I work in artificial intelligence. I should be excited about the moment we’re in. According to the industry I belong to, we are weeks away from artificial general intelligence, months away from the obsolescence of all knowledge work, and years away from a post-scarcity utopia administered by benevolent algorithms. I receive these forecasts in my inbox several times a day, sometimes from people I respect and sometimes from people who have never deployed a model to production but have very strong opinions about what models will do next.
I am, instead, mildly tired. Not cynical. Tired in the specific way one becomes when one’s profession has been kidnapped by marketing departments and held for ransom in the form of LinkedIn posts. The same field that, fifteen years ago, was a quiet corner of computer science where you tried to make a network classify a cat — and failed, often, sometimes spectacularly — has become the most over-promised, over-funded, and over-hyped human enterprise since the internet itself, which itself was the most over-hyped enterprise since the railways, which were over-hyped enough to crash the British economy in 1845. We have done this before. We will do it again. We are, apparently, doing it now, and at significantly higher resolution.
What I have learned, working in the field while watching the discourse about the field, is that the loudest voices have the weakest models. The people building the actual systems know what the actual systems can and cannot do. The people predicting the singularity have rarely had to debug a CUDA out-of-memory error at 9 AM with a deadline in three hours and a model that was working perfectly the day before for reasons no one fully understands. Both populations exist. Only one of them dominates the discourse. This is not a coincidence. This is noise.
There is also a deeper issue, which is the one that actually matters. The current generation of AI systems is genuinely impressive at producing fluent text and images by predicting the next plausible token from a vast statistical model of what humans have already written. This is real. It is also not what most people mean by intelligence. A growing number of researchers — including some of the most respected figures in the field, who happen to have built much of its theoretical foundation — argue that genuine intelligence requires something the current paradigm doesn’t have: an internal model of the world, learned through perception and interaction, that allows for prediction, planning, and reasoning about cause and effect. Not next-word prediction at scale. World modeling. The two are different in a way that the marketing department is paid not to mention.
I happen to share this view. The conviction is that intelligence — in machines or in people — begins with seeing the world, building a model of how it behaves, and updating that model when reality disagrees. Not by reading more text. The current systems can write a beautiful essay about how to ride a bicycle. None of them can ride one. There is, in this gap, the entire problem.
Through the lens of information theory
Claude Shannon, in A Mathematical Theory of Communication (1948), gave us the formal vocabulary for what we are doing wrong. A communication channel has a finite capacity. Beyond that capacity, additional signal cannot be reliably transmitted — it begins to degrade into noise. The channel is not infinite. The receiver is not infinite. Information that exceeds capacity does not become more information. It becomes interference, which is a polite engineering term for “indistinguishable from being yelled at by everyone simultaneously.”
Shannon also defined entropy in the context of messages: a measure of uncertainty. A predictable message has zero entropy. A random one has maximum entropy. Most useful communication lives between these extremes — structured enough to carry meaning, uncertain enough to be worth transmitting. A message that is entirely predictable adds no information. A message that is entirely random adds no information either. Both, despite appearances, communicate nothing. Most of modern public discourse, I suspect, lives at one of these two extremes and rarely visits the productive middle.
The paradox is precise enough to be ironic: we have access to more facts than any humans in history, and we are measurably worse at distinguishing the true ones. This is not a moral failure. It is a bandwidth problem. Shannon would have predicted it. The fact that he did, in 1948, and we still walked into it anyway, is its own kind of structural commentary on the species — the equivalent of being warned about the cliff and then running toward it slightly faster.
Through the lens of economic history
The current AI hype cycle is not unprecedented. It is, by my count, the fourth or fifth iteration of the same basic narrative structure within my lifetime, which is a remarkable rate considering that historically these things took generations to recur.
The dot-com bubble of the late 1990s promised that the internet would transform every business, eliminate intermediation, and create unimaginable wealth. It eventually did most of these things, but only after the speculative bubble collapsed in 2000 and wiped out roughly $5 trillion in market capitalization. The genuine transformation arrived twenty years later, mostly through companies that didn’t exist when the bubble peaked, which suggests that the people most confident about the future are usually pointing at the wrong part of it.
The cryptocurrency cycle of 2017–2018 promised that decentralized finance would replace banks. The NFT cycle of 2021 promised that JPEGs of pixelated apes were a new asset class. Both cycles produced enormous wealth for early entrants, enormous losses for late ones, and a small residue of genuine technological progress that was, in retrospect, indistinguishable from what serious researchers had been working on quietly the entire time.
Carlota Perez, in Technological Revolutions and Financial Capital (2002), described these cycles as a recurring historical pattern: an installation phase characterized by speculative frenzy and irrational pricing, followed by a deployment phase where the technology actually integrates with the economy. Her analysis goes back to the railway mania of the 1840s and the electrification of the 1900s. The pattern is so consistent that it should embarrass anyone who claims this time is different. This time is rarely different. The technology is often real. The hype is always disproportionate. And the people most confident about the precise shape of the future are, almost without exception, the people most exposed to its disappointment — and the people selling tickets to it.
The current AI cycle has all the textbook features. Companies with no revenue valued in tens of billions. Public claims that wildly exceed private capabilities. A vocabulary explosion — agentic, sovereign AI, embodied AI, recursive self-improvement, alignment as a service — most of which describes products that don’t exist or capabilities that haven’t been demonstrated outside a demo video filmed under conditions you’ll never replicate. The technology is genuinely transformative. So was the railway. Both can be true at the same time, and the hype is what eventually breaks. The technology survives the hype. It just does so quietly, in the background, after the conference circuit has moved on.
Through the lens of thermodynamics and infrastructure
Here is where the noise becomes literal. Every prompt sent to a large language model requires electricity, computation, and cooling. The cooling requires water. The electricity, in much of the world, requires fossil fuels. The infrastructure to support modern AI is among the most resource-intensive industrial activities ever undertaken — and the people most enthusiastic about the technology rarely discuss it, because the metaphor of “the cloud” has done its job too well.
The numbers are not metaphorical, and they come from sources that don’t have a financial stake in the outcome. According to the International Energy Agency’s Energy and AI report (2025), global data centers consumed approximately 415 terawatt-hours of electricity in 2024 — about 1.5% of global electricity consumption — and are projected to reach roughly 945 TWh by 2030. The IEA is not an environmental NGO; it was created by oil-importing nations after the 1973 oil crisis. When the IEA says energy demand is climbing, the data tends to be conservative.
In the United States alone, data centers consumed about 4.4% of total national electricity in 2024 according to a 2024 report from Lawrence Berkeley National Laboratory commissioned by the U.S. Department of Energy. The same report estimated that U.S. data centers directly consumed 17 billion gallons of water in 2023, with hyperscale facilities responsible for roughly 84% of that total. These are not advocacy figures. They are infrastructure planning numbers from a national laboratory.
A 2025 commentary in Joule — a Cell Press peer-reviewed journal — synthesized publicly available corporate data and IEA estimates and concluded that AI systems alone could be responsible for between 32.6 and 79.7 million tons of CO₂ emissions in 2025, and water consumption between 312.5 and 764.6 billion liters. The author was explicit that significant uncertainty surrounds these figures because most large AI companies do not disclose granular environmental data, which is itself a useful signal about how they expect the disclosure to be received.
To put this in human terms: every 100-word AI prompt consumes roughly half a liter of water for cooling — about a small bottle — according to research from the University of California, Riverside. A user generating images and videos with current models consumes meaningfully more. A 2024 Bloomberg News investigation, using direct surveying of recently constructed data centers, found that approximately two-thirds of new data centers built since 2022 are located in regions already experiencing water stress. We are cooling silicon with water in places that are running out of water, in order to generate text and images, much of which is used to argue on social media about whether AI is good or bad for humanity. There is a circularity here that the universe itself could not have improved on.
I am not against AI. I work in AI. I think it is, on balance, an extraordinary technology with extraordinary potential. I am pointing out that the current discourse about AI almost entirely omits the physical substrate on which it runs, which is itself a form of cognitive offloading: we have abstracted away the cost of the abstraction. The cloud is not a cloud. It is a building. The building uses electricity and water. And the people most committed to “intelligence” as a marketing concept tend to be the least curious about what makes the underlying computation possible. There is, in this, a small but unmistakable irony.
Through the lens of cognitive science
Here the evidence is more contested, and I want to be careful about what we can and cannot conclude. Several recent studies have examined whether using AI tools affects human cognition, and their results have been widely cited, often with more confidence than the methodologies justify.
The most viral example is a 2025 preprint from MIT’s Media Lab — Kosmyna and colleagues — which used EEG to measure brain activity in 54 participants writing essays under three conditions: with ChatGPT, with a search engine, and with no tools. The study reported that ChatGPT users showed weaker brain connectivity, lower memory retention, and reduced sense of ownership over their work, with effects persisting even after they stopped using the tool. This finding was reported widely, with headlines suggesting AI causes cognitive decline. The study itself, however, has not been peer-reviewed, has a small sample size, was geographically and demographically narrow, and has received published methodological critiques pointing to limitations in the EEG analysis, reporting consistency, and reproducibility. The findings are suggestive. They are not conclusive. Treating them as established science is itself a small example of the noise problem the article is about.
A 2025 paper by Michael Gerlich in Societies, surveying 666 UK participants, found a negative correlation between AI tool use and self-reported critical thinking. This study is peer-reviewed, but published in an MDPI journal — a publisher with a mixed reputation in the scientific community for review rigor — and the design is correlational rather than causal, relying on self-report. The paper subsequently required a published correction (a duplicated table). Cognitive offloading was a significant mediator in the analysis, but as several reviewers noted, offloading itself was once described in the cognitive science literature as a neutral or beneficial adaptation. The original “Google effect” paper by Sparrow, Liu, and Wegner in Science in 2011 — a more rigorously published and replicated finding — characterized offloading as functional rather than pathological.
What can be said responsibly is this: there is a growing body of research suggesting that habitual offloading of cognitive tasks to external tools may, under certain conditions, reduce the underlying cognitive engagement that develops durable memory and reasoning capacity. The mechanism is plausible. The evidence is suggestive. The headlines are running ahead of the data. This is, again, the noise problem — even research about how AI affects our thinking is being processed through an information environment that rewards confident claims over careful ones.
What we have stronger evidence for: the long-running observation, supported by multiple independent studies, that the Flynn effect — the steady rise in IQ scores throughout the 20th century — has stalled and in some countries reversed since the early 2000s. The causes are contested and multifactorial. AI is too recent to be the cause; the trend predates it. But the trend exists, and it suggests that whatever cognitive scaffolding produced rising population intelligence is no longer producing it at the same rate. Whether AI will accelerate this, mitigate it, or be irrelevant to it, we will know in twenty years. Anyone telling you they know now is selling something.
What I can say from direct experience, which is not science but isn’t nothing either: I use AI tools daily. I am writing this article with AI assistance for research. I notice, when I do this consistently, that I exercise certain mental muscles less. The atrophy is subtle. I notice it most when I’m forced into a context where AI isn’t available — a long flight, a walk without my phone — and I find that the slow, friction-laden work of figuring something out has become slightly harder than it used to be. Not catastrophic. Real. I don’t know what to do about it yet. I’m taking it seriously as a hypothesis about my own future cognition, while not pretending the science is settled.
Through the lens of sociology
We live in an age of unprecedented connectivity and unprecedented loneliness. These two facts coexist not despite each other but because of each other.
The U.S. Surgeon General’s 2023 advisory Our Epidemic of Loneliness and Isolation synthesized research showing that approximately half of U.S. adults report measurable loneliness, with health consequences including a 29% increase in coronary heart disease risk and a 32% increase in stroke risk. The mortality impact, according to a 2015 meta-analysis by Holt-Lunstad and colleagues in Perspectives on Psychological Science — a study of 70 independent samples covering over 3.4 million participants — is comparable to the harm caused by smoking 15 cigarettes per day. The Holt-Lunstad meta-analysis is one of the most rigorous and frequently cited pieces of evidence in this field. It is not a single study. It is an umbrella over decades of work. The connectivity revolution that promised to bring us together has, on the available data, made us measurably more isolated.
The same period has produced an explosion of expertise theater. Coaches, gurus, mentors, transformation specialists, manifestation experts, and influencers have proliferated to fill a vacuum left by the collapse of traditional authorities. Many of them have no demonstrable expertise in the domains they claim to teach, which is a problem they have solved by talking faster. Some are sincere. Some are not. The market does not effectively distinguish between them, because in a high-noise environment, signal detection becomes exhausting and most people stop trying. They follow the algorithm. The algorithm follows engagement. Engagement follows confidence. Confidence is uncorrelated with accuracy. We have built a recommendation engine for noise, and named it after the activity it has replaced: connection.
I notice this in myself. I notice myself receiving advice from people who have not done the things they are advising on. I notice myself mistaking confident assertion for established knowledge. I notice that the more time I spend on platforms designed to inform me, the less informed I feel — and the more I feel I should respond by consuming more information, which is the optimization strategy of someone who has not yet noticed that the strategy is the problem. If you’ve read the article on Overfitting, you may recognize the loop. It has wider applications than I originally suspected.
Through the lens of philosophy
Two books have shaped how I think about all of this, and they contradict each other in productive ways.
Robert Sapolsky, in Determined: A Science of Life Without Free Will (2023), argues with the full weight of his career as a Stanford neuroendocrinologist that human behavior is entirely determined by biology, environment, and history. His central thesis, stated unambiguously: we are nothing more or less than the cumulative biological and environmental luck, over which we had no control, that has brought us to any moment. There is no free will. The thing we call choice is a post-hoc narrative the brain constructs to explain decisions that were already made by neurons we don’t have access to. The book is not without critics — philosophers have noted that Sapolsky engages less with the philosophical literature on free will than the topic warrants, and that his definition of “free will” is broad enough to be difficult to defeat. But the science he marshals about the determinants of behavior is formidable, and the conclusion deserves serious engagement even if you don’t end up where he does.
Derek Sivers, in How to Live: 27 Conflicting Answers and One Weird Conclusion (2020), takes the opposite approach. He spends 27 chapters articulating completely contradictory philosophies of life — be fully present; be fully ambitious; pursue pleasure; pursue meaning; commit to one thing; explore everything — and demonstrates that each, taken seriously, is internally coherent and externally incompatible with the others. His weird conclusion is not which philosophy is correct. It is that no philosophy is correct, that all of them describe a partial truth about being human, and that the act of choosing between them is the only thing that resembles agency. Sivers is, in a sense, performing the very free will that Sapolsky says doesn’t exist, which is either a refutation or an illustration depending on your prior commitments.
I find both of them persuasive, which is uncomfortable but probably correct. The honest position, as far as I can tell, is that we operate as if we have agency while suspecting we might not, that we choose between incompatible philosophies while knowing none of them is fully right, and that we optimize for things we cannot fully justify in a universe that does not appear to care about our optimization. This is not relativism. It is calibrated uncertainty — knowing exactly how much we don’t know, and acting anyway.
What is genuinely ironic — and this is the irony that ties the whole article together — is that politics and religion across history have offered the same set of optimizations packaged as opposite worldviews. Liberal vs. conservative. Religious vs. secular. Capitalist vs. socialist. Western vs. Eastern. Each tradition claims to have identified the correct way to live, and each, examined honestly, optimizes for some real human need at the expense of some other real human need. They are not opposites. They are different facets of the same impossible problem: how to live well as a finite creature in an infinite universe with limited information and a brain that prefers stories to data.
The amount of human energy spent fighting between these frameworks dwarfs the amount spent acknowledging their structural similarity. Which is, I suspect, exactly what Sapolsky’s neurons would predict, and which Sivers would frame as the 28th conflicting answer that nobody bothered to write down because it doesn’t sell.
Connections
Noise is the meta-concept of this collection. Overfitting is what happens when you mistake noise for signal in your own data. Entropy is the force that produces noise in any system that isn’t actively maintained. Fragility is what becomes of a system optimized in a high-noise environment, where you cannot tell which inputs were important and which were random. Projection is the specific form of noise where your own internal signal is cast onto an external surface and mistaken for information about that surface. Every article in this collection has been, in some sense, about a different way of being fooled. Noise is the article about the fooling itself.
What I Don’t Know
I haven’t explored the political economy of attention — how surveillance capitalism, as Shoshana Zuboff argued in The Age of Surveillance Capitalism (2019), monetizes the very cognitive resources I’m describing as scarce, which would reframe the entire problem from “we are individually weak at signal detection” to “we are systematically targeted for noise injection by entities that profit from our distraction.” I haven’t engaged with the philosophy of language — how the late linguistic philosophers might frame the proliferation of confident-sounding meaningless terms (synergy, alignment, sovereign, agentic) as a specific failure mode of language under capitalism, where words drift from referents because the referents would be testable and testable claims are commercially inconvenient. I haven’t looked at the formal information warfare literature, which has spent decades formalizing exactly the dynamics I’m describing in much messier terms. And I haven’t addressed the question that haunts this entire piece: whether writing about noise is itself a form of noise, whether adding another voice to the discourse — even a careful, well-cited one — is part of the problem rather than its solution. I genuinely don’t know. I wrote it anyway. Which is, in itself, a kind of evidence about how I’ve decided to relate to the question.
Where I Stand
I don’t have a strategy for filtering noise. I have a few habits, which are different from a strategy, and I share them not as advice but as the current operating procedure of someone who has spent enough time in this environment to develop a working relationship with it.
I treat most predictions about the near-term future of AI with the same epistemic weight as predictions about the weather in three months. The technology is real. The forecasts are entertainment. I assume that anyone who is certain about what AI will do in five years is selling something — sometimes a product, sometimes a worldview, sometimes just attention. I read primary sources when I can. I read peer-reviewed papers in the fields I care about, slowly, with skepticism, with awareness that I might be wrong even when the paper survives review. I try to distinguish between what a model can demonstrably do today and what a CEO claims it will do tomorrow, because those are very different categories of claim, and one is funded by venture capital while the other is funded by an actual GPU cluster.
I limit my exposure to platforms that are optimized for engagement rather than information. Not because I’m above them — I’m not — but because I notice that I think more clearly when I’m not being algorithmically targeted by content designed to keep me scrolling. I read books, including ones I disagree with. I have conversations with people who have actually done the things they’re talking about, and I weight their input more heavily than the input of people who haven’t. This is not a sophisticated heuristic. It’s just paying attention to expertise as a real signal in a sea of performance.
I use AI tools, including for writing this article, with full awareness that doing so may be reducing my long-term cognitive capacity in ways that aren’t yet fully measured. The trade-off is real. The honest move is not abstinence — that would just be a different form of optimization theater. The honest move is awareness. I try to do at least some thinking each day that requires no tool, no search, no augmentation — just the slow, friction-laden work of figuring something out in my own head, even when I know AI could do it faster. The atrophy is plausible. The exercise is the only known counter. I’d rather be a person who can still think when the servers go down than one who can think only with assistance.
On the question of intelligence itself, I hold to a simple conviction that I share with researchers I respect: real understanding — in machines or in people — begins with seeing the world, building a model of how it behaves, and updating that model when reality disagrees. Not by reading more text. The current systems can write a beautiful essay about how to ride a bicycle. None of them can ride one. Until that changes, what we are calling intelligence is something else — impressive, useful, often beautiful, but not the thing the marketing department says it is. The gap is the work. The work is unfinished. And the people doing it quietly, who don’t know whether the next paradigm will arrive in five years or fifty, are the ones I trust about where the field is going.
And on the larger question — politics, ideology, religion, the optimization frameworks competing for my soul — I’ve stopped looking for the correct one. Sivers helped me see that they are partial answers to questions that don’t have complete answers. Sapolsky helped me see that even my dissatisfaction with them is determined. Lenses, this project, is my way of holding multiple frameworks simultaneously without committing to any of them. It is not a position. It is a practice. And the practice, imperfect as it is, has given me something more useful than certainty: a stable way to operate in a noisy environment without pretending the noise isn’t there.
The world is loud. Most of it is wrong. Some of it is true. Telling the difference is the work, and the work doesn’t end. But it can be done badly or well, and choosing to do it well — even imperfectly, even in fragments, even on a Saturday afternoon when the algorithm wants me elsewhere — is the most honest form of agency I’ve found, in a universe where Sapolsky may be right that there isn’t any. If he is, at least the determinism produced this article. Which is something.
Written: April 2026 Version: 1.0 This is how I understand this concept today. It will change.
Leave a Reply