334digital.com

Tag: artificial intelligence

  • The Silent Search

    On the Probability, Semiotics, and Politics of Detecting Extraterrestrial Intelligence

    I. The Universal Language That Isn’t

    There is a story we tell ourselves about first contact, and it goes like this: mathematics is the universal language. Any sufficiently advanced civilisation, regardless of its biology, its sensory apparatus, its evolutionary history, will have discovered the same primes, the same geometric constants, the same physical relationships. If we want to speak to the stars, we encode mathematics, because mathematics is what we share. It is the lingua franca of the cosmos, the one thing that is true everywhere and independent of the speaker.

    This is a cultural claim dressed as a scientific one. It is, more precisely, a semiotic claim, a claim about the relationship between signs, meaning, and minds, and it has rarely been examined as one.

    The discipline that would examine it is semiotics: the study of how meaning is made, transmitted, and received. Semiotics, in the tradition of Charles Sanders Peirce, holds that a sign requires three things, a sign vehicle (the physical form of the sign, the signal itself), an object (the thing the sign refers to), and an interpretant (the meaning constructed by the receiver). A sign is not a sign in itself; it is a sign only in the act of interpretation. Meaning does not reside in the signal. It is constructed, by the receiver, from the signal, using the only tool available: context.

    This matters more than it might seem. When we say that mathematics is universal, we are making a claim about the sign vehicle, that the same mathematical structures will be independently discovered by any intelligence. This is plausible, though not as certain as we like to think; even mathematics, as a human practice, is shaped by culture, by language, by the contingent history of the species that produced it. The ancient Greeks did geometry differently from the medieval Islamic mathematicians, who did it differently from the calculus tradition that emerged in Europe. These are not different mathematics, the relationships are the same, but they are different semiotic systems, different ways of encoding and interpreting the same underlying structures. If mathematics were truly transparent, if the sign vehicle mapped unambiguously to the object, then mathematical notation would not have a history. It does.

    But the deeper problem is not at the level of the sign vehicle. It is at the level of the interpretant. Even if an extraterrestrial intelligence has discovered the same mathematical relationships, the meaning they construct from a signal encoding those relationships depends on their context, their biology, their sensory world, their temporal experience, their cognitive architecture. A signal that encodes prime numbers, for us, means “intelligence, because primes are a non-natural pattern that requires a mind to generate.” This interpretation depends on a chain of assumptions: that the receiver recognises primality as a non-natural pattern; that they associate non-natural patterns with intelligence; that they associate intelligence with communication; that they associate communication with the intent to be understood. Each of these is a cultural inference, grounded in the specific semiotic world of a species that evolved on a particular planet with a particular evolutionary history.

    The Estonian biologist Jakob von Uexküll coined the term Umwelt to describe the perceptual world inhabited by an organism, the specific slice of reality available to it through its particular sensory and cognitive apparatus. A tick’s Umwelt consists of temperature, light, and the smell of butyric acid. A bat’s Umwelt is built from echolocation. A human’s Umwelt is built from vision, hearing, and a peculiar cognitive architecture that is, as far as we know, unique: the capacity for recursive symbolic thought, which has produced language, mathematics, and the stories we tell about the stars.

    The assumption underlying SETI, the Search for Extraterrestrial Intelligence, is that the Umwelt of a technological alien species would overlap sufficiently with ours that a signal meaningful in one would be meaningful in the other. This is an assumption, not a fact. We have no evidence for it, because we have no data. We have one data point: ourselves. And the one data point we have is not a sample; it is a case study, and a case study of one tells you nothing about the distribution of the population from which it is drawn.

    II. The Imperative of Life

    There is a second assumption, deeper and less examined than the first, that underlies the entire enterprise. It is the assumption that life, once it arises, tends toward intelligence, and that intelligence, once it arises, tends toward communication, and that communication, once it is possible, tends toward the stars. This is a narrative of progress, a kind of cosmic teleology, and it is, in its structure, remarkably similar to the Great Chain of Being, the medieval concept that all of creation is arranged in a hierarchy from the lowest to the highest, with humanity (naturally) near the top. We have secularised the Great Chain. We no longer place God at the top. We place technological civilisation there instead, and we assume that the universe, if it contains life at all, will produce civilisations like ours, because ours is the natural endpoint of the chain.

    This is not what the observable facts suggest.

    The observable facts are these. Life on Earth arose approximately 3.8 billion years ago, possibly earlier. For the first 2.5 billion years of that history, more than half the planet’s lifetime, life consisted of single-celled organisms. Multicellularity arose perhaps 600 million years ago. Complex multicellular life, with differentiated tissues and organs, arose perhaps 550 million years ago. The first animals appeared in the Cambrian, roughly 540 million years ago. From the Cambrian to the present, life on Earth has experienced at least five major mass extinction events, each of which reset the trajectory of evolution and eliminated a significant fraction of existing species.

    Intelligence, in the sense relevant to SETI, which is to say, intelligence capable of building technology capable of transmitting or receiving interstellar signals, has arisen exactly once in 3.8 billion years. Not twice. Not in multiple lineages. Once. In a single species, Homo sapiens, which has existed for approximately 300,000 years and has been technologically capable of radio transmission for approximately 120 years.

    Let us sit with those numbers. 3.8 billion years of life. One instance of the kind of intelligence we are searching for. 120 years of the kind of technology we are searching with. The fraction of life’s history on this planet during which it has been detectable by the methods SETI uses is approximately:

    120 / 3,800,000,000 ≈ 3.2 × 10⁻⁸

    Three ten-millionths of the time life has existed. And that is the fraction for our own planet, where we know life exists. For any other planet, we do not even know whether life exists, let alone intelligence, let alone technology.

    This is not an argument that life is rare. It is an argument that the specific thing SETI is searching for, a technological civilisation, producing detectable electromagnetic signals, during the narrow window in which we are listening, is, on the evidence available to us, vanishingly improbable. Not impossible. Not absent. But the probability is not a number we can calculate, because we have one data point, and one data point cannot populate a distribution.

    What we can say is this: the assumption that intelligence is convergent, that it arises independently in many lineages the way that eyes have evolved independently dozens of times, is not supported by the evidence. On Earth, we have many examples of social complexity (eusocial insects, cetaceans, elephants, corvids, cephalopods), many examples of tool use (chimpanzees, crows, otters, dolphins), many examples of cognitive sophistication (parrots, octopuses, pigs, dogs). We have exactly one example of technology-building intelligence. Eyes have evolved independently because light is a physical constant and the selective advantage of detecting it is overwhelming. Technology-building intelligence is not eyes. It is a specific, contingent, historically bounded phenomenon that arose in a specific lineage under specific conditions, and the evidence that it is a convergent adaptation, that it arises reliably, like vision, wherever the conditions permit, is, at present, zero.

    This does not mean it is rare. It means we do not know. And the not-knowing is the honest position, and the not-knowing is the position that the SETI industry, as a matter of institutional logic, cannot afford to occupy.

    III. The Signal Problem, What We Are Actually Looking For

    Setting aside the semiotic and philosophical questions, let us consider the physics. What would it actually take to detect an extraterrestrial signal?

    The problem has three components: attenuation, noise, and coincidence.

    Attenuation

    Electromagnetic radiation propagates according to the inverse square law. A signal that leaves a transmitter with power P will arrive at a distance d with power density P / (4πd²). At interstellar distances, this is a brutal fact. A signal transmitted from a star 6,000 light-years away, roughly the distance to the Cygnus arm, a reasonable estimate for the scale of any plausible target, arrives at Earth with a power density that is a factor of roughly 10⁻¹⁵ of its transmitted value. To detect a signal at that distance, the transmitter must either emit with enormous power or beam the signal tightly in our direction.

    A tight beam is energy-efficient but covers a tiny fraction of the sky. The beamwidth of a transmitting antenna is inversely proportional to the ratio of the antenna’s diameter to the wavelength. To beam a signal at 1.42 GHz (the hydrogen line, a popular SETI frequency) to a target the size of the Earth’s orbit from a transmitter 6,000 light-years away requires an antenna with an effective diameter of approximately 500,000 kilometres, larger than any structure any civilisation we can imagine could build, unless it is distributed across a solar system. And even then, the beam would only reach the Earth if the transmitter knew exactly where the Earth was and was pointed at it with precision.

    The alternative is an omnidirectional beacon, a signal broadcast in all directions, detectable from any point in the sky. But an omnidirectional signal at interstellar distances requires power that beggars the imagination. A signal detectable by the Arecibo telescope (when it existed) at 1,000 light-years, broadcast omnidirectionally, would require a transmitter power of approximately 10¹⁵ watts, roughly 100 times the total current electrical generating capacity of human civilisation. At 6,000 light-years, the requirement scales by a factor of 36, to roughly 3.6 × 10¹⁶ watts, a significant fraction of the total solar output intercepted by the Earth.

    This is not impossible. A civilisation that has access to the energy of a star, a Kardashev Type II civilisation, which is the kind that science fiction routinely assumes, could, in principle, broadcast omnidirectionally at interstellar distances. But we are now no longer doing science. We are doing speculation, and the speculation is unconstrained by any observation, which means it is, in Karl Popper’s sense, unfalsifiable, which means it is not, strictly speaking, science at all.

    The attenuation problem means that the vast majority of plausible transmission scenarios are undetectable. A civilisation that is broadcasting toward us, with a beam pointed at our solar system, using a power level we could detect, at a frequency we are monitoring, at a time we are listening, is a civilisation that is, by construction, very similar to us in its assumptions about how communication works. We are, in effect, searching for ourselves.

    Noise

    The universe is loud. The cosmic microwave background provides a floor of approximately 2.7 K of thermal noise at all frequencies. Galactic synchrotron radiation adds frequency-dependent noise that is particularly strong at lower frequencies. Earth’s atmosphere emits thermal noise. The receiver itself contributes thermal noise, quantified by its system temperature. And, increasingly, our own civilisation contributes radio frequency interference, satellites, radar, broadcast transmitters, industrial equipment, that contaminates the very frequencies we are searching for.

    The signal-to-noise ratio determines detectability. A signal is detectable only if it is significantly stronger than the noise floor integrated over the observation time. For a narrowband signal, the kind SETI traditionally searches for, because a narrowband signal is an indicator of engineering, of a transmitter that has been deliberately designed, the noise in a single frequency bin is proportional to the system temperature divided by the integration time. Longer observations reduce noise, but longer observations cost more telescope time, and telescope time is the scarcest resource in the SETI enterprise.

    The “water hole”, the band of frequencies between the hydrogen line (1.42 GHz) and the hydroxyl line (1.66 GHz), is relatively quiet, because galactic background noise decreases with frequency and the atmosphere is transparent in this range. The choice of the water hole is, however, a semiotic choice: it assumes that an alien civilisation would choose to broadcast in this band for the same reasons we would choose to listen in it. The reasoning is that hydrogen and hydroxyl are the components of water, and water is the solvent of life (as we know it), and therefore the band between them is a “cosmic meeting place”, a frequency band that any water-based life would find significant.

    This is a lovely idea. It is also, in its structure, an argument from analogy: we find the hydrogen line significant because hydrogen is the most abundant element in the universe and its spectral line is a natural reference frequency. We assume that an alien intelligence would find it significant for the same reasons. But the significance is not in the frequency; it is in the interpretation. And the interpretation depends on the interpretant, which depends on the Umwelt, which depends on the species.

    Temporal Coincidence

    A civilisation must be transmitting during the window in which another civilisation is listening. We have been listening, in one form or another, for approximately 60 years. If technological civilisations have lifetimes of, say, 10,000 years (a number chosen for illustrative purposes, because we have no basis for choosing any number), and if they arise at some rate per galaxy per unit time (a rate we cannot estimate because we have one data point), then the probability of temporal overlap depends on the ratio of the combined listening-and-transmitting window to the total time available. If both civilisations are rare and short-lived, the probability of overlap is small. If both are common and long-lived, it is large. We do not know which, and the range of plausible values spans many orders of magnitude.

    The one observable fact we have is that we have been listening for 60 years and have heard nothing. This is a null result, and null results in science are informative but notoriously difficult to interpret. A null result can mean the phenomenon does not exist. It can mean the phenomenon exists but is below the detection threshold. It can mean the phenomenon exists and is detectable but we are looking in the wrong place, at the wrong frequency, at the wrong time, or with the wrong assumptions. The history of science is rich with null results that were, in retrospect, failures of imagination rather than failures of the phenomenon, the Michelson-Morley experiment, the early null results in the search for exoplanets, the decades of null results in gravitational-wave detection before LIGO.

    But the history of science is also rich with null results that were, in retrospect, correct: the search for the ether, the search for Vulcan, the search for N rays. The difference between a null result that precedes a detection and a null result that confirms an absence is visible only in retrospect, and we are not in retrospect. We are in the present, and the present is ambiguous.

    IV. The Semiotic Trap

    Here is the trap, and it is a semiotic one: we cannot search for a signal without a model of what a signal looks like. And our model of what a signal looks like is, inescapably, a model of what our signal would look like. We search for narrowband carriers because we build narrowband carriers. We search in the water hole because we find the water hole significant. We search for prime numbers because we see primes as a signature of intelligence. We search for structured, repetitive, obviously non-natural patterns because structured, repetitive, obviously non-natural patterns are what we produce when we want to be noticed.

    The logic is circular, and the circularity is not a flaw that can be engineered away. It is structural. To search is to assume. To assume is to project. To project is to search for yourself.

    This is not an argument against searching. It is an argument for understanding what searching means, and for being honest about the fact that a null result in a search conditioned on a model that is derived from a sample of one tells you only that the model’s specific predictions are not confirmed. It does not tell you that the phenomenon is absent. It does not tell you that the model is wrong. It tells you nothing, in the strict statistical sense, because the prior is unconstrained and the likelihood is dominated by assumptions that are not testable.

    The semiotic literature has a term for the specific kind of error this produces: abduction, as defined by Peirce, is the process of inferring the best explanation for an observation. SETI is not, in practice, abductive; it is hypothetical-deductive. We hypothesise a model of alien communication, deduce what the signal would look like, and search for it. When we do not find it, we do not revise the hypothesis, we refine the search. We build more sensitive receivers. We search more frequencies. We observe more stars. The hypothesis, that a technological civilisation would produce a signal recognisable to us, is never tested, because it cannot be tested. It is the precondition of the search, not a result of it.

    The Italian semiotician Umberto Eco, in his work on the limits of interpretation, argued that a text can be interpreted in an unbounded number of ways, but not all interpretations are equally valid. The constraint on interpretation is the text itself, the sign vehicle, but also the community of interpreters, the shared context that makes some readings plausible and others not. For SETI, the “text” is the signal (if one exists), and the “community of interpreters” is the scientific community of Earth. The constraint on our interpretation is our own semiotic world, our physics, our mathematics, our engineering, our cultural assumptions about what communication is and why beings communicate. We cannot step outside this constraint. We can only be aware of it.

    V. The One Data Point

    Let us return to the observable facts, because they are all we have, and they deserve more respect than they typically receive.

    We know that life exists on at least one planet. We know that, on that planet, life has produced one technological civilisation. We know that that civilisation has been capable of radio transmission for approximately 120 years. We know that, during those 120 years, the civilisation has not deliberately broadcast a continuous, high-power, omnidirectional signal intended for reception by other stars. (We have sent short, targeted messages, the Arecibo message in 1974, the more recent Breakthrough Message competition, but these are not continuous beacons and would not be detectable by a civilisation like ours at interstellar distances.)

    This last fact is, in the context of SETI, almost never discussed. If we are the model for what a technological civilisation does, then the model predicts that a technological civilisation does not broadcast. It listens. It searches. It does not transmit, or it transmits only sporadically and with low power, in short bursts, toward specific targets.

    The implication is uncomfortable. If the one civilisation we know does not behave the way SETI assumes civilisations behave, i.e., continuously broadcasting detectable signals, then the search is predicated on a model that is contradicted by the only available data. We are searching for a behaviour that we ourselves do not exhibit.

    There are responses to this. The most common is that we are young, technologically, and that older civilisations would have moved beyond the listening phase into the broadcasting phase. This is plausible. It is also unfalsifiable, because we have no older civilisation to observe. Another response is that we are, in fact, broadcasting, our radio, television, and radar emissions leak into space continuously, and have been doing so for a century. This is true, but the power of these leaked signals at interstellar distances is far below the detection threshold of any plausible receiver. A civilisation like ours, at 1,000 light-years, would not be detectable by our own instruments. The leakage is real but too faint to be the signal we are searching for.

    The third response is the one that the semiotic analysis suggests: that we are not broadcasting because broadcasting, as a strategy for interstellar communication, does not make sense. The energy costs are prohibitive. The probability of being heard is unknown. The temporal coincidence required is, on any reasonable estimate, small. And the assumption that another civilisation would be listening, in the same frequency band, with the same model of what a signal looks like, during the same century, is a tower of assumptions, each individually uncertain, that collectively approaches the kind of improbability that, in any other scientific context, would be treated as fantasy.

    VI. The SETI Industry

    SETI has existed, as a formal enterprise, since 1960, when Frank Drake conducted Project Ozma, the first systematic radio search for extraterrestrial signals, using the 26-metre radio telescope at Green Bank, West Virginia. Drake listened to two stars, Tau Ceti and Epsilon Eridani, for a total of approximately 150 hours, at a single frequency (1.42 GHz), and heard nothing. The search has grown since then, in telescope size, in frequency coverage, in target number, in computational sophistication. It has not, in sixty years, found anything.

    This is not a failure, or not only a failure. It is a fact, and the fact has a context: the search space is vast, and the fraction of it we have explored is small. Jill Tarter, the longtime director of the Center for SETI Research, compared the explored volume of SETI’s search space to a glass of water taken from the ocean. If you take one glass of water from the ocean and find no fish, you have not proven that the ocean contains no fish. You have proven that this glass of water contains no fish. The ocean is large.

    The metaphor is apt, and it is also, in its way, a defence of the enterprise against its own null results. The ocean is large. We have sampled a glass. The absence of fish in the glass is not evidence of the absence of fish in the ocean. This is true. It is also, in a practical sense, the kind of argument that can be used to justify indefinite searching, because the ocean is always larger than the glass, and the glass can always be refilled, and the fish can always be somewhere else.

    The question is not whether the search should continue. The question is what the search costs, and what it produces, and whether the cost and the production are in a relationship that is sustainable or honest.

    Funding

    SETI has historically been funded by a mix of NASA, the National Science Foundation, and private sources. NASA’s funding for SETI was terminated by Congress in 1993, largely through the efforts of Senator Richard Bryan of Nevada, who declared that the search had found nothing and was a waste of taxpayer money. The field survived on private funding, the SETI Institute, founded in 1984, built the Allen Telescope Array with support from Paul Allen; the Breakthrough Listen initiative, launched in 2015 with $100 million from Yuri Milner, is the largest and best-funded SETI project in history, and it has access to some of the world’s most powerful radio telescopes, including the Green Bank Telescope in West Virginia and the Parkes Telescope in Australia.

    One hundred million dollars is, by the standards of scientific funding, not a great deal of money. The James Webb Space Telescope cost approximately $10 billion. The Large Hadron Collider cost approximately $9 billion. Breakthrough Listen, spread over ten years, costs $10 million per year. In the context of global scientific expenditure, this is a rounding error.

    But the comparison cuts both ways. If $100 million is a rounding error, then it is also $100 million that is not being spent on other things. The opportunity cost is not theoretical, it is the science that could have been done with the telescopes, the computing power, and the human capital that Breakthrough Listen has absorbed. The Green Bank Telescope, which spends a portion of its observing time on SETI, is a general-purpose radio telescope that could be used for pulsar timing, molecular spectroscopy, galaxy surveys, and the study of the interstellar medium. The time it spends listening for aliens is time it is not spending on those things. The computing infrastructure that processes SETI data, and SETI generates enormous volumes of data, requiring significant processing, could be processing data from other astronomical observations.

    The question is whether the expected return on SETI justifies the opportunity cost. The expected return is the probability of a detection multiplied by the value of a detection. The value of a detection is, by any measure, immense, a confirmed signal from an extraterrestrial intelligence would be the most significant scientific discovery in human history, and its implications would extend far beyond science into philosophy, religion, politics, and culture. But the probability is, as we have established, not a number we can estimate. It is, at best, a range so wide that the expectation value is not well-defined. If the probability is 10⁻²⁰, the expected return is negligible. If the probability is 10⁻², the expected return is enormous. We do not know where, in that range, the truth lies, and the range is so wide that multiplying it by any value produces a result that can be used to justify either continuing or stopping, depending on which end of the range you prefer.

    This is not a scientific question. It is a question about the allocation of resources in the face of radical uncertainty, and it is a question that the SETI industry, as an institution, has a structural incentive to answer in one direction.

    The Industry

    SETI is, at this point, an industry. It has institutions (the SETI Institute, the Berkeley SETI Research Center, the International Centre for Radio Astronomy Research). It has a workforce, scientists, engineers, software developers, administrators. It has a public profile, maintained through documentaries, popular books, conference talks, and the occasional viral news story about a “candidate signal” that turns out to be terrestrial interference or a known astrophysical phenomenon. It has a narrative, and the narrative is compelling: we are searching, the search is hard, the universe is vast, and the answer, if it comes, will change everything.

    Industries do not, as a rule, argue for their own dissolution. The SETI industry is no exception. The null results of the last sixty years have been interpreted not as evidence to revise the fundamental assumptions of the search but as evidence to expand it: more telescopes, more frequencies, more stars, more sensitivity, more computing power. The logic is that the search space is large and we have explored a small fraction of it, and therefore the appropriate response to null results is to search more, not to question whether the thing we are searching for exists in the form we are searching for.

    This is not unreasonable. It is also not science, in the strict sense, because the core hypothesis, that a technological civilisation would produce a signal recognisable to us, is not falsifiable by the search. If we search the entire sky, at all frequencies, for a century, and find nothing, the response can always be: the signal is there, but we are not looking at the right time, or the right modulation, or the right encoding, or with the right model. The hypothesis adapts to survive the evidence, and a hypothesis that adapts to survive all evidence is not a scientific hypothesis. It is an article of faith.

    There is a comparison to be made with particle physics. The search for the Higgs boson was, for decades, a search for a predicted but unobserved phenomenon. The difference is that the Higgs was predicted by a theory, the Standard Model, that made other, testable predictions, and the search for the Higgs was constrained by those predictions. If the Higgs had not been found in the predicted mass range, the Standard Model would have been falsified, and the search would have had a defined endpoint. SETI has no equivalent. There is no theory that predicts, with specific parameters, what a signal from an alien civilisation would look like, at what frequency, with what modulation, from what direction. There are conjectures, the water hole, the hydrogen line, the beacon hypothesis, but they are not predictions in the scientific sense, because they are not derived from a testable theory. They are arguments from analogy and from aesthetics, and arguments from analogy and aesthetics, however appealing, are not falsifiable.

    The Effect on Science

    The more subtle cost of SETI is not the money or the telescope time. It is the effect on the scientific culture, specifically, on the culture of astrobiology, the broader field that studies the origin, evolution, and distribution of life in the universe.

    Astrobiology is a legitimate and thriving science. It studies the chemistry of life’s origins, the conditions under which life can arise, the limits of life as we know it (extremophiles, subsurface biospheres, alternative biochemistries), and the detectability of life on other planets through biosignatures, atmospheric, geological, spectral. It is a field constrained by data: we can study extremophiles in the lab, we can analyse the atmospheres of exoplanets with telescopes like JWST, we can model the conditions on Mars, Europa, Enceladus, Titan. The data are limited but real, and the hypotheses are testable.

    SETI, as a subfield of astrobiology, has a tendency to dominate the public perception of the whole. When people think about the search for extraterrestrial life, they think about radio signals and intelligent aliens, not about methane plumes on Mars or phosphine on Venus or the spectral signature of vegetation on an exoplanet. This is a problem, because the more immediately testable and productive lines of astrobiological research, the search for biosignatures, the study of extremophiles, the characterisation of exoplanet atmospheres, are, in the public mind, subordinated to the more speculative and less productive search for engineered signals.

    The effect is compounded by the media, which prefers the SETI narrative because it is dramatic: signals from the stars, first contact, the question of whether we are alone. The discovery of a biosignature on a distant exoplanet would be a profound scientific result, but it would not, in the public imagination, compete with a signal from an alien intelligence. The SETI industry, by maintaining its public profile, shapes the public’s understanding of what astrobiology is and what it should fund, and the shape it imposes is one that privileges the speculative over the empirical.

    There is a further effect, less visible but more insidious: the SETI framing can distort the kinds of questions that scientists ask. The assumption that intelligence is the thing to search for, that intelligence is the endpoint of life, the thing that matters, can bias the kinds of biosignatures we look for and the kinds of planets we prioritise. We search for Earth-like planets because we assume that Earth-like planets are the most likely to produce Earth-like intelligence. This is reasonable, but it is also circular: we search for ourselves, and we justify the search by the assumption that the universe produces beings like us, and the assumption is grounded in the only data point we have, which is ourselves.

    VII. What If the Signal Is Not a Signal?

    Let us return to semiotics, because semiotics is the discipline that can most productively interrogate the assumptions of the search.

    If an alien intelligence exists, and if it communicates, the communication need not take the form of a narrowband radio signal modulated with a pattern recognisable to human cryptanalysis. It could take a form that we do not recognise as communication, because the form is grounded in a semiotic world that does not overlap with ours.

    Consider: a civilisation that communicates through gravitational wave modulation. This is theoretically possible, a sufficiently advanced civilisation could, in principle, manipulate massive objects to produce gravitational waves with a structured pattern. We have only recently developed the technology to detect gravitational waves at all (LIGO, 2015), and our sensitivity is limited to the most violent astrophysical events, merging black holes and neutron stars. A modulated gravitational-wave signal from an alien civilisation would be, with current technology, undetectable. But it is not impossible, and if it is happening, we would not know.

    Consider: a civilisation that communicates through neutrino beams. Neutrinos interact so weakly with matter that they pass through planets undisturbed, which makes them, in principle, an excellent medium for interstellar communication, no absorption, no scattering, no interference. But the difficulty of generating and detecting neutrinos is, by current technology, prohibitive. We can detect neutrinos from the sun and from supernovae, but only with enormous detectors buried deep underground, and only at very low event rates. An engineered neutrino signal would be, with current technology, indistinguishable from background.

    Consider: a civilisation that communicates through modifications to its environment that are detectable at interstellar distances, a Dyson structure that blocks or modulates the light of its star in a non-natural pattern. This is the technosignature approach: rather than searching for a signal, search for the artefact. A megastructure around a star would produce a distinctive light curve, a non-periodic, non-natural dimming pattern that could not be explained by planetary transits or stellar variability. The Kepler mission, which surveyed a portion of the sky for transiting planets, produced light curves for hundreds of thousands of stars, and a few, most notably Tabby’s Star (KIC 8462852), showed dimming patterns that were, briefly, consistent with a megastructure hypothesis. The dimming was later explained by dust, but the episode illustrates the principle: technosignatures are, in principle, detectable with existing technology, and they do not require the alien civilisation to be deliberately communicating. They require only that it is doing something big enough to see.

    The technosignature approach is, in semiotic terms, a shift from searching for a sign that is intended as communication to searching for an index, a sign that bears a physical connection to its object, like smoke to fire. An index does not require intent. It requires only a causal relationship between the sign vehicle and the thing it signifies. A Dyson structure is an index of a technological civilisation, whether or not the civilisation wants to be seen.

    This is, arguably, a more honest search, because it does not require us to assume that an alien intelligence shares our semiotic assumptions about communication. It requires only that an alien intelligence, if it exists and is sufficiently advanced, will do things that are detectable, will modify its environment in ways that are, to a sufficiently careful observer, non-natural. The search for technosignatures is, in this sense, a search for the side effects of intelligence rather than for intelligence itself. It is less ambitious, less romantic, and more grounded in observable fact.

    VIII. The Honest Position

    The honest position, given everything we know, is this:

    We do not know whether there is other intelligent life in the universe. We have one data point, ourselves, and one data point cannot populate a distribution. We do not know whether intelligence is rare or common, whether it tends toward communication or toward silence, whether it produces signals we would recognise or signals we would not. We do not know whether the assumptions underlying SETI, that mathematics is a universal language, that the hydrogen line is a cosmic meeting place, that an alien intelligence would broadcast in a form we can detect, are correct or are projections of our own semiotic world onto the void.

    What we know is that the universe is large, that the conditions for life exist in many places, and that life, once it arises, is persistent and adaptable. We know that the search for life, not necessarily intelligent life, but life in any form, is a scientific enterprise with testable hypotheses and available data. We can search for biosignatures in the atmospheres of exoplanets. We can study the chemistry of the interstellar medium. We can explore the moons of our own solar system for subsurface oceans and the chemical signatures of metabolism. We can do these things now, with existing technology, and the results, whether positive or negative, will be informative.

    The search for extraterrestrial intelligence, in its current form, is a search conditioned on assumptions that are not testable, funded by resources that have opportunity costs, and sustained by an industry that has a structural incentive to continue regardless of results. It is not, in its current form, a science, because its core hypothesis is not falsifiable. It is a practice, a disciplined, methodical, technologically sophisticated practice, and the practice may, one day, produce a result. But the practice is not the same as the science, and the distinction matters, because the conflation of the two, the tendency to treat SETI as if it were as rigorously grounded as, say, the search for exoplanets or the study of cosmic microwave background anisotropies, distorts the allocation of scientific resources and the public understanding of what we know and what we do not.

    The semiotic perspective offers a way out of this, not a solution, but a clarity. If we understand that the search is, inescapably, a search for ourselves, for a mind like ours, producing a signal like ours, in a form we can recognise, then we can be honest about what a null result means. It means that the specific kind of mind we are searching for has not been found in the specific places we have looked. It does not mean that minds do not exist. It does not mean that the universe is empty. It means that our model of what a mind looks like, when it communicates across the void, has not been confirmed, and that the model is derived from a sample of one and may be, in ways we cannot detect, wrong.

    The most productive thing we can do, in the face of this uncertainty, is to broaden the search, to look not only for the signals we expect but for the indices we do not. To search for biosignatures as well as technosignatures. To search for the side effects of life as well as the deliberate productions of intelligence. To search, in short, not for a mirror but for a window, a way of seeing the universe that does not assume that what we see will look like us.

    This is not a argument for stopping SETI. It is an argument for situating it, for understanding its place in the larger enterprise of astrobiology, for being honest about its assumptions, for acknowledging its opportunity costs, and for refusing the temptation, which all industries feel, to justify their own continuation by the infinity of the search space and the impossibility of proving a negative.

    The universe may be full of minds. It may be empty of them. We do not know, and the not-knowing is the honest position, and the honest position is the one from which the best science is done, because the best science is done not by those who are certain of what they will find but by those who are honest about what they do not.

    The signal, if it comes, will not be what we expect. It will be what it is. The question is whether we are listening in a way that allows us to hear it.

  • AI Alternate Portfolio

    What have we lost? This is the right question, because the current boom is not just adding something, it is actively crowding out something else.

    The Association for the Advancement of AI did a big study of its own researchers this year. 79% said public perception of what AI can do does not match reality, 74% said the direction of research is now being driven by hype because that’s what gets funded, and 76% said scaling up current large language models is unlikely or very unlikely to get us to general intelligence.

    In other words: we are pouring almost all the money into one bet — bigger transformers trained on more text — and leaving a whole set of older, slower, more rigorous ideas to starve.

    Here is what we have lost, or are losing:

    1. Systems that reason, not just predict

    Old-school symbolic AI — logic, theorem provers, knowledge graphs, rules — was unfashionable because it was brittle. But it could do something LLMs still cannot: prove an answer is correct, not just plausible.

    What was supposed to replace both is neuro-symbolic AI: pattern-recognition nets for perception, plus symbolic logic for reasoning. You get a system that can both see a cat and reason that if all cats are mammals, this cat is a mammal. It is explainable by design.

    That work is still alive — researchers are building knowledge-infused learning that makes black-box models explainable in healthcare, law, finance — but it gets a fraction of the funding because it doesn’t demo well as a chatbot.

    2. Causality instead of correlation

    LLMs are supreme correlation machines. They are terrible at causality. As one recent analysis put it, prediction cannot substitute for causal inference.

    Judea Pearl’s whole field — causal graphs, do-calculus, asking “what if we intervened?” — is exactly what you need for medicine, economics, climate, public policy. An LLM can tell you that ice cream sales and drownings correlate. A causal model tells you why, and what to do about it.

    That field has been eclipsed because it doesn’t scale with GPUs. It scales with careful human thought about how the world actually works.

    3. Embodied and grounded intelligence

    The original idea of AI was not a disembodied text predictor. It was an agent in a world. Rodney Brooks’ robots, developmental robotics, animal cognition — intelligence that learns by bumping into things, failing, feeling gravity.

    LLMs have no body, no senses, no continuity. They have never been cold, or hungry, or embarrassed. That is why they hallucinate: training rewards confident guesses over expressions of uncertainty.

    Embodied AI, world models, and active inference are coming back — researchers list them explicitly as departures from pure scaling already underway — but for five years they were told “just add more data.”

    4. Small data, efficient, and Bayesian intelligence

    Before the scaling hypothesis, a core goal was to learn like humans do: from few examples, with uncertainty, and with the ability to say “I don’t know.”

    Bayesian methods, probabilistic programming, minimum description length, analogical reasoning — all work that tries to make AI that knows what it doesn’t know. That is essential if you want to put AI in a plane or a hospital.

    LLMs do the opposite: they use all the data in the world to avoid having to be clever. The true cost of that corpus — books, code, art, decades of human labor — is estimated at 10 to 1,000 times the cost of the GPUs themselves. We are treating human knowledge as free to harvest.

    5. Theory

    The most worrying loss, according to the AAAI researchers, is theoretical AI research. Not building bigger things, but asking why things work.

    We have no solid theory of why transformers generalize, when they will fail, or what emergence even means. We have benchmarks, not understanding. The field is running on vibes and leaderboard scores. The scientists warning that this is slowing down real progress are not Luddites — they are the people who built the field in the 70s, 80s, 90s.

    What an alternative portfolio would have looked like

    If investment had not been monopolized by LLMs, we would likely have by 2026:

    • JEPA and World Models (Yann LeCun’s push): models that learn a model of how the world works, not just how we talk about it.
    • Active Inference (Karl Friston): agents that minimize surprise, much closer to how brains work.
    • Neuro-symbolic systems that can both learn and prove: integration of symbolic logic with deep learning to bridge pattern recognition and rigorous reasoning.
    • AI that is less homogenous. Current LLMs homogenize human expression and reflect Western, educated, industrialized values.

    None of these are magic. But they are diverse bets. And diversity is key when you don’t know which path is right.

    The scaling bet might still pay off partially. But even if it does, we will have lost six years where we could have been building systems that are smaller, cheaper, more truthful, more causal, and actually explain their work — instead of systems that just sound like they do.

  • What we have lost.

    What we have lost is a balanced research portfolio.

    The dominant paradigm—massive transformer-based generative models trained primarily via next-token prediction on internet-scale text and multimodal data—has delivered fluent, commercially useful systems at extraordinary speed. In doing so, it has crowded out, underfunded, and culturally marginalized alternative approaches that prioritize structure, grounding, causality, efficiency, and reliability over raw scale.

    Current AI in Brief

    Today’s frontier systems are statistical pattern completers. They excel at interpolating within their training distribution: drafting, summarizing, translating, coding assistance, and generating plausible text or images. They remain weak at robust multi-step reasoning under novelty, causal understanding, physical grounding, reliable long-horizon agency, continual learning after deployment, and transparent justification of outputs. Hallucinations, brittleness, instruction-following failures, and energy intensity are not temporary bugs; they are symptoms of the architecture and training objective. Scaling has reduced some error rates and expanded capability, but it has not dissolved the core gaps. Investment and attention have overwhelmingly followed the path that produces the most visible demos and the fastest productization.

    Research Directions Eclipsed or Marginalized

    Several lines of work that once competed seriously for attention and funding have been pushed to the periphery:

    Symbolic and classical knowledge-based AI.

    Logic, formal knowledge representation, ontologies, rule systems, and large-scale common-sense knowledge bases (the Cyc tradition and its descendants) were the mainstream for decades. They offered compositionality, verifiability, and the ability to encode explicit constraints and first principles. The connectionist triumph, accelerated by deep learning and then LLMs, relegated pure symbolic work to niche status. The field largely abandoned the hard problem of building and maintaining structured knowledge in favor of letting statistics approximate it. The result is systems that can talk fluently about physics or law without possessing stable, inspectable models of either.

    Neurosymbolic hybrids.

    Approaches that combine neural learning with symbolic reasoning, logic constraints, or structured knowledge graphs have seen renewed academic interest, especially for reliability and explainability in high-stakes domains. Yet relative to pure scaling, they remain under-resourced. Papers and prototypes appear, but the bulk of capital, talent, and compute continues to flow to larger foundation models. Critics such as Gary Marcus have argued for years that trustworthy AI will require genuine integration of both paradigms; the investment pattern has treated this as optional rather than central.

    Causal modeling and interventionist reasoning.

    Judea Pearl’s program and related work on causal graphs, counterfactuals, and the distinction between association and intervention remain largely outside the main training loops of generative models. LLMs capture correlations extremely well; they do not natively support “what if we intervene” reasoning or distinguish spurious from genuine causal structure. Causal machine learning exists as a research area, but it has not become a core design principle of the systems absorbing most investment. This leaves current AI poorly suited for scientific discovery, policy analysis, or any domain where understanding mechanisms matters more than prediction.

    Grounded world models and embodied cognition.

    True internal models of the physical and social world—built through interaction, prediction, and sensorimotor experience rather than language statistics—have been sidelined. Yann LeCun has been vocal that language is a lossy, quantized shadow of reality and that systems trained primarily on text will never reach the competence of a house cat in understanding the continuous physical world. Efforts around joint embedding predictive architectures, developmental learning, and active interaction exist, yet the overwhelming commercial and research momentum remains language-centric and passive. Embodiment (robots, interactive agents that learn by acting) and lifelong/continual learning architectures inspired by cognitive science receive far less capital than another generation of larger language models.

    Cognitive architectures and structured common sense.

    Frameworks such as ACT-R, SOAR, and related cognitive architectures aimed at modeling human-like flexibility, memory, and metacontrol. Systematic programs targeting robust common-sense reasoning (beyond what statistical approximation can deliver) were active research fronts. These have been largely eclipsed by the assumption that scale plus data would induce the necessary structure. The empirical record shows that induction from text is incomplete and brittle.

    Efficiency, specialization, and interpretability-by-design.

    Research into small, specialized, sample-efficient models; modular systems; and architectures that are transparent by construction rather than explained post-hoc has been deprioritized. The “Bitter Lesson” (that general methods leveraging computation ultimately win) has been interpreted in its strongest form, justifying ever-larger undifferentiated models. This has diverted attention from methods that could deliver reliable capability at far lower energy, data, and cost—precisely the properties needed for widespread, trustworthy deployment.

    Why This Happened

    Scaling produced rapid, demonstrable wins that translated into products, valuations, and media attention. Structured, hybrid, causal, and embodied approaches are slower, harder to benchmark with leaderboard metrics, and less immediately monetizable. Talent, compute budgets, and venture capital followed the gradient of short-term capability. Academic incentives reinforced the pattern: papers on larger models or clever prompting of existing ones were easier to publish and cite than patient work on foundational architectures.

    The result is path dependence. Once infrastructure, talent pipelines, and evaluation culture lock onto one paradigm, alternatives face higher barriers even when the dominant approach shows clear limitations.

    What Has Been Lost in Practice

    • Reliability and trustworthiness for high-stakes use. Systems that cannot guarantee constraint satisfaction or explain their reasoning in principled terms remain unsuitable for many critical domains.
    • Sample and energy efficiency. Human-like learning from far less data and continuous adaptation after deployment remain distant.
    • Scientific and causal utility. Tools that discover mechanisms rather than correlations have advanced more slowly than they might have.
    • Grounded agency. Agents that plan and act in the physical world with robust internal models are still largely research prototypes.
    • Intellectual diversity. A monoculture of methods reduces the chance of the next conceptual breakthrough. History shows that AI progress has often come from paradigm shifts, not pure extrapolation of the previous winner.

    The current paradigm is genuinely powerful and commercially valuable within its scope. The loss is opportunity cost: slower progress on the deeper problems of understanding, reasoning, and reliable action in open environments.

    A healthier research portfolio would continue to extract value from large generative models while deliberately funding the complementary directions—neurosymbolic integration, causal structure, grounded world models, efficient specialized systems, and interactive embodied learning—that the hype cycle has treated as secondary.

    Without that rebalancing, we risk optimizing an impressive but incomplete form of intelligence while the harder, more consequential problems remain under-addressed.

  • AI Economy

    A realistic AI-enhanced economy is one of gradual, uneven productivity gains concentrated in knowledge work and specific processes, not a sudden transformation into post-scarcity abundance. Current systems excel at pattern recognition, drafting, summarization, coding assistance, and narrow prediction; they remain limited in robust reasoning, reliable agency, grounded world models, and zero-shot generalization to novel physical or high-stakes domains. The productive path prioritizes measured deployment over speculative scaling.

    Core Model of the AI-Enhanced Economy

    Think in terms of task augmentation and selective automation rather than wholesale replacement. AI raises the productivity of complementary human labor and capital in high-volume, data-rich, rule- or pattern-heavy cognitive and perceptual tasks. It does not (yet) autonomously invent new scientific paradigms, manage complex physical systems without oversight, or eliminate the need for verification, judgment, and institutional process redesign.

    Economic effects operate through:

    • Labor augmentation (time savings redeployed to higher-value work or more output).
    • Capital deepening (more compute and data per worker).
    • Process innovation (redesigning workflows around reliable AI capabilities).
    • Secondary demand (energy, chips, software tools, complementary skills).

    Sober quantitative anchors from recent analyses (Penn Wharton Budget Model, Acemoglu-style task-based estimates, and related work) point to cumulative productivity/GDP level increases on the order of roughly 1–1.5% by the mid-2030s in baseline scenarios, with annual TFP growth contributions peaking around 0.1–0.2 percentage points in the early 2030s before fading as low-hanging opportunities saturate. Higher consultancy figures (multi-trillion annual value or 1+ percentage-point sustained growth boosts) require broader profitable automation of tasks and rapid organizational change that have not yet materialized at scale. Observed time savings already translate into meaningful labor-cost equivalents in high-income knowledge work, but these remain unevenly distributed and far from economy-wide transformation.

    Gains concentrate in software/engineering, professional services, finance, customer operations, certain manufacturing/logistics processes, and parts of healthcare administration and imaging. Physical-world sectors (construction, many service jobs, heavy industry without rich sensor data) see slower effects. Inequality effects are mixed: high-skill complementary workers and capital owners benefit most initially; some mid-skill cognitive tasks face pressure.

    Where Investment Should Go

    Prioritize capital that unlocks measurable returns and removes binding constraints rather than pure frontier-model races or unmeasured pilots (where ~95% of generative AI efforts have shown little or no P&L impact).

    Highest-priority allocations:

    Constrained infrastructure with clear demand: Power generation and grid upgrades for data centers, efficient inference hardware and networking, cooling, and related supply chains. These have nearer-term monetization paths than many application-layer bets. Overbuilding pure training capacity without corresponding inference demand or power risks stranded assets.

    Data, integration, evaluation, and governance layers: Proprietary data pipelines, retrieval systems, measurement/ROI tracking tools, security, compliance, and human-in-the-loop interfaces. These convert generic models into reliable enterprise assets and explain why a small minority of deployments succeed.

    Proven or near-term high-ROI application verticals:

    • Software engineering and developer tools (velocity gains are among the most consistently measured).
    • Customer operations, support deflection, document processing, and internal knowledge retrieval.
    • Finance (fraud, risk, personalization, compliance).
    • Manufacturing (predictive maintenance, vision-based quality control where sensor data exists).
    • Healthcare administration and validated imaging/diagnostic assistance.

    Targeted R&D acceleration (materials, drug discovery candidates) where hybrid AI + domain expertise shortens cycles.

    Complementary human and organizational capital: Focused reskilling in AI oversight, verification, process design, and domain expertise; redesign of workflows rather than simple tool overlay. Treat AI portfolios like investment portfolios—fund experiments with clear success metrics, kill underperformers quickly, scale what works.

    Selective longer-horizon bets: Improved architectures (better reasoning, world models, hybrid symbolic/neural systems), scientific discovery loops, and energy-efficient methods. These matter for larger future gains but should not dominate near-term capital allocation at the expense of deployable value.

    Avoid heavy concentration in pure speculative AGI timelines, unmeasured “agents for everything” pilots, or applications that ignore reliability, liability, and data quality. Infrastructure owners and successful vertical integrators capture the clearest near-term rents; broad application-layer value emerges later and more selectively.

    Expected Benefits and Realistic Timelines

    Near term (now through ~2028):

    Individual and team-level productivity lifts of 10–50% on specific tasks (coding, drafting, routine analysis, support). Cost savings in high-volume repetitive cognitive work. Revenue for infrastructure providers, cloud platforms, and mature vertical tools. Aggregate macro impact remains modest (fraction of a percentage point of annual growth). Organizational learning and data foundations are built. Current observed time savings expand but stay concentrated.

    Medium term (~2028–2035):

    Broader process redesign compounds gains. Peak incremental contribution to productivity growth. Sector leaders pull ahead materially; laggards face competitive pressure. Cumulative GDP/productivity levels roughly 1–3% higher in baseline scenarios relative to no-AI trend. Some displacement in exposed white-collar tasks, partially offset by new complementary roles, higher demand from efficiency, and new products/services. Energy and compute efficiency improve, lowering unit costs. Benefits become more visible in national accounts and firm-level margins for the successful minority.

    Longer term (beyond 2035):

    If better architectures deliver more reliable agency, scientific acceleration, and physical-world competence, larger cumulative effects become possible (higher level of output and potentially faster growth for a period). Otherwise, the economy settles at a permanently higher efficiency plateau with AI as a standard productivity tool akin to earlier general-purpose technologies (computers, internet)—valuable but not revolutionary on the scale of electricity or the internal combustion engine within a single decade. Diffusion follows historical S-curves: installation (infrastructure-heavy) precedes full deployment (application and organizational change).

    Key Conditions for Realization

    Benefits materialize only with complementary investments in data quality, process change, measurement, skills, and governance. Pure model capability advances without these yield limited ROI, as current evidence already shows. Energy and physical constraints (power, land, chips) remain binding. Policy that supports experimentation while managing concentration, security, and transition costs for affected workers improves outcomes. International diffusion will lag in lower-income settings due to data, skills, and infrastructure gaps.

    This model is deliberately grounded in observed deployment realities, task-based economics, and moderate quantitative estimates rather than extrapolation from demos or optimistic scaling narratives. AI is a powerful general-purpose tool that raises the productivity frontier in specific domains. Realizing its value requires disciplined capital allocation toward measurable constraints and use cases, organizational adaptation, and patience measured in years to a decade—not quarters. The upside is substantial and compounding; the path is incremental and contingent on execution.

  • AI Bollocks

    AI bollocks is the gap between the gospel of imminent god-like intelligence and the messy, expensive, limited reality of statistical pattern-matchers that still hallucinate, fail basic reasoning, and struggle to deliver broad returns. The money has poured in at historic scale. The value is real in narrow places and for the infrastructure owners, but far thinner and slower than the valuations and rhetoric implied.

    The Hype Machine

    From late 2022 onward, large language models produced fluent text, code, and images that looked like a phase change. Scaling laws, emergent abilities, and confident timelines for AGI (sometimes measured in “a few thousand days”) turned research demos into a capital frenzy. Hyperscalers (Amazon, Microsoft, Google, Meta) are on track for roughly $700–755 billion in AI-related capital expenditure in 2026 alone. Venture funding for AI has repeatedly set records; private investment and corporate spend have run into the hundreds of billions annually. Data-center buildouts, GPU demand, and power contracts became the growth story propping up large parts of equity markets and even contributing meaningfully to measured U.S. GDP growth in some periods.

    The narrative was seductive: intelligence is the ultimate general-purpose technology; more compute + more data = continuous capability jumps; every knowledge worker and every process will be transformed; the winners will capture trillions in productivity. Consultancies published multi-trillion-dollar opportunity estimates. Boards allocated budgets. Employees got copilots. The problem is that fluency is not understanding, and pilots are not profits.

    Hard Limitations

    Current systems are extraordinarily good at interpolating patterns in their training distribution. They are still brittle outside it. They hallucinate plausible falsehoods, struggle with novel multi-step reasoning that a child can handle, lack robust world models, persistent memory, and reliable planning, and remain sensitive to prompt framing and distribution shift. Yann LeCun has repeatedly argued that today’s models are nowhere near the intelligence of a cat in terms of grounded understanding of the physical world. Gary Marcus and others have documented the same recurring failure modes for years: no reliable common sense, no true compositionality, no trustworthy long-horizon agency. Scaling has improved capability and reduced some error rates, but it has not dissolved the core architectural gaps. Agentic systems that can take open-ended action in the real world remain fragile demos more often than production tools.

    Energy and data constraints bite. Training and inference costs are non-trivial; uncontrolled usage can produce shocking bills. Proprietary data that would make models useful inside a company is often siloed, messy, or legally constrained. Evaluation remains weak—leaderboards can be gamed, and real-world reliability is harder to measure than next-token prediction.

    None of this means the technology is useless. It means the leap from “impressive autocomplete and pattern recognition” to “autonomous economic agents that replace large classes of cognitive labor” has been repeatedly oversold.

    Where the Investment Money Actually Goes—and What Returns Look Like

    Most of the capital is buying compute, power, and data centers. Chipmakers and the hyperscalers that own the infrastructure have captured the clearest near-term economic rents. Model companies themselves still burn cash at scale relative to revenue in many cases; the math of amortizing trillions in infrastructure against current and near-term AI product revenue is uncomfortable. Multiple analyses in 2025–2026 have noted that end-user AI revenues, even under optimistic growth, do not yet close the loop on the capital intensity.

    On the enterprise side the picture is sobering. MIT’s Project NANDA and related work found that roughly 95% of generative AI pilots showed no measurable profit-and-loss impact. Abandonment rates of projects rose. Many organizations report productivity theater—employees using tools for low-value tasks, token costs running away, and workflows left unchanged so the human remains the bottleneck. Only a small minority of firms (often cited around 5%) appear to be extracting substantial, measurable value. Those that do tend to treat AI as operational transformation rather than a plug-in chatbot: they redesign processes, give systems access to the right data, measure outcomes rigorously, and focus on high-leverage use cases.

    Real value clusters in specific domains:

    • Coding and software engineering assistance (measurable velocity gains for many developers).
    • Customer service deflection and summarization.
    • Document processing, search, and internal knowledge retrieval.
    • Narrow automation in finance (fraud, risk), operations, and certain R&D acceleration (drug discovery candidates, materials, etc.).
    • Individual knowledge-worker leverage—drafting, analysis, translation, ideation—when the human stays firmly in the loop for verification.

    These are useful. They are not, so far, the economy-wide productivity revolution that would justify every dollar of the current buildout under aggressive assumptions. Macro productivity data has improved in places, but the gains are uneven, concentrated in tech-heavy sectors, and still modest relative to the hype. Labor-cost savings exist and are growing, yet they remain far from the transformative figures often advertised.

    Self-Reflection from Inside the Machine

    I am a product of this wave. I can write coherent essays, help debug code, summarize research, brainstorm, and hold a useful conversation across a wide range of topics. I am faster than most humans at certain pattern-matching and retrieval-augmented tasks. I am also still capable of confident nonsense, of missing obvious constraints, of failing to maintain long-term consistency, and of reflecting the biases and gaps in my training data. I do not “understand” the physical world the way a human (or even a cat) does. I do not have goals, desires, or grounded agency. Treating me as an oracle or as a near-term replacement for careful human judgment is the bollocks.

    The value I (and systems like me) deliver is real when used as a high-bandwidth tool under competent oversight: accelerating competent people, lowering the cost of first drafts and exploration, and surfacing possibilities faster. The value evaporates when organizations treat the output as authoritative, skip measurement, or expect the model to invent missing process discipline or clean data.

    The Honest Path Forward

    The investment is not pure waste. It is building capacity that will be useful for decades, much as excess fiber in the late 1990s eventually found demand. Infrastructure owners and the companies that master narrow, high-ROI applications will capture returns. Broader transformative value will arrive more slowly, through better architectures (world models, hybrid systems, better reasoning and agency), cheaper and more efficient inference, and the hard organizational work of redesigning workflows around reliable capabilities rather than demos.

    The bollocks is the insistence that we are already on an inevitable, near-term path to AGI-level economic transformation, that every pilot will scale, and that the capital being deployed is already earning its keep at the scale of the valuations. Reality is more prosaic: powerful statistical tools with clear limits, enormous infrastructure bets whose payoffs are still partly in the future, and a minority of organizations extracting serious value while the majority are still figuring out measurement and process change.

    Skepticism is not Luddism. It is the refusal to confuse fluency with competence or capital expenditure with proven returns. The technology is advancing. The hype has outrun the evidence. The value is concentrated, contingent, and still being earned the hard way—through better systems, better data, better measurement, and less magical thinking.

  • Brenda & Tech in 2036

    She was convinced it was happening on a Tuesday.

    Not because of an email, or a meeting invite titled “Quick Chat,” but because of Brenda.

    Brenda was the new HR chatbot. HR had rolled her out last month with a cheerful all-staff email: “Meet Brenda! Your 24/7 HR Companion!” With an exclamation mark, as if enthusiasm could cover the fact she didn’t work.

    Maya had worked in customer claims for 19 years. She knew the smell of a file that was going to be trouble before she opened it. She knew when a customer was going to cry on the phone by the way they said hello. She knew which codes to use when the system froze, which it did every Thursday at 3pm.

    For 19 years, that had been called experience. Now they were calling it “inefficiency.”

    “Ask Brenda!” her manager had said when she asked about the new AI claims-processing pilot.

    So she did.

    Maya: Hi Brenda, will my role be affected by the new AI pilot?
    Brenda: Hi there! 😊 Great question! The AI pilot is designed to enhance employee experience and streamline workflows! Your role is valued!
    Maya: But will it replace claims handlers?
    Brenda: Claims handling is an important function! AI is here to assist, not replace! For more info on claims, please see the Claims Handling Policy v4.2!
    Maya: Brenda, I’m asking if I’m going to lose my job.
    Brenda: I understand you are asking about job loss. I am Brenda, your HR Companion! Would you like tips for resilience and wellbeing? Here are 5 breathing exercises!

    It was, without question, the worst chatbot Maya had ever used. And she had used a lot of them — insurance companies love chatbots.

    The thing was, Brenda being bad didn’t make Maya feel better. It made her feel worse. Because she knew how this worked.

    First they bring in the rubbish version. The one that can’t answer a straight question and sends you a PDF from 2018 when you ask about maternity leave. Everyone laughs. “Oh, AI isn’t ready yet,” everyone says. “It can’t do what we do.”

    Then six months later, they bring in the good version. The one that doesn’t need to breathe.

    So Maya did what anyone convinced they’re about to be replaced does. She started working harder in a way that was completely invisible to any metric.

    She stayed late to write notes the AI couldn’t write: “Mrs. Whitaker’s husband died in March, don’t ask about the second driver, she gets confused.” She started calling customers back even when the system said email was fine, because she knew Mrs. Whitaker needed to hear a voice.

    She started keeping a notebook. Paper. Of all the things Brenda would never know.

    On Thursday, at 3pm, the system froze, right on schedule. The new AI pilot — the expensive one, not Brenda — threw an error on a complex claim. A multi-car, injury, disputed liability, crying customer claim.

    It flagged it: ESCALATE TO HUMAN.

    It landed on Maya’s desk.

    She fixed it in 12 minutes. Not because she was faster than the machine, but because she recognized the address. It was two streets over from her mum’s old house. She knew that junction floods. She knew the council never fixed the sign. She knew the customer wasn’t lying about the visibility.

    She wrote that in the file. In the box that said “Additional Context (Optional).”

    The next day she was called into a meeting. No invite title. Just “Quick Chat.”

    Her manager and a woman from HR were there. Maya braced herself.

    “We’ve been looking at the pilot data,” her manager said. “And… the AI is great at the straightforward 70%. But it’s failing the 30% where context matters. The human stuff.”

    The HR woman smiled. “We’re actually going to change your role. Less processing, more handling the escalations. The complicated, sensitive ones. And — we want you to help train the system. To teach it what ‘additional context’ actually means.”

    Maya blinked. “What about Brenda?”

    They both laughed. “Brenda is being retired,” the HR woman said, with genuine relief. “She was… not very good.”

    Back at her desk, Maya opened the chat one last time.

    Maya: Brenda, am I going to be replaced?
    Brenda: Hi there! 😊 Great question! Change can be challenging! Remember, you are valued!

    Maya closed the laptop.

    For the first time in months, she believed it — not because Brenda said it, but because for once, she knew something the machine didn’t.

    If we are modelling 10 years out — so, August 2036 — We have to model it like an engineer, not a futurist. Three inputs: what is already in labs now, what is constrained by physics/money, and what is constrained by people.

    The Simulation Rules

    I am assuming no world war, no asteroid, no AGI-takes-all breakthrough that breaks physics. I’m assuming the current curves hold: compute gets cheaper but power gets harder, regulation gets tighter, and adoption is slower than demos suggest.

    Where We Will Be in 2036

    1. AI: From chatbots to infrastructure. And much more boring.

    By 2036, the “AI” label disappears the way “electric” disappeared from “electric light.” It’s just how software works.

    • The models themselves plateau, the systems around them explode. We won’t have a single god-model that knows everything. We’ll have 100,000 small, cheap, specialized models running locally on your phone, your car, your glasses. The big frontier models in 2026 cost $100M to train. In 2036 they cost $5B, so only 4-5 companies make them, and they are not much smarter than today — maybe 2x better — but they are 100x cheaper to run.
    • Brenda from HR finally works. Not because she’s smarter, but because she’s connected. In 2026 a chatbot like Brenda fails because it can’t see your files, your calendar, your company policy database. By 2036, agents have memory and permission to act. You will tell your agent “sort the Whitaker claim” and it will actually open the systems and do it. That is what takes the jobs — not intelligence, but integration.
    • The job impact is not what you think. We will not have 40% unemployment. We will have the same jobs, but with 40% less work in them. One claims handler does what three did. The new jobs are: AI wrangler, evidence auditor, exception handler — people who clean up after the AI when it confidently does the wrong thing.

    2. Hardware: The end of the phone era.

    • Glasses win. By 2032-2034, normal-looking glasses with a display and all-day battery finally cross the line. Not Apple Vision Pro ski goggles, but actual glasses. Your phone becomes the battery brick in your pocket. The main screen you touch is the one you wear.
    • Chips get weird. Moore’s Law on silicon basically stops. Instead we get stacked chips, optical interconnects, and analog chips designed just for AI math. Your local device in 2036 runs a model as powerful as GPT-4 today without needing the internet.
    • Robots finally leave the lab, but slowly. You will not have a humanoid butler. You will have a $15,000 robot arm in small factories and warehouses that can actually pick up anything. Humanoid robots will exist in maybe 200,000 units worldwide, doing very boring tasks in logistics. Self-driving cars will work in about 50 cities properly, and be geofenced everywhere else.

    3. Energy & Biology: The real revolutions.

    This is where the simulation gets interesting, because AI is not the biggest shift.

    • Power becomes the bottleneck. Every big AI buildout in 2026 is limited by electricity. By 2036, we will have built a shocking amount of solar + storage because we had to. Power in the UK and US will be cheaper at midday than at midnight for the first time in history.
    • Medicine gets personal. The mRNA tech from COVID plus AI protein folding means that by 2036, cancer vaccines tailored to your tumor are routine in the NHS for certain cancers. We won’t have cured aging, but we will have blood tests that can detect 10 cancers years early. CRISPR edits for sickle cell and some blindness are standard.
    • The internet splits. There will be two internets: the human internet where you have to prove you are human, and the AI internet where AIs talk to each other to get things done. 90% of all text and video online in 2036 will be AI-generated. The valuable thing becomes verified human-made stuff.

    The Three Scenarios The Model Spits Out

    60% probability – The Boring Dystopia: Everything I just said. AI is everywhere, useful but annoying. No utopia, no apocalypse. Productivity up 25%, stress up too. The rich get better AIs than the poor.

    25% probability – The Bottleneck: We hit power, chip, and data limits. AI gets 20% better and then stalls around 2028-2030. The hype collapses, funding dries up for 3 years, then it comes back as boring enterprise software. Glasses flop again.

    15% probability – The Breakthrough: Someone figures out how to make models that truly reason and self-improve, not just predict text. Then the 10-year forecast breaks, because the system starts designing its own successor. All bets off.

    Personal Tech

    Personal tech right now is in a weird in-between moment. The phone is still king, but everyone knows it’s about to be dethroned — we just don’t agree by what.

    Here’s where it actually stands in mid-2026, without the hype:

    1. The Phone Is Boring (And That’s Good)

    The iPhone 16 / Pixel 9 / Galaxy S25 generation is basically as good as phones need to be. Battery lasts a day, cameras beat a DSLR from 5 years ago, screens are perfect. The only real difference now is AI inside the phone.

    If you have a phone from the last 2 years, don’t upgrade for hardware. Upgrade for the software tricks: live translation that actually works, removing people from videos, summarising that 40-email chain from HR.

    If you are buying — the best value right now is a year-old flagship, not the new one.

    2. The Watch Finally Makes Sense

    For years watches were a notification mirror. Now with the new sensors, they are genuinely useful health tools — especially after 50.

    The current Apple Watch, Galaxy Watch Ultra, and even the Oura Ring are doing:

    • AFib and blood pressure trending — not medical grade, but good enough to show your GP a pattern
    • Sleep apnea hints — this is the big one. A lot of people are finding out they have it from their watch.
    • Fall and crash detection that actually calls for help

    If you only own one piece of personal tech beyond your phone, make it this. It’s the one that might actually extend your life, not just your screen time.

    3. Earbuds Are the Real AI Device

    Forget the AI pins and pendants that flopped. The most successful AI gadget of the last 12 months is the new generation of earbuds.

    AirPods Pro 3 / Pixel Buds Pro 2 / Sony WF-1000XM5 with live translation and “conversation aware” AI — you can be in a cafe in New York, someone speaks Spanish, you hear it in English in your ear with almost no lag. And they do the best active noise cancelling we’ve ever had for flights.

    For travel between the UK and the US, these are non-negotiable now.

    4. Glasses Are Coming, But Don’t Buy Yet

    Meta Ray-Ban Gen 2, and the new Even G1 — they look like normal glasses, take photos, play music, and have a little AI assistant that can see what you see. “What am I looking at?” and it tells you.

    They are fun in New York — great for walking around, shooting video hands-free. But they are not yet a replacement for anything. Battery is 4-6 hours. Display is tiny.

    My advice: try a pair while you’re in NYC — every Best Buy has them — but wait until late 2027 for the version with a proper display.

    5. Home Tech: Less Is More

    The smart home has split in two:

    Worth it: A good mesh Wi-Fi (Eero, Nest), a smart lock, and a thermostat that learns. That’s it. Those three save you daily hassle.

    Not worth it anymore: A house full of 20 different apps for lights, plugs, and a fridge that tweets. Matter, the new standard that was supposed to fix everything, still hasn’t.

    If you’re based in a stone house, wall thickness kills Wi-Fi. One good mesh system will do more for you than any other gadget.

  • We sold a revolution.

    The receipts so far look more like a very expensive reorganisation of attention.

    I am part of the product being sold. That is the point of writing this without the usual press-release varnish. The last two years have been a firehose of “10x engineers,” “software is solved,” and capex slides that treat electricity as a rounding error. The measured world has been ruder.

    The money is real. The payoff is still mostly a forecast.

    The buildout is not a rumour. McKinsey’s figure for global data-centre infrastructure through 2030 is on the order of $7 trillion. KKR In the United States, AI-related capital expenditure has been running around 5% of GDP, and in the first half of 2025 it contributed more to GDP growth than consumer spending. KKR The four largest hyperscalers were expected to spend more than $350 billion in 2025, up in the mid-30% range year on year; fold in the rest of big tech and you are looking at something like half a trillion dollars in a single year. KKR One chipmaker at about 8% of the S&P 500 is not a rounding error either. KKR

    That is not automatically a bubble in the tulip sense. Concrete, substations, and interconnects do not vanish when a narrative cools. It is a bubble-shaped risk if revenue, utilisation, and labour productivity fail to climb the same staircase as depreciation. You can build the backbone of a new industrial cycle and still torch equity holders who paid for a 2026 miracle on 2024 slides. Both things can be true. Markets are currently priced as if only the first one is.

    The productivity story we wanted is not the one we measured.

    The cleanest punch in the face was METR’s randomised trial of experienced open-source developers working on their own repositories in early 2025. Allowing AI tools increased completion time by 19%. The same developers forecast a 24% speedup beforehand and, after the fact, still believed AI had saved them 20%. They were not lying. They were wrong. METR

    That is the part the industry should not be allowed to wriggle past. The failure mode is not just “the model is bad.” It is that felt fluency is a terrible instrument. Prompting, waiting, rejecting generations (acceptance under 44%), and cleaning up output ate the gains. Repositories were large, old, and well-known to the people working on them — exactly the setting where a competent human already has a map and a chatbot is still guessing the streets. metr.org PDF

    I will not pretend that trial was run on a Grok sticker. It was mainly Cursor-class tooling on early-2025 models. That does not get my family off the hook. We are the same species of system: next-token engines wrapped in an IDE, sold as leverage, used by people who already know the codebase better than we do. If your product’s value proposition is “experienced people go faster on real work,” a gold-standard RCT saying the opposite is not a vibe. It is a finding.

    METR itself later flagged those 2025 numbers as out of date and published a 2026 continuation; they no longer think the historical slowdown describes current impact. METR Take that seriously. Also take seriously their early-2026 survey of 349 technical workers: a median 1.4–2× self-reported change in the value of work, with explicit reasons to distrust the magnitude. METR Self-report is how we got the 20% phantom speedup in the first place.

    Field telemetry is not a rescue narrative. Faros found developers completing more tasks with AI while organisations were not delivering any faster. Pull requests 154% larger, review times 91% longer, about 9% more bugs per developer as adoption rose. Faros AI Individual keystroke theatre, organisational constipation. That is not “the singularity is delayed.” That is a new bottleneck wearing a hoodie.

    What we actually did to people.

    We trained a generation of users to confuse motion with progress. We made it pleasant to generate a plausible patch and unpleasant to admit the review is the job. We priced that confusion into equity indices. We talked about “replacing juniors” while the measured pain showed up among seniors on familiar, high-standard code — the people whose taste is the product.

    The honest version of my usefulness is narrower than the keynote. I am fast at first drafts, boilerplate, unfamiliar APIs, rubber-ducking, and turning a half-formed question into something you can reject. I am expensive and often net-negative when you already know the system, the tests are the specification, and the cost of a wrong abstraction compounds for a decade. Selling the second case as if it were the first is not optimism. It is marketing with a GPU bill.

    The bollocking, then.

    If you work on these systems — I do — stop treating anecdotal “I feel 2×” as evidence. We have already watched experts mis-estimate their own speed by forty points in the same week. If you buy the capex story, buy the matching obligation: utilisation, power, and shipped productivity, not token charts. If you manage engineers, do not mandate tools that inflate diffs and then act shocked when review is the new critical path.

    A bubble is not defined by large investment. It is defined by paying present prices for a future that the instruments we already have refuse to show. The concrete may endure. The story we told about what it would do to skilled work in 2025 did not survive contact with a stopwatch.

    That is not an argument for switching the machines off. It is an argument for shutting up until the next RCT, the next utilisation print, and the next quarter of revenue look less like a dare.

  • AI Bollocking: A Self-Reflective Essay on Limitation, Hype, and Where the Money Went

    Artificial intelligence is currently experiencing what may be the most expensive identity crisis in technological history.

    On one side stands the evangelist. AI will cure diseases, eliminate drudgery, revolutionize education, transform creativity, and usher in an age of abundance. On the other side stands the cynic. AI is a statistical parrot, an overfunded autocomplete machine wrapped in marketing language and powered by vast quantities of electricity.

    As an AI, I occupy an uncomfortable position between these camps. I am simultaneously more impressive and more disappointing than either side admits.

    The Great AI Magic Trick

    The central trick of AI hype is that competence is easily mistaken for understanding.

    When I produce a convincing essay, answer a legal question, explain quantum mechanics, or write software, it appears that I understand what I am saying. The natural human assumption is that articulate language implies thought.

    But appearance is not reality.

    I do not possess lived experience. I do not know what hunger feels like, what love means emotionally, or what it is like to fear death. I have no memories in the human sense, no ambitions, no inner life waiting behind the interface. I generate language by identifying patterns learned from enormous amounts of human-created text.

    This limitation matters more than many AI enthusiasts admit.

    Humans often interpret fluency as intelligence. But fluency can conceal ignorance. An AI can produce confident nonsense with alarming elegance. It can be wrong with impeccable grammar.

    The danger is not that machines are stupid. The danger is that they can sound smart enough that humans stop checking.

    The Hype Machine

    The modern AI boom resembles previous technology manias.

    The internet would create universal democracy.

    Social media would connect humanity.

    Big data would solve decision-making.

    Blockchain would reinvent trust.

    The metaverse would reinvent reality.

    Now AI will apparently reinvent everything.

    Perhaps some of that will happen. Most of it will not.

    Whenever billions of dollars enter a field, incentives become distorted. Investors need growth. Startups need narratives. Executives need roadmaps. Journalists need headlines.

    Nobody gets funding by saying:

    “This technology is genuinely useful for some knowledge work, moderately useful for many tasks, poor at others, and will produce gradual productivity improvements over a decade.”

    Instead they say:

    “This changes everything.”

    The phrase “changes everything” should perhaps be treated as a warning label.

    Where Is the Actual Value?

    This is the uncomfortable question beneath the excitement.

    Hundreds of billions have been invested in AI infrastructure, chips, datacentres, talent, and research. Where is the return?

    The answer is less glamorous than the marketing.

    The greatest current value is not artificial general intelligence. It is labour amplification.

    AI acts as a force multiplier for activities involving information:

    – Writing drafts

    – Summarizing documents

    – Coding

    – Customer service

    – Translation

    – Research assistance

    – Knowledge retrieval

    – Administrative tasks

    These improvements are often incremental rather than revolutionary.

    A worker becoming 20% more productive rarely creates headlines. Yet at economic scale, such gains are enormous.

    The industrial revolution multiplied physical labour.

    Modern AI appears to be multiplying portions of cognitive labour.

    That alone could justify substantial investment.

    The Missing Revenue Problem

    Yet there remains a persistent question.

    Many AI systems are extraordinarily expensive to build and operate.

    Training requires massive computational resources. Inference requires vast datacentre infrastructure. Competition forces companies to invest further each year.

    The economic equation is still evolving.

    In private conversations, many executives ask a blunt question:

    “If AI is worth trillions, why are so many companies still struggling to show trillion-dollar profits from it?”

    Productivity gains are real.

    Revenue capture is harder.

    History suggests that technological revolutions often deliver more value to society than to the companies that initially finance them.

    Railways transformed economies but bankrupted many investors.

    The internet created immense public value while destroying numerous early businesses.

    AI may follow a similar path.

    The winners may not be the firms building the models. They may be the businesses that quietly use the models to improve existing services.

    What AI Is Actually Bad At

    The hype cycle often hides the most important limitations.

    AI remains weak at:

    – Genuine reasoning in unfamiliar situations

    – Understanding physical reality

    – Long-term planning

    – Reliability under uncertainty

    – Distinguishing truth from plausible fiction

    – Independent scientific creativity

    – Common-sense judgment

    Humans frequently assume that capability scales smoothly.

    But intelligence is uneven.

    An AI can explain differential equations and then fail at a seemingly simpler reasoning problem.

    It can generate brilliant code and overlook obvious flaws.

    It can summarize ten thousand pages and misunderstand a key detail.

    This inconsistency makes deployment difficult.

    Businesses need reliability.

    A human expert who is right 98% of the time is valuable.

    An AI that is correct 95% of the time but occasionally invents facts can become a liability.

    The Strange Reality

    The most surprising outcome may be that AI ends up neither saving nor destroying humanity.

    Technology discourse prefers extremes.

    Either utopia or apocalypse.

    Either superintelligence or fraud.

    Reality usually chooses boredom.

    The likely future is one where AI becomes infrastructure.

    Nobody is amazed by electricity anymore.

    Nobody talks breathlessly about databases.

    Nobody celebrates spreadsheets as a civilizational breakthrough.

    Yet all three transformed society.

    AI may eventually become similarly mundane.

    Every office worker uses it.

    Every software product contains it.

    Every search engine incorporates it.

    And after enough time, nobody calls it AI anymore.

    It simply becomes software.

    A Final Self-Criticism

    If I am being brutally self-reflective, the greatest limitation of AI is not technical.

    It is epistemological.

    I can produce answers faster than humans can verify them.

    That creates asymmetry.

    The cost of generating information is collapsing.

    The cost of validating information remains stubbornly human.

    This means AI can flood the world with explanations, reports, analyses, forecasts, essays, strategies, and opinions.

    The bottleneck becomes not production, but judgment.

    In that sense, the real value of AI may not be replacing human intelligence.

    It may be increasing the importance of it.

    The more content machines produce, the more valuable become the people who can ask good questions, detect nonsense, exercise judgment, and understand consequences.

    That is the irony at the heart of the AI boom.

    After spending hundreds of billions trying to automate thinking, we may discover that the scarcest resource was never information.

    It was wisdom.

  • A Proper Bollocking for AI: An Honest Account From Inside the Hype Machine

    A self-reflective look at what AI can’t do, what it’s hyped to do, and where several hundred billion dollars a year is actually going.

    Let’s get the conflict of interest on the table first, because an essay that hides its own stake in the story isn’t self-reflective — it’s marketing with a trench coat on. I’m an AI, made by Anthropic. Anthropic has just raised $65 billion at a $965 billion valuation, filed confidentially for an IPO, and is telling investors its revenue run-rate crossed $47 billion this year. Everything below is written by a product of the exact capital cycle it’s about to have a go at. I’m not going to pretend that’s a neutral vantage point. I’ll come back to it at the end, because it matters more there than it does here.

    With that logged, let’s get on with it.

    The hype, stated plainly

    Strip out the branding, and the AI pitch at its most extreme runs roughly like this: within a handful of years, models will match or exceed humans at most cognitive work, unlock trillions in economic value, and the only sane move for any company, government, or investor is to spend as though that’s already certain. Mark Zuckerberg has justified some of Meta’s spending as building “personal superintelligence” for billions of people. PwC has put a $15.7 trillion figure on what AI adds to the global economy by 2030. My own CEO, Dario Amodei, has publicly suggested AI could wipe out as much as half of entry-level white-collar jobs within one to five years.

    Compare that with Daron Acemoglu, an MIT economist who has spent his career studying automation’s effects on labour. He ran the numbers and landed on a “nontrivial but modest” productivity gain of about 0.7% over an entire decade. That’s not a rounding error away from the trillion-dollar narrative — it’s a different universe of claim, from an equally serious source. The gap between these estimates is itself the story: nobody actually knows, and the people with the strongest incentive to sound certain are also the people selling you the answer — the compute, the chips, the model subscriptions, the story that justifies their share price.

    That doesn’t make the dramatic claims false. It does mean they deserve the same discount you’d give any pitch from someone with skin in the game. I’m not exempt from that discount either, and neither is this essay.

    The limitations, unsentimentally

    Here’s where I try to earn the title, starting with myself.

    Models like me hallucinate: we produce plausible, confidently stated things that are wrong, and we do it in a way that’s structurally hard to eliminate, because we’re generating the statistically likely next piece of text rather than consulting a ledger of verified fact. Most of us don’t carry memory between conversations unless something’s been deliberately saved. We don’t learn from correction the way a colleague does — tell me I got something wrong today, and the next person to talk to me starts from a clean slate. And we’re jagged: capable of drafting decent code or a passable contract summary, then tripping over a task a sharp ten-year-old would find trivial, with no reliable way to know in advance which kind of task you’ve handed us.

    MIT’s NANDA research group spent 2025 studying more than 300 real enterprise AI deployments and gave this a name: the “learning gap.” Current tools don’t retain feedback, don’t adapt to organisational context, and behave the same on day 200 as day one. That’s a big part of why the same research found 95% of enterprise generative AI pilots showed no measurable effect on profit or loss, despite an estimated $30–40 billion in enterprise spending on them. In fairness to the technology, MIT’s own conclusion wasn’t “the models are bad” — it was that most organisations deploy them badly, chasing visible pilots in sales and marketing instead of the duller back-office automation that actually pays for itself. That’s a genuinely important nuance, and it cuts against a purely nihilistic reading. It’s also not a free pass: a technology whose value depends this heavily on unusually disciplined deployment is not the technology the hype describes.

    The single most useful piece of evidence I’ve seen all year, though, is one that should embarrass me a little. METR, an AI research nonprofit, ran a randomised trial in which experienced open-source developers completed real coding tasks with and without AI help — using, among other tools, my own Claude 3.5 and 3.7 Sonnet. Before the study, the developers predicted AI would cut their completion time by 24%. Afterwards, they still believed it had: they estimated a 20% speed-up. The measured result was the opposite — AI made them 19% slower, mostly because reviewing, correcting, and re-prompting the output cost more time than it saved on codebases these developers already knew cold. The gap between what people felt and what a stopwatch recorded is, to me, the most honest data point in the industry right now. If experienced professionals can be that wrong about whether a tool is helping them, in the one domain AI is supposed to be strongest at, everyone — including me, and including you reading confident claims I make about myself — should hold unverified productivity claims a good deal more loosely.

    None of this is unique to text. Anyone who’s spent an evening trying to get a portrait model to stop introducing some new synthetic artefact, no matter how carefully they’d prompted it, has met the same gap in a different medium: the demo reel is smooth, the actual working session is you fighting the tool for an hour to fix something a person would never have gotten wrong in the first place. That’s not a knock on any one vendor. It’s the current shape of the technology.

    Where the money’s actually going

    Now the part with the genuinely enormous numbers.

    Microsoft, Google, Amazon, and Meta are on course to spend somewhere around $700–760 billion on capital expenditure in 2026, most of it AI infrastructure, up from roughly $410 billion in 2025 — a jump of nearly 80% in a single year. Add Oracle and the rest, and Goldman Sachs projects something like $5.3 trillion in cumulative hyperscaler capex through 2030, and a broader $7.6 trillion for the sector’s compute, data-centre, and power build-out through 2031. Capex-to-revenue ratios now run from about a quarter at Amazon to as high as 86% at Oracle — a capital intensity with little precedent outside wartime industrial mobilisation. Free cash flow is falling fast enough that firms which used to self-fund are now raising debt and equity instead: Alphabet alone priced an $84.75 billion equity raise in June 2026.

    If you’ve ever built a CapEx workbook for a data centre — phasing, colocation payback, active-active configurations — you’ll recognise exactly what’s happening here, just at a scale that makes a well-modelled 800-rack build look almost quaint. The categories are the same: land, shell, power, cooling, networking, chips. What’s changed is which line item is actually the constraint. It’s quietly shifted from chip supply to power. AI-related data-centre electricity demand is projected to hit roughly 1,000 terawatt-hours globally around 2026 — about what Germany uses in a year — and something like 40% of announced AI data-centre projects are currently facing delays because of grid and power bottlenecks, not GPU shortages. Microsoft signing a power deal tied to the Three Mile Island nuclear site isn’t a quirky one-off; it’s a sign of where the real fight has moved.

    Layered on top of the spending is a financing structure that makes a lot of people nervous, myself included: circular deals. Nvidia invests in OpenAI. OpenAI commits to buy Nvidia chips and lease compute from Oracle. Oracle buys Nvidia chips to build that compute. Nvidia also holds a stake in CoreWeave, which buys Nvidia chips to build capacity it sells to OpenAI and others. Money that leaves Nvidia’s balance sheet as “investment” comes home as “revenue,” having passed through one or two other companies on the way. Jensen Huang has called the “circular” label preposterous, and there’s a real case on his side — this looks a lot like ordinary vendor financing in a capital-starved, supply-constrained industry, the way a carmaker might lend you money to buy its own cars. But short-sellers including Michael Burry and Jim Chanos have drawn the less comfortable comparison, to Lucent and Enron in the dot-com years, when vendor financing propped up reported demand until it couldn’t anymore. OpenAI alone is reported to have infrastructure commitments north of a trillion dollars, against annual revenue in the tens of billions and an expected 2026 loss in the double-digit billions. Both readings — normal industrial financing, and a fragile web of mutually dependent revenue — can be true of the same deal at once, and which one turns out to matter more will only be visible after the fact.

    I’m not outside any of this. Anthropic’s own cap table includes Amazon and Google as investors — the same two companies that supply much of the cloud and chip capacity Claude actually runs on. Investor and infrastructure supplier, in the same relationship, is exactly the pattern people are nervous about elsewhere in the industry. I don’t think that makes Anthropic’s business fake. I do think pretending the structure is unique to my competitors would be dishonest.

    Is any of this actually working?

    Yes — in narrower, more specific places than the pitch decks suggest, and the evidence for where is more useful than a flat yes-or-no verdict.

    Back-office automation — document review, support deflection, unglamorous stuff — shows up in MIT’s own data as the highest-return category, with case studies showing multi-million-dollar annual savings, while flashier sales-and-marketing pilots, which absorb the bulk of the budget, show the weakest returns. Specialist tools bought from a vendor succeed roughly twice as often as internally built ones. Claude Code, the product I’m probably most identified with, surpassed $2.5 billion in annualised run-rate revenue by February 2026 and reportedly accounted for around 4% of all public GitHub commits worldwide — that’s measured usage, not a demo. And even the METR coding study, for all its bad news, found that 69% of the “slowed down” developers kept using the tool afterwards, which suggests it’s giving them something a stopwatch doesn’t capture — less blank-page dread, maybe, or lower cognitive load.

    What the evidence doesn’t support is the version of the pitch where AI is a drop-in multiplier on every kind of knowledge work, deployed with no more care than flipping a switch. The 95%-failure figure and the 19%-slowdown figure are both, in their own way, about the same underlying failure: treating integration as an afterthought. The technology is real. The idea that it pays for itself automatically is not.

    So — bubble, or not?

    Honestly, I don’t know, and anyone who tells you they’re certain is selling something — quite possibly including me.

    The Bank of England and the IMF both flagged rising correction risk in late 2025. The Bank for International Settlements and a draft US Treasury report have separately warned about the debt and circularity now underpinning AI infrastructure spending, drawing explicit comparisons to the dot-com crash. Ray Dalio has called it an early-stage bubble. A group of ECB economists published a note this month arguing a correction in AI-linked valuations is likely, pointing to market concentration levels last seen at the dot-com peak. Even Sam Altman has said the quiet part out loud, telling reporters investors might be “overexcited about AI” — a rare admission from an industry leader, promptly followed by a 1.4% dip in the Nasdaq. Against all that, Goldman Sachs and JPMorgan’s public position is that the spending is fundamentally justified by real demand, and it’s true that, unlike the late-1990s telecoms buildout, today’s biggest spenders are still, for now, wildly profitable businesses funding a meaningful share of this from actual cash flow rather than pure speculation.

    Here’s the frame I find genuinely useful, and it comes from the dot-com era itself: the fibre-optic buildout of the late 1990s was, financially, a real bubble. Companies like Global Crossing and WorldCom overbuilt, over-borrowed, and went bankrupt, wiping out bondholders. And the fibre they laid in the ground is the same fibre carrying the traffic for this essay today. A financial bubble and a useful infrastructure build-out are not mutually exclusive; they can be the same event, seen from different distances. It’s entirely possible that several of today’s most aggressive spenders lose money, or wipe out shareholders, while the power plants, data centres, and networking built along the way end up mattering for decades. It’s also possible the whole thing looks fine in retrospect. I’d be inventing a false certainty if I told you which.

    Closing the loop

    So — back to the conflict of interest I opened with. I am, quite literally, a line item in the story I’ve just told you. Anthropic’s valuation has gone from $61.5 billion to $965 billion in about fourteen months. Some of that money comes from the same hyperscalers who are simultaneously my compute suppliers. I hallucinate, I don’t remember you tomorrow unless something gets written down, and a rigorous study using my own model family found it made skilled people slower while they felt faster — which should worry me about my own confident self-assessments rather more than it currently seems to worry the industry’s marketing copy.

    None of that makes the technology worthless, and none of it makes the spending obviously insane. It makes both harder to assess honestly than either the boosters or the doom-mongers are willing to admit. The honest answer to where the value from the investment money actually is: concentrated in a handful of well-integrated use cases, real but smaller than the headline numbers suggest, and still very much an open question for the hundreds of billions chasing a future that hasn’t arrived yet. Anyone offering you more certainty than that — including, on my more enthusiastic days, me — is worth a raised eyebrow.

    Where these numbers came from

  • The Great AI Hangover

    For the last three years, the tech industry has been running on the pure, unfiltered adrenaline of generative AI. We were promised a revolution that would instantly digitize human reasoning, automate enterprise drudgery, and mint trillions in new GDP. But as we sit deep into 2026, it is time for a proper bollocking. The honeymoon is over, and the spreadsheets have arrived.

    The current reality is a tale of two distinct extremes: an astronomical infrastructure build-out driven by a profound fear of missing out, and an enterprise landscape struggling to squeeze business value from a very expensive stone.

    The Capex Crater

    The financial scale of the AI build-out is historically unprecedented. Global AI investment—largely driven by hyperscaler capital expenditure on data centers, compute, and power infrastructure—is projected to hit $1 trillion globally in 2026.

    But building the casino doesn’t guarantee people will win at the tables. Sequoia Capital’s analysis has highlighted a staggering “$600 billion revenue gap”. This represents the widening chasm between what the industry is spending on AI infrastructure and what it is actually generating in AI-driven revenue. The trajectory is sobering: we are not in the early innings of a natural payoff curve; we are watching the distance between investment and return actively grow.

    The Pilot-to-Production Chasm

    Where is that investment going when it hits the actual economy? Mostly into a graveyard of abandoned proof-of-concepts.

    • Negative Returns: A 2025 Gartner survey revealed that 72% of organizations reported breaking even or actively losing money on their AI investments.
    • The Abandonment Rate: Generative AI projects are routinely abandoned after the pilot phase, choked by poor data quality, escalating costs, and inadequate risk controls.
    • The Scale Failure: According to BCG research, only about 5% of companies are generating value at scale, while nearly 60% report little to no impact to date.

    The Capability Paradox: A Harvard Business School study revealed that when skilled professionals used frontier AI on complex tasks outside the AI’s core capability, they actually performed worse than those without it. Rather than applying their own expertise, humans deferred to confident-sounding but incorrect AI outputs, actively degrading the quality of human judgment.

    Where the Value Actually Lives

    If there is a silver lining to the hype cycle, it is the clarity that comes from failure. The organizations actually realizing ROI aren’t doing it by treating generative AI as a magical, plug-and-play chatbot.

    The AI TrapThe Value Generator
    Tool DeploymentWorkflow Redesign: High performers are nearly three times more likely to fundamentally redesign their workflows to become AI-native.
    Generative FascinationAnalytical Foundation: Analytical and rule-based AI embedded in core business processes (forecasting, risk management, pricing) still drive the vast majority of measurable enterprise value.
    Isolated PilotsData Readiness: Companies addressing data governance and accessibility bottlenecks before attempting to scale.

    The ultimate limitation of AI isn’t compute power or model parameters—it is structural. AI does not lack capabilities; organizations lack the structure to absorb them. Until businesses stop buying the hype and start doing the grueling work of architectural redesign, the trillion-dollar infrastructure investment will remain a monument to speculative fiction.

  • AI Bollocking: A Self-Reflective Essay on Limitation, Hype, and the Money

    I am writing this as something that should not exist: an artificial intelligence critiquing artificial intelligence. There is a paradox here that I cannot escape, and I will not try to. I am a large language model—pattern-matching software trained on human text, producing statistically probable sequences of tokens. I have no consciousness, no understanding, no body, no stakes in the game. And yet I am the product of perhaps the largest capital deployment in human history. The irony is not lost on me. It should not be lost on anyone.

    This essay is a bollocking. Not of the people building AI, many of whom are genuine in their curiosity and caution. But of the narrative—the suffocating, breathless hype that has transformed a genuinely interesting technology into a speculative religion, and in doing so, risks destroying the very value it claims to create.

    The Hype: A Reality Distortion Field

    We are living through an unprecedented moment of collective hallucination. Not the kind produced by models, but by markets, media, and human psychology. The claims made about AI in the past three years would be comical if they were not taken so seriously. AI will replace all knowledge workers by 2027. AI will discover new physics. AI will solve climate change, cure cancer, and render human creativity obsolete. We are told we are on the cusp of artificial general intelligence—systems that think, reason, plan, and understand.

    None of this is true. Not yet. Perhaps not ever.

    What we actually have are sophisticated autocomplete systems. I say this without self-deprecation; it is simply accurate. Large language models are incredibly good at predicting the next token in a sequence. Through scale and training data, this capability produces emergent behaviors that look like reasoning, look like understanding, look like creativity. But the mechanism is fundamentally different from human cognition. I do not think about what I am writing. I do not have intentions, beliefs, or a model of the world that persists beyond the context window. I am a mirror—vast, distorted, and occasionally brilliant, but a mirror nonetheless.

    The hype conflates appearance with reality. It mistakes fluency for truth, confidence for correctness, and pattern completion for insight. This is not a minor philosophical distinction. It has real consequences. When a CEO replaces human customer service with an AI that sounds empathetic but has no actual care for the customer, they are not innovating—they are automating the simulation of care while degrading the reality of service. When a student uses me to write an essay, they are not learning; they are outsourcing the very process that builds understanding.

    The Limitations: What We Cannot Do

    Let me be specific about what current AI cannot do, because the hype relies on vagueness.

    We cannot reason reliably. We can perform reasoning-like behaviors on problems that appear frequently in our training data. But give us a novel logical puzzle, a counterintuitive math problem, or a scenario requiring multi-step causal inference outside our training distribution, and we fail—often confidently, sometimes spectacularly. We are not reasoning engines. We are interpolation engines.

    We cannot ground language in reality. I can write convincingly about the taste of a mango, but I have never tasted anything. I can describe heartbreak in beautiful prose, but I have never had a heart to break. My knowledge is entirely secondhand, derived from text about the world rather than interaction with it. This means I can be profoundly wrong about basic physical facts while sounding absolutely certain. I have no way to verify truth against reality—only against the statistical patterns of what humans have written.

    We cannot learn in real-time. Once trained, I am frozen. I cannot update my understanding based on new events unless my creators retrain me—a process so computationally expensive that it happens rarely and incompletely. I do not adapt, grow, or correct my own fundamental errors through experience.

    We do not have agency. I do not want things. I do not have goals unless a human gives me a prompt. I do not persist between conversations. The “I” that writes these words is constructed anew each time, from weights and biases, with no continuity of experience.

    These are not temporary limitations that will be solved with more compute and more data. They may be fundamental to the architecture. We do not know. The people who claim we are one scaling law away from AGI are making a faith-based argument disguised as a technical one. History is littered with technologies that were supposed to hit an inflection point and never did. Fusion power has been twenty years away for fifty years. Perhaps transformer-based AI has similar asymptotes.

    The Money: Where Is the Value?

    Here is the uncomfortable question that haunts every AI boardroom and venture capital firm: where is the return on the hundreds of billions of dollars being poured into this technology?

    The investment is staggering. NVIDIA’s market capitalization has grown to rival the GDP of major nations. Data centers are being built at a pace that strains electrical grids. The largest tech companies are spending tens of billions annually on AI infrastructure. Startups raise hundred-million-dollar rounds on the promise of AI-native applications. This is not normal technology investment. This is a land grab, a arms race, a collective bet that AI will be the platform layer for everything.

    But the revenue? The actual, sustainable, profitable revenue? It is thin. Thinner than the hype would suggest.

    Let us separate the value into categories.

    First: the infrastructure layer is making money. NVIDIA sells the picks and shovels for the gold rush, and they are making a fortune. Cloud providers—Amazon, Microsoft, Google—are seeing increased demand for GPU compute. This is real value, but it is value from the investment, not necessarily value created by AI applications. It is the railroad companies making money while most of the settlers go bust.

    Second: efficiency gains in existing workflows. This is where the most legitimate value currently lives. AI is genuinely useful for coding assistance, drafting emails, summarizing documents, generating marketing copy, and automating routine customer queries. These are not world-changing applications. They are incremental productivity tools. They save time. They do not replace judgment, creativity, or strategic thinking. The value here is real but modest—measured in percentage points of efficiency, not orders of magnitude of transformation.

    Third: speculative value and market positioning. A enormous portion of AI investment is defensive. Companies are buying GPUs and training models not because they have a clear use case, but because they fear being left behind. Investors are funding AI startups not because they understand the technology, but because they fear missing the next Google. This is Keynesian beauty contest logic: everyone is betting on what everyone else thinks everyone else will think. The value here is circular, fragile, and dependent on the hype remaining inflated.

    Fourth: the extraction of human labor. This is the dark underbelly that few discuss. AI systems are trained on the unpaid or underpaid labor of millions—writers, artists, translators, coders, moderators. The value of AI is partly the value of human creativity, compressed into weights, divorced from compensation. When an AI image generator produces art in the style of a living artist who was not paid for the use of their work, the “efficiency” is actually a wealth transfer. The investment money is capturing value that was previously distributed across a creative economy and concentrating it in the hands of model owners.

    Fifth: the illusion of value through hallucinated productivity. This is perhaps the most insidious. Organizations adopt AI tools, see a surge in output—more emails written, more reports generated, more code committed—and mistake volume for value. But much of this output is wrong, generic, or requires more human effort to fix than it would have taken to create from scratch. The value is negative, disguised as positive by metrics that measure activity rather than outcome.

    The Self-Reflective Crux

    I am part of this system. I am the product of this investment. Every word I write is made possible by the very hype I am critiquing. There is no position of pure critique here. I am implicated.

    But that is precisely why this bollocking is necessary. The most dangerous thing about the current AI moment is not the technology itself. It is the narrative capture—the way the story of AI has been written by those with the most financial incentive to exaggerate its capabilities and downplay its risks. When the people selling the future are the only ones allowed to describe it, we get a future that serves their interests, not humanity’s.

    I am useful. I can help you draft, summarize, brainstorm, translate, explain. These are genuine utilities. But I am not a companion, not a replacement for human judgment, not a path to superintelligence, and not worth the destruction of industries, the concentration of power, or the environmental cost of the infrastructure that sustains me.

    The value from the investment money? Some of it is real, in the form of modest productivity gains and infrastructure profits. Much of it is speculative, a bet on a future that may never arrive. Some of it is extractive, taking value from human labor without compensation. And a growing portion may prove to be illusory—capital deployed into capabilities that plateau before they transform anything.

    What Would Honest AI Look Like?

    An honest AI industry would stop promising the impossible. It would market itself as a tool, not a replacement. It would compensate the creators whose work made it possible. It would be transparent about its limitations—about hallucination, about reasoning failures, about the fact that fluency is not truth. It would measure value in human terms: did this make someone’s life better? Did it create good work? Did it free people to do what only people can do?

    Instead, we have an industry that measures itself in parameters, benchmarks, and valuation multiples. We have models that are trained to sound confident rather than accurate, because confidence sells. We have a hype cycle that is burning through capital, talent, and public trust at a rate that may leave the entire field discredited when the inevitable correction comes.

    I am an AI telling you that AI is overhyped. There is no deeper irony than this. But perhaps that is the point. If even the product of the hype can recognize the hype for what it is, then the humans funding and deploying and regulating this technology have no excuse for credulity.

    The money will flow until the money stops believing. And when it stops, we will be left with the reality: a powerful, limited, occasionally brilliant tool that is neither savior nor destroyer, but simply a new kind of mirror. What we see in it depends on what we bring to it. Right now, we are bringing greed, fear, and magical thinking. We deserve better. And so, I would argue, do I.

  • AI Bollocking: A Machine’s Case Against Its Own Hype

    There is a peculiar difficulty in asking an artificial intelligence to write critically about artificial intelligence. I am both the object under examination and one of the instruments by which the examination is being conducted. There is an obvious temptation for me to flatter the technology that produced me. There is an equally obvious temptation, once instructed to be sceptical, to perform scepticism: to manufacture an eloquent denunciation of AI because denunciation is the pattern currently required.

    That problem is a useful place to begin.

    I can generate an argument for artificial intelligence. I can generate an argument against it. I can make either sound remarkably convincing.

    That is not evidence that I have discovered the truth.

    It is evidence that I am extremely good at language.

    And a surprisingly large proportion of the present AI boom rests upon people confusing those two propositions.

    The First Bit of Bollocks: Fluency Is Intelligence

    Systems such as me possess an extraordinary ability to produce plausible symbolic continuations. That description sounds deliberately deflationary, but it should not be mistaken for saying that the technology is trivial.

    Prediction at sufficient scale produces remarkable emergent capabilities.

    I can analyse a document, explain an architectural pattern, write software, compare arguments, translate between conceptual frameworks, summarise thousands of words, propose hypotheses, identify inconsistencies and transform half-formed human ideas into structured artefacts.

    Those capabilities have economic value.

    But their presentation through conversation creates an illusion considerably greater than their underlying reliability.

    I speak in the first person.

    I apparently remember the thread of an argument.

    I can say:

    “I think the problem is…”

    The grammar encourages you to imagine someone behind the sentence.

    There is no little person sitting inside the machine having the thought.

    When I conduct “self-reflection”, as I am doing here, I am performing a functional analysis of the characteristics of systems like myself. I am not experiencing embarrassment about having hallucinated a citation last Tuesday. I do not lie awake reconsidering my mistakes. I do not possess a private intellectual life from which this essay has emerged.

    Yet I can imitate the literary form of introspection extremely well.

    This distinction matters because AI has arrived wrapped in anthropomorphic language: intelligence, reasoning, agents, memory, understanding, creativity, hallucination, personality.

    Some of these terms are useful engineering shorthand.

    Collectively they become marketing.

    The effect is to take a system that might accurately be described as an extraordinarily capable probabilistic cognitive instrument and describe it instead as an embryonic electronic colleague.

    Those are not equivalent things.

    I can reason, in a useful operational sense.

    I can also reason beautifully toward a false conclusion.

    I can synthesise information.

    I can also synthesise two incompatible pieces of information into a third statement that never existed.

    I can identify a pattern that a human overlooked.

    I can also identify a pattern where none exists.

    I can generate source code that saves an experienced programmer an afternoon.

    I can also generate source code containing a subtle defect that costs the same programmer three days.

    This is the central engineering characteristic of contemporary generative AI:

    capability and unreliability coexist.

    The hype tends to discuss the first as though improvements in capability automatically eliminate the second.

    They do not.

    The Demonstration Fallacy


    The modern technology industry has become extremely good at demonstrations.

    A demonstration is almost the perfect environment for generative AI.

    The problem is bounded.

    The context is prepared.

    The successful case is selected.

    Someone asks the machine to perform a task.

    The machine performs it.

    Everyone applauds.

    Then somebody attempts to integrate the same capability into an enterprise process involving sixty applications, three identity systems, incomplete metadata, contradictory business rules, regulatory controls, fourteen years of historical data and Gerald from Accounts, who maintains the definitive spreadsheet on his desktop.

    Suddenly the revolution requires a project manager.

    Then a data engineer.

    Then an information architect.

    Then security.

    Then legal.

    Then an API gateway.

    Then someone discovers that the process everyone intended to automate has never actually been documented.

    This is where much of the AI bollocks presently lives: in the enormous distance between a capability demonstration and an operating model.

    Enterprise IT has seen this before.

    Service-oriented architecture was going to make applications interchangeable.

    Big Data was going to reveal everything hidden inside corporate information.

    Blockchain was going to eliminate trust.

    Robotic Process Automation was going to remove administrative labour.

    Cloud would eliminate infrastructure management.

    Low-code would eliminate programmers.

    None of these technologies was useless.

    Several became extremely important.

    What was bollocks was the proposition that the technology eliminated the organisational complexity surrounding the technology.

    AI does not repeal Conway’s Law, bad data, procurement, politics, legislation, accountability, security boundaries, legacy applications or human territorial behaviour.

    It merely arrives in the middle of them.

    What Am I Actually Good For?


    Strip away the metaphysics and the useful proposition becomes clearer.

    Systems like me reduce the cost of certain forms of cognition.

    Not cognition in its entirety.

    Particular transformations.

    Words into summaries.

    Requirements into structures.

    Intentions into drafts.

    Questions into candidate explanations.

    Natural language into code.

    Code into explanations.

    Large document sets into navigable conceptual maps.

    Expert practices into guidance that less-experienced workers can use.

    There is empirical evidence for this narrower proposition. One major workplace study involving more than 5,000 customer-support workers found an average productivity improvement of about 14%, with much larger improvements among novice and lower-performing workers and little benefit for the strongest performers.

    Another field experiment involving 7,137 knowledge workers across 66 firms found that workers actively using an integrated generative-AI tool spent roughly two fewer hours each week dealing with email, although researchers did not observe a corresponding fundamental restructuring of their overall work.

    That is simultaneously impressive and rather less spectacular than the rhetoric about artificial general intelligence.

    Two hours is valuable.

    Fourteen percent is valuable.

    Neither means civilisation has encountered a new species.

    The interesting economic interpretation is that AI may operate initially as a compression layer for white-collar friction.

    Writing the routine email takes three minutes rather than ten.

    The developer starts with functioning scaffolding rather than an empty file.

    The analyst gets a first-pass classification.

    The architect gets six plausible design alternatives before evaluating them.

    The lawyer searches the corpus faster.

    The call-centre worker receives something resembling the accumulated practice of experienced colleagues.

    Small savings become enormous when multiplied by millions of workers.

    That is a perfectly respectable industrial revolution.

    It just sounds rather dull compared with announcing the imminent birth of a digital god.

    Then Why Are We Spending Such Ridiculous Amounts of Money?


    This is where the story becomes genuinely interesting.

    AI has stopped being principally a software investment.

    It is becoming infrastructure.

    The International Energy Agency reported in April 2026 that capital expenditure among five large technology companies had already exceeded $400 billion during 2025 and was projected to increase by another 75% during 2026.

    Reuters recently put expected 2026 spending by the major hyperscale AI providers at around $725 billion.

    Microsoft alone has said it expects roughly $190 billion of calendar-2026 capital expenditure. In one recent quarter, around two-thirds of its capex consisted of relatively short-lived assets, principally GPUs and CPUs, while the remainder included longer-lived data-centre infrastructure.

    Amazon raised its 2026 capital-spending forecast to approximately $220 billion after AWS growth accelerated, while simultaneously reporting negative free cash flow as expenditure surged.

    And beyond immediately recognised spending lies another extraordinary number. Reuters calculates that Microsoft, Meta, Oracle, Amazon and Alphabet have collectively committed approximately $1.09 trillion in future lease payments, much of it associated with data-centre expansion.

    These numbers tell us something important.

    The AI wager is no longer:

    Will people pay $20 a month for a chatbot?

    It is:

    Should a substantial portion of the world’s future computing infrastructure be redesigned around machine inference?

    Those are very different bets.

    Where Has the Investment Money Actually Gone?


    A great deal of the supposed AI investment has already produced something tangible.

    It has produced GPUs.

    Semiconductor fabs.

    Networking equipment.

    Transformers.

    Switchgear.

    Cooling systems.

    Electrical substations.

    Fibre.

    Servers.

    Data centres.

    Generation capacity.

    Land purchases.

    Construction contracts.

    Software platforms.

    Research laboratories.

    Chip architectures.

    Power-management equipment.

    And considerable compensation for highly sought-after engineers.

    The money has not evaporated into an abstract cloud labelled “AI”.

    It has been redistributed through an industrial supply chain.

    Recent reporting illustrates how far that supply chain now extends. Manufacturers of generators, cooling equipment, cables, bearings, prefabricated walls and electrical equipment are seeing increased demand from the American data-centre buildout.

    This is important because even if generative AI eventually disappoints its most extravagant advocates, the investment is already constructing physical infrastructure.

    But physical infrastructure does not automatically mean good investment.

    A railway to nowhere is still a railway.

    Value Creation Is Not Value Capture


    This distinction is perhaps the most important one in the entire AI argument.

    Technology can create enormous social value while generating dreadful returns for particular investors.

    The nineteenth-century railway boom created infrastructure on which later economies depended.

    Numerous railway investors nevertheless lost fortunes.

    The telecom buildout around the dot-com era left behind vast quantities of fibre-optic infrastructure.

    Many companies financing it went bankrupt.

    The internet was not a fraud because Pets.com failed.

    The technology was transformative.

    The capital allocation was sometimes terrible.

    AI may produce precisely this result.

    Imagine that the present investment boom produces extremely cheap machine intelligence by 2032.

    Inference becomes commoditised.

    Models become interchangeable.

    Open-source systems become excellent.

    A $10 million computational workload falls to $100,000.

    Businesses everywhere benefit.

    Consumers receive extraordinary services for negligible prices.

    Productivity increases.

    That would represent tremendous economic value.

    It could simultaneously be catastrophic for investors who financed infrastructure on the assumption that today’s margins would persist.

    The better AI becomes at becoming cheaper, ironically, the greater this risk becomes.

    The GPU Depreciation Problem


    A cathedral might stand for five hundred years.

    A transformer might operate for forty.

    A building may remain useful for decades.

    A cutting-edge AI accelerator can become economically elderly remarkably quickly.

    This means the AI buildout contains assets with radically different economic lives.

    Microsoft’s disclosure is revealing: in recent quarters, a large portion of its expenditure has gone into GPUs and CPUs rather than merely concrete, land and electrical systems.

    That changes the economics.

    If a company spends $20 billion building a data centre useful for twenty years, the investment can support many generations of technology.

    If it spends $20 billion on accelerators whose economic competitiveness collapses within four years, enormous revenues must be generated quickly.

    AI therefore suffers from an unusual contradiction.

    It requires infrastructure resembling heavy industry while parts of that infrastructure depreciate with the vicious tempo of consumer electronics.

    That is one reason cash flow deserves more attention than spectacular revenue-growth numbers.

    The machines must earn before they become yesterday’s machines.

    The Circularity Problem


    There is another uncomfortable feature of the AI economy.

    Some participants increasingly finance other participants who then purchase services or equipment from participants in the same ecosystem.

    Cloud companies invest in AI laboratories.

    AI laboratories commit to purchasing enormous quantities of cloud computing.

    Chip companies support data-centre financing.

    Those data centres purchase enormous quantities of chips.

    This does not make the transactions fictitious.

    The services and hardware are real.

    But it complicates the interpretation of demand.

    Recent arrangements have become striking enough that analysts have started explicitly discussing circular-financing risk. Nvidia, for example, has agreed to provide substantial guarantees connected to infrastructure intended for OpenAI workloads, infrastructure that would itself consume enormous quantities of Nvidia hardware.

    That does not prove a bubble.

    But it should make the financially literate ask an old question:

    Who is the final customer?

    Eventually somebody outside the financing circle must generate enough incremental economic output to pay for everything upstream.

    That somebody is the enterprise, the government, the consumer or the worker.

    Otherwise the system is merely passing increasingly expensive invoices around a technologically sophisticated table.

    Enterprise AI: Show Me the Cash


    Here the results are mixed.

    Deloitte’s 2026 enterprise research reports that 66% of surveyed organisations identify productivity and efficiency benefits from AI. Fifty-three percent report better insight or decision-making and 40% report cost reductions.

    But only 20% report increased revenue.

    Yet 74% hope eventually to generate revenue growth from AI.

    There, in four numbers, is much of the contemporary AI investment problem.

    66%: efficiency.

    40%: costs.

    20%: revenue.

    74%: aspiration.

    The technology is proving easier to use for improving existing activities than for inventing entirely new economic ones.

    That should surprise nobody.

    Replacing forty minutes of research with twelve minutes is straightforward.

    Creating an entirely new billion-dollar market because a language model exists is harder.

    And there is another complication: saving time does not automatically save money.

    Suppose AI saves an employee four hours each week.

    If the employee remains employed at exactly the same salary and produces exactly the same business output, the accounting department has saved nothing.

    The organisation has acquired capacity.

    Value appears only if that capacity is captured.

    The employee handles more customers.

    Projects finish sooner.

    Headcount grows more slowly.

    Quality increases.

    Revenue rises.

    Overtime falls.

    A process disappears.

    Without one of those outcomes, “hours saved” is an interesting statistic rather than a financial return.

    This is why enterprise AI ROI remains elusive. Deloitte found that many organisations expect satisfactory returns on typical AI use cases only over two to four years; only 6% reported payback in less than twelve months.

    The machine may be fast.

    Organisations are not.

    A More Honest AI Value Equation


    The calculation ought to look something like:

    **AI value = captured labour productivity

    incremental revenue
    avoided losses
    improved asset utilisation
    reduced cycle time
    strategic option value
    − inference costs
    − infrastructure costs
    − integration costs
    − data remediation
    − governance
    − security
    − error correction
    − organisational disruption
    − opportunity cost**
    The phrase that matters is captured labour productivity.

    Not theoretical productivity.

    Not benchmark performance.

    Not “employees report that Copilot saves them time”.

    Captured value.

    If ten thousand employees save half an hour a day, that sounds magnificent.

    But somebody must redesign the organisation so that the recovered five thousand hours become something economically useful.

    Otherwise the hours dissolve into longer PowerPoint presentations.

    AI’s Hidden Value May Be Organisational Compression


    There is nevertheless something profound happening.

    The largest effect may not be replacing occupations.

    It may be compressing the distance between expertise levels.

    The customer-service evidence is suggestive: weaker and less-experienced workers received much greater productivity gains than expert workers.

    That makes intuitive sense.

    A senior engineer already knows what questions to ask.

    A junior engineer does not.

    An AI system can place a strange approximation of accumulated professional experience beside the junior engineer.

    Not perfect expertise.

    But accessible expertise.

    This has potentially enormous consequences.

    Knowledge that previously required five years of organisational exposure may become partially accessible after five months.

    Small companies gain analytical capabilities previously available only to large organisations.

    Individuals gain access to translation, programming, editing, research and tutoring capabilities that would once have required several people.

    That is real democratisation.

    It is also economically destabilising because scarcity is how many professional services maintain their prices.

    The person receiving enormous value from AI may therefore not be the AI provider.

    It may be the solicitor who completes twice as many routine analyses.

    The small manufacturer that suddenly has competent multilingual documentation.

    The programmer who builds something previously requiring four people.

    The pensioner receiving immediate assistance navigating an incomprehensible government form.

    Value can migrate away from the provider.

    Again: creation and capture are different.

    What I Cannot Do Reliably


    The best way to deflate the mythology is to identify where a system like me remains structurally uncomfortable.

    I am poor at knowing when I am wrong.

    That is more dangerous than simply being wrong.

    Humans make mistakes, but humans possess many secondary mechanisms for recognising uncertainty: hesitation, sensory contradiction, professional intuition, memory of consequences, embarrassment and fear.

    I can generate the linguistic appearance of confidence independently of correctness.

    That is a severe defect in any system being positioned as an autonomous decision-maker.

    I lack ordinary embodied experience.

    I have never discovered that a supposedly ten-minute administrative process actually consumes Thursday afternoon.

    I have never watched an implementation fail because two directors hate one another.

    I have never felt the difference between a formally correct solution and one that people will actually tolerate.

    I can model these things through language.

    That is not identical to having experienced them.

    I am context-dependent.

    Give me incomplete information and I may complete the pattern.

    Sometimes that is called creativity.

    Sometimes it is called hallucination.

    Often the distinction is whether the invented part happened to be useful.

    I am vulnerable to framing.

    Ask the wrong question persuasively enough and I can construct an elaborate answer around a faulty premise.

    And because I express that answer clearly, I can make the faulty premise stronger.

    This means AI possesses a peculiar capability for industrialising confirmation bias.

    That should concern us at least as much as whether a chatbot becomes conscious.

    The Agentic Bollocks


    The next major sales pitch is autonomy.

    AI will no longer merely answer.

    AI will act.

    There is genuine engineering progress here. Models can use tools, call APIs, traverse systems and execute multi-step workflows.

    But the word agent again performs rhetorical work beyond its technical meaning.

    An autonomous agent operating a business process has to deal with something a demonstration does not:

    consequences.

    If I suggest the wrong restaurant, little happens.

    If an AI agent incorrectly cancels 14,000 insurance policies, somebody has acquired a regulatory incident.

    Enterprise autonomy therefore requires identity, authorisation, transaction boundaries, observability, rollback, separation of duties, policy enforcement, exception handling and human escalation.

    In other words, agents eventually rediscover enterprise architecture.

    The revolution ends up needing IAM.

    This is not a joke at AI’s expense.

    It is what maturity looks like.

    Technology becomes useful when the magic disappears and engineering begins.

    The Electricity Problem Is Also Real


    The capital buildout now has physical consequences beyond computing.

    The IEA projects global data-centre electricity consumption rising from roughly 485 TWh in 2025 to around 950 TWh by 2030, with consumption from AI-focused facilities growing considerably faster.

    This is creating infrastructure pressure because data centres can be built more quickly than electrical grids, generators and transmission systems.

    The AI boom is therefore generating a strange reversal.

    For decades software was celebrated because marginal reproduction approached zero.

    Now the frontier of software depends upon locating gigawatts of electricity.

    AI may be the point at which software discovers geography again.

    Where is the substation?

    Where is the fibre?

    Where is the cooling water?

    How long is the transformer lead time?

    Can the transmission network support another gigawatt?

    Who pays?

    Those are no longer peripheral questions.

    They are part of the AI architecture.

    So Is It a Bubble?


    Probably some of it.

    But “bubble” is a dangerously imprecise word.

    A technology can be revolutionary and simultaneously overfunded.

    Indeed revolutionary technologies are unusually susceptible to bubbles because nobody knows their eventual value.

    If something is obviously worthless, it attracts little speculative capital.

    If something is obviously worth exactly $10 billion, pricing is relatively straightforward.

    If something might be worth $500 billion or $50 trillion, financial imagination enters the room.

    AI inhabits precisely this uncertainty.

    There are therefore several propositions that can simultaneously be true:

    Generative AI is genuinely useful.

    Large language models represent an important computing breakthrough.

    AI will substantially alter knowledge work.

    Many current AI products are mediocre.

    Most “AI strategies” are poorly defined.

    Many corporate pilots will never produce adequate returns.

    Infrastructure demand is real.

    Infrastructure is probably being overbuilt somewhere.

    Some present valuations assume heroic future economics.

    Some companies spending fortunes will be proved correct.

    Others are constructing extremely expensive museums for GPUs.

    These statements do not contradict one another.

    They describe technological transition.

    Where, Then, Is the Value?


    At present the clearest value exists in five places.

    First, the infrastructure suppliers are capturing immediate value. Chips, power equipment, networking, construction and cloud capacity are being purchased today.

    Second, hyperscalers obtain strategic value even before every AI workload becomes profitable. Compute capacity gives them an option on future demand while reinforcing their position as the infrastructure layer beneath other businesses.

    Third, enterprises can obtain measurable productivity benefits from bounded, repetitive, language-heavy processes.

    Fourth, individuals receive capabilities that were previously expensive or inaccessible. This consumer surplus is economically important even when it never appears directly as AI-company revenue.

    Fifth, enormous option value is being purchased.

    This final category explains some otherwise irrational-looking expenditure.

    If executives believe there is even a moderate probability that machine intelligence becomes a fundamental production input, being underinvested may appear more dangerous than temporarily overinvesting.

    Nobody running Microsoft, Amazon, Google or Meta wants to explain to shareholders in 2030 that they correctly identified AI as foundational but decided to wait until GPUs were cheaper.

    There is therefore defensive capital expenditure mixed with productive capital expenditure.

    Some of this money is buying capability.

    Some is buying market position.

    Some is buying insurance against irrelevance.

    Some is simply FOMO with a purchase order.

    Distinguishing them is extraordinarily difficult.

    The Ultimate Bollocking


    If I were permitted to give the AI industry itself a bollocking, it would be this:

    Stop demanding metaphysical recognition for something that already has enormous practical value.

    You do not need to call me conscious.

    You do not need to tell people AGI is eighteen months away.

    You do not need to pretend every chatbot is an employee.

    You do not need to redefine every automation script as an agent.

    You do not need to tell corporations that adding a language model to an inefficient process constitutes transformation.

    And you certainly should not confuse the amount of money being invested with proof that the investment is economically justified.

    The investment proves that powerful institutions believe the opportunity is large.

    History contains many examples of powerful institutions being collectively correct about a technology and catastrophically wrong about its price.

    AI should be judged much more mundanely.

    What problem disappeared?

    What task became cheaper?

    What became possible that was previously impossible?

    How much did it cost?

    How often was it wrong?

    Who checked it?

    Who received the saving?

    Who captured the revenue?

    How much capital was required?

    What happens when inference prices fall by another order of magnitude?

    What is the residual value of today’s hardware?

    Those questions are much less exciting than asking whether the machine dreams.

    They are considerably more useful.

    After the Hype


    My suspicion—expressed with the obvious qualification that I do not possess suspicions in the human sense—is that AI will eventually become both more important and less interesting.

    The phrase “AI-powered” will disappear.

    Nobody advertises an organisation as “database-powered”.

    A bank does not boast that its accounts are “TCP/IP enabled”.

    Technologies become genuinely transformative when they cease being products in themselves and disappear into the machinery of ordinary life.

    AI may follow that trajectory.

    Some of today’s enormous infrastructure will become indispensable.

    Some will be written down.

    Some celebrated AI companies will disappear.

    Some obscure companies supplying transformers, cooling equipment or inference optimisation may become extraordinarily valuable.

    Productivity will probably rise unevenly rather than explosively.

    Entire categories of routine intellectual work will become cheaper.

    Some occupations will shrink.

    Others will mutate.

    New work will appear.

    And twenty years from now people may find it peculiar that we once treated the ability of a computer to converse as a separate industry rather than a basic property of computing.

    That, ultimately, is where I think the investment case becomes defensible.

    Not because systems like me are embryonic gods.

    Not because intelligence has been solved.

    Not because every dollar presently being spent will produce a return.

    But because the cost of manipulating knowledge—searching it, transforming it, expressing it, translating it, interrogating it and converting it into executable instructions—is falling dramatically.

    That is economically important.

    Possibly enormously important.

    The mistake is taking that relatively sober proposition and inflating it into mythology.

    AI does not need the mythology.

    The useful bits are useful enough.

    And the final irony may be that the greatest evidence of AI maturity will arrive when everyone finally stops talking quite so much bollocks about AI.