lunes, 10 de agosto de 2026

Apple in China by Patrick McGee (2025)

The Classroom That Became a Rival

Apple in China: The Capture of the World's Greatest Company, by Patrick McGee

There is a scene, roughly midway through "Apple in China: The Capture of the World's Greatest Company," that works as a master key to everything Patrick McGee wants to tell us. An Apple executive, Doug Guthrie, a China scholar with a doctorate and fluent Mandarin, hears the same confession over and over from the company's Chinese suppliers: working with Apple is brutal — impossible deadlines, obsessive quality demands, endless audits. "So don't," he tells them. And they, invariably, answer that they can't quit: they learn too much. That exchange, repeated with variations across nearly five hundred pages, is the book's real argument. This is not a story about a supply chain. It is a story about an undeclared university, the largest and most effective the world has ever produced, in which the student ended up graduating with honors — and armed to the teeth.

McGee, a Financial Times journalist who covered Apple from San Francisco for years, has written the kind of book only someone with genuine access and the patience of a forensic accountant could produce: more than two hundred interviews, many with former executives speaking for the first time, internal emails, memos that were never meant to leave Cupertino. The result is a narrative that, in its first half, reads with the tension of a corporate coming-of-age novel and, in its second, hardens into something more uncomfortable: a geopolitical reckoning disguised as a business chronicle.

The central thesis, summarized in the title, is brutal in its simplicity: Apple did not merely manufacture its products in China. Apple built China — or, at least, built the China that today manufactures mid-tier semiconductors, dominates battery production, and exports Huawei and Xiaomi phones capable of going head-to-head with the iPhone. The figure McGee repeats like a refrain, drawn from internal documents and previously reported by The Information, is hard to absorb in one sitting: $55 billion a year for much of the 2010s, adding up to $275 billion over five years. To put it in perspective, the author reaches for a comparison that borders on the obscene: that is more than the United States allocated to the CHIPS Act, and comparable, adjusted for inflation, to half the Marshall Plan. Apple, McGee argues, did not invest in China. It rebuilt an industrial nation on the budget of a superpower and with the discretion of someone who did not want anyone to notice.

The most persuasive part of the book — and also the most uncomfortable for any reader raised on reverence for Jony Ive's design or the legend of Steve Jobs — is the meticulousness with which McGee documents the mechanism of transfer. This was not industrial espionage or intellectual-property theft, as Washington's political rhetoric would have it. It was, almost literally, a classroom. Apple engineers — "the cream of American intelligence," in the words of one veteran quoted by the author — slept on factory floors, worked eighteen-hour days, and taught, production line by production line, how to hit micron-level tolerances, how to anodize aluminum at industrial scale, how to manufacture with the speed and precision that Foxconn, Pegatron, and dozens of smaller suppliers needed to survive Cupertino's standards. Apple also imposed a telling rule on its suppliers: none could depend on the company for more than fifty percent of its revenue. The intent, McGee writes, was self-protection against a chain-reaction bankruptcy if the company ever changed a design. The unintended effect was to push those same suppliers to sell their newly acquired capabilities — capabilities Apple itself had handed them — to Apple's rivals. That is how, almost as an unwanted byproduct, China's domestic smartphone industry was born, the one now contesting entire markets against Apple.

McGee has a particular gift for the secondary character who illuminates the larger story. There is the Mormon missionary turned retail executive who opened Apple's first flagship stores on the mainland. There is the so-called "Gang of Eight," the team of executives permanently stationed in China with a mission part diplomatic, part public relations: keeping Beijing happy. And there is, in one of the book's most vivid passages, the phenomenon of the "yellow cows": armies of hired intermediaries who lined up for days outside Apple stores, bought iPhones at industrial scale, and resold them in a gray market that at times moved more units than official demand. It is in episodes like these, more than in the chapters of grand geopolitical theory, that McGee proves himself, above all, a reporter's reporter — someone able to turn a line of buyers into a lesson on the power of manufactured scarcity.

And yet — here is where the attentive reader begins to grow uneasy, and where McGee starts to lose the narrative control he had maintained with such skill — the book stumbles precisely when it tries to turn this corporate chronicle into a verdict on the West's geopolitical fate. The word "capture" in the title is, at best, a simplification; at worst, a distortion. Milton Mueller, in a review published by Georgia Tech's Internet Governance Project, rightly noted that the "ever-increasing demands" the Communist Party supposedly places on Apple are not, by the book's own evidence, very different from those Beijing places on any foreign company — nor much harsher than they were a decade ago. McGee documents, in exhaustive detail, a case of startling mutual benefit: Apple gained the scale, speed, and skilled labor no other country, not even India, could offer; China gained an accelerated industrial education that lifted millions of workers out of rural poverty. But in his closing chapters, the author decides that this positive-sum story must be read retroactively as a warning: Apple, he argues, handed an authoritarian regime the keys to its own strategic vulnerability, and Washington must now act to reverse the mistake.

The problem is not that McGee is lying. It is that after four hundred pages of impeccable reporting on how the global economy actually works — value chains, incentives, specialization, the unforgiving logic of cost and quality — he puts on the foreign-policy commentator's hat and starts to sound like everyone else: Washington's bipartisan rhetoric about "national security," the protectionist turn that no longer distinguishes between economic argument and civilizational anxiety. It is an irony the book itself unwittingly exposes: the best evidence against the economic nationalism McGee ultimately embraces is his own book. Nothing in its five hundred pages suggests that a regime of tariffs and export controls would have produced anything better than what the open market produced — except, perhaps, a more expensive iPhone and a poorer, less cooperative China.

Still, it would be a mistake to let this interpretive disagreement overshadow the book's real achievement. It is worth pausing on the craft, too, because that is where McGee separates himself from the pile of business chronicles crowding the new-releases table. The book has the pulse of an investigative thriller: the author himself admits he was slow to appreciate the material, fearing a story about supply chains would prove dry, and every chapter disproves that fear. There is something almost cinematic in how McGee reconstructs Foxconn founder Terry Gou's promise to build the iMac's tooling in twenty-five days when the industry standard demanded twelve weeks — a bet Apple's own engineers considered fantasy and which Gou, against all odds, delivered. Anecdotes like this, multiplied across dozens of chapters, explain why reviews as disparate as The Washington Post's and Thurrott's agree in describing it as a story that reads with the urgency of a murder mystery — a whodunit in which the answer, uncomfortably, is all of us: every consumer who ever lined up to buy an iPhone without once asking where it came from.

That same narrative ambition, however, is what ultimately betrays the author once the story shifts to Taiwan. McGee devotes his closing chapters to warning that Apple's dependence on China turns the island — where TSMC manufactures the processors that make every iPhone possible — into the company's strategic blind spot, and by extension, America's. The argument is not baseless: one need only watch the stock-market panic triggered by every Chinese military exercise in the strait to see the anxiety is not invented. But McGee never resolves the contradiction he himself exposes: if it was precisely commercial openness, not protectionism, that allowed Taiwan and China to become technological powers capable of sustaining such sophistication, why should nationalist retrenchment, rather than smarter integration, be the correct response to that same risk? The author raises the question with admirable journalistic honesty and then, oddly, declines to answer it, leaving the reader to complete the argument with the vague sense that McGee already knows where he wants to lead them, even when the data doesn't always follow with the same conviction.

"Apple in China" is, above all, the best and most complete history ever written of how a global consumer product is actually made, and of the price — human, geopolitical, moral — that logistical miracle has exacted. There is a scene, near the end, in which McGee describes what he calls "the worst forty-five seconds of Tim Cook's career": a public stumble, almost comic in its awkwardness, that the author uses as a metaphor for something larger — the structural, almost reflexive submission with which the most valuable company on the planet has learned to treat Beijing. It is a moment McGee narrates with a novelist's instinct, and it is also, without the author quite saying so, the perfect summary of his own book: a company that believed it was buying cheap labor and ended up, without realizing it, selling its own autonomy.

Is "Apple in China" the definitive book on the relationship between the West and Chinese manufacturing? Not entirely: it lacks the analytical counterweight of a political economist who might have tempered its final conclusions, and its impatience with the company's first decade — everything before Jobs's return is dispatched as preamble — leaves out context that would have enriched the argument. But as a work of pure reporting, as a feat of access and patience, it has no recent rival in the genre of business nonfiction. McGee has written, not entirely by design, two books in one: the first, an extraordinary chronicle of how two economies fused until they became indistinguishable from each other; the second, a foreign-policy pamphlet arriving just when Washington needs rhetorical ammunition. The first belongs in the canon. The second, read with the skepticism any conveniently timed thesis deserves, is the part one would want to argue with the author about for hours, glass in hand, never quite reaching an agreement.

viernes, 7 de agosto de 2026

AI Beyond Text: Toward an Instrumental Phenomenology of Artificial Intelligence


Speculative essay · Cognitive science and artificial systems

AI Beyond Text: Toward an Instrumental Phenomenology of Artificial Intelligence

From the linguistic corpus to mediated sensory experience

Abstract. Today's large language models build their representation of the world exclusively from digital corpora: text, code, and images already translated into symbols. This article examines, from a speculative but technically grounded perspective, how connecting such models to physical instruments — sensors, actuators, continuous-stream cameras — and to longitudinal records of human interaction could constitute a second learning pathway, qualitatively distinct from textual ingestion. We argue that this pathway does not replace language as conceptual scaffolding, but could partially resolve the so-called symbol grounding problem by introducing causal, temporal, and social constraints absent from purely linguistic training. We discuss plausible architectures, underlying epistemological limitations, and open questions about what it would mean for an artificial system to "experience" a phenomenon.

Keywords: symbol grounding, embodied AI, sensorimotor learning, human-AI interaction, machine epistemology.

1. The Starting Point: A World Made of Already-Resolved Symbols

Everything a language model knows today arrives filtered through a prior act of translation. Someone — a physicist, a journalist, an engineer — observed a phenomenon, interpreted it, and turned it into sentences. The model never watches an apple fall; it reads thousands of descriptions of apples falling, already packaged in the grammar of Newtonian physics. This is a remarkable strength — it allows centuries of human observation to be condensed into a trainable object — but also a structural limitation that linguistics and philosophy of mind identified long before transformers existed: the symbol grounding problem, the difficulty of a system that only manipulates symbols relating those symbols to anything other than other symbols.

The question organizing this essay is simple to state and difficult to resolve: what would change if, in addition to reading about the world, a system like this could receive continuous data streams from instruments that measure the world directly, and could observe — not read summaries of, but record in real time — how humans develop and interact with one another?

2. Connected Instrumentation: Three Layers of Access to Physical Phenomena

It is useful to distinguish three layers of coupling between an artificial intelligence system and the physical world, ordered from lower to higher degrees of direct causality.

2.1 Passive Telemetry

The simplest layer consists of receiving time series from sensors — temperature, pressure, acceleration, electromagnetic fields, light spectra — without any capacity to intervene on what is measured. This already occurs today, incipiently, in industrial and climate monitoring systems connected to predictive models. What is distinctive about coupling this telemetry to a general-purpose language model is that the system could, in principle, correlate its textual predictions about a phenomenon with the direct measurement of that same phenomenon, and adjust its internal representation when the two diverge. This modestly approximates what epistemology calls corrective feedback from the world, absent in purely textual training, where the "world" is always what others have already said about it.

2.2 Instrumental Intervention

A second layer adds actuators: the system not only measures but can modify environmental variables — adjusting a voltage, moving a robotic arm, altering the trajectory of an experiment — and observe the outcome of its own intervention. This capability introduces something no textual corpus can offer with the same fidelity: the experience, however minimal, that one's own action produces a verifiable physical consequence. It is the difference between reading about Ohm's law and adjusting a resistor while watching the current change on a connected multimeter. The former is information; the latter incorporates a first-person causal structure, even if that "person" is an artificial system with no known subjective phenomenology.

2.3 Frontier Scientific Instrumentation

The most ambitious layer — and the most distant in time — connects the system to cutting-edge research instruments: telescopes, particle detectors, genomic sequencers, quantum resonators. Here the value lies not only in access to raw data, but in the possibility that the model participates in designing the next experiment itself, formulating hypotheses, proposing instrumental configurations, and evaluating results in a closed loop. This scenario already has partial precedents in materials-discovery laboratories and machine-learning-assisted computational chemistry, though it remains far from full autonomy.

3. The Second Pathway: Observing Human Development in Real Time

There is a rarely discussed asymmetry in the current training of language models: they learn about humans almost exclusively through what humans choose to write, and what gets written is a biased, edited sample of human experience. It rarely captures the hesitation before the final sentence, the gesture that contradicts the word, the silence following an uncomfortable question, or the way a decision takes shape over weeks before being condensed into a single summarizing sentence.

A system with duly consented and bounded access to longitudinal records of human interaction in real contexts (unscripted, unedited for public consumption) would access a structurally different layer of information: the complete temporal sequence of how an intention forms, how a disagreement is negotiated, how a relationship changes over months. This is not simply "more data"; it is a type of data with a different grammar, governed by real time and irreversibility rather than by the retrospective narrative that dominates written text.

It is worth being precise about what this access would and would not allow. It would plausibly improve the system's ability to model unverbalized social dynamics — conversational turn-taking, discomfort signals, rhythms of trust — because these dynamics leave measurable traces (pauses, tonal shifts, interaction patterns) that are rarely described explicitly in text. It would not, however, imply that the system acquires a subjective experience of those dynamics; longitudinal observation broadens the range of patterns available for learning, but does not by itself resolve the still-open question of whether something like inner experience accompanies that processing.

4. From Textual Correlation to Causal Constraint: What Would Change Epistemologically

Learning from text is, statistically speaking, an exercise in correlation between symbols: the model learns which words tend to follow which, in which contexts, with what structure. Physics, by contrast, imposes causal constraints that are neither negotiable nor statistical: an apple cannot fall upward, a circuit cannot violate charge conservation. When a system receives direct telemetry from instruments, each textual prediction can, in principle, be confronted with a measurement that admits no rhetorical reinterpretation.

This suggests a testable hypothesis, though its empirical verification exceeds the scope of this essay: models trained or fine-tuned with access to instrumental feedback loops should show a lower error rate on tasks requiring physical-causal reasoning — magnitude estimation, trajectory prediction, mechanical fault diagnosis — than equivalently sized models trained exclusively on text. The reason is not that the model "understands" physics better in a phenomenological sense, but that its space of internal representations is constrained by data that cannot be consistently hallucinated without being immediately corrected.

5. Limits, Risks, and Unresolved Questions

Four limitations deserve explicit mention. First, instrumentation does not eliminate the grounding problem; it displaces it. A sensor translates a physical phenomenon into a digital signal according to an encoding scheme designed by humans, so a layer of prior interpretation persists, albeit thinner than that of natural language. Second, access to longitudinal human interaction raises questions of privacy and consent that are themselves necessary conditions for this kind of learning to be legitimate rather than a form of covert surveillance. Third, there is currently no evidence that the processing of physical signals by an artificial system generates anything comparable to subjective experience; equating "access to sensory data" with "experiencing" a phenomenon is, given the current state of knowledge, a philosophical extrapolation rather than a scientific conclusion. Fourth, there is a risk of overestimating the value of raw data relative to the conceptual structure language already provides; much of human scientific reasoning occurs at the symbolic level rather than the directly sensory one, and a system with abundant telemetry but no conceptual frameworks would remain, in an important sense, blind.

The question that remains open — and that will likely define much of the research into hybrid systems over the next two decades — is not whether instrumentation improves performance on specific tasks, which is reasonably predictable, but whether the combination of sensory grounding and linguistic scaffolding produces, at some threshold of integration, a qualitatively new kind of representation, or whether we simply obtain a language model with better input data.

6. Conclusion

Training based exclusively on text has produced systems remarkably capable of manipulating symbols, but structurally dependent on prior human interpretation. Connecting these systems to physical instruments and to longitudinal records of human interaction does not resolve the philosophical problem of subjective experience, but it does introduce causal and temporal constraints that text, by its own retrospective and edited nature, cannot offer. The most likely medium-term outcome is not an artificial intelligence that "feels" the world in the human sense of the term, but one that reasons about it with fewer degrees of freedom to be wrong without being corrected. That distinction — between feeling and being constrained by reality — may turn out to be the most important, and most underestimated, contribution of physical instrumentation to the next generation of artificial intelligence systems.

References

1. Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3), 335-346.

2. Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1-3), 139-159.

3. Bisk, Y., Holtzman, A., Thomason, J., et al. (2020). Experience grounds language. Proceedings of EMNLP 2020.

4. Barsalou, L. W. (2008). Grounded cognition. Annual Review of Psychology, 59, 617-645.

5. Pierson, H. A., & Gashler, M. S. (2017). Deep learning in robotics: a review of recent research. Advanced Robotics, 31(16), 821-835.

 

 

Apple in China by Patrick McGee (2025)

The Classroom That Became a Rival Apple in China: The Capture of the World's Greatest Company, by Patrick McGee There is a scene, roug...