📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=AREWYbVtX64
Beyond the AGI Delusion: Why the Future of AI is Economic, Not Just Computational
Current discourse around AI is trapped in a binary of utopian superintelligence or existential doom, a framing that Michael Jordan argues is more science fiction than science. By shifting our focus from mimicking the individual brain to building intelligent, multi-agent markets, we can create technology that actually serves human needs.
Core Question: How can we move past the “superintelligence” hype to build AI systems grounded in the principles of economics, statistics, and collective welfare?
Highlights
- The term “AGI” is a PR distraction that demoralizes young researchers by suggesting the “pure” work of intelligence is already solved.
- Intelligence is not just a “brain in a box”; it is a social, collective phenomenon that has existed in the form of markets for millennia.
- Current AI business models “steal” data without returning value to creators, leading to broken ecosystems in music, journalism, and medicine.
- A robust “Third Way” for AI requires a synthesis of three distinct thinking styles: Computational (CS), Inferential (Statistics), and Economic (Incentives).
⏱️ Reading time: approx. 8 minutes · Saves you about 69 minutes vs. watching.
Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇
The Engineering Crisis and the AGI PR Trap
Science Fiction vs. Principled Design
We are currently building massive statistical “boxes” that we don’t fully understand, yet we treat them as if they are nearing god-hood. This level of detachment from reality is unusual in the history of engineering, where disciplines like chemical or electrical engineering relied on grounded principles like Maxwell’s or Newton’s equations to ensure safety and utility.
The current generation of “thought leaders” is obsessed with the idea that we have figured out the brain because gradient descent works at scale, which is an intellectual leap that most neuroscientists find laughable. By promoting the narrative that “superintelligence” is just around the corner, we are telling 20-year-olds that there is nothing left for them to build except for a “secretary on their shoulder” that they will eventually want to turn off. This outlook is fundamentally demoralizing for young folks who want to use technology to help their families and their countries through tangible, human-scale innovation rather than chasing sci-fi ghosts.
AGI is a PR term, not a scientific milestone.
It creates a false choice for the next generation: you are either an “exuberant” believer in the digital god or an “alarmist” terrified of extinction. This binary ignores the massive middle ground of building specific, helpful systems for healthcare, transportation, and finance that don’t need to “understand” anything to be transformative.
💡 Digging Deeper
Q: Why does Jordan dislike the term AGI?
A: He views it as a distortion that confuses young people and distracts from the goal of building systems that solve specific societal problems through data flows and economic incentives.
Q: Is building systems we don’t understand inherently bad?
A: No, but we must build “guardrails” around them using statistical inference and economic rules of thumb to ensure they remain predictable and beneficial to the humans interacting with them.
Q: What is the “First Step Fallacy”?
A: It is the mistaken belief that because a system can do something impressive (like generate fluent text), we are only one step away from it being able to do “anything” or having general human capability.
Intelligence as a Collective Economic Equilibrium
The Three-Layer Data Market
Most current AI models are built by scraping the internet—essentially taking the creative output of billions of people without returning any value to the originators. This is not just an ethical issue; it is a “dumb business model” that fails to treat humans as producers and consumers within a larger, intelligent network.
If we look at a “three-layer data market”—consisting of users, platforms, and third-party data buyers—we see that current systems are under immense stress because they don’t account for privacy or the shifting value of information. A user might provide data to a platform for a service, but when that platform sells the data to a third party, the user loses privacy. To fix this, we need a system where privacy levels (like differential privacy) are tunable, allowing the market to reach an equilibrium where users are compensated for the quality and privacy-risk of the data they provide.
Markets are a form of collective intelligence that existed long before capitalism.
When we think of AI as an economic system, we stop trying to build a single “super-brain” and start trying to build an ecosystem where individual agents—humans and algorithms—can cooperate and compete effectively. For example, in the music industry, services like Spotify function as monopolies that pay artists very little, but an “AI-native” market could connect musicians directly to brands and audiences, setting prices based on actual competitive value rather than top-down algorithms.
💡 Digging Deeper
Q: How does economics improve AI safety?
A: Economics focuses on incentives. If you design a system where agents are incentivized to provide truthful, high-quality data, the overall system becomes more robust and predictable than one relying purely on black-box predictions.
Q: What is the difference between optimization and equilibrium?
A: Machine learning people excel at optimization (finding the best single answer), but society is an equilibrium problem where multiple self-interested parties interact, requiring fix-point algorithms and game theory.
Q: Why are “metadata” and “providence” important for LLMs?
A: LLMs treat all data as equal, but in the real world, data decays. Knowing that medical data is ten years old should automatically increase the uncertainty of the model’s prediction—something current “black box” models fail to do.
The Triangle of Modern Intellectual Thought
Synthesizing CS, Statistics, and Economics
To move forward, we must stop viewing AI as a sub-branch of Computer Science and start seeing it as a synthesis of three distinct pillars: Computational thinking, Inferential thinking, and Economic thinking. Computational thinking gives us modularity and APIs; Inferential thinking gives us the ability to quantify uncertainty; and Economic thinking ensures that the systems we build respect human incentives and social welfare.
Statistics is the “Liberal Arts” of the data era.
Without statistical “anytime inference” and uncertainty quantification, AI models are just making confident guesses that can lead to catastrophic failures in high-stakes fields like medicine or regulatory drug discovery. We need “prediction-powered inference” that merges massive biased foundation models with small amounts of ground-truth data to create error bars that scientists can actually trust. This isn’t science fiction; it’s a mathematical necessity.
We must build systems that allow bottom-up preferences to be expressed.
The “God-view” from Silicon Valley—where a few leaders decide the “human value function” for everyone—is a recipe for broken social systems. Instead, we should be building “market-making” mechanisms that allow people to express their own ephemeral, contextual preferences in the moment. This is how we avoid the “blind men and the elephant” problem, allowing diverse human perspectives to inform a collective hole that no single algorithm could ever fully comprehend.
💡 Digging Deeper
Q: What is “conformal prediction”?
A: It is a statistical method that puts a “confidence region” around black-box AI outputs without requiring complex assumptions, ensuring the model’s answer falls within a specific range of certainty.
Q: How do “E-values” differ from “P-values”?
A: P-values are a one-shot measure of how improbable a result is, often leading to “p-hacking,” while E-values are martingales that allow for “anytime inference,” meaning you can stop and check the data at any point without losing statistical validity.
Q: What is “Mechanism Design”?
A: It is the “inverse of game theory.” While game theory predicts what will happen in a given game, mechanism design asks: “What game should I build to ensure a specific, fair outcome occurs among self-interested people?”
Key Takeaways
We are at a turning point where AI must evolve from a collection of “smart” algorithms into a robust engineering discipline. Michael Jordan emphasizes that the current focus on “understanding” and “intelligence” is largely a distraction for the media; the real work lies in building systems that manage uncertainty and incentives at a massive scale. By treating AI as a component of an economic ecosystem, we can move away from the “labor vs. capital” displacement fear and toward a future where technology creates new markets for human creativity and talent.
The path forward requires a new “Liberal Arts” core for the next generation—one that balances the ability to code with the ability to reason about social welfare and statistical truth. We don’t need digital gods; we need better “autopilots” for society that allow humans to remain in control of the high-level decisions while algorithms handle the complex, data-heavy logistics of a world with 8 billion people.
Q&A
Q1: Is the brain really just doing gradient descent?
A1: Most neuroscientists would say no. While gradient descent works incredibly well at scale for machines, calling it “intelligence” is a cartoonish metaphor that ignores the biological and social complexity of the human mind.
Q2: Why is the current AI dialogue considered “harmful”?
A2: It frames the future as a binary between superintelligence and extinction. This is demoralizing for young developers, suggesting that the “hard work” of intelligence is over and there is no room for human-scale innovation.
Q3: How can AI help fix broken political systems?
A3: By improving information flow and signaling. Much human conflict arises from a lack of understanding of the “other’s” motivations. AI could aid in creating better “game-theoretic” mechanisms for negotiation and decision-making in democracies.
Q4: What is the “Duck Example” regarding uncertainty?
A4: If a duck knows food is at one side of a lake 2/3 of the time, it doesn’t just go there 100% of the time (the Bayesian choice). It hedges its bets. In a group, ducks distribute themselves 2/3 and 1/3, which is a Nash equilibrium. Intelligence is often about how individuals act within a population.
Q5: Can AI replace teachers or doctors?
A5: It can aid them significantly, especially with routine tasks, but “good” teachers and doctors operate at the “edge of knowledge,” where data is sparse and human judgment is required. AI is least effective at this edge.
Q6: What is “Statistical Contract Theory”?
A6: It is a way to handle interactions where one party (like a doctor) has more info than another (the patient). It uses statistical “evidence gathering” to create menus of options that incentivize both parties to be truthful and collaborative.
Q7: Is Michael Jordan “bullish” on AI?
A7: Yes, but only if we stop building “computational artifacts” for their own sake and start building systems that respect human labor, data ownership, and the messy reality of social interactions.
