Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria) | 20 May 2026 | Staff Favourites | Machine Learning Street Talk
"I think this anthropomorphizing of intelligence and understanding all that is not necessary not appropriate and is is a distraction for many many problems why say it understands i think it's science fiction" - Prof. Michael I. Jordan [00:22:57]
"super intelligent arrive soon so there's nothing left to do that's in your lifetime that is so demoralizing so demoralizing and that thing I think that bothers me the most i mean the second part that bothers me is there's no economic thinking going on there" - Prof. Michael I. Jordan [00:49:31]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"human culture creates abstractions individual humans create abstractions too that work for them and and when those abstractions are kind of useful enough and they can communicate and kind of get promoted into the culture and that flows up and down all the time" - Prof. Michael I. Jordan [00:41:57]
"just putting a super intelligence behind the wheel of a car dumb dumb way to think about technology" - Prof. Michael I. Jordan [00:58:10]
"the poor LLM has none of the above and um so what should it say when you ask how sure are you... it just mimics that those kind of assertions but that's not reasoning under uncertainty" - Prof. Michael I. Jordan [01:14:49]
"I don't think it's bad to build systems you don't understand but I think this level of detachment from reality is unusual for human history" - Prof. Michael I. Jordan [00:01:55]
Speakers & Credentials
Host (Tim Scarfe / Machine Learning Street Talk): Guides the highly technical, philosophical, and systemic discussion, framing questions around complex systems, historical tech narratives, and statistical methodologies.
Prof. Michael I. Jordan: Professor at UC Berkeley / Inria. Renowned as one of the world's leading figures in machine learning, statistics, and artificial intelligence. Recognized by Nature as the most influential computer scientist in the world. He approaches AI not through a biological or neuro-mimetic lens, but through statistics, economics, and contract theory.
1. Executive Summary
The prevalent Silicon Valley narrative surrounding Artificial General Intelligence (AGI) and autonomous superintelligence is a scientifically ungrounded distraction that demoralizes young engineers by falsely implying all major problems are solved or soon will be by recursive algorithms.
True intelligence is inherently collective, contextual, and economic; it emerges not from scaling gradient descent on static datasets, but from human aggregation, cultural abstraction, and the dynamic exchange of value across billions of agents.
Modern AI development is suffering from a catastrophic lack of economic thinking and systems engineering, often extracting massive amounts of data without providing localized value, transparency, or proper incentive structures to the original creators.
While tools like AlphaFold represent massive engineering triumphs (predicting 200 million protein structures), they still fail at uncertainty quantification at the edge of knowledge without external statistical frameworks like "Prediction-Powered Inference."
To transition from creating parlor-trick chatbots to foundational societal infrastructure, the field must adopt a "triangular" academic and engineering discipline: fusing computational modularity (Computer Science), robust uncertainty and error control (Statistics), and incentive/market design (Economics).
Future technological success relies on embracing mechanism design and contract theory—building transparent data markets and hybrid socio-technical ecosystems that respect information asymmetry and empower human creativity, rather than seeking to replace the human element entirely.
2. Chronological Table of Contents
[00:00:00] - The AGI Distraction and Demoralization of Young Engineers
[00:06:01] - A Collectivist, Economic Perspective on AI
[00:15:15] - Beyond the Soup: Predictable Systems vs. Mechanistic Interpretability
[00:20:25] - AlphaFold, Hypothesis Testing, and Prediction-Powered Inference
[00:22:51] - Anthropomorphizing AI: The Fallacy of Machine "Understanding"
[00:28:46] - Real-World Markets, Drug Discovery, and Asymmetric Information
[00:32:46] - The Three-Layer Data Market: Privacy, Equilibria, and Social Welfare
[00:45:43] - The Spotify Problem: Misaligned Incentives in AI Generation
[00:48:28] - Utopian Visions vs. The Historical Realities of Silicon Valley
[01:01:46] - Game Theory, Mechanism Design, and Contract Theory
[01:08:17] - The Liberal Arts of the Era: CS, Economics, and Statistics
[01:11:44] - The Statistician Duck: Contextual Uncertainty vs. LLM Mimicry
3. Detailed Thematic Summary
The Fallacy of AGI and the Silicon Valley Narrative
The term "AGI" is fundamentally a PR mechanism that distorts reality and creates a binary environment for 20-to-25-year-old engineers, forcing them to choose between irrational exuberance or existential alarmism [00:02:46].
The Silicon Valley ethos currently relies on brute-forcing gradient descent over exabytes of scraped data without returning value to creators—a methodology Jordan categorizes as lacking deep intellectual rigor and operating closer to science fiction than systems engineering [00:13:54].
The narrative of "recursive self-improvement" leading to a "superintelligence" taking over is explicitly dismissed. Jordan argues it is an unhelpful metaphor that completely ignores the microeconomic realities of how complex systems function in human society [00:55:13].
Intelligence as a Collective and Economic Phenomenon
Intelligence is inherently a social construct. Human intelligence is derived from the aggregation of millions of interactions, opinions, and cultural memories. A decision that appears intelligent in one context may be foolish in another [00:07:20].
Systems cannot be evaluated merely as statistical boxes mapping inputs to outputs; they must be viewed as ecosystems. Modern AI systems serve billions of users but fail to act as true markets because they do not process game-theoretic signals (cooperation, signaling, exploitation) that naturally exist among billions of human agents [00:08:30].
Instead of building a "secretary sitting on your shoulder," AI should facilitate large-scale human coordination. For instance, creating systems that optimize healthcare, transportation, and finance by connecting producers and consumers securely and efficiently [00:09:38].
Historical Context & Deep-Time Systems Evolution
The Origins of Markets (Millennia Ago): Jordan emphasizes that markets evolved thousands of years ago as decentralized, bottom-up systems to solve resource allocation before the concept of capitalism even existed. Markets mitigate human uncertainty (e.g., ensuring a stable supply of tomatoes for a pizzeria) by providing incentives for exploration and exploitation [00:10:19] and [01:16:19].
The 1950s AI Terminology Genesis: The very term "Artificial Intelligence" was coined by John McCarthy in the 1950s with specific, logic-based inference goals in mind. When those failed, the 1970s and 1980s birthed "machine learning" in operations research and statistics, which actually led to the industrial supply chain miracles we see today [00:03:37].
Historical Engineering Parallels: Just as early chemical engineers in the 1940s and 50s couldn't just "throw a lot of stuff together" without causing explosions, today's AI builders cannot simply mix billions of parameters without creating societal damage (e.g., algorithmic mental health crises among teens). Jordan notes a stark detachment from reality in the current AI narrative, which is historically unusual; legacy engineering disciplines were founded on rigorous, stabilizing equations (like Maxwell's or Newton's) to prevent catastrophic failures [00:12:12] and [00:14:29].
The Evolution of Human Autonomy: The implementation of autopilots in commercial aviation dramatically reduced plane crashes over the last few decades. This represents a historical shift toward human-AI hybrid autonomy that succeeds because the operating parameters (3D space) are mathematically clean, unlike the chaotic, 2D realities of autonomous driving where tens of thousands still die annually [00:56:54] and [00:57:50].
Systems Engineering, Statistics, and Uncertainty
While AlphaFold represents a massive leap forward by predicting over 200 million protein structures, these base predictions lack reliable confidence intervals when probed with novel scientific queries (like analyzing quantum fluctuations in proteins) [00:19:36].
To fix this, Jordan champions "Prediction-Powered Inference"—a statistical method where a small amount of gold-standard ground truth data is merged with massive, biased foundation model outputs to produce high-power, mathematically trustable bounds [00:20:29].
Current LLMs fail completely at uncertainty quantification. When an LLM expresses "confidence," it is not calculating epistemic uncertainty; it is merely predicting the next word based on internet text where a human expressed confidence. It lacks the multidimensional nuance of human uncertainty, which factors in provenance (data age) and information asymmetry [01:14:49].
The desire for machines to "understand" is fundamentally misplaced. Amazon's supply chain algorithm, routing billions of products to 100 million people a day, doesn't "understand" logistics in a human sense, but it optimizes effectively to allow human engineers to build robust infrastructure around it [00:23:27].
The Three-Layer Data Market and Mechanism Design
In the modern internet architecture, users trade data to a platform (Layer 1 to Layer 2) for a service. However, platforms often sell this data to third-party data buyers (Layer 3), breaking the user's privacy equilibrium [00:33:40].
By treating this as a Stackelberg game, systems can be designed where users select their preferred differential privacy levels (e.g., 0.3 vs. 0.7). While high privacy protects the user, it introduces noise that makes the data less valuable to the third-party buyer. Calculating the equilibrium allows engineers to maximize total social welfare quantitatively [00:35:12].
This shifts the paradigm from forward-looking science (predicting what happens) to inverse-problem engineering (Mechanism Design). By deciding the desired social outcome first, we can mathematically design the game (the contract or the auction) that ensures agents act in ways that produce that specific result [01:03:19].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Cyber Fund Capital
$2,000,000
The funding amount offered per team to AI native founders entering the "monastery" incubator.
Prediction-Powered Inference [00:20:41]
This statistical framework resolves a fundamental vulnerability in massive AI models: they confidently output biased answers at the edge of human knowledge. By taking the massive output of a foundation model (e.g., 200 million AlphaFold predictions) and mathematically tying it to a small, pristine dataset of ground-truth examples, researchers can shift the error bars. This hybrid approach ensures that the model covers the actual scientific truth while maintaining the narrow confidence intervals necessary for high-power hypothesis testing, transforming an untrustworthy black box into a rigorous scientific instrument.
The Three-Layer Data Market (Stackelberg Game) [00:33:12]
Traditional machine learning views data exchange as a two-player game: User to Platform. Jordan introduces the reality of the third layer: Platform to Data Buyer. Modeled as a Stackelberg game, this framework maps the conflicting incentives where users demand privacy (reducing data value), platforms want better services, and buyers want raw signal. By calculating the mathematical equilibria of these competing forces, engineers can map the "Pareto frontier," allowing regulators and technologists to optimize overall social welfare quantitatively rather than relying on blunt, reactive legislation.
Behavioral Organization [00:44:00]
An economic and sociological framework utilized to understand how human groups maintain, prune, and evolve knowledge over time. While individual strategies decay, organizations serve as an adaptive hard drive. Jordan leverages this to argue that AI systems should be engineered to augment the behavioral organization of humans—creating ecosystems where collaborative meaning is derived naturally—rather than attempting to build a singular, top-down omniscient model.
Mechanism Design and Contract Theory [01:03:19]
If Game Theory is the physics of economics (predicting how actors will behave in a given system), Mechanism Design is its engineering (designing the system to force a specific outcome). Contract Theory applies this to asymmetric information—when one party knows more than the other. In the AI context, this framework demands that instead of building an omniscient, centralized intelligence, we build the digital "contracts" and incentive structures that compel decentralized, self-interested agents to reveal truthful data and cooperate safely at global scale.
The Liberal Arts of the AI Era (The Triangular Foundation) [01:08:17]
A pedagogical and structural framework arguing that pure computer science is insufficient for modern AI. The ideal architecture requires three equal pillars: (1) Computational Thinking (APIs, modularity, scale), (2) Inferential Thinking (Statistics, controlling uncertainty, p-values, e-values), and (3) Economic Thinking (Incentives, game theory, social welfare). Systems built without this trinity inevitably devolve into chaotic, ungrounded technical artifacts that destabilize human labor and privacy.
Anytime Inference (E-Values vs. P-Values) [01:05:55]
A departure from classical Fisherian statistics (p-values), which are highly vulnerable to "p-hacking" when data is sampled repeatedly over time. Anytime inference utilizes e-values (expectations of non-negative super martingales) to allow scientists and algorithms to continuously stream, peek at, and gather new data without breaking the statistical validity of the test. This continuous evidence-gathering mathematical structure is perfectly mapped to AI systems that live in dynamic, updating environments, directly linking game-theoretic probability to real-world incentives.
6. Anecdotes
The Amazon Supply Chain Epiphany [00:23:09]
Context: Jordan shares a story about visiting Amazon around the year 2000. He witnessed them using massive Random Forest neural networks to predict shipping delays in the Indian Ocean, orchestrating billions of products to 100 million people.
Purpose: He uses this to destroy the anthropomorphic "understanding" argument. No human could understand the routing math happening in that black box, and the system itself didn't "understand" the concept of a boat—but it was an immensely successful engineering tool that reduced real-world uncertainty.
Dick Fosbury and the Evolution of Abstraction [00:25:38]
Context: Jordan recalls his youth as a high jumper, noting that everyone used the "barrel roll" until Dick Fosbury tried jumping backwards (the Fosbury Flop), raising the Olympic bar by half a meter instantly.
Purpose: He uses this to explain that optimization and human intelligence are often driven by A/B testing ("try it out and see what works") rather than deep, top-down theoretical "understanding." It proves that the most powerful abstractions often emerge bottom-up through experimentation.
The Pharmaceutical Regulatory Game [00:28:25]
Context: Jordan details the asymmetric relationship between drug companies (who want to push drugs to market for profit) and regulatory agencies (who want to minimize false positives/negatives in society).
Purpose: He uses this as a prime example of why machine learning needs contract theory. If a regulator just passively looks at data, pharma companies will flood them with bad drugs hoping for a false positive. The system must actively structure incentives to ensure companies only submit high-confidence data.
The Airline Ticket Pricing Mechanism [00:29:48]
Context: A thousand people arrive at an airport wanting to go to London. The airline has no idea what any individual's true budget or desperation level is (information asymmetry).
Purpose: Jordan explains how airlines offer bracketed price tiers to force consumers to organically reveal their internal preferences. This exemplifies how we can build robust economic AI systems that function flawlessly despite massive ignorance about the individual actors inside them.
The Statistician Duck at the Lake [01:11:44]
Context: A hypothetical duck calculates that the left side of the lake has twice as much food as the right. A pure Bayesian duck would go left 100% of the time to maximize expected value. Real ducks go left 2/3 of the time and right 1/3.
Purpose: Jordan tells this to illustrate contextual uncertainty and Nash Equilibria. The duck isn't making a math error; it is operating in a multi-agent system. If all ducks went left, the resource would deplete instantly. True intelligence hedges based on the existence of a collective population.
7. References & Recommendations
People & Thinkers
John McCarthy: Computer scientist who coined the term "Artificial Intelligence" in the 1950s. Cited to show how the original pursuit of AI was rooted in logic and symbolic inference, which largely failed before machine learning took over. [00:03:37]
Hubert Dreyfus: Philosopher. Referenced by the host regarding the "first step fallacy"—the human illusion that because a machine can do one amazing thing, it is only one step away from doing everything. [00:08:52]
David Deutsch: Physicist. Mentioned by the host to contrast high-level physics abstractions with the abstractions generated by economics. [00:11:42]
Ilya Sutskever: Co-founder of OpenAI. Mentioned by the host to contrast the Silicon Valley "multi-agent LLM" approach with Jordan's formal economic models. [00:11:54]
John Jumper: AlphaFold lead at DeepMind. Referenced by the host regarding his reluctance to say AlphaFold "understands" biology, which Jordan heavily agrees with. [00:18:10]
David Krakauer: Complexity scientist. Brought up by the host to discuss intelligence as the adaptation and synthesis of coarse-grained representations. [00:24:40]
Geoffrey Hinton & Stuart Russell: Prominent AI figures. Mentioned as voices propagating the existential risk / superintelligence narratives that Jordan finds demoralizing and unscientific. [00:48:28]
Sam Altman: CEO of OpenAI. Jordan points to him as an example of someone skimming the top off decades of open-source research without clear social utility. [00:51:25]
John von Neumann: Mathematician. Referenced as the originator of Game Theory in the 1920s, laying the groundwork for mathematical predictions of strategic interaction. [01:01:00]
Vladimir Vovk: Professor and co-inventor of conformal prediction. Praised by Jordan for his foundational work connecting game-theoretic probability to statistical evidence gathering. [01:04:58]
Phil Dawid & David Blackwell: Esteemed statisticians. Listed by Jordan to contextualize Vovk's lineage in the rigorous application of probability and decision theory beyond standard parameters. [01:05:29]
R.A. Fisher: Statistician. Referenced to define the origins of classical p-values and tail probabilities that break down under repeated sampling (p-hacking). [01:05:38]
Jeannette Wing: Computer Scientist. Acknowledged by Jordan for her seminal paper on "Computational Thinking," promoting computer science abstractions across all disciplines. [01:08:31]
Companies & Organizations
Cyber Fund: Investment firm. Mentioned in the intro ad read, offering $2M to AI native founders in an incubator called "the monastery." [00:02:11]
Airbnb: Technology platform. Used as an example of human choices that are fundamentally inexplicable but predictably modeled through economic options. [00:15:31]
Anthropic: AI research company. Jordan praises them explicitly for pioneering the practice of paying people for their data, signaling a move toward healthy data markets. [00:31:53]
Mastercard: Financial services company. Discussed as an organization that legitimately needs to sell behavioral data for market research, illustrating the real-world three-layer data market. [00:34:24]
Meta / Facebook: Technology conglomerate. Criticized directly for creating structures that damage mental health and operate without returning direct economic value to users. [00:45:04]
Spotify: Audio streaming platform. Used as a cautionary tale of misaligned incentives, where the platform's monopoly dynamics currently incentivize them to generate cheap AI music to avoid paying human artists. [00:45:43]
UnitedMasters: A music distribution company for independent artists. Jordan mentions he is a scientific advisor here, helping build systems where artists retain ownership and bypass monopolistic platforms. [00:46:02]
Google & YouTube: Tech giants. Jordan criticizes their original business model shift toward third-party advertising, arguing they missed a historical opportunity to build a direct, incentivized producer-to-consumer micro-economy. [00:47:38]
Historical & Scientific Concepts
Maxwell's and Newton's Equations: Foundational laws of physics. Jordan uses them to show that true engineering disciplines require mathematical guardrails and predictable models, unlike current neural network brute-forcing. [00:14:29]
Conformal Prediction: A statistical method for determining confidence intervals. Discussed as a vital tool for the future of AI, moving away from black-box guessing to mathematically guaranteed bounds. [01:04:51]
8. The Bottomline (by AI)
The hyper-fixation on an autonomous, "brain-like" AGI is a dangerous distraction that obscures the actual path to technological value: integrating AI into complex human economic systems. To prevent societal destabilization and monopolistic extraction, the AI industry must immediately pivot from pure computer science scaling to mechanism design—building transparent data markets, utilizing prediction-powered statistics, and aligning incentives mathematically. Watch for the emergence of "Contract Theory" applications in AI architectures and a fracturing between companies focused on synthetic superintelligence versus those building decentralized, human-in-the-loop microeconomic infrastructure.
"Brookfield's the largest infrastructure owner in the world... We drew a pipeline and we showed all the different components of the payments ecosystem on a pipeline and said it's like a pipe that moves any commodity except what it's moving…
Consumer Scale
100,000,000
Daily users served by Amazon's early predictive supply chain models.