People who are trying to save money don’t go to nightclubs every weekend.

A friend said that to me. We were discussing another friend of ours who had been cut off by his conservative parents because they didn’t approve of his girlfriend, who he intended to marry. Not one to back down, he promptly moved out of his parents’ house, got a less-than-ideal job and started living on his own. The plan was to save up enough money to get as far away from his extended family (imagine a large, wealthy, patriarchal family you’d typically see in an Ekta Kapoor serial) as possible, which meant moving abroad.

We were, of course, sympathetic. But my friend did not seem to have a good grasp on his own financial reality. His lifestyle did not change. Maybe, underneath all the drama, he knew that he could always count on his parents (if not his uncle—who we had so far pictured like an Alok Nath, but now he was an Amrish Puri). Nevertheless, he could at least have acted like someone who’s been cut off. And we judged him for it. It wasn’t until much later that I realized that perhaps the judgment was a little hasty. I myself had, more than a few times, acted just like that, although not for reasons that had—thank heavens—anything to do with family drama on a daily soap scale.

I studied in Pune and my first job took me to Mumbai. After years of staying in a hostel (my parents would have preferred “pigsty”, which was a dig at me, not the hostel), I would finally live like a respectable adult in a metro. I had learned to live frugally in college, and now, suddenly there was guilt-free money. It did not bother me that I was paying a criminal rent for half a bedroom in Powai. After all, I had managed to get a job at Enthought—the organization that was the Times Square of my area of interest. If you stayed there long enough, you’d run into everyone who mattered in the field of scientific computing. It was three years before we realized that we didn’t have to stay in Powai—our MD was a professor at IIT Bombay, and he agreed that we could easily move to a much cheaper place. So Enthought India packed up and left for Pune.

But even after three years of Mumbai, I had little to show for it financially. Other than the annual PPF contributions, I had no savings or investments. I had been reluctant to ask for a raise. Honestly, I was thankful to even have the job I did—there were exactly three people in the entire company who had less than postgrads, and I was one of them. The other two were legendary programmers in their own right, and I was merely a dabbler. So ultimately, after rent, food, transport and the utilities, I had very little left over. The only thing I could do with a few thousand rupees a month was... nightclubs every weekend. Even a couple of years after I left that job and started freelance consulting, when clients were hard to come by there wasn’t a regular paycheck, I was often careless with whatever little money I had. It was my partner who housed, clothed and fed me—and I was happy to let her do so. My own money went into comic books, video games and... nightclubs every weekend. It would be some time before I got wiser with money.

None of this is to say that I understand destitution—not by a long shot. As I write this, I am reminded of a particularly tone-deaf 2016 article on the “Urban Poor”—one the pitiable condition of millennials who don’t make enough money to be able to keep up appearances. Once you become aware of your privilege, you can only laugh at the idea.

I always knew I had a thick and durable safety net and a ton of privilege. But I have a distinct feel for the financial shortsightedness and irrationality that comes with not having disposable income.

It is precisely this kind of irrationality that breaks economic models. As money gets tighter, people seem to spend less and less reasonably. But what seems like unreasonable spending to the untrained eye often has very interesting, and sometimes, heartbreaking explanations... as we shall see.

दरिद्रता में जो एक प्रकार की अदूरदर्शिता होती है, वह निर्लज्जता के तकाजे, गाली और मार से भी भयभीत नहीं होती

—Munshi Premchand, Godaan

The Moral Hazard

The idea for this edition of Being Bayesian came from a recent column by Dr Shamika Ravi, in which she wrote that increasing public healthcare might be at odds with the increasing consumption of intoxicants. Drawing from HCES, she concluded that the poorest 40% households in India spend only 2.5% of their budget on education, but 4% on pan, gutkha, tobacco and other intoxicants. She quotes more numbers on how the consumption of intoxicants has increased by 33% in just over a decade. And sure enough, resources that were published before the 2023-2024 survey mention that paan and tobacco consumption has been declining1.

When we additionally consider the steadily rising enrolments in the Ayushman Bharat Yojana, Ravi and her colleagues have a good point. A naked-eye reading of the situation seems straightforward:

• the poor are more likely to be insured, and

• they are also more likely to eat more gutkha, and

• since tobacco leads to non-communicable diseases whose costs increasingly fall on the state,

we have a classic case of moral hazard.

One line in the column, however, threw me off.

What makes the current moment particularly fraught is the interaction between rising tobacco use and expanding public healthcare coverage.

The problematic word here is “interaction”. On closer inspection, we see that the article simply juxtaposes two trends, assumes an interaction and invokes moral hazard. Thankfully, the HCES also records for each respondent household whether they have a member who is enrolled in PMJAY or a state-specific healthcare scheme, and whether they received healthcare benefits.

So I went looking for the interaction between intoxicants and public health coverage.

There’s barely any interaction. The difference between the consumption of intoxicants by beneficiaries of public healthcare, and that of non-beneficiaries, is negligible. For gutkha in particular, the urban beneficiary spends an average of 2 rupees per month more than the urban non-beneficiary, whereas the rural beneficiary spends 4 rupees less than the rural non-beneficiary. The entire chart spans only a difference of +/- Rs 15.

Of course, this does not mean that the moral hazard concerns are invalid. But it does suggest that we need some nuance. People’s worries about recipients of welfare becoming freeloaders often pop up in public discourse, but often without the slightest nuance2. Examples range from news anchors exasperatedly asking how long the middle classes can be expected to subsidize the poor3; to the chronic criticism of pre-election “revdis”, to entrepreneurs claiming that the gig economy will finally free the poor from handouts.

The simplest explanation for why this is such a bone of contention is that welfare in general doesn’t seem to work, much less for a country like ours. The implementation of a welfare scheme must navigate the labyrinth of social, cultural and systemic barriers to make its results visible. Abhijit Banerjee and Esther Duflo wrote an entire book, Poor Economics, showing that there’s no silver bullet to solve poverty4. As such, large, nationwide results are often disappointing. But with a little nuance, and if we know where to look, we can see small, but interesting effects.

Looking Inside the Basket

Of all the indicators of welfare schemes that appear within the HCES, I’ve chosen five: free electricity, LPG subsidy, state or central health insurance, midday meals and PDS. For each of them, I’ve looked at their effect on food and education. This choice is motivated mostly by convenience and simplicity. The survey questionnaires of both food and education are simple enough that I can construct a consumer basket for each household, and then compare the baskets of beneficiaries and non-beneficiaries. I’ve added details of the analytical method in a later section.

Broadly, for each type of aid, we first find how the food budget of a beneficiary household differs from that of a non-beneficiary household. Then we do the same for education. If we were to look at nationwide averages, the results are unsurprising, if a little disappointing.

As expected, beneficiaries of almost every scheme spend less on food and education than their non-beneficiary counterparts. This is not surprising because the schemes are targeted at the poor, who spend less on almost everything. In fact, only free electricity gives beneficiaries an edge in both food and education. But the effects are small: households that receive free electricity spend only a few hundred rupees more on food, and less than a hundred rupees more on education. LPG subsidy, too, gives a tiny boost to education, amounting to less than a hundred rupees per student. In a sense, the differences are somewhat obvious. If beneficiaries were spending more than non-beneficiaries, that would perhaps be an indication of failed targeting.

Held at this level of national abstraction, the numbers don’t tell us anything we don’t already know. But if we dig into the finer details of food and education, we see something interesting. For example, consider the pair of bars denoting the food expenditure associated with PDS (in the left column, the second pair of bars from the bottom). The difference in expenditure here is only Rs 168. But let’s look at how this difference is distributed between people who get subsidized rations against those who don’t.

The average PDS beneficiary consumes nearly Rs 100 worth of dairy less than their counterpart, and also less cereals. But they also eat Rs 50 worth more of eggs, fish and meat. This pattern of less dairy and a little more meat repeats in LPG subsidies as well as public healthcare. Moreover, the positive spend on meat survives even when we break it down by religion and caste. It’s not like a subset of beneficiary households are eating more meat and pushing the average up. Among every caste and religion (with the exception of Sikhism), recipients of aid eat more meat than non-recipients.

At first, I thought this might be a calculated reallocation of the food budget—when people don’t have to worry about carbs, they can afford more protein. But there’s a more interesting explanation. Banerjee and Duflo write,

... the money that people spend on food is not spent to maximize the intake of calories or micronutrients. When very poor people get a chance to spend a bit more on food, they don’t put everything into getting more calories. Instead, they buy better-tasting, more expensive calories5.

They further cite a study6 which found that households which receive subsidies for cereals consumed less of them, and ate more meat.

In education, too, we see a peculiar trend. Note that LPG subsidy is one of the only two interventions that gives a modest education spend among beneficiaries. For every student from a household which receives LPG subsidy, nearly Rs 60 is spent more on education than for a student from a non-beneficiary household. But of those 60 rupees, two-thirds goes to private coaching. Similar effects are seen for students who get midday meals at school, even if the effect is smaller in size.

We could keep on doing these experiments endlessly—create treatment and control groups and look at the differences in their spending patterns. For public welfare, we’d expect the treatment group to spend less on almost everything, but there would be some interesting buckets where the difference flips. It’s very tempting to come up with stories on why the sign flips for particular items; but rigorous explanations are best left to people who do field work. I may try to hunt down some such explanations in a later edition.

For now, the takeaway is simple: the effects of welfare rarely show up in the aggregate, but they do appear in specific pockets of consumption.

More than the Daily Bread

In the very first chart, beneficiaries of public healthcare spend slightly more on country liquor than those who are not beneficiaries. The difference is tiny—barely a few rupees a month—but it exists.

Years ago, a friend of mine worked at a microfinance firm. In his spare time he volunteered to teach the children of migrant labourers who lived near construction sites and factory grounds. Many of the parents worked long days carrying cement, bricks and steel—backbreaking labour that left their bodies permanently bent and scarred. Their food was poor, often lacking even basic nutrients. By nightfall their shoulders and backs would be in excruciating pain.

Sleep was almost impossible. The only thing that could knock them out was alcohol. It was not an indulgence—it was something that helped them work the next day. And they knew they were killing themselves. They had no choice.

Surveys or even field work alone cannot tell us everything about how welfare works. But they do offer a glimpse of something that public debates often miss: when even a small margin opens up in a household budget, people do not suddenly become irresponsible or extravagant. They simply make choices—sometimes careful ones, sometimes imperfect ones—much like anyone else would.

And sometimes, when the budget loosens just a little, those choices include something slightly nicer than before—better food, a little coaching for a child, or simply a way to get through the night and return to work the next morning.

Not cake, but something slightly better than bread.

Acknowledgements

I am highly indebted to Prof Vimal Kumar for asking me to ensure that comparisons are made only between households that are comparable, to Prof Ashwini Deshpande for showing me how to do so, and to Prof Reetika Khera, for convincing me that looking for the impact of welfare is going to take more than a few weeks and a Substack post.

Notes on the Analysis

The findings here are obtained by loosely imagining welfare interventions as randomized controlled trials—“loosely” being the operative word. Actual RCTs are expensive and take a very long time to conduct. The Household Consumption Expenditure Survey (HCES), on the other hand, gives us observations for sample households. No matter how we spin the data, it can’t exactly mimic the conditions of an RCT. The HCES contains nationally representative observations of households, but participation in welfare programs is not randomly assigned. The household characteristics certainly are not controlled, and the outcomes (which are, in this case, expenditure on and consumption of various goods and services) cannot be meaningfully linked to a trial intervention. As such, we lose causality, and limit ourselves to infer only associations. It’s true that correlation isn’t causation, but it’s not nothing. A friend used to say that throwing correlation away just because it’s not causation is like throwing the baby out with the bathwater.

So we end up with tenuous simulations of RCTs. We imagine welfare schemes as interventions, with each scheme having a fixed eligibility criterion. Based on this, we create a cohort of eligible households which we divide into “treatment” and “control” groups. For example, if we consider PDS as the intervention, then every household that has a ration card is eligible. Households that actually used the card to purchase subsidized goods end up in the treatment group. Instead, if the intervention is midday meals, then the eligibility changes to households that have children enrolled in public schools. Thankfully, much of this information can be found readily in the HCES.

The eligibility for healthcare interventions, however, was a little different from the other interventions. I have included households that had at least one member who was enrolled in Ayushman Bharat or a state-specific healthcare scheme, provided that nobody in the household had been hospitalized in the last year. In this sense, we only get to see if people view health insurance as a safety net. An actual health shock would have rendered households uncomparable, since I would have no way of controlling for its severity and duration, or its financial and emotional impact on the household.

However, eligibility alone is not enough; households must also be comparable in the socio-economic sense. Even without any intervention, consumption patterns vary widely. Household size, number of children and elderly members, religion, caste, and geography all influence what people buy. For example, a Dalit household in Vidarbha is bound to eat very differently from one in coastal Maharashtra, even if both are in the same state. So the eligible cohort needs to be further filtered down to only include households that are similar along observable socio-economic characteristics to each other, except for the intervention itself. We use a method called propensity score matching, which estimates the likelihood of a household receiving an intervention based on various attributes. Using the estimated propensity scores, we limit the analysis to households that have more than a 60% chance of being selected in a program (the top two quintiles of the PSM scores), and then we divide them between those that received the treatment and those that did not.

All numbers you see here are the differences in average expenditure between such cohorts, and the comparisons are statistically significant at the 5% level. The code and data needed to reproduce these results are available here.

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.

Subscribe now

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

I was walking her in the lane behind my house when she saw something, probably a cat. With one tight jerk, she lunged, her collar snapped and she disappeared. I ran after her, but even as a puppy with legs barely six inches long, she was disproportionately faster. Even today, she can be caught only if she wants to be caught.

I asked bystanders if they saw her. They pointed in the direction of my house. She's smart, I thought, she can find her way home. Then I heard distant barks - from at least two other dogs. I haven't run faster in my life.

What I saw on my doorstep is seared in my mind.

There was a large bale of black and brown fur. My brain took a moment to resolve it into three huge dogs.

These three monstrously large1 dogs had their backs to me, hackles raised and backs arched, all growling menacingly. One of them was Bobo, who we have otherwise loved as our own - but in that moment I wanted to kill him. Another was a very old dog who used to frequent the service lane behind my house. I never saw him again. And the third dog was the resident stray of the next lane. I've always very uncharitably called him the "asshole dog" because he stealthily attacked Cookie once later too.

Not one of them moved, nor made the slightest change in their stances as I stomped towards them, making every effort to stomp as loudly as possible.

And then I saw her over their shoulders.

She was on her hind legs, almost standing erect on them, her back glued to my front door. She was snarling with teeth bared, completely cornered and desperate. Her pursuers had had her to themselves for at least a minute before I showed up - and in that minute, they could have torn her apart.

Even when she was standing erect, she was just about a foot tall. But there was an indignation, an anger in all six kilos of her. I can’t imagine something barely larger than a football having so much ferocity and menace. I could not be more proud. After some frantic shouting and stomping, I managed to shoo the dogs off. Cookie had held her ground honourably.

But since then, Cookie has hated stray dogs with a vengeance. She can easily pass off as an RWA uncle. She would even make an excellent amicus curiae in the Supreme Court's quest for a Delhi free of strays.

Cookie has a very different relationship with Bobo now. Over the years they have been playmates, enemies and friends. Now, as age catches up with Bobo, they are simply tolerant of each other. I saw Cookie's anger fade away gradually, as she started regarding Bobo with curiosity - sometimes cautious and sometimes playful.

The more I think of it, the more I realize that that's exactly how I write editions of Being Bayesian2. My first draft is, without exception, an angry rant at something outrageous. It's never published. I just let it cook and gestate for a few weeks. Then, when the anger has faded away and I've developed a healthy degree of detachment from the original draft - I go to work. It also helps that I'm a better editor than a writer.

After the anger and indignation have evaporated, all that's left is curiosity - cautious and playful. I have no choice except to be level-headed. This is the right time to look for solid, reliable data which could illuminate whatever it is that I was earlier mad about. The analysis which follows often breaks my own biases (called "priors" in Bayesian inference) - and if I'm lucky, it helps me make my point about how some arbitrary opinion was laughable.

Unfortunately, I have not been able to follow my process in this instance. I can't let my rant sit for weeks, because honestly, there's a lot which I want to get off my chest. I tried very hard looking for reasonably reliable data, but such data as there is, is mired in contradictions. In fact, unlike my other posts, there are no charts in this one. I'll make do with pictures of dogs, instead.

Substituting images, feelings and memory for evidence is precisely the sort of thing that would otherwise send me flying off the handle. But the recent order about the stray dog menace from the Hon'ble Supreme Court is so devoid of rationality and the scientific method, that by rambling on as I am now, I hope to even ingratiate myself with the bench.

Here, in a nutshell, is what they want:

That was, at least, the impression they gave in their oral observations on the 11th of August. After the following outrage, when the official order was released, some of the most blatantly impractical aspects were rethought. For example, they talk of first building capacity for 5000 dogs and then expanding progressively. Starting with 5000 dogs and covering up to potentially a million dogs is a tall order, and not only because of the thousand-to-million order-of-magnitude difference.

Since there's no other reliable data, let us constrain ourselves to look at only these two numbers: 5000 and 1 million. As you'll see, that alone is baffling enough. Nobody knows where either of those numbers came from. Apparently, nobody knows much at all about the size of the problem, as Vignesh Radhakrishnan of the Hindu says here:

So what makes the court think that 5000 is a good number to start with? For one, The MCD has only 20 animal birth control centres in Delhi, most of which are run in partnerships with NGOs. They can barely hold 2,500 dogs at a time. Moreover, they are predominantly medical facilities, not long-term shelters. There's also the question of cost - the cost of feeding, vaccines, staffing, construction and logistics - which the order does not consider. A quick back-of-the-envelope calculation suggests that unless the MCD can somehow conjure double the space, much more staff, and a steady operating budget of ₹20+ crore within weeks, even these initial 5000 dogs will go to the dogs.

I'll hasten to add that all of these numbers are built on sand. Even the figure of 1 million stray dogs in Delhi NCR is a wild guess. The last census happened in 2009 which counted approximately half a million strays. That number supposedly grew to 8 lakhs in 2019, again based only on an estimate. For the 1 million figure, there's no data at all.

Most baffling of all, is the fact that the SC's order mandates an outcome, not a process. Not only does it ignore basic math, it assumes a budget that doesn't exist - all while not pulling up the governing bodies for not doing their jobs.

For a decade, Bobo has been a resident stray of my lane. Pretty much everybody here loves him - which is remarkable in itself. On the other hand, since August 11, some members of the RWA have become somewhat trigger-happy - encouraging residents and guards to photograph and report stray dogs all over the colony to the MCD. But O Judgment! Until thou art fled to brutish beasts, and men have lost their reason, Bobo will stay with Cookie.

I love dogs. I have been bitten by three of them (and a cat), and I have had the misfortune of having to search frantically for rabies vaccine - even in a place like Delhi. So while I admit that I have a personal stake in this matter, it also bears mention that none of this really has to do with how you feel about dogs.

It's about how the math isn't mathing in the first place. Piyush Goyal might be okay with that - responsible citizens should not be.

Disclaimer: The Hon’ble Supreme Court has since the time of writing this, referred the matter to a larger bench, which was heard on the 14th of August 2025. The official order is yet to be published.

My wife and I recently had a... discussion. I don't exactly remember what it was about. But it was something to do with home improvement... something like doorknobs or curtain rods. Whatever it may have been about, at the core of it was an ideological difference. My wife believes that even the smallest inconvenience should be immediately addressed - immediately being the operative word. I, on the other hand, think that most small inconveniences can wait. In fact, I firmly believe that just because something is doable doesn't mean it ought to be done. There's an infinite number of doable things. Undoable things, on the other hand, are very few. There's nothing inherently special about most things that are considered doable.

Before I realize it, I'm actually saying all of this out loud with a lot of sincerity. She initially finds it amusing: I can't be serious! How does something as trivial as a broken doorknob push me into delivering a deep metaphysical thesis? As I ramble on, her amusement becomes frustration. She reminds me that the whole issue is a lot simpler than I'm making it - it's all about a few stupid doorknobs. Why can't I just fix them today? I don't know what to say.

The answer is actually pretty straightforward - interruptions have a cost. And for programmers (I've been working from home for the last seven years), interruptions are unaffordable. It takes programmers 10-15 minutes to resume work in earnest after being interrupted1. My wife and my cook both know that even the smallest, most straightforward question will garner an unfocused, glassy-eyed stare, accompanied by a "huh...?!" So when I can't explain why I don't particularly care about what I eat or wear - it's the same basic principle. Decisions, no matter how trivial, have a cost.

But there's no way I can say this to someone, let alone my significant other, without sounding like a pompous, self-important jerk.

So the path of least resistance is to just fix the doorknobs. But close on the doorknobs' heels are doorstoppers the dog chewed off, the sofa cushions with only the smallest visible tears, the printer cartridge refills and the endless stream of tiny decisions...

... but how about little faults, little pains, little worries. The cosmic ulcer comes not from great concerns but from little irritations. And great things can kill a man, but if they do not, he is stronger and better for them. Man is destroyed by the duck nibblings of nagging, small bills, telephones, athlete's foot, ragweed, common cold, boredom. All of these are the negatives, the tiny frustrations and no one is stronger for them.

- John Steinbeck, Journal of a Novel: The East of Eden Letters

I shared this quote with my wife once. She said, "It's almost as if you dictated this to Steinbeck." It was easily the most flattering thing anyone had ever said to me. But as I visibly gloated, she added that she didn't mean it as a compliment.

Over time, I’ve come to think that while Steinbeck meant to write about the gender-neutral “man”, but it landed squarely in the male sense. There exists an opposite perspective of “the negatives, the tiny frustrations” - one that calls “not errands, but love letters to the world.”

Then we’d tackle whatever needed fixing that week: the mixer grinder’s stubborn motor, new batteries for a dying watch, shoes with broken straps. Finally came my favourite part: long walks in the park nearby, followed by sugarcane juice that painted green moustaches on our faces as we giggled in the golden light. In that park, peacocks would strut between the trees, and I’d always return home clutching a perfect fallen feather – my weekly treasure, proof of magic found.

- Kanchan Balani, It's the Small Things | Everyday Errands

For me, the cosmic ulcer didn’t always exist. A decade ago, life was just sorted. Both of us had a fairly predictable job, and a fixed routine. We would see each other only at dinner and on weekends. Even the time spent on leisure was fixed - reading on the train, Netflix during dinner and video games and socializing over the weekends. Doorknobs were strictly the landlords' headache. It was all very comfortable. It never occurred to me that the routine could be suboptimal. No thoughts of optimization or efficiency crossed my mind.

And then something broke. I didn't quite know what.

I took a new job. It was a dream job - the kind that you think you'll spend the rest of your life at. There was only a small team in Delhi and no office. So I worked from home. For the first time in life, I had more time than I knew what to do with. Initially all the extra time felt like a great luxury. Like a dutiful worker, most of the leftover time went into doing extra work. I quickly emerged as the new guy who gave more than 100%. I soon realized that being that guy has diminishing marginal utility. But by then I already had a reputation - and upholding it meant that I had to squeeze every last droplet of time and siphon it into work. This, in turn, meant that there was less time for everything else - especially leisure and relationships. And, naturally, all concerned parties grew irritable, bitter and unhappy - in that order. This was classic burnout, but, in my infinite hubris, I refused to admit it even to myself. Finally, as a self-respecting hacker, instead of solving the root cause of burnout, I avoided it altogether with productivity hacks.

What had broken was Time - my idea of it and how I used it.

Personal productivity wasn't yet the widely influential global enterprise it is today. Sadly, the people wielding that influence today seem to suffer from the Kardashian effect - they seem to be only productive at creating content about productivity2. My journal entries from that period (journalling, too, was something that productivity gurus recommended - I didn't start enjoying it until much later) are full of whining about how I couldn't get enough done. Ultimately I realized that my ideal day was getting more and more ridiculous and impractical. The day would never come when I get up early, do enough exercise, crack off a major work problem before lunch, read a book in the late afternoon, work on a side project in the evening, take a leisurely walk in the park with the dog, and top off the day with a deep conversation over dinner. When I look back, I see that that was the last time I whined about not getting enough done. So, eventually, many books, articles, podcasts and videos later, I gave up on productivity hacks3.

This was the second time Time broke, coinciding with the COVID-19 lockdown. All hacks broke. Routine broke. Habits, albeit with a lot of effort, broke. The only "hack" that remained was to work with depth and focus at a slow and steady pace. Cal Newport would eventually call this slow productivity. It worked well enough for me to do some of my best work.

After all that breakage, what remains with me is a small set of ideas - don't multitask, deal with interruptions the best you can, sleep well, etc. I no longer claim or aim to be the master of my time. It's not even like I'm cured of the cosmic ulcer - far from it. I just care less now.

Every now and then I see success stories of slow productivity - small doses of vindication for my rediscovered approach. A friend heard my rant about doorknobs. He could not have been more sympathetic. He admitted that non-programmers rarely understand flow and uninterrupted immersion, and how, in social conversation it is difficult to defend not fixing a doorknob. He then shared with me a Quanta profile of June Huh. "Somewhere here there's a story of a Fields medalist stapling together a makeshift blanket...", he said. June Huh is a late blooming mathematician who dropped out of high school to write poetry, then got enrolled in a math course, took six years to graduate and finally went on to win the Fields medal. He is described as extremely slow and methodical. He hacked together a blanket from scrap cloth because going to buy a proper blanket was too mentally exhausting. I somehow just get that.

There's also the story of the great Richard Hamming. In his classic lecture You and Your Research, he advocates exceptional ambition, attitude and passion. His idea of knowledge and productivity being like compound interest has stayed with me since my college days.

Both Hamming and Huh are poster boys of deep work and slow productivity. But there's also something else that's common between Huh and Hamming - both their wives had to put up with their shenanigans. Huh admits that his wife was disappointed at the lack of balance in his life. Hamming, too, admits that he "sort of neglected" his wife sometimes, "You have to neglect things if you intend to get what you want done."4

It's infuriating how, even with the best of intentions on your part, there's always someone else paying the price of your productivity - a parent, a spouse, a child or a pet.

So now imagine my indignation when a captain of industry tells me that I ought to be working 90 hours a week, through weekends, and that I've been staring at my spouse for too long.

The War on Time

N R Narayana Murthy wants people to work 70 hours a week. And frankly, that doesn't sound too bad. But S N Subrahmanyan dials it up to 90 hours a week. Bhavish Aggarwal claims to work 140 hours a week. He believes that weekends and work-life balance don't apply to Indians because they are a western concept. He may be right, but only a little homework would have revealed that the concept of leisure isn't foreign at all. It is advocated in something quintessentially Indian - the Constitution.

The State shall endeavour to secure, by suitable legislation or economic organisation or in any other way, to all workers, agricultural, industrial or otherwise, work, a living wage, conditions of work ensuring a decent standard of life and full enjoyment of leisure and social and cultural opportunities...

- Article 43, Directive Principles of State Policy

Now, credit where it is due, building and running an industrial empire must come with a deep understanding of how people do their best work. But when exhortations to work longer hours are accompanied by news of stress-induced diseases, deaths and even suicides in the workplace - it's clear that this fanatical quest for workplace productivity is not just misguided, but also tone-deaf. It is also, as we shall see, inevitably sexist.

How the Salaried and Skilled India Works

Between January and December of 2024, the National Statistical Office conducted the second national Time Use Survey (TUS) in India (The first was conducted in 2019, before the pandemic). It offers precisely the kind of data that helps us address the question of how feasible long work-hours are, and who they affect the most. As I wrote the following section, the central guiding question in my analysis was,

If people had to work 90 hours a week, what would they have to give up?

The 2024 TUS covers the entire country apart from some inaccessible parts of Andaman & Nicobar, and is administered over the whole year to account for seasonal variations. For each respondent, it records the time spent on everything in the last 24 hours in 30 minute intervals. As such, it presents an exceptionally detailed view of how people spend their time. The survey data is rich beyond an analyst's dreams... after losing myself for weeks in the labyrinth of the various possibilities of filtering the data and designing sample cohorts, I've settled down on a ruthlessly simplified but still rigorous approach to the data.

For one, I will be writing about that subset of the population which is most likely to be at the receiving end of productivity-shaming: salaried white-collar workers. The TUS helpfully includes a field which records whether each respondent was involved in an economic activity, and if so, whether they were self-employed or salaried. In addition, the survey also records the industrial classification of the relevant economic activity. With a combination of these two and the typical working age (15-60 years), I created filters which indicate whether someone can be asked to work for 90 hours a week with a straight face5.

Even this modestly-sized sample of the white-collar workforce is still quite fertile. Fortunately, there's no demographic which is excluded from it. People of all religions, castes, educational backgrounds, ages and genders find decent representations in the sample. Of these, some are more indicative (meaning, correlated with) the use of time in specific activities than others, as we shall see. The effects of some of these demographic variables are subtle, they show up only in very nuanced conditions and with smaller effects. Others jump right off the computer screen. The biggest example of the latter is gender. The fact that time use is gendered wouldn't surprise anyone, but the extent of the effect of gender is staggering. Gender beats all other attributes of the respondent by far, and there's no way you can manipulate the data which makes this effect disappear6.

If we simply look at the average time men and women spend on different types of activities and adjust it for the week, the picture doesn't look bad at all. Both men and women spend about 85 hours a week on self-care and maintenance. This includes activities like sleep, personal care, hygiene and non-social eating and drinking. Men work nearly 56 hours a week, and women work 48.6 hours a week. This translates into women working only an hour less than men every day. If you've wondered if your female colleagues seem to consistently leave the office earlier, here is statistical proof.

It is in the next category of activities that things get interesting. Women spend four times as much time on chores (they're designated as "unpaid domestic services for household and family members" under ICATUS) than men7. Activities under this category involve working in the kitchen, doing the laundry and other miscellaneous household management. In our quest for the 90 hour workweek, it's pointless to look at anything but the largest component of chores, and that is the kitchen.

Here, too, the magnitude is more surprising than the direction. We'd expect that men spend much less time in the kitchen than women overall in the country - and we'd be right. But to see that happen even in a small, presumably emancipated section of the population is shocking. Even among the salaried class with skilled jobs, men barely spend an hour cooking. Moreover, cooking is only a part of kitchen work - there are also the allied activities of serving food and cleaning up after, where men are insultingly absent. This reminds me of a scene from the Malayalam movie The Great Indian Kitchen (which was also captured well in its Hindi remake) where an overbearing relative announces that he's going to make dinner, that the women can rest, but ends up leaving the kitchen in a mess.

Other factors associated with kitchen time - like marital status, education and household wealth - also influence it in expected ways. But the relationship between them and time spent in the kitchen isn't just monotonic - it is extreme. The women who come from the wealthiest quartile spend nearly three hours less in the kitchen than those from the second wealthiest quartile. Women with graduate degrees spend nearly 12.5 hours per week in the kitchen, and as education drops to the higher secondary level, the time spent almost doubles. Going from being single to being married increases the time spent in the kitchen by a whopping ten hours per week.

The time spent in chores (particularly in the kitchen and otherwise in cleaning and laundry) is inversely correlated to the presence of labour-saving devices and methods in the household. This is not surprising, but it _is_ statistically significant. What is surprising is that these devices don't have as large an effect as we'd expect. Come to think of it, washing machines, refrigerators, microwave ovens and vacuum cleaners are technologies that are old enough to have revolutionized chores. They are generally reasonably priced, too. Given that we're talking about a relatively better off section of the population - why don't we see a larger effect? I don't have a clear answer, but I did get a very interesting perspective from Krish Ashok, in one of his most interesting podcast moments:

“...optimizing cooking time with appliances is a progression. My grandfather didn't like rice cooked in a pressure cooker. A generation before they didn't like stoves, and preferred wood fires. Nowadays they don't like microwaves and air fryers and say that refrigerated food loses nutrition. Whenever there's a new tool for improving kitchen productivity, there will be an entire generation of men and women resisting it...”

Anyhow, for the sake of argument, let us allow ourselves, for the time being, to dream of a world where both men and women spend significantly less time on chores. Let's say, optimistically, that both genders spend no more than an hour a day on chores, and pump all the remaining time into work. Even in this ridiculous world, women's work time would grow only to 67 hours a week. The situation for men would not change since they already work less than an hour per day on chores. To hit the 90 hour / week target, there's still a large deficit - 22 hours for women and 34 hours for men. At this point, work has no choice but to eat into sleep and leisure.

This is probably the only chart in my entire analysis which looks gender balanced. In fact the only small imbalance I see is that men watch TV for a couple of hours longer than women (presumably when their wives are in the kitchen). Otherwise both genders spend comparable amounts of time sleeping, talking, eating and otherwise chilling. It must be noted, however, that I've wrought an anomaly here by cherrypicking a specific cohort - otherwise, in general, time spent on leisure and self-care is also heavily gendered. For instance, women take 2.3 times as long as men to use the toilet8 - there's no way that personal care and hygiene costs both genders the same amount of time across the country. Even socialization and relaxing can mean different things to both genders. In her book Whole Numbers and Half Truths, Rukmini S writes some very interesting examples of how leisure means different things to different sets of Indians, particularly to men and women. Even something as innocuous as commandeering the TV remote can affect how women feel about their freedom of leisure.

But as far as this cohort of salaried white-collar workers is concerned, this chart tells us that there's a lot of productivity still to be squeezed out. For example, both men and women could do with an hour less of sleep everyday. For one, seven hours a night is by no means cruelly short. For another, every productivity guru worth their salt will tell you that it's the quality of sleep that matters more than the duration. That's seven hours added to the respective deficits right there!

When it comes to talking, chatting and texting, 15 hours a week looks downright excessive. All 10-12 hours of watching TV could be cut down. There's certainly some creative incentive that managers can come up with to get people to cancel their Netflix subscriptions and delete their Instagram accounts. In fact I'll go so far as to say that even some personal care and hygiene can be sacrificed. Ask Bhavish Aggarwal - if he can work 20 hours a day and still afford to look somewhat well groomed, what’s your excuse? Of course, haters could ask why working 20 hours a day isn't enough for him to make decent scooters. If we're going to make reasonable arguments like time spent on work is no guarantee of quality, this whole post is moot.

A colleague of mine once took a picture of a couple of his teammates sitting with their laptops open in the back of an autorickshaw, in the middle of a busy Hyderabad street. The marketing team picked it up and used it in a social media campaign with hashtags like #hustle, #dedication, #productivity, etc. All I could think of was what a potential client might think about the quality of our work if we were so busy hustling.

Another colleague once posted a picture of himself sitting on his desk, hunched up over his laptop with the caption, "burning the midnight oil after everyone has gone home." Someone asked him who took the picture if everyone had gone home. He didn't reply.

I don't honestly think that people actually confuse time with productivity. Nobody is that stupid. But knowledge workers aren't turning out widgets on an assembly line. Time spent is perhaps the only lead metric of productivity we have.

The single most enjoyable thing about being a programmer is the obsessive pursuit of a complex problem. Every good programmer has a lot of experience with spending hours and days holed up in a room with an unhealthy disregard for food, sleep and general well-being. It is the cost of what we do. It is why we're able to take any pride in our craft. It's the rite of passage of every true hacker. And they suffer from it, too. I recently ran into some old friends at a conference. A decade ago, we could talk of nothing but the hustle. This time we only complained about our backs, our necks and our cholesterol levels. So, I understand only too well the appeal of working longer hours. It's not that I'd never sell long hours to my younger colleagues.

But I know bullshit when I see it. Long hours were never sold to me on a platter of nation building. It took me a while, but I recognized precisely what enables me to work long hours, and I acknowledged my privilege. I can think of no better acknowledgement than this quote from Stephen King:

The combination of a healthy body and a stable relationship with a self-reliant woman who takes zero shit from me or anyone else has made the continuity of my working life possible.

- Stephen King, On Writing

But when people ask you to work ridiculously long hours, healthy bodies and stable relationships are exactly the kinds of things they want you to give up.

It better be worth it.

Subscribe now

Acknowledgements: Special thanks are due to Prof Ashwini Deshpande for helping me navigate the Time Use Survey microdata, and to Sameera Khan for her perspectives on gender equality in the workplace.

I started driving when my dog was a year old, and some of her contempt for electric scooters started creeping into me. I came to understand why she hates them so much. Maybe getting mad at them for quietly invading her space is perhaps not her fault at all. It's not just the noise - it's the riders too. Like my dog, even the riders are rarely looking at the road. They're always looking at their phones, at house numbers, at people standing on the first and second floor balconies - anywhere but the road. In traffic, they're straight-up reckless. They accelerate too fast and brake too hard. They sway from side to side, jump read lights and don't hesitate to drive on the wrong side of the road. If you indicate you want to turn left or right, they'll end up in your blind spot with remarkable precision. Every time I yell an obscenity at them while driving, I can feel my dog rolling her eyes at me. There was a time when dogs had a peculiar disdain for postmen, now they have it for delivery boys1 executives.

Strangely, all of that becomes irrelevant when they bring me stuff. It doesn't matter how many vehicles they've cut off or how many dogs they've annoyed. I'm never anything but polite to a stranger who brings me food. I make it a point to address them by name, I thank them and tip them. If it's too hot, I ask them if they want some water. If it’s raining, I tell them that they're welcome to wait the bad weather out in my verandah. This irritates my dog even more, she’s the sovereign of the verandah. Her eyes show a clear sense of betrayal.

The point is that my attitude towards gig workers is a fairly uniform distribution of emotions - scattered evenly between generosity and irritation. And I'm sure that's how most people treat them. Most people are... just regular humans: kind, generous, irritable and angry in equal measure. The polite will look them in the eye and smile at them. The sociable will strike conversations with cab drivers in a new city; the tardy will ask them to jump traffic lights. The indecisive will ask waiters for recommendations and the large-hearted will tip generously.

But one particular species of humans, the 'influencers', are a different breed entirely. They can't let so much as a fly buzz past without photographing it and captioning it on social media with a motivational quote.

Social media has a peculiar relationship with gig workers. On the one hand, it’s a good place to complain about the workers' unprofessionalism, laziness or downright criminality. On the other hand, social media virtually worships their hustle. We all have this fantasy of being a liberated, empowered gig worker - someone who can decide their own hours and make theoretically infinite money. For those of us who take our job security, labour rights and social security benefits for granted - the fantasy is all the more appealing.

But the most interesting reactions towards gig workers are about how they are emancipated by startups. It's not enough for LinkedIn influencers that companies exist for profit - even though, arguably, that alone is a morally and economically justifiable reason for startups to exist. Even if startups are successful, their unappointed champions must find a reason to justify their existence, preferably one that works well with unnecessary jingoism.

One such example came from a gentleman who's very happy that Vijay Shekhar Sharma prefers to use flamboyant designations like "Indian Digital Services Members" instead of "gig workers". I wouldn't entirely blame VSS either - whatever image the phrase "gig worker" invokes in your mind, it's certainly not likely to be dignified. Notice the hashtags in the post. Words like “#viksitbharat”, “#aadhar”, “#digitaleconomy”, etc, have been woven into the post with such flourish that it almost seems like all of India's problems are solved with startups.

However, when you put the chest-thumping aside, the OP does seem to have a good point. They say that UPI, Aadhar and the India Stack have changed the way we transact and work. That's true - they have. They say that there are pros and cons to everything, and that we shouldn't always dwell on the negatives. There are, and we shouldn't. But for some inexplicable reason, these, apparently, are things that cannot be said without embellishment. And it's in this tadka of nationalism that even the most well-meaning people get carried away.

And there, as the Bard says, is the rub.

The Newer Colossus: India’s Platform Economy

Give me your tired, your poor,

Your huddled masses yearning to breathe free,

The wretched refuse of your teeming shore.

(From the sonnet by Emma Lazarus, inscribed on the Statue of Liberty)

The OP mentions that they lived in Dehradun between 2009 and 2015, where they witnessed "frequent strikes at Gandhi Park... mostly by unemployed youth, including fresh graduates and dropouts", and that "today such strikes are rare." Sadly, a quick search on Google News reveals that protests at Gandhi Park are all but rare. And the most prominent of them in the recent past was indeed held by "unemployed youth" who were demanding a CBI probe into recruitment scams and action against paper leaks.

Poor students... if only they'd known that the gig economy has been waiting for them with open arms! Sure, they'll have to wear uncomfortable uniforms and deal with the lack of a written contract, job security or other benefits that come with being "employees", but that's still better than being eyesores in Gandhi Park. Unemployed youth need to get with the times.

I don't blame the OP for not knowing what's happening in Gandhi Park since 2015 - especially since they moved out of Dehradun, and availability bias is something that affects the best of us. I wouldn't even blame them for moving out of Dehradun precisely a year before demonetisation, after which VSS's Paytm rocketed in popularity, and changed many lives including that of the OP. Remember how petrol pumps used to discourage the use of mobile phones because they were thought to be unsafe? But, as soon as the demonetization was announced, Paytm rendered all petrol pumps in the country safe from mobile phones overnight. Masterstroke! Changing lives, truly…

The OP's other claims are not so easily falsifiable2. For example, they claim that "many who once faced barriers like lack of education or technical skills, now earn ₹15,000 to ₹45,000 or even more." For claims like these, which fall squarely in the category of not even wrong, we can't use anecdotal evidence or newspaper reports. We need solid primary data. Unfortunately, it's nearly impossible to get anything very reliable about the gig workforce from existing surveys. For starters, there is no legal definition of what constitutes gig work, or that of a gig worker. Sources like the Periodic Labour Force Survey provide proxies based on which individual respondents may be singled out as gig workers, but that too would be approximate. In fact, until recently there have been very few enthusiastic moves by the government, corporate lobbyists or even the civil society to codify the gig workforce within India's labour law.

A bill was introduced in the Lok Sabha in 2020 - The Code on Social Security, 2020 which has an entire chapter on gig work, which proposed that the central government shall formulate and notify welfare schemes for gig workers. Labour experts like Prof Babu Mathew allege, however, that what's crept into the bill is the aggregators' definition of a gig worker. For instance, the bill states that the aggregator should pay no more than 5% of the amount payable to a worker towards social security or other benefits. Prof Mathew calls it a "half-hearted and inadequate attempt". Rajasthan remains the only state to enact a law that provides for the welfare of gig workers, but the law is yet to be brought into force. The Telangana and Karnataka governments appear to be making progressive steps towards this cause too3.

As much as gig workers stand to lose from lobbying, they stand to lose more from the misclassification of their work. No aggregator wants them as employees. Most don't even qualify as contractors. It's because of this that we find ourselves in a strange situation - surveys, policies and studies do include gig workers, but only in the aggregate. There's virtually no way to filter them out from the rest of the workforce. For example, the PLFS divides the population into 4 categories - self-employed, regular wage/salaried employees, casual labourers and finally those that are not a part of the workforce. Where in these categories do gig workers fit? They could self-report as being self-employed or even as casual labourers. There are also a lot of people who have full time jobs but perform gig work for side-income.

All in all, there's no way to disambiguate a gig worker. Until disaggregated estimates of the gig workforce are not obtained, social media will continue to adore their hustle and praise platforms for saving them from eternal damnation - all while remaining oblivious to their reality.

Estimating India’s Gig Workforce

Even if surveys like the PLFS don't give us clear indicators of the gig workforce, they're still the most comprehensive source we have on the matter. PLFS employs a rotating sampling panel design which collects data from individuals and households throughout the year, accounting for seasonality and bolstered by the very robust methods of the NSS.

In addition, there's a 2022 report from NITI Aayog which provides helpful guidelines as to which PLFS respondents are likely to be gig workers. These guidelines allow us to apply filters of age, education, occupation, industry to the PLFS data in order to approximate the gig workforce. For example, gig workers are likely to be largely urban, between 18 to 45 years old, and come from households that lie under the 75th percentile of monthly per-capita expenditure (in other words, the poorest three-fourths of the population). The basic categories of work (self-employed, regular wage/salaried employees, casual labourers) are still included in this filtered subset of the data, but we can more easily cross-reference them against education and monthly income.

This view of PLFS (filtered down to the individuals likely to be representative of the gig workforce) allows us to move the OP's claim about education and income away from an unfalsifiable claim to something that can be very broadly verified. And credit where it's due - the OP's estimate of the lower bound of income (Rs 15,000) does fall comfortably within the range of incomes. The upper bound of Rs 45,000, however seems quite out of reach. And even those who "faced barriers like lack of education or technical skills" are not all that worse off, at least among men. Regardless of education, there doesn't seem to be much evidence that digital platforms are lifting people out of poverty.

Trolling as a Public Service

The OP aren't the first person to make tall claims about the gig economy with a side of jingoism, and they won't be the last. There seems to be a pattern to how people like these come swarming out of the bushes to defend a misplaced ideal. That pattern, I can't help but think, is Kunal Kamra.

Kamra made a documentary about the troubles of gig workers4 and released it last August. Then, five months later on New Years Day, Albinder Dhindsa tweeted that lakhs of condoms, thousands of cans of tonic water and hundreds of lighters were en-route to be delivered in the next ten minutes. Kamra quoted him and asked how much the delivery partners were making. Dhindsa later deleted his tweet, but his unappointed champions were still hard at work.

Kamra did it earlier to Bhavish Aggarwal, too. On both occasions, people were quick to point out that Kamra was just trolling. We're not prepared for a reality where satire is a better delivery mechanism for news than whatever abomination passes for mainstream media. A comedian defending workers' rights is more bizarre to us than a newsperson’s naked and blatant destruction of journalistic standards. Trolls should have no business doing anything but trolling (except when the troll is Piyush Goyal - if you’re him, you can say what Kamra says and get away with it).

Let’s not forget the other kind of troll - the dim-witted, but dangerous mythical creature. The Witcher video games and books feature some very entertaining troll encounters. Geralt’s interactions with trolls (and other sentient “monsters”) are the most enjoyable part of the game. Even in the Witcher books, trolls are very complex creatures. This conversation from The Last Wish shows the role trolls play in the world of the Witcher:

…under that bridge sits a troll and demands every passerby pays him. Those who refuse have a leg injured, sometimes both. So I go to the alderman: “How much will you give me for that troll?” He’s amazed. “What are you talking about?” he asks, “Who will repair the bridge if the troll’s not there? He repairs it regularly with the sweat of his brow, solid work, first rate. It’s cheaper to pay his toll.

I grew fond of the supermarket, as it slowly replaced the neighbourhood grocery store. I forgot the name and the face of the local grocer, who was fast approaching the designation of "my grandfather's grocer". Supermarket visits were exciting because it was fun to push the shopping cart. It was easy to sneak packets of chocolate and potato wafers into the cart. They'd be discovered only at checkout, by which time a parent would be too hassled to not buy it.

Today, the only time I go to a grocery store is if I'm visiting something near one. I've stopped going to supermarkets too - I'm completely dependent on BigBasket and Blinkit. What's really surprising is that my parents use these apps too. When they first visited me in Delhi, one of the first things they asked me was directions to the nearest grocery store. I didn't know what to tell them. Their eyebrows disappeared when I told them where the groceries really came from.

"Even the fruits, vegetables, milk and eggs?!"

"Yes."

"How do you know they're not stale?"

"Well, as long as they don't stink..."

They gave each other the hopeless look that all children know so well.

I come from a long line of late adopters. Simon Sinek described late adopters as the people who'd only buy a touch-tone phone because nobody would repair a rotary phone. That's exactly who my parents were. But now, post-COVID, they're different people. They're older, they've lost some of their mobility, and their sons don't feel too enthusiastic pushing carts down the supermarket aisle. Now, they are quicker to use Blinkit than me. My mother even talks about instant delivery apps as if she discovered them.

If a bunch of apps can change my mother's shopping habits - they can do anything.

In the past year alone, we have enough evidence of e-commerce changing the market. In 2024, Swiggy went public, Zepto raised half a billion dollars, and Flipkart and Bigbasket started operating in a visibly better manner. In April 2024, Blinkit announced that they'd deliver a PlayStation 5 in ten minutes. To me the idea was outrageous. As anyone from the erstwhile pre-liberalisation middle class knows, buying expensive items is a ritual. Buying costly appliances in just ten minutes - without months of research and penny pinching - is just disrespectful.

I asked a lot of people about the most expensive thing they'd bought off an insta-delivery app. I couldn't find anyone who'd spent more than a few thousand rupees. A few people said they'd ordered protein powder or a video game controller or a game (not the console itself). Relatively the most expensive item I encountered was a UPS my wife bought for her office.

But who would actually need a PlayStation delivered in ten minutes?

I spoke to Deepak "Chuck" Gopalakrishnan about this. According to him, extravagant offers are more marketing signals than revenue drivers - if someone claims they can deliver gaming consoles in ten minutes, then they can certainly deliver groceries in time. Perhaps, it takes signals like these to reach late adopters like my parents. Many other customers like myself will sell their loyalty to a brand for a lot less. Quick commerce to me is no longer a luxury, but an unequivocal default, a habit.

So when someone on LinkedIN said that e-commerce apps are destroying traditional kirana stores, I wasn't surprised. On the contrary, I was sympathetic. The post claimed that nearly 2,00,000 kirana stores had shut down in the last year because of the 'predatory' practices of e-commerce companies. That's not a small number. It means that the lives of at least 2,00,000 proprietors have been affected - not to mention those of their staff and families. It won't be a stretch to say that at least a million people could have been affected by this, either directly or indirectly.

But in a swift change of tone, the post went on to offer some solutions which, at best, seemed like a rehash of management consulting jargon, and at worst, sounded condescending. The post suggested that kirana stores can bridge the gap with "tech integration", "omni-channel presence", "customer experience", "training and upskilling", etc. Not only does this sound disingenuous by itself, but when used in the context of grocery stores, it also comes across as downright unprincipled. I wish someone would say these phrases out loud in a kirana store and record the reactions. The post was clickbait to begin with. The caption had the words "SHOCKING DATA" in the thumbnail, in a large font, complete with the picture of an old man sitting behind the counter of a customer-less store - only his face visible between jars of confectionery and overhanging packets of chips.

Clearly, something was being sold to the reader. When a less-than-sincere sale is coupled with clickbait, there's always something interesting under the hood.

Under the Surface

The OP had cited no sources, even though they were a quick Google search away. In October 2024, the All India Consumer Products Distributors Federation (AICPDF) reportedly wrote to various government bodies, including the Competition Commission of India (CCI), complaining of the (allegedly) predatory practices of e-commerce companies. Their letters or press releases aren't directly available, but many media outlets have independently reported their claim of 2,00,000 kirana stores shutting down in the last year. It's safe to say that this entire discussion stems from an unverified claim.

But when we read the details of this claim, the picture becomes less bleak. The country apparently has 13 million kirana stores. So 2,00,000 of them shutting down is a 1.5% reduction. Further, if we were to limit the numbers only to metros, the closure rate rises only up to 5%. It would be tempting to think that this is a small fraction, but that may be a mistake, since we don't know what the usual rate of closures under healthy market conditions is - we don't know what the default benchmark is. It's quite likely that the OP is right, and that 5% kirana stores shutting down in metros is indeed "SHOCKING DATA".

But the claim that this is happening solely because of e-commerce companies is certainly not beyond reasonable doubt. So I went looking for something that would help me find the general, average success rates of kirana stores.

Sadly, I could find no such resource. What I did find was that kirana stores are no strangers to threats from e-commerce.

There are enough reports dating back to the 2010s which show that there have been concerns about supermarket chains damaging Kirana stores (it was the heyday of Big Bazaar and D-Mart). Then in the late 2010s, Amazon and Flipkart became a concern, but they didn't manage to displace more than 1-2% of local stores. Even when the lockdown came, e-commerce managed to displace about only half a percent of kirana stores. In this light, it's reasonable to assume that while the figure of 2,00,000 looks big - it's certainly not an outlier. It is also noteworthy that as much as there is concern towards e-commerce, there are also stories of the triumph of kirana stores! PhonePe claimed, in 2021, to have digitized 25 million "kirana stores and small merchants."

By now it's clear that grocery stores have survived many threats. In business (especially one as price-sensitive as FMCG retail) as well as in technology, nearly every decade brings a new supposed threat. Morgan Housel wrote in is book, Same as Ever, "It's good to always assume the world will break about once per decade, because historically it has."1 But decade after decade, kirana stores seem to have been remarkably resilient.

But no matter how much of a threat or opportunity anyone may believe e-commerce to be, the discourse so far has been purely theoretical - based only on claims, counterclaims and conjecture. There are bigger issues to resolve, like identifying the true impact of e-commerce on grocery shopping. To answer this, we turn to the Household Consumption and Expenditure Survey (HCES).

Digging Deeper

HCES 2023-24 was conducted between August 2023 and July 2024. The survey interviewed nearly a million individuals across almost 2,50,000 representative households2. A major part of the survey deals with groceries3, including some auxiliary information on payment methods and use of welfare schemes.

It is only in the context of the HCES microdata that I found the caption "SHOCKING DATA!" to be truly justified. For instance, only 7% of households in the country shop online for groceries!4 And of those households, more than half are located in metros and a few Tier-I cities, across only four states.

Online grocery shopping is a predominantly urban phenomenon, and saying even that is a stretch. A whopping 83% of even urban households don't shop online! Recall AICPDF's claim of 5% of kirana stores shutting in metros. The numbers are far too lopsided on the side of traditional grocery shopping to consider e-commerce as a serious threat. Other than traditional grocery stores, the survey also contains details of other ways in which consumers can procure groceries, like subsidized rations, the Public Distribution System and the Pradhan Mantri Garib Kalyan Yojana.

In short, simply shopping online or offline for groceries, with cash or with UPI isn’t the end of the story. The survey reveals such a great diversity in what and how Indians eat, that it's almost impossible to pinpoint the habits of an average Indian grocery shopper. For those of us who, like Amit Tandon says, order the bread after they break the eggs, who instinctively look for a QR code whenever we buy something from a brick-and-mortar store, it's impossible to imagine who the "offline consumer" really is.

The HCES allows us to divide the respondents into cohorts of online and offline shoppers. The imbalance between the sizes of these cohorts is staggering (7% online vs 93% offline!), but when we adjust for size, we find some significant markers of difference. For example, living in a home which has access to piped cooking gas makes you four times more likely to shop online than if you cooked with firewood, biogas or even LPG. When we move from the cohort of graduates to postgraduates, the odds of finding online shoppers are doubled.

In this light, the conflict between e-commerce and kirana stores seems to be insignificant. In fact, it seems like the real conflict is between e-commerce and lack of urban infrastructure, lack of quality education and massive income inequality.

Consider this: if e-commerce can target no more than 7% of Indian households, and if most Indians have no discretionary spending power - then who's the one who is really in trouble?

Certainly not the kirana store.

The well meaning folk who urge traditional stores to "bridge the gap" with "tech integration" and "omni-channel presence" might be better off directing their wisdom elsewhere.

Zooming Out

Even with all the number crunching, we're still looking at consumers simply as weighted sums of statistical attributes. The real Indian consumer still remains elusive, if there even is such a thing. The common consumer, then, becomes a mere metaphor in public discourse.

The Marathi movie Half Ticket shows such a metaphor. It's a story of two young slum-dwellers who steal eggs from crows' nests, because they can't afford chicken eggs. The park from where they get these eggs is sold off and turned into a pizza joint. The kids then get obsessed with eating a pizza. They work odd jobs for weeks, trying to save up the 300 rupees for a pizza. Finally when they're ready to order one, they realise that it can't be delivered to their home, because their address is simply the name of the jhuggi where they live. They have no house number and no street name5. When they visit the store, they're humiliated and thrown out because of their appearance. They go back to work and save more money to buy better clothes. When they return to the store they're thrown out once again - the guard knows they are slum dwellers. The movie does have a happy ending, the kids do get their pizza, but only because someone thinks that feeding a couple of poor kids is good PR. They have baggage that, even with money and clean clothes, cannot be jettisoned.

It's like there's an entirely different country living within India. I myself am squarely in the 7%, and there's a whole world out there which I'll never see, no matter how hard I look.

But I certainly wouldn't presume to sell anti-e-commerce solutions to kirana stores. Much less on social media of all places, with clickbait and alarmist captions.

Note: The HCES datasets, their analysis and the code used to generate the visuals here are available for scrutiny and reuse. Please leave a comment if you'd like to get your hands on them.

Subscribe now