Intelligence Is Collective, Not Artificial
written by Stefan Christoph
A Run, a Podcast, and a Different Perspective
Long weekend, public holiday, sunshine. I went for a run and put on Machine Learning Street Talk, a podcast I have been following for a while now. The episode: “Intelligence Is Collective, Not Artificial” with Professor Michael I. Jordan [1]. He recently published a paper formalizing these ideas as “A Collectivist Economic Perspective on AI” [3].
I expected a technical deep dive. What I got was a fundamental challenge to how our entire industry thinks about AI. Not from a critic or a philosopher, but from someone who helped build the mathematical foundations we all stand on.
It was exactly the kind of perspective shift you need when you spend most of your time deep in the weeds of agent architectures, model selection, and customer deployments. A reminder to zoom out.
Who Is Michael I. Jordan?
This is not a random contrarian. Michael Irwin Jordan is the Pehong Chen Distinguished Professor Emeritus at UC Berkeley, with a joint appointment across Electrical Engineering and Computer Science, and Statistics. In 2016, Science magazine named him the world’s most influential computer scientist [2]. In 2022, he won the inaugural World Laureates Association Prize [5] “for fundamental contributions to the foundations of machine learning and its application.”
His academic path is unusual: a Bachelor’s in Psychology from Louisiana State, a Master’s in Mathematics from Arizona State, and a PhD in Cognitive Science from UC San Diego (1985), where he studied under David Rumelhart in the Parallel Distributed Processing group. He was a professor at MIT’s Department of Brain and Cognitive Sciences from 1988 to 1998 before moving to Berkeley.
His research spans Bayesian nonparametric analysis, probabilistic graphical models, variational inference, and mechanism design. He is a member of the National Academy of Engineering, the National Academy of Sciences, and the American Academy of Arts and Sciences. He holds honorary doctorates from Yale and multiple other institutions [4].
The point: when Jordan says the AI framing is wrong, he is not speaking from the outside. He helped create the field. He just never called it AI.
The Core Argument: Equilibria, Not Optimization
Jordan’s thesis is deceptively simple: most of what we call AI is optimization. Minimize a loss function. Find the best single answer. One agent, one objective.
Optimization, in the sense Jordan contrasts here, means a single decision-maker trying to find the best possible outcome according to one metric. Think of it like a GPS navigation system: given a destination, find the shortest route. There is one objective (minimize travel time), one agent (the driver), and a clear “best answer.” In AI, this is how we train models (minimize prediction error), how we tune prompts (maximize accuracy), and how we design most agent workflows (maximize task completion rate).
Example: you build a customer service agent. You optimize for resolution time. The agent learns to close tickets fast. It works, for that single metric.
Equilibrium means you have multiple decision-makers, each pursuing their own interests. A state where no one can unilaterally improve their position is an equilibrium, whether the system actually finds and settles into one is a separate question. There is no single “best answer”, there is a stable configuration that can emerge from competing forces. In economics, this is the textbook competitive-market model: buyers want low prices, sellers want high prices, and, under assumptions like price-taking participants and a well-defined good, the market price is the equilibrium where supply meets demand.
Example: you deploy that same customer service agent, but now the customer wants thorough answers (not fast ones), the support team wants manageable workloads, and the business wants low cost. Optimizing for resolution time alone makes customers unhappy and creates repeat tickets. The real challenge is finding a stable balance between all three parties, an equilibrium, not an optimum.
| Dimension | Optimization | Equilibrium |
|---|---|---|
| Agents | One decision-maker | Multiple with competing interests |
| Objective | Single metric to minimize/maximize | Stable state where no party can unilaterally improve |
| Assumption | There is a “best answer” | There is a “sustainable balance” |
| Typical toolkit | Gradient descent, search, RL | Game theory, mechanism design, market design, often computed with those same algorithms |
| Failure mode | Goodhart’s Law (metric gaming) | Incentive misalignment (one party exploits others) |
| AI example | Train a model, tune a prompt | Design a multi-agent marketplace |
| Real-world analogy | GPS finding shortest route | Traffic flow settling into a rush-hour pattern, Wardrop’s user equilibrium, the traffic-specific cousin of the Nash equilibrium [6] |
| When it breaks | When other agents react to your optimization | When you assume cooperation but get competition |
Optimization solves for one agent’s best answer. Equilibrium analysis asks where multiple competing agents settle, a different problem formulation, often computed with the same optimization machinery.
“My machine learning colleagues don’t know much about fixpoint algorithms and finding equilibria. They’re really good at optimization, but this is not an optimization problem.”
He illustrates this with a three-layer data market: users send data to platforms, platforms provide services, third parties buy aggregated data. When the third layer appears, the equilibrium shifts. Users lost privacy. The system needs rebalancing. You can model how equilibria move as parameters change (privacy budgets, pricing, regulation) and find which equilibrium maximizes social welfare.
This is not the same as “multi-agent systems” in the way our industry currently uses the term. Most multi-agent architectures today are still optimization-centric: an orchestrator assigns tasks, workers execute, results aggregate. That is a command economy. Jordan’s collective intelligence is a market economy: agents have competing interests, information asymmetries, and the system reaches stability through mechanism design, not top-down orchestration.
This is not abstract theory. It is mechanism design [7]: instead of asking “given this game, what happens?” you ask “what game do I design so that a desired outcome is realized?” Contract theory, auction theory, market design. A rich mathematical tradition that mainstream ML practice has mostly left on the shelf, the economics-and-computation research community aside.
The Business Model Critique
Jordan does not hold back:
“A secretary sitting on your shoulder whispering things to you. It’s just a dumb business model. I don’t think many people really will want that. They’ll turn the damn thing off.”
His alternative: healthcare systems, transportation networks, financial markets. Systems with billions of interacting agents where the real value lies in coordination, not individual prediction. Systems that already use machine learning but are “ripe for thinking in a more economic way.”
The current approach, he argues, is “big statistical boxes that do inputs and outputs.” That is not systems thinking. Real engineering asks: what ecosystem does this belong to? Who is interacting? What values are being created? What jobs emerge?
The Human Cost of Hype
The part that stuck with me most was his concern for young researchers:
“25 and 20 year olds are watching [thought leaders] and saying: am I going to be exuberant or alarmist? Those are the two choices. And I hope this conversation makes it clear to young people that there are other ways to approach life and technology.”
He calls AGI a “PR term” that confuses young people. The exuberant voices say “we solved intelligence.” The alarmist voices say “it’s dangerous, nothing left to do.” Both are demoralizing for people who want to build things that help their families and communities.
“This level of detachment from reality is unusual for human history.”
Where I Agree, and Where I Keep Building
Jordan’s critique is sharp and, I think, largely correct. The industry does over-index on individual model capabilities and under-invest in systems thinking. The economic dimension, who creates value, who captures it, how incentives align, is genuinely underdeveloped.
But here is where I land differently: there is still enormous practical value in what this technology can do today. I work with enterprise customers every day who are generating real business outcomes with AI. Not by chasing AGI, but by solving specific problems: automating document processing, accelerating code reviews, building intelligent search systems, creating agent workflows that save hours of manual work.
The technology is real. The value is real. What Jordan is arguing, and I agree, is that we need better frameworks for thinking about how these systems interact with humans and with each other at scale. Optimization gets you a good model. Equilibrium thinking gets you a good system.
I keep investing in understanding what I can do with this technology, and I remain positive that my customers can generate significant value from it. But Jordan’s perspective is a healthy corrective: better understanding of the technology, its limitations, and its systemic effects is a good and important goal. Not every problem is a loss function to minimize.
The Tomato Market Analogy
Markets reduce uncertainty. A stable tomato supply lets you build a pizza restaurant. A stable agent marketplace lets you build reliable AI systems.
Jordan ends with a beautiful example. If you run a pizza restaurant and have to forage for tomatoes every day, your uncertainty about having pizza tonight is high. But because a market exists where someone else did the foraging, there is a stable supply. Your uncertainty dropped. You can build on top of that stability.
Markets mitigate uncertainty, Jordan argues, less through optimal experiment design or multi-armed bandits than through incentive structures that encourage exploration and exploitation at scale.
The parallel to AI systems is clear: instead of building one omniscient model, build markets where specialized agents trade capabilities, where data creators get compensated, where the collective intelligence of the system exceeds any individual component.
The analogy has limits. Physical markets had centuries of institutional infrastructure to develop. But digital markets (ad exchanges, cloud spot pricing, API marketplaces) emerged in years, not centuries. The infrastructure for AI agent markets is being built now: x402 micropayment protocols, agent registries, tool marketplaces. The trajectory is there, even if the destination is not.
That is a vision worth building toward. And it does not require AGI. It requires economics, mechanism design, and systems thinking. Skills that most ML engineers have never studied.
What This Means for Practitioners
Jordan’s framework becomes tangible when you apply it to specific industries. The same three questions, who are the agents, is this an equilibrium problem, and what value flows back, reveal different challenges depending on where you sit.
Media and Entertainment: Broadcasters
A broadcaster building an AI-powered content recommendation engine typically optimizes for watch time. One metric, one objective. But the real system has competing agents: viewers want relevance without filter bubbles, content creators want fair exposure for their work, advertisers want attention, and regulators want diversity. Optimizing for watch time alone risks the engagement trap widely blamed for eroding trust in social platforms, one factor among the product and governance choices involved there. The equilibrium question: what recommendation mechanism balances viewer satisfaction, creator compensation, advertiser reach, and content diversity, without any party being able to game the system at others’ expense?
Media and Entertainment: Publishers
A publisher deploying AI agents to summarize or redistribute their content faces Jordan’s data market problem directly. The agents consume articles, platforms serve summaries to users, and no value flows back to the content creator. In Jordan’s three-layer market the third layer is third parties buying aggregated data; the publisher’s missing piece is its counterpart, a compensation flow back to the people whose work feeds the other layers. The equilibrium has shifted against the content creator. This is why micropayment protocols like x402 matter [8]: they reintroduce the economic feedback loop. The question is not “how do I optimize my paywall?” but “what market mechanism ensures my content retains value when consumed by agents?”
Advertising
Ad tech is already an equilibrium system, auction mechanisms, real-time bidding, supply and demand. But AI agents are changing the game. When an AI agent shops on behalf of a consumer, who does the ad target, the human or the agent? The agent may carry its owner’s encoded preferences, but it has no impulse purchases of its own, no emotional triggers, no brand loyalty. To the extent buying shifts to agents like that, the advertising equilibrium built on human psychology starts to break down. The question becomes: what new mechanism design allows advertisers to reach purchase intent when the decision-maker is an algorithm with a budget constraint?
Retail
A retailer optimizing pricing with AI typically minimizes margin loss or maximizes conversion. But customers, suppliers, and competitors are all agents in the system. Dynamic pricing that optimizes for the retailer can trigger price wars (competitors react), customer backlash (trust erodes), or supplier squeezes (margins shift upstream). The equilibrium question: what pricing mechanism creates sustainable margins while maintaining customer trust and supplier relationships? Amazon’s own flywheel is a self-reinforcing feedback loop, lower prices attract customers, which attracts sellers, which enables lower prices. That is designing a system around incentives, not optimizing a single metric.
If you are building AI systems in any of these domains, the pattern is the same: stop asking “what single metric do I optimize?” and start asking “what system do I design that is both stable and worth participating in for every party?” Stability and willing participation are separate properties, a design has to earn both.
This is a perspective piece, not a technical guide. The practical implications, mechanism design for agent systems, data market architectures, incentive-aware orchestration, deserve their own posts with concrete implementations. But the lens matters. You cannot build the right system if you are asking the wrong questions.
If You Are Running This on AWS
Jordan’s equilibrium thinking maps to concrete AWS services. Here is how the patterns from each industry example translate to what you can build today.
Multi-agent orchestration with competing objectives: Amazon Bedrock Agents and AgentCore let you deploy multiple specialized agents that interact through defined protocols. Instead of one monolithic agent optimized for a single metric, you can build agent ecosystems where each agent is a stakeholder’s proxy (a content quality agent, a cost optimization agent, a compliance agent) and orchestrate their interactions through Strands Agents or Step Functions. That is the deployment plumbing, the mechanism itself (each stakeholder’s objectives, the rules of interaction, how a stable outcome is computed) is a design layer you still have to supply.
Advertising: Agents for Amazon Ads Solution Guidance. Amazon Ads already operates as an equilibrium system (auction-based bidding). The Agents for Amazon Ads Solution Guidance [9] takes this further: AI agents that help advertisers work through campaign strategy, budget allocation, and audience targeting. The agent does not just optimize for one metric, it balances advertiser goals, audience relevance, and platform health. This is an agent operating inside a designed mechanism: it works within the rules of the advertising marketplace, not above it.
Publisher monetization: AgentCore Payments. Amazon Bedrock AgentCore Payments (Preview) supplies the compensation flow missing from the publisher example above, not Jordan’s third layer itself (third parties buying data), but the mechanism that lets value flow back to content creators. AI agents can make micropayments via the x402 protocol when consuming publisher content. Budget controls (per-session, per-agent, per-day caps) and cryptographic mandates enforce the spending rules: agents cannot overspend, and publishers get compensated per access. Whether the resulting market settles into a healthy equilibrium is a separate, economic question that spending controls alone don’t answer.
Retail pricing as equilibrium: Amazon Personalize and Amazon Bedrock can power pricing systems that account for multiple agents. Instead of optimizing price for conversion alone, you can model customer price sensitivity, competitor reactions (via market data feeds), and supplier margin constraints as inputs to a pricing equilibrium, the services supply the prediction and data plumbing, while the equilibrium model itself (whose reactions you assume, and how a stable price is computed) is the economics you bring. The flywheel is the architecture pattern: lower prices → more customers → more sellers → lower prices.
Data markets and privacy: AWS Clean Rooms lets multiple parties collaborate on data analysis without exposing raw data to each other, privacy infrastructure a Jordan-style data market would need, though the market itself (pricing, compensation, incentives) still has to be designed around it. Combined with AWS Entity Resolution for identity matching and Amazon DataZone for governance, you can build data marketplaces where the equilibrium accounts for privacy budgets.
Jordan’s three-layer data market mapped to AWS services. The missing piece in most AI architectures: a compensation flow where value returns to data creators.
Sources
[1] Machine Learning Street Talk, “Intelligence Is Collective, Not Artificial”, Prof. Michael I. Jordan (UC Berkeley / Inria), May 2026. https://www.youtube.com/watch?v=AREWYbVtX64
[2] Science magazine, “Who’s the most influential computer scientist?”, Michael I. Jordan ranked #1 by h-index across computer science, 2016. https://www.sciencemag.org/news/2016/12/who-s-most-influential-computer-scientist
[3] Michael I. Jordan, “A Collectivist, Economic Perspective on AI,” arXiv:2507.06268, July 2025. https://arxiv.org/abs/2507.06268
[4] Michael I. Jordan, Wikipedia, career, awards, publications. https://en.wikipedia.org/wiki/Michael_I._Jordan
[5] World Laureates Association Prize in Computer Science (2022), “For fundamental contributions to the foundations of machine learning and its application.” https://www.thewlaprize.org/
[6] John Nash, “Non-Cooperative Games,” Annals of Mathematics, 1951, foundational paper on Nash equilibrium.
[7] Alvin Roth, “Who Gets What, and Why: The New Economics of Matchmaking and Market Design,” 2015, accessible introduction to mechanism design by the Nobel laureate who built kidney exchange markets.
[8] Stefan Christoph, “Agentic AI Payments: When Agents Pay for Content,” schristoph.online, May 2026. https://schristoph.online/blog/http-402-agents-pay/?utm=linkedin
[9] “Agents for Amazon Ads Solution Guidance,” AWS Solutions Library. https://aws.amazon.com/solutions/guidance/agents-for-amazon-ads/
About the Author
Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.
This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.