The Compounding Fallacies of AI Hype

Some of the most beautiful mathematics and engineering of our age is buried under hubris, shifting goalposts and pretending that AGI is inevitable.

After a decade of not knowing what I do for a living, friends and family have woken up to the fact that I work in AI. When elders ask me what AI is, I’m tempted to say it’s nothing that hasn’t been around for decades. When friends ask if they’re going to lose their jobs to AI, I’m tempted to tell them that the layoffs have nothing to do with AI, and that I can’t even get ChatGPT to outperform the least productive person on my team. But saying this, even to myself, makes me sound cynical.

I think I suffer from what Steven Pinker calls the Curse of Knowledge: “a difficulty in imagining what it is like for someone else not to know something that you know.” Being intimately familiar with the inner mechanisms of what passes for “AI”¹, I’m usually unimpressed by it.

Underpinning the cynicism is a mindset: as someone who’s not a salesman, but a student, a practitioner and occasional teacher of AI—my outlook towards it is bound to be reductive. I fail to see something like a large language model (an LLM, of which ChatGPT is one application) as greater than the sum of its parts. All AI models are the result of a handful of relatively simple foundational concepts, usually taught in undergraduate mathematics. That they take years to master is another matter. But once grokked, they blow your wonder-fuse.

It is said that to explain is to explain away. This maxim is nowhere so well fulfilled as in the area of computer programming, especially in what is called heuristic programming and artificial intelligence. For in those realms machines are made to behave in wondrous ways, often sufficient to dazzle even the most experienced observer. But once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away…²

—Joseph Weizenbaum

This is why I find it very difficult to explain why LLMs have taken the world by storm. On the one hand, there’s no question that the world has irrevocably changed; but on the other hand, the world breaks every decade or so anyway³. This post is an attempt to illuminate this question: what is it about LLMs and AI that’s different this time?

(a large language model) is the internet remixed… it is language atomized by tokens and vectors. A million-layered neural network sandwich that is held together by calculus, linear algebra, and probability—also known as wild hope.

—Rohit Gupta, aka Compasswallah

Part of the answer lies in the fact that few things are as human as language, and LLMs have learnt language. But a much larger part of why it’s different this time is attributable to disingenuous, fallacious, unfalsifiable hype.

The History of AI Hype

The idea of thinking machines and automatons in mythology and fiction dates back to antiquity. Cultivating some form of intelligence in a mechanical, controllable object has long captured humanity’s imagination. This impulse persists across ages and cultures, as persistent in us as the search for meaning or the urge to control and master. The pursuit of artificial intelligence is also a somewhat spiritual pursuit—an exploration of what it means to be intelligent, to be human. The philosopher Mark Rowlands argues that the only thing that makes humans unique is the desire to separate ourselves from other animals. And we haven’t been particularly successful at it either. Over the years, we realized that language, tool use, rationality are all displayed by various other animals. We then thought we’re unique in our ability to self-reflect. But we found that primates, elephants and—to our disappointment—even pigeons can identify themselves and know their place in the world⁴.

All anthropomorphism, thus, is a matter of degrees of similarity to human behaviour. There isn’t a qualitative threshold beyond which a being is deemed human. So when it comes to human-level intelligence, emotion or behaviour, AI researchers probably should just call it a day and move on to better-defined problems.

Nevertheless, a programmer who sees their machine learn how to see, write, speak and read for the first time feels like Hephaestus creating Pandora. Algorithms become alchemy. It’s only natural that we love it.

In the last few decades, AI has become an umbrella term, referring to a broad set of technologies and methods. But in its current usage in the media, “AI” refers almost exclusively to generative AI—stuff that learns from patterns in language, vision and speech to generate human-like content. Generative AI itself is a specialized application of artificial neural networks (ANNs)—algorithms that use a collection of artificial neurons, which in turn are inspired by biological neurons. The artificial neuron was first proposed 80 years ago—by neuroscientist Warren McCulloch and logician Walter Pitts. The history of hype in AI is almost as old as the artificial neuron, as we shall see.

A biological neuron, or a nerve cell, has dendrites, axons and a soma; which respectively perform the functions of input stimuli, output stimuli and an aggregation process on the inputs. McCulloch & Pitts recognized that this simplified interpretation of a biological function was still powerful enough to be programmed into a computer. A machine could be instructed to accept a bunch of numerical inputs, aggregate them (like summing or averaging them), and if the aggregate value was higher than a fixed threshold, the neuron “fires”, relaying an impulse to a neighbouring neuron. This idea of interpreting a neuron as a miniature calculator could be generalized further by arranging multiple neurons in a network—thus, neural networks.

But this model was constrained to explicit programming. To perform arithmetic or logic, you had to tell it exactly what to do. McCulloch & Pitts hadn’t suggested how it could be trained. That flash of genius came in 1957 from the psychologist Frank Rosenblatt. He called his invention the perceptron—an artificial neuron that could be programmed to train itself.

The perceptron was… beautiful, almost magical. Few algorithms are as elegant as the one used to train the perceptron. It is a binary classification algorithm—inputs to it are categorized into one of two classes. The algorithm itself can be written in no more than three sentences. And best of all, it came with a guarantee: if the categories presented to it while training can be separated by drawing a straight line (or plane) between then, then the perceptron will necessarily find that separating line. The implications are far reaching: if any two statistical patterns are presented such that they have a straight gap between then, the perceptron will, eventually, but automatically, find that gap. When I was a student I was obsessed with how something so simple and straightforward, something that could be programmed with only a few lines of code, could wield such power⁵.

In a public demonstration in 1958, Rosenblatt showed how the perceptron learnt to distinguish between paper sheets marked on the right versus those marked on the left. In a later experiment, it could tell circles apart from squares. The media lapped it up, Rosenblatt would say in hindsight, “with all of the exuberance and sense of discretion of a pack of happy bloodhounds”. The New York Times carried a piece titled “Electronic Brain Teaches Itself”. The New Yorker magazine went further and claimed that the perceptron is capable of original thought. Rosenblatt himself thought that his invention would one day be capable of conscious thought, but would later admit a certain “lack of mathematical rigor in preliminary reports.”⁶

Unfortunately, this was also the first instance of overpromising and under-delivering in the history of AI. As powerful as the perceptron’s guarantee of separating patterns was, it did not take long for people to realize that there are enough classification problems that cannot possibly be expressed as separable patterns. Indeed, most interesting problems in machine learning were of the sort where it’s not possible to separate patterns by drawing straight lines between them. In 1969, Marvin Minsky and Seymour Papert highlighted some of the most apparent limitations of the perceptron. Their work showed how these gaps could be plugged, but it was the criticism that went viral. The frequent mis-citing of their work so definitively punctured the AI bubble that interest and funding dried up, and the field plunged into the AI Winter. The next few decades were punctuated with incremental, standalone innovations. In 1986, Geoffrey Hinton with his collaborators Rumelhart and Williams figured out how to train neural networks with a process known as error backpropagation. A decade later, Yann LeCun created the convolutional neural network—an architecture that allows neural networks to “see”.

It was in the early 2010s that the field saw a series of fortuitous events. An evolved CNN architecture named AlexNet was developed, which broke records in visual recognition. Nvidia published a software platform named CUDA, which allowed programmers to use GPUs for not just graphics, but general purpose computing. GPUs are massively parallel computers. As such, they were exceptionally well suited to the task of training neural networks (neurons can process their inputs independent of what most other neurons are doing at the same time). Coursera was founded in 2012—and one of their two introductory courses was on machine learning. This meant that anybody could study machine learning, practically for free. Quietly in the background, the popularity of open source software was growing, resulting in various open-source deep learning frameworks that are thriving to this day.

For me, too, the stars had aligned. I was in the final year of engineering school in 2011, and I took a course in artificial neural networks. The most expensive lab on campus (to which only a few of us had exclusive, unfettered access), contained six Nvidia workstations armed with top-of-the-line GPUs. Although nobody could have predicted what was on the horizon, the early 2010s were a great time to be studying deep learning (it wasn’t called AI yet).

Since then, the field has seen an insane growth. As an undergraduate, I could have sincerely believed that AGI (artificial general intelligence—AI that matches or surpasses human performance) was our manifest destiny—it was obvious and certain. It’s the sort of thing that only a naive student would believe. Today, this naivety and hubris are the default. And once again, eager packs of happy bloodhounds are back on the scent.

The Prophecy of Superintelligence

The seeds of AI hype and misinformation started sprouting in popular discourse long before Instagram was awash with AI slop and every other post on LinkedIn preyed on AI FOMO. As we’ve seen before, there was nothing essentially new about it, but this time, the hype virus found a new ecosystem to infect: social media.

In 2015, Tim Urban, the writer of the popular blog “Wait But Why”, wrote a two-part series on AI, titled “The AI Revolution: The Road to Superintelligence”. Even if he wasn’t the first AI booster⁷ and wouldn’t be the last, few people have been as influential in popularizing AI for a layperson. A fantastic writer and explainer of complex topics, Urban did a great job of collating many common AI fallacies in one place. What follows might read like a hit piece against Urban, but it’s really a case study. A critique of his posts serves as a microcosm of much that is wrong with AI hype.

It’s an intimate mixture of rubbish and good ideas, and it’s very hard to disentangle the two, because these are smart people; they’re not stupid.

—Douglas Hofstadter, American Scientist

Nearly every bizarre claim about AI can be traced back to the so-called “Law of Accelerating Returns”, proposed by Ray Kurzweil. Kurzweil is a colourful figure in the history of technology: a pioneer of optical character recognition, speech synthesis and electronic keyboards, he has long since emerged as a prophet of technology, complete with his share of acolytes and critics. He introduced the Law of Accelerating Returns in his 1999 book The Age of Spiritual Machines. Here’s an excerpt from a 2001 essay by Kurzweil:

… the history of technology shows that technological change is exponential… So we won’t experience 100 years of progress in the 21st century — it will be more like 20,000 years of progress (at today’s rate). The “returns,” such as chip speed and cost-effectiveness, also increase exponentially. There’s even exponential growth in the rate of exponential growth. Within a few decades, machine intelligence will surpass human intelligence, leading to The Singularity — technological change so rapid and profound it represents a rupture in the fabric of human history.

If it sounds like unfalsifiable nonsense, that’s because it is. It’s not a “law” of nature, but simply an absurd extrapolation of an observed pattern. Just like the current AI hype can be traced back to this ’law’, the law can in turn be traced back to another non-law: Moore’s Law.

Moore’s Law is the observation that the number of transistors (electronic switches) in an integrated circuit (like a microprocessor in our computers) doubles every two years. This means that the number of arithmetic and logical operations a chip can perform in a fixed time doubles every two years—without growing in size. Kurzweil’s proposal is predicated on the idea (or at least the bastardization of it) that things—economies, technologies, collective human knowledge, etc—all grow over time in similar ways. Even if Moore’s Law is the cornerstone of his proposal, he expects everything to show “exponential growth”. Moore’s Law stopped being true years ago. But even if were to hold ad infinitum, the assumption that every desirable component of technological change will continue to grow exponentially is absurd. The economist Daniel Susskind writes,

… our apparent success created the impression that an expanding economy was the norm, with any slowdown to be regarded as an unfortunate but temporary exception. Today, that assuredness feels misplaced. Almost every country has slumped its way into the 21st century, though the timings differ… Most economies, battered by two decades of crises—including the dot-com bust, the 2007-8 financial crisis and the COVID-19 pandemic—are sluggish shadows of former selves. We increasingly realize we cannot take growth for granted.⁸

For the sake of argument, let’s grant Kurzweil his premise—that growth is preordained, unstoppable. The deeper fallacy still remains: the conflation of all growth with exponential growth.

The adjective “exponential” has a precise mathematical meaning. When we say that something grows exponentially, we mean that the rate at which it grows is proportional to its current size. Compound interest is a classic example of exponential growth. If your wealth were to grow at 100% (i.e it doubles every year), the resulting growth would be described as massive, staggering, astronomical—pick an adjective. But that doesn’t mean that everything which grows massively, staggeringly, or astronomically is exponential. Quite often, the correct adjective is simply ‘more’. Every other chart and visual in Urban’s post—like stick figures standing on a time-series curve, seemingly oblivious to the walls next to them, and a GIF of how fast an empty lake fills up if the water pouring in it doubles every second—reinforces this idea of the exponential curve.

Even if we concede that something does grow exponentially in the true sense of the word, we still need to ask whether such growth is observed in specific time-windows, or whether it is predestined. The exponential decay of a radioactive element is inevitable. But the growth of an economy or that of human population are not. They only look exponential if we hold our time-windows to convenient periods. It is all too easy to confuse the observed effect of something with the primal cause of something else.

In a 2005 interview⁹, Gordon Moore himself, with humility and amusement, said that the semiconductor industry “made it (Moore’s Law) a self-fulfilling prophecy”, and that it was, on his part, “lucky extrapolation”. But that did nothing to temper Kurzweil’s enthusiasm. Many of his predictions were predicated almost exclusively on Moore’s law—and only a few of them have come true. He has repeatedly time-shifted and revised his predictions, but never fully retracted them. At best, his predictions can be called inaccurate. At worst, they are unfalsifiable.

And yet, in as late as 2015 Urban stands squarely on Kurzweil’s shoulders to reiterate the same predictions. Urban’s obsequiousness to Kurzweil’s ideas could not be more evident. He quotes Kurzweil more than any other source by far, and appeals to his authority every step of the way. Nearly half of the footnotes in the first part of his post are from Kurzweil’s book: The Singularity is Near¹⁰.

There’s a sprawling introduction to Kurzweil’s law, followed by justifications of why artificial superintelligence is an inevitable outcome of it. It’s peppered with quotes like

This isn’t science fiction. It’s what many scientists smarter and more knowledgeable than you or I firmly believe…

By the time Urban comes to talk about the Singularity, or an “intelligence explosion”, he begins by reminding the reader that

… every single thing I’m going to say is real—real science and real forecasts of the future from a large array of the most respected thinkers and scientists. Just keep remembering that.

What in the world are “real forecasts of the future”? Not once does Urban turn a critical eye towards these ideas—which, even in 2015, would have been doable and revealing. For instance, he calls Moore’s Law a “historically reliable rule”. He takes cheaper and faster computation for granted, which to him means that we will be able to “reverse engineer the brain”, and then reminds us that “if emulating the brain seems hopeless, remember the power of exponential growth”. The logic is obviously circular.

Now, the argument that simply scaling a computer makes it smarter, certainly has merit, as we will see in the following section. But to suggest that scaling alone will make it as smart as the human brain, and that in turn will somehow bring forth the rapture, is reductive. It presupposes that the human brain is nothing more than a giant, parallel computer which can only do arithmetic and logic operations; and that intelligence is a function of only these operations.

Neil Postman addressed¹¹ the problem with this line of thought quite succinctly (his commentary applies all the more to modern chatbots):

What is most significant about this line of thinking is the dangerous reductionism it represents. Human intelligence, as Weizenbaum has tried energetically to remind everyone, is not transferable… It is meaning, not utterance, that makes the mind unique.

Tim Urban, after repeating the same few arguments over and over, with a lot more hubris (“we have a lot of advantages over evolution”, and we can build a superintelligence by trying “to do what evolution did, but this time for us”), climaxes with “An Intelligence Explosion—the ultimate example of the Law of Accelerating Returns.” If that is indeed true, then the explosion will be as full of gas as the law itself.

In short,

• if Moore’s Law continues to hold long enough,

• and if the Law of Accelerating Returns rides well enough on it,

• and if highly scaled computers indeed mimic the human brain,

• and if we actually reverse engineer the human brain,

then we will have superintelligence on our hands.

It’s such an exceptional compounding of fallacies that I’m tempted to slip up myself and declare that instead of machine intelligence, it is human stupidity that is growing exponentially. Nearly all of it is based on appeal to authority and hasty generalization. Princeton researchers Arvin Narayanan and Sayash Kapoor came to a similar conclusion in their book AI Snake Oil. In an entire chapter dedicated to such fallacies, they write, “…we’ve seen in the history of AI research, that once one aspect gets automated, other aspects that weren’t recognized earlier tend to reveal themselves as bottlenecks.” Every link in the chain of “ifs” above has its own bottlenecks¹².

But credit where it is due, Tim Urban did write that a “growth spurt might be brewing right now”, and he was right. In 2017, researchers from Google introduced the transformer architecture—the next big leap in machine learning. It soon became the workhorse of the LLM revolution of the early 2020s.

Bigger, Faster… Smarter?

Before transformers, language models were typically recurrent, meaning they processed text one word (or token) at a time, each prediction depending on the one before it. That worked, but it made it hard to handle long passages: the farther back the context went, the harder it became for the model to remember and compute efficiently. The transformer solved this bottleneck by replacing recurrence with attention — a way for the model to look at all words in a sequence at once and decide which ones matter most to predicting the next. This made training faster, more parallel and scalable.

The implication was that researchers were now ready to test what they had suspected for ages—the question of whether sheer scale can produce intelligence. Could we systematically scale training data, computing power and sizes of models to see if perhaps intelligence emerges? In 2018, OpenAI published their results from the first of such experiments. The resulting model was called GPT—Generative Pretrained Transformer. It was this model that would one day become ChatGPT. OpenAI’s series of GPTs, Anthropic’s Claude, Google’s Gemini models are all, essentially, massively scaled transformers.

And that’s how we arrived here. Dwarkesh Patel, the host of a popular podcast on AI calls the period from 2019 to 2025 the Scaling Era. In the early 2020s the sense was that now that we could, it was time to finally put exponentialism to the test.

The capacity of learning in an ANN (on which most modern AI models are based) is measured in parameters—a bunch of adjustable numbers, like knobs and dials, which when carefully tuned, trains the network for a given task. The more knobs, the more tunable the network becomes, and the better it can learn complex patterns. It stands to reason that more parameters means a better ability to learn¹³. And when more parameters are coupled with higher quality data and faster compute… Hallelujah!

GPT-1 had 117 million parameters. Its 10x scaled successor, GPT-2, could actually generate coherent text! GPT-3, with 175 billion parameters, was made receptive to prompting and called ‘InstructGPT’; which in turn formed the basis of ChatGPT. GPT-4 was rumoured to have nearly 1.8 trillion parameters. Between 2018 and 2024, exponentialism played out impressively. In late 2022, AI researcher and educator Andrej Karpathy said on the Lex Fridman podcast that the zeitgeist was “don’t touch the transformer, touch everything else; scale up data; scale up evaluations”. Even experts were fascinated by scaling leading to emergent capabilities. Many felt that there could, after all, be a magical number of parameters on the horizon, which would unlock true intelligence. The journalist Karen Hao has dedicated an entire chapter in her book, Empire of AI, to scaling. She writes,

The scaling doctrine had become so ingrained that some are even beginning to view it as something of a natural phenomenon. Scaling compute is the way, not just a way, to reach more advanced AI capabilities. …the neglected paths of improving the neural network itself or even the quality of its training data can significantly reduce the amount of expensive compute needed to reach the same performance. That’s not even considering the approaches that move away from deep learning.¹⁴

Drawing a parallel to Moore’s Law, she further writes,

OpenAI’s Law, or what the company would later replace with an even more fevered pursuit of so-called scaling laws, is exactly the same (as Moore’s Law). It is not a natural phenomenon. It’s a self-fulfilling prophecy.

Even if this outright equivalence to Moore’s Law, and Kurzweil’s Law is accepted, there’s a more fundamental idea which needs to be questioned. Is deep learning—as pioneered by McCulloch & Pitts, improved by Rosenblatt, and generalized by Hinton—a good approximation of the human brain? To what extent can hyperconnected, massive networks of artificial neurons be grown to match human intelligence? They certainly lend themselves to scaling, but how far can we stretch this approximation?

Yann LeCun, one of the champions of deep learning, doubts that scaling LLMs alone will get us to human-level intelligence¹⁵. But another pioneer of scaling, and co-founder of OpenAI, Ilya Sutskever, appears to be religiously committed to the idea. Hao wrote in The Atlantic,

Sutskever began to behave like a spiritual leader… His constant, enthusiastic refrain was “feel the AGI”… At OpenAI’s 2022 holiday party, Sutskever led employees in a chant: “Feel the AGI! Feel the AGI!”

His religious zeal for scaling was clearly infectious. Dario Amodei, previously VP of research at OpenAI and then founder of Anthropic said on Dwarkesh Patel’s podcast¹⁶,

One of the first things Ilya Sutskever said to me was, “Look. The models just want to learn. You have to understand this. The models just want to learn. It was a bit like a Zen koan. I listened to this and I became enlightened.

Perhaps, this kind of religious fervour is necessary to drive a movement, anything less would not do. But I wonder if it couldn’t have been done without jingoism and hyperbole. It’s not like Sutskever isn’t capable of restraint. In a 2023 interview with Dwarkesh Patel, when pushed to speculate on AGI timelines, Sutskever carefully, and correctly, hesitates¹⁷. When asked about how seriously he takes scaling laws, he admits¹⁸ that there’s a fundamental gap between the accuracy of models predicated on scaling laws and the emergence of actual reasoning capability.

As of today, however, there is reason to believe that the music might have stopped.

In November 2024, Sutskever reportedly told Reuters that results from the scaling laws have plateaued. It was around the same time that Satya Nadella proposed the emergence of a new scaling law, based on letting models take longer to think before responding. Going a step further, in March 2025, Jensen Huang suggested that the entire world got the scaling laws wrong, and the amount of compute needed now was “easily a hundred times more than we thought we needed this time last year.”

If you’re tempted to believe either of them, remember only Upton Sinclair’s famous line,

It is difficult to get a man to understand something when his salary depends on his not understanding it.

Take this tweet, for example, by someone who sells AI courses:

Already in the fall of 2025, GPT-5 released to lukewarm reviews. What people were led to believe would be a vindication of scaling turned out to be not significantly better than its predecessors¹⁹. Meta delayed the release of their AI model named “Behemoth” or “Llama 4”, citing scaling limitations. xAI’s grok series of models, too, saw delayed releases and disappointing results.

Nadella and Huang seem just to be putting a positive spin on all of it—not to mention the multiple times Sam Altman has turned on a dime after the release of GPT-5.

In short, it’s worth asking whether we’re seeing the tail-end of exponentialism. For all we know, LLMs today are as good as they’re going to be for a long, long time.

Not With a Bang, But With Slop

In August, in the Rajya Sabha, Aam Aadmi Party MP Raghav Chadha demanded free access to generative AI tools for all Indians. Undeterred by the question of who will pay for them, social media hailed Chadha’s demand as revolutionary. It was called “bold”, “heroic”, “a crucial step towards digital democracy” and a way to wrest the AI revolution away from Silicon Valley. The irony of how any of this will happen by purchasing subscriptions from companies already in Silicon Valley is lost on LinkedIn’s ‘Top Voices’. Influencers who would otherwise balk at affirmative action or public welfare schemes loved the idea of free AI subscriptions. As much as Chadha’s own party has been criticised for appeasement and “freebies”, his proposal on AI subscriptions was unanimously applauded.

Notably, an earlier speech of Chadha from March, one in which he highlights the need for infrastructure and funding needed for indigenous AI research, went relatively unnoticed. In India, where fewer than 10% of households own a laptop or a computer²⁰ and (even with smartphones and mobile data) internet usage is concentrated among younger, urban, educated and wealthy Indians²¹, the economic utility of free ChatGPT or Gemini subscriptions is dubious.

It is not only this kind of collective, misty-eyed delusion that obfuscates the promise of AI. Even the most earnest estimates of AI’s economic impact are dubious. OpenAI’s own charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”; but we don’t have consensus, even between aligned parties, on what constitutes economically valuable work or outperforming humans. Hinton famously said in 2016 that we should stop training radiologists because it was “completely obvious” that deep learning would be able to outperform them in a decade. If that had happened, it would have been a good example of both AI outperforming humans as well as creating economic value. Today, however, the field of radiology is thriving; in no small part because they have managed to integrate AI tools. Hinton has acknowledged his error. But many in the community who’ve made similar erroneous predictions show barely a soft backpedal—insisting instead that AI was always meant to be a companion. This pattern shows up distressingly often—claim that AI will replace humans at something or the other, and when it turns out that there’s no real replacement, insist that it was never meant in that sense.

At the root of this confusion is the fact that terms like AI, AGI, autonomous systems, superintelligence, etc. are all ill-defined. They all refer to a very large set of technologies and computing paradigms that are loosely similar to each other. Asking questions about the economic impact of AI is like asking questions about the impact of computers or the internet—they certainly start with a bang, but take decades to diffuse through the economy and show tangible impact. Deep down, we understand this, but the FOMO around AI makes us slip up. Take, for instance, Andrej Karpathy’s response when asked about how best to measure AI progress—he said that he was “almost tempted to reject the question entirely” because it was so ill-posed. He admits that there will always be jobs that are “amenable to automation sooner or later”, but that there’s no interpretation of “AI” or “work” that leads us to a clear and significant economic impact. He also says that self-driving cars are “nowhere near done”, and getting them to be economically feasible will be hard.

To his credit, Karpathy has refined his outlook to adjust to reality. Only three years ago, he was having a very different conversation with Lex Fridman. He said he was bullish on our ability to build AGIs. Three years later he admits (to Dwarkesh Patel) that he gets triggered by the noise about AI on social media, and that AI agents are slop. Earlier, he said that he thinks of aging as a disease; but he’d rather chase AGI and let the AGI solve the problems of aging. Today, he admits that there’s no magical generalization.

At the end of the Fridman interview, he predicts (correctly) that the cost of content creation will fall, and excitedly muses what would happen when we can generate art on demand, and when there’s infinite of it. Today we have arrived at Studio Ghibli-style image generation. All it has done is generate some short-lived amusement, and a whole lot of long-term consternation. All on-demand video generators have shown us are fake videos of house cats chasing lions and stuff that need not have been faked at all, like news anchors polishing politicians’ boots.

For years, we’ve been sold nothing short of the transformation of civilization. AGI, they said, would cure cancer, fix the climate, free us from work and usher in an era of unprecedented prosperity. All of this is perpetually around the corner; has been for years. It is an extraordinary mismatch: the loftiest ambitions in the history of technology, followed by outputs that collapse under the gentlest scrutiny. As Karen Hao shows, the scaling doctrine now resembles an arms race more than a scientific project, and its results—despite the marketing—are looking more like diminishing returns.

And if there is a lesson here, it is the same one that we keep rediscovering: every time intelligence is reduced to a technical problem, the technology eventually reveals its limits. We ought to remember that intelligence requires judgement. And judgement begins with refusing to fall in love with your own legend.

---

I’m finishing this piece on a Sunday afternoon. Tomorrow when I go to work, I’m going to have to deal with many problems which are, by popular AI narrative, solved. I’m going to be writing code with an LLM. I’m going to review my colleagues’ code with an LLM. I’m going to tweak ML models which can comprehend images and language—which I shouldn’t really have to do because both problems are handily addressed by generative AI. Indeed, every once in a while, my colleagues and I ponder over why AI hasn’t yet taken our jobs.

The truth is that AI isn’t nearly as ubiquitous, inexpensive or reliable as we’re led to believe. By refusing to let more of it into our workflows, we’re supposedly doomed to forever lag behind the cutting edge. But we’re okay with that, especially if being on the cutting edge means ceding control, agency and privacy (ours and that of our users). The power-hungry GPUs which state-of-the-art AI models need are a far cry from the consumer-grade hardware used by my customers. I could, of course, expose cutting-edge AI functionality to my users over the internet, but the latency and storage costs alone would kill the profit. Like self-driving cars, it’s yet to reach the doorstep. And until that happens, there’s plenty of room at the bottom.

If all that makes us Luddites, we’ll happily reclaim the term.

Those who resist the imposition of technology are disparaged as technophobes behind the times or incompetent, sometimes even Luddites. But in fact, Luddites is exactly the right term, even as those using it as an insult don’t realize it. In the tradition of the original Luddites, writers, actors, hotline workers, visual artists and crowd workers alike show us that automation is not a suitable replacement for their labor. We don’t have to accept a reorganization of the workplace that puts automation at the centre with devalued human workers propping it up.

—Emily Bender & Alex Hanna, The AI Con²²

1. The term “AI” is not very strictly defined. Any and everything that is concerned with automation, language, intelligence, robotics, etc can be called AI. It does not refer to a specific, or even coherent set of technologies. There are countless fields of study which are considered relevant within the field of AI.

2. Joseph Weizenbaum, ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine

3. Morgan Housel, Same as Ever (Harriman House, 2023), p. 47.

4. Mark Rowlands, The Happiness of Dogs: Why the Unexamined Life is Most Worth Living (Granta, 2024), p. 78.

5. It would take me a while to learn that elegance and simplicity are hallmarks of mathematical wisdom. George Polya often said, “Simplex sigillum veri—simplicity is the seal of the truth.”

6. Brian Christian, The Alignment Problem (Atlantic Books, 2020), p 31.

7. That dubious honour goes, sadly, to Frank Rosenblatt himself.

8. Daniel Susskind, Growth: A Reckoning (Allen Lane, 2024), p. xiv.

9. Excerpts from A Conversation with Gordon Moore

10. Kurzweil later wrote another book, called The Singularity is Nearer. Since the Singularity is perpetually just around the corner, it makes sense that Kurzweil took a serious interest in immortality. At the time Urban wrote his article, Kurzweil was taking 100 pills every day.

11. Neil Postman, Technopoly: The Surrender of Culture to Technology (Vintage Books, 1993), p. 112.

12. Arvind Narayanan & Sayash Kapoor, AI Snake Oil (Princeton University Press, 2024), p. 152.

13. This is not without caveats. Models that are too large are prone to “overfitting”; the learning task and the data they are given must both be commensurate with their capacity. Otherwise they either simply memorize their training data, or don’t learn anything useful at all. Overly capable models with poor training are a bit like awkward geniuses: like Sheldon Cooper’s social skills.

14. Karen Hao, Empire of AI (Allen Lane, 2025), p. 115

15. “Yann LeCun… had a particular distaste for OpenAI and what he viewed as its bludgeon approach to pure scaling.” Ibid., p. 159.

16. Dwarkesh Patel & Gavin Leech, The Scaling Era: An Oral History of AI, 2019-2025 (Stripe Press, 2025), p. 24

17. Ibid., p 168.

18. Ibid., p 26.

19. Even as of today, other than in web development tasks, GPT-5 is outperformed by other OpenAI models, according to LMArena.

20. Computer ownership and usage by Abhishek Waghmare, Data For India (July 2025): https://www.dataforindia.com/computers/

21. Access to phones and the internet by Abhishek Waghmare, Data For India (February 2024): https://www.dataforindia.com/comm-tech/

22. Emily M. Bender & Alex Hanna, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want (The Bodley Head, 2025), p. 66