Feb 24, 2020, 8:47 AM ETInterview, Podcast

Michael I. Jordan: Machine Learning, Recommender Systems, and Future of AI | Lex Fridman Podcast #74

Michael I. Jordan rejects the current era as "Artificial Intelligence" in the philosophical sense (mimicking human thought), instead defining it as the emergence of a new engineering discipline analogous to chemical or electrical engineering.
This new field is currently a "proto-form" that combines statistical theory and computer science to build systems that make decisions at planetary scale, rather than focusing on fundamental understanding of human intelligence.
The term "AI" originated as a strategic rebranding by John McCarthy to distance his work from Norbert Wiener's "Cybernetics," a decision Jordan argues created unrealistic promises about understanding human cognition that the field cannot currently fulfill.
Jordan distinguishes between "scientific understanding" (which is lacking in neuroscience and language) and "engineering" (building functional systems like database management or recommendation engines that provide immediate value).

Jordan asserts that our understanding of the human brain is currently at a "clueless" stage, comparable to Greek speculation about space travel, and that fundamental neuroscience is a task spanning centuries.
He views current Brain-Computer Interface (BCI) efforts, such as Neuralink, as valid engineering attempts to find patterns but notes they lack a scientific basis for understanding deep brain algorithms.
Jordan disagrees with Elon Musk's timeline for deep brain-computer integration, stating that creating systems that truly understand biological algorithms is not feasible for this generation or even this century.
Current medical applications (e.g., using lithium for mood swings) often work without a clear understanding of the underlying biochemical mechanisms, highlighting the gap between engineering efficacy and scientific explanation.

Jordan and Yann LeCun (the "Miles Davis" of ML) generally agree on building concrete systems, but Jordan emphasizes a critical divergence: LeCun prioritizes prediction and pattern recognition, while Jordan argues for decision making under uncertainty.
Jordan critiques the industry's over-focus on perfect prediction, noting that humans constantly face novel situations where prediction is impossible and risk assessment, error bars, and causal reasoning are required.
He argues that prediction is insufficient for consequential real-world decisions (e.g., medical treatments, driving), which require understanding counterfactuals ("what if") and the economic/risk context surrounding a choice.
The core problem with current AI is that it ignores the "market forces" and "risk evaluations" inherent in human decision-making, treating data sets as static rather than embedded in a dynamic, strategic environment.

Jordan identifies a missing market in the digital creator economy where long-tail creators (e.g., independent musicians) cannot monetize their work despite having audiences, as current platforms rely on ad models and highlight only a few superstars.
He proposes a "two-way market" model where AI connects producers and consumers directly (e.g., a dashboard showing exactly where music is being streamed to facilitate local shows), creating jobs and economic value through transparency.
Jordan argues that the advertising business model is the root cause of issues like misinformation and user manipulation, as it incentivizes click-through rates rather than genuine producer-consumer value.
He suggests that a transition to a micro-payment or direct transaction model (where platforms take a small cut of direct value exchange) is necessary to solve "fake news" and align company incentives with societal health.
While acknowledging advertising has a role in signaling product value (e.g., a company spending money to launch a new vacuum cleaner), he believes the current scale of ad-based tracking is "creepy" and erodes trust.

Jordan distinguishes between "creepy" surveillance (companies like Facebook knowing personal plans without a caring relationship) and trusted utility (personal assistants like Alexa that help with specific tasks with user control).
He posits that privacy is not a binary legal concept but a relational one, where individuals need agency to manage which data is shared with whom, similar to historical village dynamics versus modern urban anonymity.
He predicts that the future of trustworthy technology will involve a complex "onion" of oversight, standards, and institutions (like Underwriters Laboratory for electricity) to manage privacy and safety, a process that will take decades.
Jordan expresses skepticism that companies like Facebook will pivot to a direct producer-consumer model on their own, as their current business model is too lucrative and entrenched.

Jordan describes optimization as a limited tool for human life due to its complexity but notes that game theory (specifically equilibria like Nash and Stackelberg) is essential for modeling multi-agent systems and markets.
He highlights Nesterov acceleration as a profound mathematical idea in optimization that achieves faster convergence rates by utilizing momentum, even if the algorithm sometimes moves "uphill" temporarily.
Jordan explains the tension between Bayesian and Frequentist statistics as analogous to wave-particle duality in physics:
- Frequentist: Averages over all possible data sets to provide guarantees for software shipped to the public.
- Bayesian: Conditions on the specific data set obtained, incorporating prior human expertise.
He favors Empirical Bayes methods and False Discovery Rate (FDR) as practical frameworks that blend these philosophies to manage error rates in large-scale decision-making.

Jordan advises aspiring researchers to treat the field as an apprenticeship requiring humility, hard work, and broad collaboration rather than relying solely on "brilliance."
He emphasizes that the field is highly international and cooperative, countering the narrative of fierce nationalism or cutthroat competition often seen in public discourse.
He encourages students to study mathematics, poetry, history, and languages to become well-rounded thinkers, noting that his own fluency in French and Italian deepened his empathy and understanding of diverse human experiences.
Jordan identifies deep natural language understanding as the most significant unsolved scientific challenge, admitting he missed this opportunity in his own career but remains an admirer of those pursuing it.