Jordan, Michael I.

professor at UC Berkeley; highly cited researcher in machine learning and statistics; mentor to Andrew Ng, Zoubin Ghahramani, and others

Michael I. Jordan: Machine Learning, Recommender Systems, and Future of AI

Lex Fridman Podcast · Feb 24, 2020

• Michael I. Jordan
• machine learning as engineering
• decision theory
• recommender systems
• market design
• Bayesian vs frequentist statistics
• Nesterov acceleration
• stochastic gradient descent
• AI hype critique
• intelligent infrastructure

Core ideas

Jordan frames what's currently called "AI" not as the realization of John McCarthy's original philosophical dream of machine intelligence, but as a nascent engineering discipline analogous to chemical engineering emerging from chemistry, or electrical engineering from electromagnetism — a proto-field combining statistics and computer science that builds real, valuable systems without yet having a coherent theoretical foundation for doing so safely and predictably at scale. He argues neuroscience remains extremely far from understanding the brain's basic computational principles (comparing the field's current state to ancient Greeks speculating about spaceflight) and is skeptical of near-term brain-computer interface ambitions like Neuralink, distinguishing legitimate narrow applications (e.g., cursor control) from unfounded claims of deep brain-computer integration, which he considers unlikely "for centuries." He similarly dismisses demos like AI phone-call assistants as engineering feats mimicking intelligence rather than evidence of genuine understanding, arguing true breakthroups (in his view, comparable to Kolmogorov-level insight) haven't happened and that hype from both companies seeking funding and public figures like Elon Musk (whom he criticizes by name for overstating natural-language understanding progress) misleads the public and young researchers.

Jordan rejects the word "AI" itself as a poor historical branding choice by McCarthy (needed mainly to differentiate from Norbert Wiener's "cybernetics") that overpromises understanding of "intelligence" the field doesn't have, preferring to describe the field's real subject as making large-scale, consequential decisions under uncertainty involving networks of many decision-making entities — encompassing both pattern recognition/prediction (Yann LeCun's emphasis) and full decision-making under real-world constraints (risk, cost, causal reasoning, market forces), which Jordan argues is comparatively neglected despite being where consequential real-world stakes (like medical decisions) actually lie. He illustrates this decision-making layer with a detailed example of an unrealized "music market": streaming platforms extract value from millions of independent creators (many young, talented, uncompensated) without connecting them to consumers in an actual economic relationship, whereas a well-designed market (citing his board role at United Masters, whose music appears on NBA broadcasts) could give creators real income streams, tour bookings, and direct merchandise sales — framing this as AI-enabled job creation rather than automation-driven job loss, contingent on solving genuinely new market-design and mechanism-design problems at unprecedented scale.

Jordan offers a sustained critique of the advertising-based business models of Google and especially Facebook, arguing they inherited television's ad-supported playbook without ever developing genuine two-sided producer-consumer markets, which he sees as the root cause of problems like fake news (driven by click-through-rate-optimized monetization) — while defending a legitimate, limited role for advertising as a costly, credible signal that a producer believes in their product, distinct from surveillance-based "creepy" personalization that consumers widely mistrust. He argues the discomfort with Facebook isn't really about privacy in the abstract, but the absence of any caring or reciprocal relationship justifying what a company knows about a person, and expresses skepticism (in explicit disagreement with the host's optimism) that Facebook will discover a better model absent a fundamental business-model pivot, contrasting this with Amazon (where he spends one day a week and sees genuine attempts at real seller/buyer market-building) and a partially-improving Microsoft (crediting Satya Nadella's shift toward user trust). He argues good recommender systems add real value (praising his own experience with Amazon's book recommendations as social learning from peers) but pushes back against fully AI-anticipated personalization, arguing humans are far too rich and idiosyncratic to be "figured out" at scale, and that valuable serendipity (discovering unexpected interests) requires bounded, transparent, human-controlled spaces rather than exhaustive behavioral surveillance.

On mathematical foundations, Jordan discusses Occam's-razor-adjacent themes through statistics and optimization: he explains the Bayesian/frequentist distinction as a genuine, unresolved duality (analogous to wave-particle duality in physics) rooted in decision theory's two-argument loss function (data and unknown parameter) — frequentists average over hypothetical data given a fixed parameter (yielding guarantees like "95% of the time this procedure works," suited to shipped software), while Bayesians average over the unknown parameter conditional on the actual observed data (suited to incorporating domain expertise) — and describes "empirical Bayes" and "false discovery rate" methods as elegant middle-ground syntheses. On optimization, he highlights Nesterov acceleration as a genuinely surprising, still not fully understood algorithmic idea (combining two gradients, sometimes moving locally uphill, to beat gradient descent's convergence rate), and explains why stochasticity in optimization (e.g., stochastic gradient descent) is not merely practically convenient but mathematically principled — helping avoid pathological features of a loss surface (like points of non-differentiability) that a fully deterministic algorithm could deterministically hit and get stuck on.

Jordan's most conceptually distinctive argument, developed in his blog post "Artificial Intelligence — The Revolution Hasn't Happened Yet," is that "intelligence" should not be understood as fundamentally about replicating human cognition: markets themselves — decentralized systems of millions of independent decisions (matching supply and demand, adapting to shocks, functioning robustly across scales from villages to megacities for millennia) — constitute a genuinely different, non-human form of intelligence with real underlying principles (supply/demand curves, auctions, matching theory) that a lifetime of studying human psychology would never reveal. He argues future computational systems should draw on these non-anthropomorphic forms of "intelligence" (markets, economic mechanism design) as much as on human-cognition-inspired approaches, and dismisses speculative science-fiction questions about superhuman AGI or brain-uploading scenarios as fun but ultimately unproductive distractions from more pressing, tractable engineering and scientific challenges facing the present generation. He closes with advice for aspiring researchers emphasizing apprenticeship, humility, and years of sustained community-embedded practice (comparing graduate school to becoming a musician or artist) over the myth of individual genius, and reflects on his own lifelong interest in learning languages (French, Italian) as expanding his appreciation for how differently human experience can be expressed — tying this to his view that deep natural language understanding, done properly, remains one of AI's most profound and still largely unsolved scientific challenges.

Excerpts