There is a dangerous myth around platforms: that they are neutral places where people gather, exchange, create, and compete. We like to imagine them as digital public squares or modern bazaars, open spaces where opportunity naturally multiplies. The seller finds the buyer, the worker finds the job, the creator finds the audience, the citizen finds the conversation. Tim Wu’s The Age of Extraction asks us to look more carefully at what happens after the square becomes privately owned.

The question is not whether platforms create value. They clearly do. The question is who owns the infrastructure, who sets the rules, who captures the upside, and who slowly becomes dependent on the system.

Wu’s argument is sharp because he refuses to treat digital platforms as entirely new. His deeper claim is that platform power belongs to a much older story about land, ownership, dependency, monopoly, and political instability. The technology is new. The power pattern is not.

Mental Models Acquired

The most useful mental model I took away from this book is that platforms are not just products; in fact, they are economic environments.

A platform is not merely something people use. It is a place where other people’s activities take place. When a platform works well, it enables participation, lowers barriers, creates access, and allows many people to build, trade, speak, and experiment.

Once a platform becomes essential, its character changes and the owner is no longer just providing a service. They have the ability to govern the conditions of participation such as who gets visibility, who pays what fees, whose work is discoverable, whose business survives, and what kinds of behavior are rewarded. At that point, the platform is not simply creating opportunity; it is administering an ecosystem of dependency.

A true public square belongs to civic life. It is messy, open, and collectively important. A privately owned digital square is disguised to appear public while operating according to private incentives, and that gap matters.

What This Book Gets Right

Wu’s core move is to connect platform power to feudalism, and I take up that argument in full in the signals below. What I want to credit here is the historical craft that makes the argument land.

The history around IBM is especially useful because it reminds us that today’s platform economy did not arrive out of nowhere. The relationship among hardware, software, operating systems, and market dominance has been shaped by decades of technical, legal, and business decisions. What looks inevitable in hindsight was actually designed, contested, and shaped.

Wu captures how the rise of digital software depended on its separation from hardware, and how that separation created a new field of power. Software could scale differently; it could sit above hardware. It could become the layer through which other systems operated.

I also appreciated the examples of Japan and other cultures developing different models of innovation. These examples matter because they remind us that capitalism has more than one form. Innovation does not have to mean winner-take-all extraction. Markets can be organized differently. Firms can relate to workers, suppliers, and society differently. The American platform model is powerful, but it is not natural law.

The strongest part of the book is this: Wu helps us see extraction as an old power problem wearing new technological clothing.

Skeptic’s Corner

My skepticism is not with the diagnosis; it is with the scope and the solution. The book is framed through a very Western lens. The historical arc is powerful, but it leaves me wanting more attention to China and India, especially because both countries complicate the platform story in different ways.

China raises difficult questions about state power, platform power, surveillance, industrial policy, and innovation under centralized control. India raises another set of questions around digital public infrastructure, informal economies, leapfrogging, identity systems, and scale. If the argument is about the future of platforms, then China and India are not side cases; they are central. I will come back to India, because it turns out to matter for Wu’s own solutions.

The solutions at the end also felt more utopian than operational. Wu’s five points are directionally right, but they only scratch the surface: anti-monopoly action, neutrality rules, utility-style regulation, countervailing power, and line-of-business restrictions all make sense as principles.

The book names the right doors. It does not fully walk through them. The missing layer is implementation: who governs the infrastructure, who audits it, who pays for it, who can contest decisions, and what happens when public systems fail?

Top 3 Signals I’m Taking Forward

1. The link between land ownership, feudalism, the industrial age, and the evolution of the platform
   The most important signal is the historical continuity. Feudalism was built around control of land. Lords owned the essential resource. Peasants and serfs depended on access to that resource in order to survive. The system was not only unequal but also binding; it shaped mobility, work, dignity, and political power.
   
   Wu shows how the same pattern reappears in different forms: land ownership gives way to industrial monopoly, which gives way to software platforms and then ultimately to data, attention, and algorithmic control. Each era has its own tools, but the underlying question remains the same: who controls the infrastructure that everyone else depends on?

   
   This changes how I think about the platform story. The argument shifts from “platforms create opportunity” to “platforms can become private governments if left unchecked.” That shift matters; it moves the conversation from innovation to governance. It asks not only what platforms make possible, but what they make unavoidable.

2. Extraction, attention, data mining, and the rise of convenience culture
   The second signal is the link between extraction and convenience. Platforms rarely feel oppressive at the start; they win by removing friction. They remember preferences, streamline tasks, and collapse multiple activities into a single, seamless interface.

   
   That convenience has a cost. As we outsource effort, memory, and decision-making, platforms learn from us: capturing attention, shaping behavior, and reinforcing passive use. Extraction becomes not just economic, but behavioral.
   
   You can see it in how people adapt to the systems they depend on. Sellers optimize for Amazon’s rankings. Creators anticipate algorithmic preferences. Drivers respond to pricing signals they do not control. Participation gradually becomes dependence.
   
   Convenience begins to stand in for progress, and the risk is not just that platforms take value, but that they reshape how we act and diminish our ability to maintain our agency.

3. What if technology platforms were utilities?
   The third signal is the question of utility. What would an economy look like if some technology platforms were treated less like private empires and more like public infrastructure?
   
   Wu reaches back to France’s pre-internet Minitel system from the 1980s. Fascinating, but a museum piece, tied to the technological conditions of its time. Here is where the book’s Western frame costs it something real: the strongest living answer to Wu’s own question is running right now in India.
   
   India’s Unified Payments Interface (UPI) is not a perfect public-good story, but it is a radically different sequencing model: build interoperable rails first, then allow private competition on top of them. It processes billions of transactions a month. Aadhaar and the broader India Stack raise their own hard questions about surveillance and exclusion, and those questions deserve scrutiny, but together they represent the most ambitious real-world experiment in treating digital infrastructure as a public good rather than a private empire.
   
   Notice what India did: it reversed the sequence. The American pattern has been innovation first, platform dominance second, regulation after the fact. The Indian model builds public infrastructure first, then allows competition on top of it. That ordering matters because it constrains extraction at the level of system design, not policy. Wu asks what utility-style platforms might look like. India is not speculating. It is operating.
   
   That reframes the timeless question underneath. What digital infrastructure should be public, neutral, or utility-like? Identity, payments, search, marketplace access, communication protocols, data portability, AI infrastructure. Not every platform should be a utility. But some systems become so foundational that leaving them entirely to private control looks less like innovation and more like abdication.
   
   The real design question is this: how do we create digital infrastructure that enables participation without turning every participant into a captive user?

Signal Worth Your Time?

Yes.

The Age of Extraction is worth your time because it gives language to something many of us feel but struggle to name: the unease of living inside systems that are useful, convenient, and quietly controlling.

Wu’s contribution is not simply that platforms are powerful. We already know that. His contribution is showing that this power belongs to a much longer history of extraction. Land, factories, software, data, attention: each becomes dangerous when a small group controls what everyone else needs.

The real risk is not that platforms become powerful. It is that they become unavoidable. The future of technology will not be decided only by what we can build. It will be decided by what we allow to become infrastructure, who owns that infrastructure, and whether the people who depend on it have any real power inside the system.

That is the question I am carrying forward: Are we building platforms that expand human agency, or are we building better-designed forms of dependency?

Mental Models Acquired

The most useful shift this book prompted was reframing AI as a context problem, not a capability problem. Technology does not arrive as good or bad. It arrives neutral, and gets shaped by whoever is closest to the problem, has the strongest conviction, and can survive long enough inside an institution to see it through. That last condition carries more weight than most people admit.

The real protagonist here is the translator. Not the algorithm, but the person standing between what is technically possible, what the institution will tolerate, and what the human in front of them actually needs. That is the hardest job in any AI deployment, and it is almost never formalized, funded, or named.

What This Book Gets Right

Each chapter is a story about someone who recognized that the gap between current AI capability and meaningful human benefit is primarily a design and deployment problem and then proved it. That is not trivial.

What Khanmigo and the Cleveland Clinic’s Bayesian Health platform share is not the technology. It is the specificity of the problem framing. Neither began with “apply AI to education” or “apply AI to healthcare.” They began with a precise, human-centered question and worked backward to the tool. That discipline is rarer than the technology itself and significantly harder to scale.

The book also makes one of the more compelling cases I have seen for AI as a mechanism for expanding access and participation, not as an aspiration, but as a demonstrated practice.

Skeptic’s Corner

The structure, chapter by chapter success stories, also introduces a quiet survivorship bias. We do not hear from the ideas that died in a committee meeting, got defunded in Q3, or worked beautifully in the pilot and then collapsed the moment someone had to actually change their workflow. Those stories are at least as instructive. Possibly more.

There is also an implicit narrative running through many of these examples: that an individual with enough conviction can act as a change agent. While directionally true, it leans on a familiar myth. Many of the people featured are not just determined, they are also highly educated, risk-tolerant, and operating with forms of institutional access that are not widely available.

That matters.

It raises a harder question, the book does not fully engage with it: what happens when the people closest to the problem lack the authority, resources, or technical fluency to shape the solution? Conviction without leverage rarely survives contact with institutional reality.

The book makes a strong case for what is possible. It is less explicit about the structural conditions that make those possibilities achievable, or the absence of those conditions in most environments.

Top 3 Signals I’m Taking Forward

1. The translator is the most important role nobody is hiring for.
   The people in this book are not techno-optimists. They are not the move-fast-and-break-things types. Across the book, the heroes are not abstract technologists. They are cardiologists expanding access to critical care, educators trying to meet students where they are, and public servants doing the difficult work of modernizing institutions from the inside. They decided the gap between what AI can do and what it is actually doing for people was unacceptable, and then did something about it. That orientation, fixing rather than disrupting, produces fundamentally different outcomes than most AI deployment stories you read about.

2. Specificity of problem framing is the actual competitive advantage.
   Not the model. Not the budget. Not the roadmap.
   The organizations doing this well started with a precise human question. Khanmigo did not start with "let’s apply AI to education.” Sepsis prevention did not start with "let’s apply AI to healthcare." They started with a specific failure point and worked backward to the tool. The ones struggling started with a technology and looked for a use case. The difference compounds fast.

3. Underserved populations are not an edge case. They are a design stress test.
   When you design for the people the mainstream ignores, you build something that actually works. A scientist building AI-powered communication tools for nonverbal autistic children was not running an altruism project. She was not running an altruism project. She was solving a real, specific, urgent problem with no acceptable workaround, and the urgency was personal. That combination produces the clearest brief in product design.

Signal Worth Your Time?

Yes, with conditions.

If you lead a team that deploys AI, this book is a useful reality check against the assumption that harder problems are out of reach. The examples are not theoretical. They are proof that the gap between current AI capability and meaningful human benefit is mostly a design and deployment problem, not a technology problem.

What the book does not do is help you think about the costs. Every success story here sits downstream of enormous institutional effort, personal conviction, and risk tolerance that most organizations cannot sustain. That is not a reason to dismiss it. That is the harder question it leaves you with: who in your organization has the proximity, the authority, and the judgment to do this work when the incentives are pointed elsewhere?

Read it as a catalog of what is possible. Then ask yourself what conditions would need to be true for any of it to happen inside your walls.

Mental Models Acquired

One way I’ve started measuring whether something was worth reading is simple:
Did it give me language I didn’t have before?

The most durable mental model I took from AI Snake Oil is the distinction between generative AI and predictive AI.

Generative AI creates. It writes, speaks, paints, and composes. When it fails, it produces misinformation or convincing nonsense.

Predictive AI forecasts. It attempts to determine who will succeed, what will happen, and where resources should go. When it fails, those errors shape decisions about people’s lives.

This distinction matters more than most conversations acknowledge. “AI” is often treated as a single category, flattening fundamentally different capabilities and risks.

I suspect I’ll keep using this distinction long after I’ve forgotten the specifics of the book.

What This Book Gets Right

The authors situate today’s AI boom within a longer history of technologies that promised more certainty than they could deliver. Their argument is not that AI doesn’t work—it’s that institutions repeatedly ask these systems to do more than they can reliably support.

What makes the book compelling is its restraint. It acknowledges the real capabilities of generative AI while remaining sharply skeptical of predictive systems deployed in hiring, policing, education, and healthcare.

At its core, this is not a book about AI.

It is a book about claims:
Who makes them, how they are validated, and what happens when statistical outputs are mistaken for judgment.

Skeptic’s Corner

My main critique is one the authors couldn’t fully avoid: the book was written at a moment the technology is outrunning.

Some examples will age. Models are improving faster than the cautionary cases the authors build around them, and readers coming to this in two years may find the specific claims less grounded than the underlying argument deserves.

The generative-versus-predictive distinction is genuinely useful, but agentic systems are already complicating it. A system that reasons, recommends, and acts is not cleanly one or the other. The authors don’t fully reckon with what happens when the boundary moves.

The deeper limitation is structural. This is a book written for skeptics who need permission, not for practitioners who need guidance. If you are already inclined to question AI claims, the book confirms your instincts. If you are inside an organization trying to figure out where to draw the line, you will finish it knowing what to distrust without knowing what to do instead.

That’s a real gap. And it’s not unique to this book. It’s the limitation of the genre.

Top 3 Signals I’m Taking Forward

1. AI Is Not One Thing
   Conversations about AI often collapse radically different systems into one category. The generative vs. predictive distinction helps explain why different applications carry different risks and deserve different levels of trust.

2. Overconfidence Is a Recurring Pattern
   The arc of AI is not new. Many technologies have promised objectivity and certainty before their limits became clear. Today’s moment fits a familiar cycle: optimism, adoption, and eventual reckoning.

3. The Hard Problem Is Delegation
   Every AI system encodes a decision about what gets automated and what remains human. Organizations optimize for speed and scale, but context, accountability, and judgment are harder to operationalize.

The real design challenge is not the model; it is the boundary.

Signal Worth Your Time?

Yes. Not because it made me more skeptical, but because it sharpened the questions I ask, especially about claims, trust, and where judgment belongs.

Subscribe now

Building Judgment

I write about what it takes to lead, decide, and develop people during times of transition

The Mental Model

Bloomers, Doomers, Zoomers, and Gloomers. Hoffman’s taxonomy earns its keep as a shorthand for the dominant AI positions, which too often collapse into either uncritical enthusiasm or reflexive panic. It is portable, legible, and useful enough to structure a real conversation. Worth stealing.

What It Gets Right

The historical analogies do meaningful work. The printing press, automobile, and GPS are used to argue that precaution can delay benefits without fully preventing harm, and that competition, when paired with credible benchmarking, can become a form of distributed regulation.

The “cognitive exoskeleton” frame is also pointed in the right direction. AI extends human capability; it does not simply replace it. That is a necessary baseline for any serious conversation about AI at work.

Skeptic’s Corner

“Tech billionaire thinks AI is good” is not incidental; it is the epistemic frame. The book underweights climate infrastructure costs and is notably casual about data privacy. These are not edge concerns. They are structural externalities of the system being advocated for.

More importantly, the argument for AI as an amplifier is underdeveloped. If AI amplifies human decision-making, the critical question is: what is it amplifying?

Hoffman treats human judgment as a reliable input. The evidence suggests that the assumption deserves more scrutiny. Judgment is not static. It is trained through repetition, degraded by disuse, distributed unevenly across systems, and shaped by incentives. When AI removes the low-stakes repetitions that build intuition, it may also erode the very capability it depends on.

The exoskeleton metaphor assumes the body remains intact. That assumption is doing more work than the book acknowledges.

Three Signals I’m Taking Forward

1. Benchmarking as “Regulation 2.0”
   The interesting move here is shifting governance from rules to measurement. Benchmarks, if designed well, do not just evaluate systems; they shape behavior upstream. They determine what gets optimized, what gets deprioritized, and what becomes visible enough to govern.
   

   In that sense, benchmarks are not neutral instruments. They are policies encoded as metrics. For product and organizational design, this raises a sharper question: who defines the benchmark, and what failure modes does it make invisible?

   If benchmarking becomes the de facto governance layer for AI, then designing benchmarks is designing the system. That work is still immature and wide open.
   

2. Taxonomies as a coordination infrastructure
   The Bloomer/Doomer/Zoomer/Gloomer framework works not because it is precise, but because it travels. It compresses a messy landscape into something a room can quickly align around, allowing for a real conversation instead of a definitional argument.
   

   The sharper question is who builds the vocabulary and what that choice makes invisible. Every shared frame has a center of gravity. This one is optimist-adjacent by construction, which means the failure modes it surfaces are the ones that fit inside a growth narrative. The risks that do not fit that shape tend to disappear into the “Gloomer” bucket and no longer receive scrutiny. Legible taxonomies are not neutral. They are governance. Design them accordingly.
   

3. Amplification requires substrate design
   “AI as amplifier” is only meaningful if the thing being amplified is intentionally maintained. This is where the book stops short. In real systems, human judgment is not a fixed asset. It is trained, degraded, distributed, and unevenly held. If AI systems remove the small, repetitive decisions that build intuition, they may quietly weaken the very capability they rely on.

   
   So the design problem shifts. It is not just about how to integrate AI into workflows. It is how to preserve and regenerate human judgment inside them. That may mean designing for friction, rehearsal, visibility into reasoning, and moments where humans still have to practice, not just approve. Amplification is not a property of the tool. It is a property of the system in which the tool is embedded.

Signal Worth Your Time?

Yes, with calibrated skepticism.

Read it as one of the more coherent versions of the optimist case. Then pressure-test its central premise: what exactly is being amplified, and has your organization done anything to ensure it still exists?