Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria)

Jordan’s core message: stop treating models as artificial minds and start engineering socio-technical systems. The work that matters now is microeconomics, incentives, and governance—how models, people, and organizations exchange data, value, and decisions at scale. AGI talk is a distraction; the meaningful frontier is mechanism design around ML components.

Practically, this means building systems around models, not inside them. Use models for prediction/optimization, then wrap them with: (a) error bars and cohort-style explanations users can act on, (b) small, targeted ground-truth collection to de-bias foundation model outputs (prediction-powered inference), and (c) incentive-compatible data contracts so producers are paid and data quality is rewarded. He highlights a concrete “three-layer data market” (users → platform → data buyers) where privacy budgets, payments, and data utility interact as a Stackelberg game—an equilibrium problem, not an optimization one.

Organizationally, this reframes AI programs: hire or partner with economists/market designers; add fixpoint/equilibrium analysis to the ML toolbox; budget for data acquisition and payments (not only GPUs); implement tunable privacy (e.g., differential privacy menus); and measure social welfare, regret, and equilibrium stability—not just model accuracy. Replace “AI as a personal assistant” with AI-augmented markets and workflows in regulated domains (finance, healthcare, supply chains), where human producers and consumers must be part of the loop.

Tradeoffs: equilibrium-centric design is harder, slower, and messier than shipping a chatbot. Privacy adds noise, reducing data buyer value; incentives reduce free data; small ground-truth studies cost time/money; and governance becomes first-class engineering. But these constraints are precisely how you achieve robustness, adoption, and regulatory fit.

Weak evidence/hype to discount: claims that multi-agent LLMs “discover” economics for free; anthropomorphic “understanding”; AGI timelines. What’s underweighted: aligning incentives yields larger system-level gains than chasing marginal model wins; explanation should be actionable (cohort-based), not mechanistic interpretability for its own sake.