “real uncertainty”, not “a little bit of error bars like in a statistical setting” [10:52], and gives the finance example of a quant firm: “if you don’t really take uncertainty seriously, you lose.”

📚 Deeper theory.

Calibration is measurable. Early research on LLMs (Kadavath et al. 2022, “Language Models (Mostly) Know What They Know”) found base models fairly calibrated when asked for probabilities, and that RLHF-tuned models tended to be less calibrated (OpenAI’s GPT-4 technical report, 2023, shows a calibration plot degrade after post-training). From memory; not checked. This fits Jordan’s observation and may explain the 0 / 1 / 50-50 pattern. His plot is from an unnamed study; I could not identify it.
Sycophancy is the technical name for his “very agreeable” point. Anthropic and others have published on it (Perez et al. 2022; Sharma et al. 2023, from memory). RLHF rewards answers that human raters like.
Aleatoric vs epistemic uncertainty (Der Kiureghian and Ditlevsen 2009): randomness in the world vs lack of knowledge. A system that predicts text copies the verbal form of uncertainty. Whether it has epistemic uncertainty internally is an open research question (probing and conformal methods).
Conformal prediction (Vovk; Angelopoulos and Bates’s tutorial) is one statistically valid approach to add uncertainty to a black box. It is of the same family as PPI (2.3, 2.8).
Knightian uncertainty (Frank Knight 1921): risk (known odds) vs uncertainty (unknown odds). “Real uncertainty” in Jordan’s sense is closer to Knight’s.

Where others push back.

Some labs report improved calibration in newer models and show models that can say “I don’t know” or use tools to check. The talk’s evidence is from a particular model and date (2024).
Cognitive science: humans themselves are often badly calibrated (overconfidence, Kahneman and Tversky). Jordan says people are not bad at uncertainty in daily life [13:13], which is a claim that differs from the heuristics-and-biases tradition (see deep dive 4).
“Agreeable” behaviour is partly a product design choice and could be changed.

What they leave out. Whether systems of models plus tools plus human review can be calibrated even if single models are not. That is close to his collective argument, but he does not make it here.

Check yourself.

Why is “copying forward a sensible-sounding answer” not the same as having uncertainty?
What does his doctor example show about embedding such systems in decisions?
What is the difference between error bars and “real uncertainty” in his usage?

4. Collectives, ducks and probability matching

🎙 In the talk.

Humans and uncertainty: “The vast uncertainty that all of us live in … But nonetheless, I get up every morning and I live.” [13:13]
“we create science and culture because we’re very good at collaborating relative to mitigating uncertainty” [13:13]; “coming together in collectives … mitigates uncertainty” [13:53].
The setup: the duck “has figured out that 2/3 of the food can be found on this side of the lake and 1/3 over here” [13:53]. The Bayesian optimal decision is “to go over here with probability one”. Ducks and humans instead “hedge their bets”, “probability matching”, “viewed as suboptimal” [14:32].
Failed explanations: “ducks are stupid”, “humans are stupid too”, or exploration vs exploitation, which is “too big to be explained by those phenomena” [15:08].
The answer, from the audience: “There’s other ducks.” [15:43] Under a Nash equilibrium each duck goes “with probably 2/3 and probably 1/3”, and “more food gets eaten than any other algorithm” [16:21].
Decentralised does as well as central allocation: “you could design an algorithm where you tell ducks where to go … but the randomized— decentralized algorithms works just as well.” [16:21]
Lesson: “We evolved in worlds where we’re collective, and we should be living in worlds where collective thinking is actually favorable.” [16:21]
The networking turn: connecting people gives more data and better predictions, “But we also know this leads to social network chaos.” [16:55] Then: “why are not people— what prevents them from lying? Nothing prevents them from lying.” [17:29] “We’ve got to start to bring microeconomics into this picture. It’s not just a networking problem.”

📚 Deeper theory.

Why probability matching can be rational in a group. If every duck went to the 2/3 side, that side would be overcrowded and each duck would get less than on the thin side. The equilibrium puts the per-duck payoff equal on both sides. This is the ideal free distribution (2.6). In game terms, it is a congestion game, and in a congestion game the Nash equilibrium is typically a mixed or proportional split.
A decentralised mixed strategy matches a central plan. With the randomised rule, the group distribution is right on average, but individual ducks can still end up unlucky. A planner could do better for the group by assigning ducks and avoiding random clumping. Jordan says the decentralised rule “works just as well”; in expectation this is true for the split but not for variance. From memory / my reasoning; not checked against the experiments.
What this says about human “irrationality”. The heuristics-and-biases tradition (Kahneman, Tversky) treats deviation from single-agent expected-utility maximisation as bias. The ecological rationality programme (Gigerenzer; Simon’s “bounded rationality”) argues behaviour should be judged against the environment it fits. Jordan’s version adds that the environment includes other agents.
The scale of the claim. The duck result shows that an apparent individual error is a collective optimum. It does not show that collective decisions are always better. Wisdom of crowds fails when judgments are correlated or when information cascades (Bikhchandani, Hirshleifer, Welch 1992). That is the “social network chaos” he mentions in passing.
The lying problem is mechanism design. The revelation principle (Myerson, Maskin, Hurwicz) asks how to make truth-telling the best strategy. Jordan’s wish for a market where “nothing prevents them from lying” is closed by contracts and scoring rules (see deep dive 8 and the e-value result).
Network and social-media research (Watts, Centola, Bak-Coleman et al.) studies how information flows and goes wrong; Jordan’s “all the wrong people are connecting” is a short form.

Where others push back.

A single explanation for probability matching is hard to prove; learning-rule explanations (reinforcement, win-stay/lose-shift) also fit the data.
People often match probabilities even when alone in a lab, with no other agents present. The “other ducks” story explains why it might have evolved, not why it persists alone. Jordan uses evolution as the answer: “even if we’re not explicitly thinking about it, that’s how we evolved” [15:43].
Critics of collective intelligence worry about herding and polarisation.

What they leave out. Hierarchy and power within collectives; who gets excluded from them.

Check yourself.

Why would “always go to the 2/3 side” be worse for the group of ducks than matching?
What does the duck example show about the claim that “humans are not Bayesian”?
Why does he say connection alone creates “social network chaos”?

5. Against the “thing on the hill”, for markets

🎙 In the talk.

The fever dream: “there’s going to be this thing on the hill, which is the super entity. It’s like a search engine plus plus, and it will soak up all of humans data.” [17:29]–[18:08] “If it weren’t sort of so crazy, I wouldn’t even put a slide about it.” [18:08]
The replacement: “let’s embed in this whole network, all the computing … some of those entities are computers. Who cares? … it’ll be part of the collective.” [18:42]
The objection of historians: “markets are terrible things … Market’s destroying the world.” His reply: markets “have been around for longer than humans have been around … And they have problems and we have to think about that. But everything has problems. Everything that’s powerful has problems.” [19:19]
Three goods: equilibrium (“everybody gets a piece”), reduced uncertainty, stability [19:19]–[19:54].
The tomato market [19:54]: thousands of individual decisions “leads to a market, which brings food into the city every day”. The pizzeria: “because the market produces stability, that I can now build my pizzeria. And on top of that, other things can be built.” [20:34]
On regulation: “Oh my God, that scares me. We have to regulate it. How do we regulate? We don’t know.” [21:06] The economist’s rule: “you got to regulate the level of the equilibrium, not the level of the mechanism.” [21:06] Bring economists “in the room, at least have them be part of the discussion.”
“predictions are not enough. We start to have to use language out of economics; incentives and prices.” [21:37] “Prices are ways to decentralize computation.” [21:37]

📚 Deeper theory.

The “thing on the hill” is the central planner of the calculation debate (2.4). Hayek’s argument is that no planner can hold the dispersed, tacit, changing knowledge of millions; prices carry it. Jordan’s airline example [30:22]–[31:36] is the same point: “in our own heads, we don’t even know what to do until we think about it” [31:36].
Markets as computers. Modern work treats price systems as distributed algorithms (Arrow–Hurwicz gradient dynamics; “tatonnement”). Jordan, who works on optimisation, hears prices as the messages in a distributed optimisation. The link between dual variables in optimisation and prices is classical (Lagrange multipliers = shadow prices).
“Regulate the equilibrium, not the mechanism.” This is close to the logic of a Pigouvian tax or a cap-and-trade system: set the outcome-level lever (price or quantity) and let the market find the mechanism. It is also an echo of the Lucas critique (policy changes the equilibrium). He labels it as hearsay from economists. I did not verify a source.
Why historians object. Karl Polanyi’s The Great Transformation (1944) argued that treating land, labour and money as commodities destroys social fabric. Marxist and feminist critiques add unpaid labour. The more recent critique of platform capitalism (Srnicek 2016; Zuboff 2019) says today’s data “markets” are not competitive markets but extraction. Jordan’s quick reply (“they’ve been around for longer than humans”) is a rhetorical, not a philosophical, answer.
Stability enabling investment is a real topic in development economics (Hirschman, “backward and forward linkages”). Reliable supply chains are the basis of specialisation.
Designed markets (Al Roth: kidney exchange, school choice, medical residency; Nobel 2012) are the closest precedent for what he is proposing: matching and market design that is deliberately engineered with theory.

Where others push back.

Strong critique: markets price what can be priced and leave out what cannot (ecology, dignity, future generations). Market design can entrench the powers that run the market.
AI-specific critique: ordinary markets need property rights; “data” is non-rival and hard to own. Whether one can have markets in training data is exactly what the copyright fights are about.
A pragmatic critique: designing a market for each domain takes years and large firms may prefer not to adopt it.

What they leave out. Who sets the rules of the market (the “market maker”)? In his music example it is a company. Governance is not discussed. Also unaddressed: how a market handles non-participants who are still affected.

Check yourself.

Why does Jordan say a market “yields equilibrium” and why does he care that it is not just an optimum?
How does the tomato and pizzeria example link markets to prediction?
What does “regulate the equilibrium, not the mechanism” mean, and what would you need to know to do it?

6. The UnitedMasters music market and unraveling

🎙 In the talk.

“I’m in a company that has designed a three-way market that’s now thriving. It is a market for music. It’s an alternative to the record companies” [21:37]. It “aims to serve not the famous big superstar musicians … regular musicians” [21:37].
Three sides: musicians, listeners, and brands. Listeners “don’t get any money for it. It aims to help them make some money” [22:15] (the captions seem to mean the musicians). Stoute “signed up the NBA so that the music coming from these musicians is the sole source of music that’s being played on NBA website now” [22:48].
Brands can see this: “Another brand can look at that and say, I like that demographic associated with that company. I’m going to reach out to that musician” [22:48].
“it’s not enough just to make predictions” [23:23]: “under the hood … a pretty sophisticated … AI engine that has market mechanisms together with statistics.”
Why three sides: “If you didn’t have a three-way market, if it was just two, this would fall apart. It would unravel.” [23:23] Spotify “take music from people, they stream it to listeners, and then they make money by selling subscriptions”, and “They are incentivized to get rid of these people altogether. If they could replace them with generative AI, it would be in their interest to do so.” [24:00]
Q&A [1:00:00]: “it’s called unraveling” and links to “Market For ‘Lemons’“. “our more endogenous organic model where the only money coming in is via the willingness of the brands to pay for a product that is in their interest, then the market does not unravel.”
Consumers: “Individual users don’t pay. Now, I would love if individuals consumers do pay. I would pay for … YouTube. Sadly, the Google and Facebook and all people decided that everything should be free. And I think that’s been very distortionary.” [1:00:41]
The aim: “I want to find more mechanisms like that for other domains.” [1:01:20]

📚 Deeper theory.

Multi-sided platforms: Rochet and Tirole (2003), Armstrong (2006): platforms must get several sides on board; pricing on one side subsidises another; a side may be priced at zero. Jordan’s market is a design where the brand side pays.
Unraveling in the lemons model (Akerlof 1970): when buyers can’t judge quality, they pay the average, high-quality sellers leave, the average falls, and so on. His claim is that a Spotify-style model, funded by subscriptions or ads, unravels in his model if the platform can substitute generated music. He says his group wrote the analysis [1:00:00]; I did not find the paper. Mark: unverified.
Advertising model critique: “if you’re not paying, you are the product” (a slogan). Advertising makes the user’s attention the good being sold (see the Citton page on attention). Jordan’s complaint is about price distortion: with free access, no market signal says what people value.
Direct-to-artist platforms: Bandcamp, Patreon, Substack are other attempts at tying payment to value. Whether the brand-funded model scales or just moves the dependence to brands (and their taste) is an open question.
Competing with generative AI. Spotify’s reported use of low-cost “ambient” tracks and the arrival of AI-generated music services (Suno, Udio, 2024) make his claim that platforms are “incentivized to get rid of” artists concrete. Streaming fraud and AI tracks are live issues in 2025–26. From memory; not checked.
Relation to Part 1: the book-end of the talk. He built the product after the theory, which he uses as the proof of concept.

Where others push back.

It is a commercial venture and he is an involved party (“I’m in a company”); the performance claims (“thriving”, revenue, “scaring the record companies”) come from him and are not independently verified here.
Brand-sponsored music may shape what music gets made (toward brand-safe, demographic-targeted styles).
Creators in the AI-rights debate argue for licensing and opt-in rules rather than a new market. These are complementary, not opposed.

What they leave out. Numbers: how many artists, what they earn, how the split is set, who owns the masters. Also what happens to musicians who don’t fit any brand’s demographic.

Check yourself.

Who are the three sides, and who pays whom?
Why would Spotify want to replace the artists, in his account, and why would the three-way market not have that incentive?
What is “unraveling” and how does the lemons paper explain it?

🎙 In the talk.

The structure: users supply data to platforms and get a service back; “More data arrives at this company, they can build a better service and we’re happy” [25:14]. A company like Mastercard “does not make enough money off of this service to stay in business”, so it sells data “to a third party data buyer” who wants “market research” [25:14]–[25:48].
The question: “are all the players actually incentivized to be part of this game? … they lose their privacy.” [25:48]
Two options he rejects or notes: government-mandated privacy “is going to break this market”; or platforms set privacy themselves. “Differential privacy is a mechanism where you add a little bit of noise to data in a controlled way.” [26:25] “You decide.”
The interlocked incentives: users go to the more private platform; buyers prefer less noise and pay more for it; but the platform with more users has a better service and more data [26:25]–[27:00]. “It’s not clear what’s going to happen here.”
Method: “calculate the equilibrium of this system” [27:00], with data “valued under a statistical model” [27:32]. “this is exactly a blend of statistics and economics.”
The Stackelberg order: “First, platforms decide how much privacy they want to offer. Users decide whether to opt in or not. Platforms decide a price to charge, and then buyers decide from which platforms to purchase data” [34:27]–[35:00].
Results: there are two thresholds of beta; above one, “all platforms enter the market”; below, “only certain platforms will enter the market, what we call low-cost platforms” [35:00].
GDPR: “The big firms are not being hurt by GDPR. The small to medium enterprises are being hurt” [35:40]. The full ban “only maximizes user utility in a certain setting where all the platforms are low cost” and “a non-uniform privacy mandate actually increases user utility” [35:40]–[36:20].
To regulators: “that’s the kind of thing I want to show to a regulator.” [34:27] “I’ve gone to the government in France and talked to them and they like this language.” [36:20]

📚 Deeper theory.

What the model is. A multi-stage game: platforms choose privacy level; users opt in; platforms set prices; buyers purchase. Solved backward (subgame-perfect equilibrium). The “statistical” part is that the value of data to the platform and buyer is derived from an estimation problem (linear regression in the paper, per [33:15]), so more noise lowers accuracy in a computable way.
Differential privacy as a price/quality dial. The epsilon parameter trades privacy against statistical accuracy, so it can be put in a utility function. This is why DP fits economics better than “anonymisation”, which has no dial.
Why a mandate “breaks” the market. If every platform must use the same noise, platforms that depend on selling data (high-cost, small) can’t differentiate or may exit; large platforms that earn elsewhere (low-cost, “Google like things where they don’t need this market”) stay. That is consistent with the GDPR-and-concentration debate. His theory gives a formal route to that intuition (from the claims; I did not read the paper).
Empirical literature on GDPR. Studies (for example on app developers, venture funding, web technology vendors) find reduced entry or funding in the EU after 2018, while others find modest effects. The truth is mixed. From memory; not checked.
Data markets in theory. Bergemann and Bonatti on selling information; Acemoglu, Makhdoumi, Malekian and Ozdaglar on “too much data”: data externalities, where one person’s data reveals others’ (genetics, social graphs), so individuals sell privacy too cheaply. This is a standard objection to treating privacy as a personal market good. From memory.
“Regulator-facing diagrams.” The plot with two beta thresholds is policy analysis: a map from parameters to regimes (good/bad equilibria). Jordan’s point is that economists in government can read such maps.

Where others push back.

The externality problem above; a model with individual opt-in can be wrong when data about me reveals data about you.
Rights-based privacy advocates reject putting a price on privacy.
Whether Mastercard and similar firms do sell data this way is an empirical question he treats as given (“that kind of exists”).
He is a critic of GDPR here; others (and the European Commission) defend it, and the later European “digital omnibus” proposals show the debate is live. From memory; not checked.

What they leave out. Users’ knowledge: the model assumes users can judge what a given noise level means for them, which empirical privacy research doubts. Also the cost of exiting: in practice, “opt out” often means no service.

Check yourself.

What are the four decisions in the Stackelberg order, and who makes each?
Why might a privacy mandate help big firms more than small ones?
What does beta represent, and why is it the number he wants to show a regulator?

8. Contract theory, airlines and statistical contracts

🎙 In the talk.

The map: game theory is the “forward discipline” [28:13]: “write down the rules of some game, have players come and play”. Mechanism design goes backwards: “I’d like a certain outcome like a certain building stands up. What equations do I design to achieve that outcome?” [28:44]
“Nash equilibria are ones in which you have a symmetric game”; “the inverse of that is an auction” [29:17]. Stackelberg: “agents who have different knowledge and they’re not symmetric. And one of them often goes first and the other one follows.” The inverse: “contract theory” [29:50].
His own conversion: “I learned about this about three years ago by reading a book about it, and it’s changed everything we’ve been doing ever since.” [29:50]
The airline story [30:22]: “if John walks onto the airline tomorrow … they’ll offer him a price of $1,000”, Eli gets “$100”, but “the next day … John will arrive and he’ll be dressed like Eli. So they can’t make this discrimination.”
“that Sam Altman thing on the hill, knowing everything and knowing how to price everything … that doesn’t exist because in our own heads, we don’t even know what to do until we think about it.” [31:36]
The solution: “set up a set of contracts which are service price, service, price service price, and we give that list to everybody, the same list to everybody. Some people will pick that one, some people will pick this one.” [31:36] The result is “high revenue, but it’ll also lead to high social welfare because everyone found something they like.”
“weirdly, this mechanism is hardly studied at all outside of a very small branch of economics.” [32:09]
The statistical turn: “A good statistician would look at that problem and say, hey, can I actually learn the contract?” [32:43]
Roles: “the platform is a principal from the point of view of this interaction or this is an agent, but the platform is an agent from the point of view of this thing being a principal. So you can be both in these networks.” [32:43]

📚 Deeper theory.

Screening and self-selection. In a menu, each option is built so that customers with different willingness to pay prefer the option meant for them. Two conditions: participation (the customer’s utility is at least what they’d get by walking away) and incentive compatibility (they don’t prefer another row). This is Mussa and Rosen (1978) and Maskin and Riley (1984). The “distribution of willingness to pay” Jordan refers to [31:36] is the type distribution the designer knows only statistically.
What “learn the contract” means. If the type distribution is not known, the designer can learn it from choices, as in bandit problems (learning to price; “dynamic pricing” and “online contract design”). So the statistics is in estimating the distribution while trading. From memory.
Airlines as the origin. The airline case is the standard introduction to yield management after the US Airline Deregulation Act of 1978. His telling skips the nuance but captures the logic.
Why the principal–agent tag matters. The principal (designer) can’t observe the agent’s private information or effort (adverse selection and moral hazard). Contract theory is the toolkit for both. The Hart–Holmström Nobel (2016) was for this.
“Hardly studied outside a small branch” is an overstatement: contract theory is a core topic in economics and finance. It is rare in computer science and statistics, which is probably what he means. My reading.
Nash vs Stackelberg for regulation. Regulation is naturally Stackelberg: the regulator moves first and firms respond. Hence contract theory is the right toolkit for regulatory design (Laffont and Tirole, A Theory of Incentives in Procurement and Regulation, 1993, from memory).

Where others push back.

Real contracts are incomplete; parties can’t foresee all cases (Grossman, Hart, Moore). Optimal contracts in textbooks are often complex and rarely used.
Price discrimination via menus can still feel unfair to consumers; the welfare story depends on the setting.
Behavioural economists say people choose badly from menus (inattention, default effects).

What they leave out. Enforcement. A contract needs a court or a platform that can enforce it.

Check yourself.

Why did selling every seat at the same price fail for airlines, and why did “price by how someone is dressed” fail?
In a menu of (service, price) pairs, why does self-selection raise both revenue and welfare in his telling?
How can one entity be both a principal and an agent?

9. Statistical contracts for clinical trials (FDA, pharma, e-values)

🎙 In the talk.

The setting: trials are “causal inference experiments”, e.g. “10,000 people in the treatment, 10,000 people in the control randomly” [44:06]. But the hypotheses “are coming out of a pharmaceutical company that’s got a vested interest” [44:39].
The statistician’s guarantee: “if the drug is actually not effective … the probability of approval is less than 0.05 … if the drug is actually effective, the probability of approval is 0.8, the power.” [44:39]–[45:18]
Pharma’s view, case 1: “It costs 20 million to run the trial … if they were approved, they would make 200 million”. Not effective means “minus 10 million”, so they screen: “only send in candidates if you’re really convinced they’re a good drug.” [45:55]
Case 2: “20 million to run the trial and 2 billion if you’re approved. Ibuprofen or something.” Now “if the drug is actually not effective, the expected profit will be 80 million. And why? Well, because 0.05 is not a minuscule.” The CEO says “send every drug up that’s possible.” [46:30]
Consequence: “people do die not because the drug is hurting people, but because it could have been a better drug.” [47:00]
Fix: “this is a contract theory problem. There is a principle [principal]. That’s the FDA … And there is an agent who has private information.” [47:00]
“It needs to design not just a price, 20 million, it needs to design a contract … adaptively so when you put it out there and people actually use it, you control type I and type II error. So that latter statement is what’s new.” [47:32]
Mechanics: agent “opts in or opts out”, pays the reservation price, “pick a payout function from a menu”; then “you sample nature and you run the clinical trial” [48:05]. A confident firm “would have picked a favorable payout function”; a less confident one a “less risky choice” [48:38].
Result: “A contract is incentive aligned … if and only if all the payoff functions are what are called e values. E values are non-negative supermartingales.” [49:13] “If you want to design good contracts, they have to be collections of non-negative supermartingales.” [49:52]
Why it helps: it gives regulators “a vocabulary” and “we know how to design a little vocabulary here” [49:52].

📚 Deeper theory.

What the numbers show. Expected profit of a bad drug = (probability of approval) x (gain) minus trial cost. With alpha = 0.05: 0.05 x 200M - 20M = -10M. With 2B: 0.05 x 2000M - 20M = +80M. So the false-positive rate that is “small” in statistics is a price in economics. The numbers check out. This is a neat arithmetic point: the same alpha deters in one market and invites in another.
The paper. The work appears to be Bates, Jordan, Sklar and Soloff, “Principal-Agent Hypothesis Testing” (2022, arXiv), and a follow-up on e-value contracts. I am naming this from memory and did not check that these are the papers he describes.
E-values and why they are the natural currency. An e-value E satisfies E[E] ⇐ 1 under the null. A “bet” against the null that pays E for each unit staked is fair or worse for the bettor if the null is true. So a payout function that is an e-value is a bet that cannot make expected profit when the drug is useless. That is exactly the incentive-compatibility condition: bad drugs can’t expect profit; good drugs can. In the talk: “a certain expectation is less than or equal to one” [49:13]. Supermartingale gives the sequential version, which allows early stopping (Ville’s inequality). The connection between e-values and betting is the “game-theoretic probability” of Shafer and Vovk (2019, from memory).
Link to Neyman–Pearson. Rejecting when E >= 1/alpha controls type I error at alpha (Markov’s inequality). So e-values generalise the standard threshold test, and the contract is a menu of such tests with different stakes and rewards.
Pre-registration and the FDA reality. In actual FDA practice, the agency also sees the preclinical data, requires phase I–II evidence, and often requires two trials. The talk’s model is a stylised single-threshold decision. Jordan says as much: “in the simplest case version of this” [44:39].
The ethics. Jordan’s remark “is it unethical? I don’t know, but it’s what happens” [47:00] sets the model as descriptive of incentives, not of intent.

Where others push back.

The model assumes that a firm’s private information is about the probability that the drug works. In practice, firms also decide trial design, endpoints, and which results to publish (publication bias, selective reporting).
Regulators might not accept a menu: fairness between sponsors, legal procedure, and the need for transparent rules limit contract-like flexibility.
A pharma-focused critique: most drugs cost far more than 20M to test; real costs and gains differ, and the approval pipeline has several gates.

What they leave out. Safety (the model is about efficacy and error rates); patients’ own interests; competition among firms.

Check yourself.

Recompute the CEO’s expected profit for each case. Why does a small alpha not protect you in the big-profit case?
What does “incentive aligned” mean, and what is the role of e-values?
Why is “a statistical criterion for a contract” something that “you didn’t have … in classical contract theory”?

10. Prediction-powered inference: when AI-made data lie

🎙 In the talk.

“this is one where there’s not really a story about incentives. It’s a story about bias and very, very large bias.” [37:30] AlphaFold “is really, really highly accurate” and “a collective mechanism” [37:30].
The temptation: “there are still only about 100,000 sequences whose structure is known from lab work. On the other hand, AlphaFold can generate hundreds of millions” [38:12].
The test case: a two-by-two table (disorder yes/no vs phosphorylation yes/no); “calculate the odds ratio … and then you ask for the standard deviation” [39:21]. Lab data alone: the interval “covered the value one, which means there no association.” AlphaFold data: “200 million data points … the error bar is really tiny” and “did not cover the value one so they said, yeah, there’s an association … it’s bad statistics” [39:21]–[40:00].
The check: held out half the data; ground truth “a little over two” [40:00]. AlphaFold: “very narrow confidence interval and it’s not covering one, but it’s not covering the truth either … very sure of yourself and you’re very far wrong. You’re very biased.” [40:32]
The fix: “prediction-powered inference … a way you put to blend these two … It doesn’t throw away AlphaFold. … it also adds a little bit of extra data for the specific hypothesis that you have in mind” [41:06]. The green intervals “provably cover the truth” and are “almost always much smaller” than classical ones [41:06].
Diagnosis: AlphaFold “just missed the quantum fluctuation completely” [41:47]. “Scientists are not interested in the old data … They’re interested in brand new data” [42:19].
“if you add statistical reasoning to the top of it, you can actually get the best of both worlds.” [42:19]
On the last example: “It’s honest. It says, I can correct that, but I’m not going to over promise.” [42:54]
Framing: “AlphaFold knows a lot. It’s like the tower on the hill, but I got a particular problem and I know a little bit more about my particular problem. Can I put the two together?” [43:27]

📚 Deeper theory.

The basic estimator. Let f be a model. On N unlabelled points, the mean of f is an estimate (cheap, possibly biased). On n labelled points, the mean of (true y minus f) estimates the bias. Add the two: the corrected estimate is unbiased, and its variance is much smaller than the labelled-only estimate if f is accurate. This is what “rectifier” means in the PPI paper (from memory). The same idea sits in survey sampling as “model-assisted estimation” and in “control variates”.
Why it is honest. The interval widens automatically when the model’s errors are large. A bad model never gets you a wrong answer, only a wide one. That is why Jordan says the last example “is honest”.
Why bias is worst in discovery. Models perform best on cases like their training data. Research wants new cases. This is a distribution shift problem. The deeper point: confidence from model output scales with sample size N, but bias does not shrink with N. More synthetic data make you more certain of a wrong number.
Parallel with LLM-as-judge and synthetic respondents. The same problem arises when AI model outputs are used as survey answers or labels. The corrections are analogous. (This is the connection that matters for synthetic-respondent research. My comment, not his.)
Relation to the collectivist thesis. AlphaFold is built from a global archive of lab work, hence “a collective mechanism”. PPI is a rule for how a user of a collectively built model can combine it with her own small data without being misled: the “principal–agent” asymmetry between model-maker and model-user (Jordan frames it as information asymmetry, but not as an incentive problem, “there’s not really a story about incentives” [37:30]).

Where others push back.

PPI needs some labelled data, drawn at random from the same population as the unlabelled data. Often you only have convenience samples. Violating that assumption reintroduces bias.
Protein-structure specialists note that AlphaFold also reports per-residue confidence scores (pLDDT) and that predicted disorder correlates with low confidence. Whether the odds-ratio failure is a surprise to structural biologists, I can’t judge. From memory; not checked.

What they leave out. Costs of labelling and how many labels are enough. (The paper handles it; the talk skips it.)

Check yourself.

Why does the AlphaFold-only interval get narrower and more wrong as the sample grows?
What are the two data sources that PPI combines, and what does each contribute?
Why is the problem worst for new science?

11. Price discrimination and privacy: the shifted triangle

🎙 In the talk.

“A lot of people in government care about price discrimination. Should companies be allowed to discriminate in price or are we getting price gouging?” [50:27]
The Bergemann et al. result: “the resulting set of all possible utilities is a triangle in a certain space. This is a very important result. It may get a Nobel Prize itself.” [51:12]
His addition: privacy. The movie-theatre discount: “to get that discount, I’ve got to reveal my age. And so I lose a little privacy, but I get something in return.” [51:43] “If we reveal something and get nothing in return, who wants to do that?”
The result: “The classical triangle shifts down and shifts over a little bit. And it’s not even a triangle anymore. It’s a convex shape, and it actually is a projection of a high dimensional polytope.” [52:19] “consumer utility is bounded away from zero. You always get positive consumer utility at the expense of a little drop in producer utility. That’s maybe something government people will be happy with.” [52:19]

📚 Deeper theory.

The triangle. The axes are consumer surplus and producer surplus. The “no-information” point gives the monopoly outcome under uniform pricing. The perfect-discrimination corner gives the seller all surplus (consumer zero). The third corner: the seller gets only the uniform-price profit and consumers get the rest (efficient). Every point in the triangle can be reached by some segmentation. Consequence for policy: price discrimination can help consumers or hurt them, depending on the segmentation, so “allow or ban” is the wrong question; who controls the segmentation is the right one. From memory of the 2015 paper.
Privacy as a limit on segmentation. If the seller only sees a noisy signal about each buyer, they can’t perfectly segment, so the worst corner (perfect discrimination, zero consumer surplus) is cut off. That is the “bounded away from zero”. DP by design guarantees a floor for consumers. This is a neat link between a technical privacy property and a distributional outcome.
Policy reading. It turns privacy from an obstacle into a lever in competition policy: it caps how well firms can personalise prices. Whether regulators would accept it depends on political attitudes to personalised pricing (the EU, the FTC “surveillance pricing” inquiries in 2024, from memory).

Where others push back.

Price discrimination with consumer-friendly outcomes depends on the segmentation being chosen with the consumer’s interests in mind. In practice the seller chooses it.
DP noise lowers the accuracy of both personalised discounts and personalised markups. Whether consumers prefer that depends on their share of discounts vs markups.

What they leave out. Dynamic pricing and learning over time; competition between sellers.

Check yourself.

What are the corners of the Bergemann triangle?
How does adding privacy change the set of reachable outcomes?
Why is “bounded away from zero” attractive to a regulator?

12. Barriers to entry for LLM firms

🎙 In the talk.

“there’s a worry among many people, including government people, that it’s all going to reduce to one LLM.” [53:28] “are there also reasons why it wouldn’t reduce to a single LLM? And there are.” One is “reputation damage”: “If you’re a very big company and you have a glitch in your LLM, you’re going to suffer a big reputation hit. If you’re a tiny little company … no one notices.” [53:28]–[54:04]
Quantify it: “how does the probability of a glitch relate to the amount of data you have, you can talk about these scaling laws” [54:04].
The slide shown to the French government: a very large firm with “infinite data” and a small entrant. The gap in quality between them is the axis [54:36]. “If that gap is really, really small, then I need a lot of data to enter into this market, but not infinite.” And as the gap grows, “the amount of data I need to enter into this market gets to be very, very small.” [55:12] “So this is a very favorable result for market entry.”
“Behind this is some random matrix theory.” The officials asked “Is that Marchenko-Pastur?” [55:12]
Method point: “you can analyze the overall scaling law and put these in as economic entities” [55:47].

📚 Deeper theory.

Scaling laws: Kaplan et al. (2020), Hoffmann et al. (Chinchilla, 2022): loss falls as a power law in data and compute. In an economic model, this gives a cost function for quality.
Random matrix theory (Wigner, Marchenko–Pastur 1967): spectra of large random matrices. It explains double-descent and generalisation in high-dimensional regression, which is the toy model for scaling laws in theory work (from memory).
Barriers to entry in economics: fixed costs, network effects, learning by doing. The classic worry is a natural monopoly: if the average cost falls forever with scale, one firm wins. His result says: when the reputational cost of a glitch scales with size, bigger is not always better, and a small entrant can survive in a niche.
Competition policy angle. The EU, UK and US competition authorities all opened reviews of foundation-model markets in 2023–24 (from memory). A model that says “entry is feasible” is relevant to whether to break up or mandate access.
Is this plausible? In practice, the 2024–26 market shows many strong entrants (open-weights models, fast-follow labs), which is consistent with his claim of low barriers, though capital costs for frontier training are high. This is my observation; not a test of his model.

Where others push back.

Compute and capital are the main barriers, not data volume, and these are not in the slide as Jordan describes it.
Reputation damage may matter less than distribution (bundling with operating systems and cloud contracts).

What they leave out. Cost of compute; regulation; open-source dynamics; lock-in.

Check yourself.

Why would a bigger firm face more reputation damage from the same glitch?
What does the slide say happens to the data an entrant needs as the quality gap grows?
Why did the French officials ask about Marchenko–Pastur?

13. The engineering close, and the Q&A

🎙 In the talk.

“I think our era is an era where a new engineering field is emerging that is not about a superintelligence that is threatening us all, but rather about all of us being connected in new ways.” [55:47]
Possible benefits: “more money for musicians”, “new Democratic principles”, “better healthy world” [56:26].
The analogy: “There was chemistry, and then 20 years of effort led to chemical engineering where you could build factories that really worked. Electrical engineering. You had Maxwell’s equations, and then 20 years of effort led to electrical engineering.” [56:26] For AI: “we need about 20 years to start to build systems that actually, for good reasons, lead to human happiness.” [56:26]
“I think I am an engineer.” [57:05] Statistics emphasises “science and mathematics”, but “can’t we also design good engineering artifacts with statistics? … only if we start to bring in economics and put it all together.” [57:05]
Q1 (assumptions): “Most economic theory, you wrote down a bunch of rationality … parametric curves … We’re statisticians. We bring in data. And all of our curves are informed by the data.” [58:12] “Here, the rationality is just that are you incentivized to play the game or not? That’s it.” [58:46] And “behavioral economics is really a big part of this story” [59:23].
Q3 (France): “That’s absolutely one of the reasons I’m in France. I think the dialogue here in the United States is completely broken.” [1:01:53]
Q4 (operations research): “it’s totally true what you just said”; but “Where they got it wrong was they weren’t being statisticians” [1:02:31]. “Jack Kiefer, who was here, wrote one of the first papers on stochastic gradients” [1:03:07]. “AI is basically gradient algorithms that everybody knows, and the archive is full of every new last idea and trick on this. It’s not that you can have any country dominating it” [1:03:07]. “a lot of these fields kind of lost because they didn’t have statistics in play … And then statistics lost it because they didn’t really want to think about these big scale issues.” [1:03:41]
Q5 (more sides): “it does … We’re going to see brokers arising … they’ll be kind of new statistical agents.” [1:04:12]–[1:04:45]

📚 Deeper theory.

On “data inform the curves”. Replacing assumed utility functions with estimated ones is the structural econometrics and empirical industrial organisation project (Berry, Levinsohn, Pakes 1995). It reduces but does not remove assumptions: the functional form, the equilibrium concept, and the data’s coverage still matter. Jordan’s own gloss: “a little bit of a glib answer, but I actually believe it” [58:46].
Minimal rationality. “Are you incentivized to play” is the participation constraint. It does not require maximisation of anything else. But mechanism design also uses incentive compatibility, which is stronger. He is simplifying for the audience.
Unraveling vs lemons: see deep dive 6.
Who will dominate AI? His point that the archive is public echoes the “no moat” debate (2023 leaked Google memo, from memory) and also the counter-view that compute and data access create real advantages.
Brokers. Intermediaries that aggregate many small suppliers and users (data trusts, data cooperatives, collective bargaining for creators) are a live design space. The “new statistical agents” line suggests brokers that use statistics to price and match. Compare data unions and the proposals of Lanier and Weyl for a “mediator of individual data” (MIDs). From memory.
Twenty years is a rough rhetorical figure; nothing in the talk supports the number.

Where others push back.

The engineering-analogy view assumes AI will be dominated by a few well-understood principles. Critics say AI systems are not well-specified or well-understood enough for engineering in the chemical or electrical sense, and that some risks are not engineering questions at all.
Safety researchers worry about capability risks that he mostly sets aside (“superintelligence that is threatening us all”).

What they leave out. Safety/alignment as a technical field; military applications; labour displacement; the global South (his world is the US, France, EU regulators).

Check yourself.

In what sense does his rationality assumption differ from the textbook one?
What does he think went wrong for operations research and what for statistics?
Why does he expect brokers, and what makes them “statistical agents”?

Putting it together

Mental model. Treat AI as an economy, not a creature. The creature picture asks how smart the one big mind is. The economy picture asks who contributes data and knowledge, who gets paid or protected, what each party knows that the others do not, and whether the rules make it worth joining and honest to report. Statistics tells you what the data can and can’t say (and how wrong a model’s confidence is when it faces new cases). Economics tells you whether people will bother to take part and what stable states the system will settle into. Computer science supplies the machinery. Jordan’s programme is to write the rules of such systems as mathematics: menus of contracts, equilibria across platforms, payout functions that are e-values. You can then show a regulator a map of good and bad regimes and design the rules rather than guess.

Open tensions.

Incentives vs rights. He treats privacy, data and participation as things to be traded at the right price. Rights-based views say some things should not be tradable.
Collective vs concentrated. He describes a collective world, but real AI is built by a few firms. His market-entry result is a hopeful argument that this need not last; the evidence is mixed.
Descriptive vs normative. Equilibrium tells you what will happen, not what is good. “Everybody gets a piece” says nothing about how big the pieces are.
Where does the market-maker come from? UnitedMasters is a company. A designed market needs a designer; who governs the designer?
Safety. The talk puts the “superintelligence” story aside. Whether that is right is exactly what safety researchers dispute.
Talk vs paper. Several results are given as headline claims without detail (“we prove theorems”, “we wrote a whole unraveling” analysis). I haven’t read the papers.

Reading path.

Start: Jordan, “A Collectivist, Economic Perspective on AI” (arXiv 2507.06268); and his 2019 HDSR essays (from memory; not checked: “Artificial Intelligence — The Revolution Hasn’t Happened Yet”).
Then: Hayek, “The Use of Knowledge in Society” (1945); Akerlof, “The Market for ‘Lemons’” (1970); Bergemann, Brooks, Morris, “The Limits of Price Discrimination” (2015); Angelopoulos et al., prediction-powered inference (Science 2023).
Deeper: Laffont and Martimort, The Theory of Incentives; Dwork and Roth, The Algorithmic Foundations of Differential Privacy; Shafer and Vovk, Game-Theoretic Foundations for Probability and Finance; Al Roth, Who Gets What — and Why; Breiman, “Statistical Modeling: The Two Cultures”.
Counterweights: Zuboff, The Age of Surveillance Capitalism; Polanyi, The Great Transformation; Narayanan and Kapoor, “AI as Normal Technology” (all from memory; not checked).

Talks

Explorer

A Collectivist, Economic View of AI — Michael I. Jordan

Speakers

Summary

Map — talking points

Part 1 — Before you read

Key terms

People named (a reading list)

Jordan’s own frame

The current debate (2025–26)

Part 2 — The intellectual map

2.1 Three visions of intelligence — control, symbols, augmentation

2.2 Machine learning as engineering, and the two cultures

2.3 Statistics: evidence, partial observation, honest uncertainty

2.4 Economics I: markets, prices, equilibrium

2.5 Economics II: game theory, mechanism design and contract theory

2.6 Evolution, ecology and collective behaviour

2.7 Privacy, data markets and price discrimination

2.8 Bias of AI-generated data and prediction-powered inference

2.9 Engineering disciplines and “how fields are born”

Part 3 — Deep dives

1. Why the AI dialogue is “lost” — history and the reframe

2. Three ways of thinking, and what is missing

3. ChatGPT, agreeableness and real uncertainty

4. Collectives, ducks and probability matching

5. Against the “thing on the hill”, for markets

6. The UnitedMasters music market and unraveling

8. Contract theory, airlines and statistical contracts

9. Statistical contracts for clinical trials (FDA, pharma, e-values)

10. Prediction-powered inference: when AI-made data lie

11. Price discrimination and privacy: the shifted triangle

12. Barriers to entry for LLM firms

13. The engineering close, and the Q&A

Putting it together

Graph View

Table of Contents

Backlinks

Talks

Explorer

A Collectivist, Economic View of AI — Michael I. Jordan

Speakers

Summary

Map — talking points

Part 1 — Before you read

Key terms

People named (a reading list)

Jordan’s own frame

The current debate (2025–26)

Part 2 — The intellectual map

2.1 Three visions of intelligence — control, symbols, augmentation

2.2 Machine learning as engineering, and the two cultures

2.3 Statistics: evidence, partial observation, honest uncertainty

2.4 Economics I: markets, prices, equilibrium

2.5 Economics II: game theory, mechanism design and contract theory

2.6 Evolution, ecology and collective behaviour

2.7 Privacy, data markets and price discrimination

2.8 Bias of AI-generated data and prediction-powered inference

2.9 Engineering disciplines and “how fields are born”

Part 3 — Deep dives

1. Why the AI dialogue is “lost” — history and the reframe

2. Three ways of thinking, and what is missing

3. ChatGPT, agreeableness and real uncertainty

4. Collectives, ducks and probability matching

5. Against the “thing on the hill”, for markets

6. The UnitedMasters music market and unraveling

7. Three-layer data markets, differential privacy and GDPR

8. Contract theory, airlines and statistical contracts

9. Statistical contracts for clinical trials (FDA, pharma, e-values)

10. Prediction-powered inference: when AI-made data lie

11. Price discrimination and privacy: the shifted triangle

12. Barriers to entry for LLM firms

13. The engineering close, and the Q&A

Putting it together

Graph View

Table of Contents

Backlinks