Michael I. Jordan says the AI revolution hasn't happened yet—we're just in an early engineering phase, far from understanding the brain. He criticizes current AI for focusing on prediction without managing risk, like recommender systems that guess your likes but miss real needs. He blames Facebook and Google's ad-driven model for fake news, since it rewards clicks over value. His fix: a micropayment system that directly connects creators and users.
At a Glance Michael I. Jordan (UC Berkeley professor, with over 170,000 citations) discussed machine learning, recommendation systems, and the future of AI on the Lex Fridman podcast. His core argument is that the AI revolution has not truly occurred yet, and the field needs to broaden its scope—tre
Michael I. Jordan (UC Berkeley professor, cited over 170,000 times) discussed machine learning, recommendation systems, and the future of AI on the Lex Fridman podcast. His core argument is that the AI revolution has not truly occurred yet, and the field needs to be broadened, viewing AI as a deeply human-centered endeavor—not only engineering algorithms and robots, but also understanding and empowering humans across all levels of abstraction, from the individual to civilization. Jordan, called "the Miles Davis of machine learning" by Yann LeCun, is known for his constant self-reinvention. He has mentored top researchers such as Andrew Ng and Zoubin Ghahramani, emphasizing the intersection of statistics and computer science, and advocates that AI should move beyond its current narrow technological perspective.
Michael I. Jordan argues that the current field of AI is mislabeled as "artificial intelligence" and is actually a new engineering discipline born from statistics and computer science, akin to how chemical engineering emerged from chemistry and electrical engineering from electromagnetism.
Jordan builds his argument through historical analogy: In the 1930s–40s, chemistry and fluid mechanics already existed, but "chemical engineering" had not yet taken shape as an independent discipline—people wanted to build factories and scale up chemical production, but lacked systematic engineering principles; factories sometimes exploded or failed to turn a profit. Similarly, Maxwell's equations already encompassed all knowledge of electromagnetism, but safely wiring circuits, modular integration, and long-distance power transmission required the discipline of electrical engineering. Jordan concludes: "I think this is what is happening—we have a precursor field, which is the more theoretical, algorithmic part of statistics and computer science. That is enough to start building things. But what are we building? Systems that deliver value to humans, use human data, and incorporate human decision-making. And the engineering side is currently all ad hoc."
He further emphasizes that human understanding of the brain remains at an "ancient Greek level": "We have absolutely no idea how the brain performs computations. We are more ignorant than the ancient Greeks were about any interesting scientific question." Real neuroscientists study electron microscope images of individual synapses—"that is a city"—and a synapse is just a tiny point on a dendritic tree. Jordan criticizes the industry's excessive marketing: "Many young people think we are on the verge of a breakthrough. People who will stop at nothing—or even those who just need lab funding—will overhype."
Falsification condition: If, within the next 20 years, an interpretable, general-purpose learning algorithm based on genuine neural mechanisms emerges—rather than the current gradient-optimization-based engineering approach—then Jordan's assessment would be falsified.
The core disagreement between Jordan and Yann LeCun lies in this: LeCun emphasizes prediction (pattern recognition), while Jordan emphasizes decision-making — the latter being "where the rubber meets the road," involving risk, causality, dialogue, and market forces.
Jordan points out that current recommendation systems and neural networks only make predictions, not risk assessments: "They have no error bars." Using a medical scenario as an example, he states: "I wouldn't undergo heart surgery just because a neural network's output exceeds 0.7. Even if you had all the world's heart disease data — more than any doctor — I still wouldn't trust that output. I want to ask 'what if' questions, see causal data I haven't collected, and have a conversation with the doctor." He further breaks it down: real-world decision-making requires handling scarcity, others' decisions, multi-step decision chains, and economic costs — none of which pure prediction can solve.
Jordan argues that the industry's excessive focus on prediction has led the public to overlook the challenges of decision-making: "Prediction plus decision-making is everything — both are equally important. But the field overemphasizes prediction, at the cost of people not realizing that decision-making is where the real crux lies — involving human lives, risk-taking, data collection, error bars, consequences of others' decisions, and the economics surrounding decisions."
Unique insight: Jordan clearly separates "prediction" from "decision-making" and points out that the current AI boom is essentially "prediction engineering," not "decision engineering" — the latter being the domain that truly impacts human well-being.
Jordan argues that the fundamental issue with platforms like Facebook and Google is not algorithmic flaws, but the advertising-driven business model—it incentivizes click-through rates rather than genuine value creation, leading to misinformation, privacy violations, and market failure.
Jordan uses the music market as an example to illustrate the problem: there is currently no real music market—record labels prop up a handful of superstars, but a vast number of talented creators (such as a 16-year-old making hip-hop music on a laptop) cannot make a living from it. Platforms like Spotify generate revenue through subscriptions or advertising, yet creators receive inadequate compensation. Jordan envisions a better system: creators should have a dashboard showing a map of where their songs are being played across the U.S., with transparent and verifiable data; local event organizers, upon seeing this data, would invite the creator to perform, earning them $20,000 per show. With three such shows a year, they would have a professional income. The platform could take only a 5% cut and still be profitable.
He criticizes Google and Facebook for never seriously considering a direct producer-consumer connection: "They went back to the TV-era advertising playbook. No one wanted to pay for the signal, so ads filled the gap. But Google did it so well and made so much money that it never stopped to think, 'Wait, could we build a producer-consumer relationship here?'" Jordan believes that the problem of fake news is essentially a result of advertising incentives: "Fake news rides on click-through rates. You have to remove that core problem."
Unique Insight: Jordan does not believe that AI can "fix" recommendation systems under the advertising model—"You can tweak it with smart AI algorithms, but I think it's basically hopeless." He advocates for a complete overhaul of the business model, shifting toward micropayments and direct connections.
Jordan opposes the current paradigm of recommendation systems that "collect all data → predict user behavior," arguing that humans are too rich and complex to be preemptively anticipated by algorithms; a better direction is to create open spaces that allow humans to discover surprises on their own.
Jordan illustrates data noise with a personal experience: half-asleep in the morning, he clicked on a news story about the Queen of England, even though he "couldn't care less about the Queen of England" — it was clickbait. He warns: "The system will think I care about the Queen of England. Any reasonable system would." He further points out that humans "each have their own unique little quirks, each have the potential to suddenly fall in love with something they themselves didn't know about — with no prior indication. I don't want companies trying to predict that."
He argues that recommendation systems should be "more limited": "Create vast spaces where human creativity and style can flourish and be expressed. More transparency. Don't let people comment anonymously and arbitrarily based on facts they know about me." Jordan uses music as an example: walking down the street, he heard Chilean music and discovered he liked it — this kind of serendipitous discovery is the true value of recommendation systems, not predictions based on browsing history.
Falsification condition: If a future algorithm can accurately predict long-term shifts in human interests (such as suddenly falling in love with Chilean music), and user trust in it exceeds that in current systems, then Jordan's judgment would be partially falsified.
Jordan defines statistics as "principles that allow you to make inferences and decisions you can reasonably believe in," and notes that the divide between Bayesian and frequentist approaches resembles the wave-particle duality in quantum physics — they are physically distinct, sometimes converge in practice, and at other times yield entirely different answers.
Jordan explains this through the lens of decision theory: the loss function is a function of both data \( x \) and parameter \( \theta \), both of which are unknown. The frequentist approach averages over \( x \) (the data) to obtain "risk" — suitable for software products, ensuring correctness 95% of the time across large datasets of users. The Bayesian approach averages over \( \theta \) (the parameter), leveraging prior knowledge — suitable for scientists focusing on a specific dataset. Jordan argues that both are necessary and advocates for "empirical Bayes" as a middle path: starting from a Bayesian framework, estimating unknowns from data, and then ensuring performance through mathematical guarantees.
He specifically highlights the False Discovery Rate (FDR) as "one of the most beautiful ideas": FDR starts from the data and asks, "Among the results you declare as discoveries, what proportion are false positives?" — this is a Bayesian direction (inferring hypotheses from data), rather than a frequentist direction (projecting data from hypotheses). Jordan notes that this idea was developed by Robbins (1960), Brad Efron, Benjamini, and Hochberg, calling it "a beautiful set of ideas."
Jordan argues that markets (such as urban food supply systems) possess all the characteristics of intelligence: robustness, adaptability, self-healing, and scalability. This represents a "third type of intelligence" distinct from human intelligence, and is more relevant to current computer systems than understanding the human brain.
Jordan contends: "A market that moves goods into a city and supplies restaurants every day is a system — a set of decentralized decisions. From far enough away, it looks like a set of neurons: each neuron makes its own small decisions, unaware of the overall goal, but something emerges at the aggregate level. The same is true for economic systems: people eat in cities, it is robust, it works from small villages to large cities, has been operating for thousands of years, and adapts regardless of weather." He believes that even after thousands of years of studying human psychology, one might not discover market principles such as supply-demand curves, matching, and auctions — these are real principles that constitute a form of intelligence.
Unique insight: Jordan liberates "intelligence" from anthropocentrism, arguing that market intelligence should be the primary focus of current AI research, rather than mimicking the human brain — the latter being "a task for centuries from now."
| Position | Guest Attitude | Key Data |
|---|---|---|
| Amazon | Positive (partially affirming) | Half of employees make decisions, half perform pattern recognition; Jordan goes there one day a week |
| Neutral to critical | Advertising model is successful but has not established a producer-consumer market | |
| Critical | Advertising model leads to fake news; user trust is "at rock bottom" | |
| Spotify | Neutral | Aims to enable 1 million creators to live comfortably, but the current model has not achieved this |
| YouTube | Neutral to positive | "More potential than imagined" |
| Neuralink | Critical | "From electrodes to understanding brain algorithms—this is not for this generation, nor even for this century" |
| United Masters | Positive (Jordan serves on the board) | 100,000 artists already signed; music featured in NBA promotional videos |
| Microsoft | Positive | "Transforming into a trustworthy old uncle" |
| Uber | Positive (as a breakthrough case) | Creates new markets and new value |
1. "The AI revolution has not yet occurred" (Jordan): The current moment is akin to the birth of chemical engineering or electrical engineering as disciplines, not the moment of understanding the human brain—humanity's grasp of the brain remains at an "Ancient Greek level."
2. "Prediction is not decision-making" (Jordan): Current AI overemphasizes pattern recognition while neglecting risk assessment, causal reasoning, and multi-step decision chains—"Prediction plus decision-making is everything."
3. "The advertising model is the root of fake news" (Jordan): The advertising incentives of Facebook and Google drive click-through rates, and AI cannot fix the business model problem—"You can tweak it with smart AI algorithms, but I think it's basically hopeless."
4. "The market itself is a form of intelligence" (Jordan): Urban food supply systems exhibit intelligent features such as robustness, adaptability, and self-healing—this is a "third type of intelligence" distinct from human intelligence, and more relevant to current AI than understanding the human brain.
5. "Don't try to predict humans—you can't" (Jordan): Humans are too rich and complex for algorithms to anticipate sudden shifts in interest (e.g., falling in love with Chilean music)—recommendation systems should create open spaces rather than predict users.
6. "Bayesian and frequentist approaches are like wave-particle duality" (Jordan): They are physically different, and in practice sometimes converge, sometimes yield completely different answers—empirical Bayes is the middle path.
7. "False discovery rate (FDR) is one of the most beautiful ideas in statistics" (Jordan): Inferring back from data to hypotheses, asking "what proportion of declared findings are false positives"—this is a Bayesian direction developed by Robbins, Efron, Benjamini, and others.
8. "Learning language is a path to understanding the core of AI" (Jordan): Natural language understanding is "the most interesting scientific challenge," but "it will not happen in our lifetime"—Jordan himself trains "that part of the brain" by self-studying French and Italian.