The problem was never the salad. The problem was that I had no model for what a salad is supposed to do. Once I started treating the bowl as an engineering problem instead of a moral one, everything changed. A good salad keeps me full for hours at 300 to 500 calories with enough protein to actually count as a meal. For a vegetarian this is harder than it looks, because the easiest protein and fat sources in most salad advice are meat, and the moment you remove that you have to be deliberate about replacing it.
This is the model I use.
The standard salad fails because it optimizes the wrong thing. People obsess over the dressing being low calorie or the leaves being organic, when the actual constraint is satiety per calorie. You want maximum fullness and maximum flavor for the fewest calories, and you want enough protein that your body registers the bowl as food rather than as a snack.
There are two distinct fullness mechanisms at work, and the standard salad only uses one of them. The first is volume. High water and fiber foods stretch the stomach and trigger the mechanoreceptors that signal fullness, which is why a large low calorie bowl feels satisfying while you eat it. Researchers have a clean name for this, energy density, meaning calories per gram, and the lower it is the more food you get to eat for a fixed calorie budget. In one well known set of experiments by Barbara Rolls, people who started a meal with a large low energy density salad ate fewer total calories and reported feeling just as full as people who skipped the salad entirely. That is the volume mechanism doing its job.
The catch is that volume fades. Once the food clears your stomach and the stretch receptors stop firing, hunger comes back, and a pile of lettuce has almost no protein and no fat to carry you past that point. This is the second mechanism, and it is the one most salads ignore. Protein is the most satiating of the three macronutrients, and it works through hormones rather than stomach stretch. A high protein meal raises the satiety hormones GLP-1, PYY, and CCK while lowering ghrelin, the hormone that drives hunger, and that hormonal shift is what keeps you full for hours instead of minutes. Fiber adds to this by slowing digestion, and soluble fiber in particular forms a gel that delays gastric emptying, so the food lingers and the fullness signal lasts longer.
The fix is to build the bowl from distinct functional layers, where each layer does one job, and where the protein and fat layers carry the weight that meat would normally carry in a non-vegetarian version. Volume gets you through the meal. Protein and fiber get you to the next one.
The first decision is the theme, and it matters more than any single ingredient. Italian, Mexican, Caesar, taco, pizza, even a Chipotle bowl reimagined as a salad. The theme is the constraint that makes the rest of the choices easy.
Without a theme you end up with a bowl of random healthy things that fight each other. Apple next to feta next to salsa next to peanuts tastes like a mistake. With a theme, every ingredient has to earn its place by fitting the flavor world you already chose, and the bowl tastes like a dish someone designed rather than a fridge someone emptied.
Pick the theme first. Everything downstream gets simpler.
A complete salad has seven layers. The first four build volume and flavor for almost no calories, which is the energy density mechanism doing the work. The last three are where a vegetarian has to be deliberate, because protein and fat are the layers that decide whether you stay full past the meal itself.
Lettuce, rocket, spinach, kale, cabbage. This is the bulk that fills the bowl and your stomach without spending calories, and it is the cheapest fullness you will ever buy. To put the energy density point in concrete terms, for the calories in a small order of fries you could eat ten cups of spinach, which is the entire argument for building on a base of greens. Shredded cabbage and kale hold up better than spinach over time, which matters if you meal prep. Be generous here.
Cucumber, carrots, peppers, onions, courgette, radish, tomato. Chop them small. Small pieces distribute through the bowl so every forkful has texture, and the chewing itself is part of what makes a salad feel like a meal, since longer eating time and more oral exposure both feed into the fullness signal. A bowl of large clumsy chunks eats like three separate foods. A bowl of small dice eats like one dish.
A handful of apple, grapes, blueberries, pomegranate, or peach. Sweetness against savory is what keeps the bowl interesting enough to finish. Save a few sweet bites for the end and the salad closes like a small dessert instead of a chore. Keep the quantity controlled and lean toward lower glycemic load fruit like berries and apple over very starchy additions, because fruit is where the calories creep up quietly and starchy choices spike blood sugar in a way that can bring hunger back sooner.
This layer is optional and it is also the difference between a salad you tolerate and a salad you crave. Pickled onions, pickled jalapeños, capers, fresh herbs, nutritional yeast, salsa, a spoon of kimchi. Everything here is low calorie and high impact. For a vegetarian, nutritional yeast pulls double duty by adding a savory, almost cheesy depth that meat would normally provide.
Fat is non-negotiable and it doubles as the dressing, which means you stop pouring calories from a bottle and start getting them from food that also fills you. For a vegetarian the holy trinity is cottage cheese, cheese, and avocado. Cottage cheese deserves special attention because it is high protein and high fat at once, so it covers two layers in one spoon, and the plain or pineapple versions both work depending on your theme. Greek yogurt, olive oil, and crushed nuts round out the options.
This is the layer that keeps you full past the next five minutes, and it is the layer where most vegetarian salad advice quietly gives up. You have more options than people think. Paneer, grilled or pan seared, takes a theme beautifully and holds its shape, and it carries plenty of leucine, the amino acid that triggers muscle protein synthesis. Tofu, pressed and crisped, does the same for lighter flavor worlds, and soy is one of the few plant proteins that is complete on its own. Chickpeas, lentils, and black beans add protein and a little carb at once. Edamame, halloumi, or a generous portion of cottage cheese all qualify.
One vegetarian specific point worth knowing. Most individual plant proteins are limited in at least one essential amino acid, lentils run low on methionine, for instance, so combining sources covers the gaps, which is why chickpeas with a grain or tofu with mixed vegetables and seeds works so well. Aim for roughly 25 to 40 grams of protein in the bowl and you cross the threshold that meaningfully stimulates muscle protein synthesis and triggers the satiety hormone response. Pick the protein that fits the theme. Paneer tikka spice for an Indian bowl, chipotle tofu for a Mexican one, halloumi and herbs for something Mediterranean.
Croutons, roasted potato, rice, roasted root vegetables, toasted chickpeas. This layer adds crunch and staying power and rounds the bowl out to a real meal. Keep the portion measured, since this is the other place calories climb fast, but do not skip it, because the crunch is part of what makes the bowl satisfying rather than rabbit food.
The dressing is where most diets lose the plot, because the standard move is to pour fifteen grams of fat from a bottle that adds nothing but calories. The fix is to make the dressing out of the fat layer itself, so the same spoon that flavors the bowl also fills you. Every option below is vegetarian and built around protein or whole food fat rather than oil alone.
The cottage cheese base is the one I reach for most. Blend half a cup of cottage cheese until smooth and you get a creamy, high protein dressing that coats everything and adds almost no extra calories beyond the cheese you were already counting. Thin it with a splash of water, then season to the theme. Lemon and dill for a Mediterranean bowl, chipotle and lime for a Mexican one, garlic and black pepper for a Caesar feel.
The Greek yogurt vinaigrette is lighter and sharper. Whisk Greek yogurt with red wine vinegar or lemon juice, a little Dijon, salt, and pepper, and you have something tangy that still carries protein. A teaspoon of honey rounds it out if the bowl leans savory.
The tahini lemon dressing brings whole food fat and a nutty depth. Loosen tahini with lemon juice and warm water until it pours, then add crushed garlic and salt. This one suits roasted vegetable and chickpea bowls, and the sesame fat keeps you full without any animal product, so it works for vegans too.
The avocado green goddess is the richest of the set. Blend half an avocado with herbs, lemon, a spoon of yogurt, and water until it turns into a thick green sauce. Use it where you want indulgence, and keep the portion measured since avocado is calorie dense.
The salsa and pickle approach is the leanest option of all. A few spoons of salsa, a splash of the brine from your pickle jar, and a little olive oil dress a Mexican or taco bowl for almost nothing, while the acid does the work of waking up every other ingredient.
Three complete bowls, each built from the seven layers, each landing in the 300 to 500 calorie range with 25 to 40 grams of protein. Quantities are a starting point, so adjust to your own targets.
A North Indian themed bowl where spiced paneer carries the protein and a yogurt dressing cools it down.
A bright, smoky bowl where black beans and chipotle tofu stack protein and a cottage cheese dressing stands in for sour cream.
An herby, lemony bowl where halloumi and chickpeas share the protein load and a tahini dressing brings the fat.
Start with a theme, then build the seven layers in order, leaning on cottage cheese and paneer to carry the protein and fat that meat would otherwise carry. A Mexican bowl becomes shredded lettuce and cabbage, diced peppers and onion, a little corn, pickled jalapeños and salsa, a scoop of cottage cheese standing in for sour cream, chipotle tofu or black beans for protein, and toasted tortilla strips for crunch. That bowl lands around 400 calories with 35 grams of protein and tastes like something you would order rather than something you are enduring.
The whole point is that the bowl works for you instead of against you. The volume keeps you full while you eat, and the protein and fiber keep you full long after, which is the combination the research keeps pointing to for staying lean without feeling starved. Pick a theme tonight and build one. The model is simple enough that after two or three bowls you will stop measuring and start improvising, which is exactly when it becomes sustainable.
]]>Mandana Mishra, the greatest living Mimamsaka, was losing. The boy across from him, barely twenty, from a village in Kerala nobody had heard of, was dismantling his arguments. Mandana Mishra’s own wife was judging. The boy’s name was Shankara. He would leave this hall with Mandana Mishra’s students, his library, and his legacy. Those were the stakes, and everyone had agreed to them before the first syllable.
A handwritten credit note from a small town in Tamil Nadu was being honoured in Rangoon. The Chettiar merchant house that issued it had 75 crore rupees deployed across Burma. The hundi was backed by nothing but the family’s reputation. Two of India’s major banks would later trace their founding capital to this network.
A guy I grew up with watches porn on his phone every night. In the morning he calls Valentine’s Day pashchimi sanskruti and curses young people for holding hands in malls. Between his world and theirs, someone performed a specific act of surgery. The scar tissue is everything we now mistake for tradition.
I grew up three hours from Khajuraho. On the school trip the guide walked fast and the teachers studied the ceiling, and we thirteen-year-old boys giggled at the mithuna panels where a woman arched her back against a man’s chest while two attendants held a lamp and her anklet bell caught the light. The panels sit at the entrance for a reason. You walk through kama on the way to the garbhagriha.
Back in Pataliputra, the ganika had more to teach the young man on the mattress than positions. She pressed her thumb two fingers below his navel and waited until his breath deepened and his spine released in sequence from sacrum to skull. The Kamasutra names sixteen varieties of kisses and documents which pressure on which nerve returns which response. She knew where his breath caught, which muscle let go when she bit his shoulder, how long to wait before moving her hips. When he came she held the back of his neck and made him stay in his body until the tremor passed, because a man who spilled and rolled away had learned nothing. Amrapali of Vaishali charged fifty karshapanas a night, hosted the Buddha in the same week, and donated her mango grove to the Sangha. Her legal standing in 600 BCE exceeds what a sex worker has in 2026 India.
Macaulay stamped Henry VIII’s sodomy law onto the subcontinent in 1862 as Section 377. Missionaries criminalized the devadasi’s dance and the Criminal Tribes Act registered hijras for gradual extinction. Free India’s first cultural act after August 1947 was abolishing its own temple artists. The Manusmriti had been restrictive long before any Englishman arrived, and colonial law gave that old anxiety the prestige of modernity.
The guy on his phone thinks he is guarding dharma. He is performing Victorian Christianity in Hindi. Aur usse koi bataye to maane bhi nahi.
Jacks studied electrical engineering at IIT Roorkee while his father mortgaged four hectares outside Bhopal, so that his son could sit for the UPSC in a ten-by-ten hostel room off Karol Bagh. Eight years went by in that room. The topper list came out every year with other names on it.
The people who clear are brilliant and disciplined, and a father’s bet on a collector’s stability is rational in a country where the state remains the safest paymaster. Jacks’ great-grandfather would have walked into a shreni instead. The merchant guild would have apprenticed him and written his working capital against its collective treasury, and by his third year he would have been drawing a share of the syndicate’s returns. Kautilya opened the Arthashastra with arthasya moolam rajyam, the state exists for the generation of wealth. He would not recognize a republic where the state itself is the prize.
On the second of February, 1835, Macaulay told Parliament he intended to produce “a class of persons Indian in blood and colour, but English in tastes, in opinions, in morals and in intellect,” and he used that sentence to gut the indigenous knowledge economy and replace it with the Indian Civil Service, a priesthood of generalists who distrusted commerce and worshipped the file. Nehru extended the logic through the License Raj until 1991, when India flew 67 tonnes of physical gold to London to avoid default. Pichai runs Google and Nadella runs Microsoft, and both of them had to leave the country of Kautilya to play the artha game he wrote down. Macaulay ke bachche, abhi tak naukri khel rahe hain.
When Khilji’s cavalry sacked Nalanda around 1200, the manuscripts burned for three months. Eight centuries of accumulated argument rose as smoke over Bihar. The university had rejected four of every five applicants at its entrance exams. In a village in fourteenth-century Kerala, Madhava of Sangamagrama sat under a coconut tree and derived an infinite series for pi that Leibniz would stumble onto two hundred and fifty years later. The network of village schools that Dharampal documented in The Beautiful Tree taught one boy or girl per household, and the majority of the students were Shudras and non-Brahmins. Macaulay defunded that network without knowing a word of Sanskrit or Arabic, and the pandits lost both their students and their reason to keep arguing.
What survives is the edge case consumed as the whole system. I track my Vimshottari dashas and the tradition carries serious depth in the hands of someone trained. What most Indians encounter is a pandit in a television studio selling remedies for Kaal Sarp Dosha, a concept that appears in no classical text and was fabricated sometime in the last century. Nalanda’s entrance exams became WhatsApp forwards about Rahu Kaal.
The guy on the phone observes his Rahu Kaal every Tuesday and would not recognize Panini’s name if it fell into his lap. He thinks he is guarding a tradition. He is guarding the outline of a tradition that the British drew for him.
My grandmother disappeared somewhere every evening with her mala. Her lips moved and her eyes closed, and something in her face went still in a way that nothing else made her still. I was six years old and I could not follow her there, but I could see the place she had gone was real.
Her lineage was wilder than anything her grandson would be taught in school. Abhinavagupta, a Kashmiri householder, wrote that sexual climax and the terror of imminent death were the same doorway, since both were moments where consciousness caught itself watching. The Aghoris ate from human skulls on cremation grounds because if Brahman is everything then the burning ground cannot be less holy than the temple. Kabir told the pandits of Kashi and the qazis of Jaunpur that both of them were lying, and when he died in Maghar both traditions showed up to claim his corpse. Mirabai said no to a Rajput king for the sake of Krishna and was probably poisoned for the refusal. My grandmother touched that whole lineage quietly every evening in a house in Guna.
James Mill wrote The History of British India in 1817 without setting foot in the country and without knowing a single Indian language. The English-educated Indian swallowed his diagnosis and called his own grandmother’s mala village nonsense. The vacuum filled with criminals in saffron selling miracles on television. My grandmother’s pranayama goes for two hundred dollars an hour in a studio in Brooklyn now, and India earns six percent of the global yoga tourism market it invented. Aur humne khud hi toh chhoda tha.
Jacks is thirty and starting over. The eight years are gone and the engineering is still in his hands, and the system that ate the years is two centuries old and thinning. The guy on the phone will post something about Indian culture in the morning in English, using a grammar shaped by the man who wanted to manufacture him. My grandmother is still going somewhere every evening with her mala, and my own fingers have started to learn the count.
]]>I’ve always picked Spiderman. And for a long time I thought that was a lesser choice, the way you might think preferring chai over single malt is somehow less sophisticated. It took me years to realize I wasn’t picking the easier hero. I was picking the harder philosophy.
There are two ways to think about what you owe the world.
The first says: detach. Let go of outcomes, let go of relationships, let go of the personal and become pure instrument. This is the Stoic path, the ascetic path, and in its most refined form, a certain reading of the Gita where कर्मण्येवाधिकारस्ते becomes a mandate to act without feeling. Batman lives here. His parents die in an alley and he transmutes that grief into a system. Every relationship he forms afterwards is structured around his mission. Alfred serves the mission. The Robins are recruited into the mission. Commissioner Gordon is a tool of the mission. Love exists, but only as fuel, never as a force that might redirect him.
The second way says: your duty arises from your relationships, your station, your particular entanglements with the people around you. This is dharma in its classical sense, where what you owe depends on who you are to whom. Spiderman lives here. His heroism is born from a specific guilt, Uncle Ben dying in an alley because Peter chose not to act, and sustained by specific loves, Aunt May worrying about his safety and MJ furious that he’s always late, and tested by specific failures that actually change him. He keeps getting pulled away from the hero work by the human work, and that tension is where the real story lives.
I think people pick Batman because detachment dressed as discipline looks like strength from the outside. Staying in the mess of your relationships while trying to save the world looks like weakness. That framing deserves questioning.
Batman’s appeal contains a specific promise worth examining: that if you suffer enough and control enough and refuse enough vulnerability, you become untouchable. His trauma becomes his superpower and his isolation becomes his clarity. Every contingency is planned, every variable accounted for, every weakness transformed into a weapon.
This is deeply appealing, especially if you’ve been hurt. The promise is that you can build a Batcave inside yourself where nothing reaches you, and from that fortress of solitude you can save the world on your own terms. Founders love this story. So do high-achievers, competitive exam toppers, anyone who learned early that self-reliance is the highest virtue and needing people is a structural weakness.
And the tricky part is that it works just well enough to be dangerous. You can actually build a life on Batman’s philosophy. You can isolate, optimize, control, and produce extraordinary results for a surprisingly long time. What becomes harder to do, over enough years, is sustain it, because the human parts you amputated to become the machine eventually demand their due.
Spiderman never gets the option of invulnerability. He keeps failing and losing people, and he makes choices that haunt him, and he has to sit with the consequences in a way Batman never does, because Batman’s wealth and infrastructure absorb the cost of his decisions while Spiderman absorbs them in his body, his relationships, his rent.
And every single time, after every failure, he has to choose connection again. Go back to Aunt May’s apartment knowing she’ll worry. Call MJ knowing she’s furious. Swing back out into the city knowing he’ll probably get hurt again, and that this time there’s no Batcave to retreat to, only a cramped Queens apartment and a biology textbook.
Choosing vulnerability repeatedly, with full knowledge of its cost, is the more difficult position. I’m starting to think every serious spiritual tradition knows this, which is why the greatest figures in devotional literature are rarely the renunciates who left the world but the ones who stayed in it and let it reshape them. Mirabai stayed in her body. The Gopis lived inside their longing for Krishna, and the living was the practice.
I keep seeing this play out in how people build companies.
You can build infrastructure. Clean, elegant, technically brilliant systems that solve hard problems. This is Batman energy: controlled, capital-efficient, indispensable. The work is pure because it answers to engineering constraints, and engineering constraints are honest in a way human needs rarely are.
Or you can build something that requires you to stay embedded in the mess of what people actually want, actually share, actually care about. Consumer products, creative tools, social platforms. This work is closer to Spiderman: you are constantly being pulled by forces you cannot fully control, constantly failing in public, constantly discovering that the thing you built is different from what people needed, and adjusting from a position of listening rather than mastery.
The infrastructure path is seductive because it feels like competence. The consumer path is terrifying because it feels like exposure. One scales through precision and the other through resonance, and resonance requires staying human enough to feel what others feel, which means the Batcave, however tempting, is exactly the thing you have to give up.
I’ve tried both. I’m still learning which one I’m actually built for, but I know where the energy comes from, and it comes from the mess, from the people, from staying close enough to hear what someone wants to create before they fully know how to say it.
There is a figure in Hindu tradition who I think about more than any Western superhero when I’m trying to work this out. Hanuman is stronger than Ram and Ravana both, powerful enough to carry a mountain across the sky, and he chooses to serve. He fights his own son Makaradhwaja and feels pride when his son fights well, because the boy is following his dharma. He forgets his own strength until Jambavan reminds him, which means his power lives in relationship, not in isolation. He could justify solitude more than anyone and he refuses it every single time.
That is the version of strength I keep returning to. Capable of everything Batman dreams of, choosing to live like Spiderman, embedded in love, answerable to someone, inside the mess.
I don’t think I’ve figured this out. I’m twenty-seven, building a company while caring for family while learning how to love someone while writing about dharma for people my age, and most days the tension between the Batman impulse and the Spiderman instinct is a live wire. The temptation to retreat into pure competence, to build the Batcave and lock the door, is real and it shows up more often than I’d like to admit.
But every time I’ve followed that impulse, something essential went quiet. And every time I’ve stayed in the mess, stayed in the relationships, let people change me even when it was uncomfortable, the work got better and the life got richer in ways I couldn’t have planned for.
I pick Spiderman, and I think I always will, and I’m still figuring out everything that means.
]]>Nobody buys his “busy lately.” Everyone knows what the warmth was buying. There’s an entire internet vocabulary for this man: nice guy, covert contract, emotional manipulation. We’ve learned to see the ledger when a man holds one.
This essay is about the ledger everyone else is holding too.
A friend helps you move apartments in July heat, forearms streaked with dust, laughing about your book collection. Two weeks later, life swallows you whole and you don’t call. “I was there for you and you can’t even check in?” The help was genuine. The resentment is also genuine. Both breathe in the same sentence because the help was never free, it carried an expectation of reciprocity on a timeline you never negotiated.
A relative calls every Diwali, every birthday, sends mithai wrapped in three layers of cloth like your grandmother used to. Your family makes a financial decision that doesn’t favor her side. Next Diwali, your phone stays silent, and you find yourself checking it twice before you understand what you’re checking for.
The warmth was real, the sweat on the forearms, the three layers of cloth. And all of it had a price that was never spoken aloud.
The internet taught us to name the ledger when a man runs it. Rightly so. But flip the gender and watch the language soften. When a woman invests months of emotional availability into a man and goes cold after he starts dating someone else, the same people who spotted the covert contract in the opening reach for gentler words. “She’s protecting her heart.” “You can’t expect her to stick around and watch.”
Same vanishing act, same silent phone, same confused person on the other end. Only the words changed.
A principle that bends based on who’s performing the behavior is a preference wearing the clothes of a principle.
I’ve sat in a living room at 1 AM with a friend falling apart over the same person for the third time, his voice hoarse, cycling between anger and grief. I stayed until his breathing slowed. I’ve sat across from a friend after his fifth UPSC attempt, watched him stir his chai without drinking it, the spoon going around with the slowness of someone deciding whether to feel the thing or postpone it one more day. In business, I’ve shared sales strategies with competing founders, written recommendation letters for people whose names would never appear next to mine again.
All real. And underneath, a quiet accounting system was running. When a founder I helped didn’t return my call six months later, I felt it in my jaw before my thoughts, a tightening, a recalculation. When a friend I’d carried through his worst months didn’t show up for mine, I caught myself scrolling our old messages with the energy of someone tallying receipts.
That was the moment I saw it clearly, sitting on my bed at 2 AM, assembling evidence for a case I hadn’t admitted I was building.
Fixing this meant separating giving from expecting, while simultaneously learning when to say no. Transactional kindness and unlimited availability are two roads to the same place: resentment, withdrawal, and the quiet closing of a door.
Before I give my time now, I hold one question: can I sustain this without a ledger forming? If yes, I show up completely. If no, I say so clearly and early. These conversations sometimes land you in someone’s bad graces. That’s a cost worth paying. A clean no is infinitely kinder than a yes with invisible conditions, because when you say yes while keeping score, you become the man in the opening paragraph.
Your right is to the action alone, never to its fruits. This gets quoted so often its radical demand has been domesticated into a motivational poster, printed on mugs, stripped of its teeth. But sit with it in the context of kindness and it becomes almost unbearable: give everything, hold nothing, and when the person you gave to walks away without looking back, let the giving be enough.
I think about the friend with the cold chai, the spoon still going around. I think about the living room at 1 AM, the hoarse voice finding its way back to steady breathing. Those moments are mine. They live in the giving and they don’t need a response to be real. The day I understood that, the ledger closed, and my hands were lighter than they had been in years.
]]>Michael Burry sat alone in his office in San Jose, headphones on, reading subprime mortgage bond prospectuses that no one else on Wall Street had bothered to open. He noticed a pattern buried in thousands of pages of fine print that made the entire American housing market look like a house of cards. The ability to look at a system everyone takes for granted and see what’s actually happening inside it is the rarest quality a human being can possess.
Bob Noyce had this quality in 1959 when he looked at a silicon wafer and realized you could etch an entire circuit onto a single chip instead of cutting it apart and wiring the pieces together. Bret Victor had it when he realized that creators disconnected from the immediate effects of their work were being robbed of their best ideas. Burry had it when he saw that the entire mortgage bond market was built on garbage. All three looked at the world, saw something broken that others couldn’t see, and felt compelled to act. The difference is that Noyce and Victor were inside games that let them build something with that insight, and Burry was inside a game that only let him bet.
Bret Victor’s “Inventing on Principle” talk has been passed around among builders for over a decade. The core idea: find something in the world that feels like a moral wrong to you, something specific enough to act on, and spend your life inventing against it. Larry Tesler hated that users got trapped in software modes, so he killed modes. Stallman believed software must be free, so he built GNU. Victor believed creators need an immediate connection to what they create, so he built tools that honored that connection. Each of them found their principle, and the principle became their life’s work.
I think the talk contains a silent assumption that, if you get wrong, makes the entire framework collapse. Every one of Victor’s examples was working inside a positive-sum game. The things they created did not come at anyone’s expense. Tesler’s modeless editing made every subsequent interface better for everyone. Stallman’s GPL powers the infrastructure of the modern internet. Their effort expanded the total value in the world.
Inventing on principle requires a game where your creation adds something that wasn’t there before, where the sum is positive.
Naval Ravikant named this distinction clearly: “Wealth creation is an evolutionarily recent positive-sum game. Status is an old zero-sum game. Those attacking wealth creation are often just seeking status.” The word evolutionarily recent carries the whole insight. Wealth creation, making something from nothing through technology and knowledge, is a new trick for the species. Status games, hierarchical jockeying for position within a tribe, are ancient and wired into us from before agriculture, when survival depended entirely on rank. For number three to move to number two, number two has to move out of that slot.
In a zero-sum game, there is nothing to invent. The total value is fixed and the only moves available are capturing, extracting, and redistributing what already exists. You can be brilliant at it, you can see patterns nobody else sees, but every dollar you made is a dollar someone else lost. The structure of a zero-sum game turns even the most principled person into an extractor.
This is what haunted me about the film. Every protagonist has the one quality Victor says matters most: they look at the world, see something deeply broken, and feel a moral compulsion to act on what they see.
Burry reads the loan-level data and sees that millions of families are going to lose their homes because the bonds their mortgages were bundled into are junk. Mark Baum walks through the wreckage and the wrongness of the system lands on him like a physical weight. They both see a principle being violated, that people should not be lied to about the biggest financial decision of their lives, that banks should not package garbage and sell it as gold.
And the only tool the game gives any of them is a credit default swap.
Baum’s disgust is real, Burry’s conviction is real, their principles are real, but the game they’re inside has exactly one mechanism for acting on those principles: bet against the people who are going to get hurt, and profit from the collapse. The zero-sum architecture of trading takes their genuine moral clarity and channels it into extraction. They see the wrong and the game says: good, now get paid when it explodes.
Imagine Burry in a different game. Same mind, same ability to see what nobody else sees, same compulsion to act. But instead of reading bond prospectuses, he’s reading research papers on neglected diseases, or examining how software tools fail creators, or studying why housing construction in America is so grotesquely slow and expensive. In a positive-sum game, Burry’s gift becomes invention. In the zero-sum game of Wall Street, it becomes a wager. The game ate the principle whole.
Here is where it gets uncomfortable. The careers that society wraps in the most prestige, the ones your parents dreamed about, the ones your relatives bring up at weddings, are disproportionately zero-sum.
Finance, law, politics, and consulting recruit from the top of every class at every elite university in the world. McKinsey, Goldman Sachs, Sullivan & Cromwell, the Senate, these institutions have first pick of the most talented young people alive, and every one of them, at its structural core, is a game where the total value is fixed and the work consists of capturing share. A trader’s profit is another trader’s loss, a litigator wins when opposing counsel loses, and a politician holds office only if another candidate does not. Even the language tells you the shape of the game: you “win” cases, “beat” competitors, “capture” market share, “gain” ground, all combat verbs for a combat structure.
Peter Thiel recognized this when he wrote that competition is for losers. All happy companies are different, each earning a monopoly by solving a unique problem, and all failed companies are the same, having failed to escape competition. Thiel spent his early career at a law firm, making serious money, and later described it as one of the worst periods of his life because everyone was competing ferociously for the same fixed set of outcomes.
Compare that with the people who were working at Fairchild Semiconductor and Bell Labs and Xerox PARC during those same decades, deploying comparable intelligence in a game where every invention expanded human capability without reducing anyone else’s. Robert Noyce, the preacher’s son from Iowa who co-founded Intel, used to tell his employees to “go off and do something wonderful.” Warren Buffett, who knew Noyce well, said he was an extraordinarily smart guy who didn’t need to let you know he was that smart. That casualness about brilliance is only possible in a positive-sum game, because in a zero-sum game you have to constantly signal your rank. Rank is the only currency.
The trap of prestige is that it steers the best minds into zero-sum kitchens and hands them all forks.
The question that precedes where you live, who you’re with, and what you do is which game you play. Most people make this choice unconsciously and spend decades living inside it.
If you choose to build products, to do science, to make art, to engineer things that serve people in ways that didn’t exist before, you are choosing the game where your principle can breathe. If you choose finance of the trading variety, or litigation, or politics, or consulting that exists to help one company beat another, you are choosing a game where your principle, however sincere, will be structurally bent toward extraction. The game will take your clarity and hand you a credit default swap, take your conviction and hand you a short position, take your intelligence and hand you a fork.
Every principled life happens only once. The next Victor will not care about immediate feedback in creative tools. The next Stallman will not care about software freedom. The next you will not care about what you care about.
Noyce told his employees to go off and do something wonderful. Victor told students that inventing on principle is a path available to them that no career fair will ever mention. Both of them were saying the same thing from inside the same kind of game: the positive-sum kind, the kind where building is possible, the kind where your principle gets to become an invention instead of a bet.
The game you pick is the life you live.
]]>I’ve been watching my parents like a student.
My mother is recovering. My father still buys her biscuits from the shop that closes at nine. She still complains about the chai, still asks if I’ve eaten, still wants to know about that nice girl. The hospital room smells like phenyl and jasmine. The jasmine drifts in from the temple next door. Thirty years of marriage is holding steady.
I used to think the Ship of Theseus was a riddle. Now I think it’s a love story. The Greeks had wine and syllogisms. We have gulab jamun and kundali. But the question is the same: replace every plank and something still sails. Someone asks kuch khaya? and someone lies and says no, and that is the thread, that is what survives.
My father is an instrumentation engineer at a fertilizer plant. He has spent his career calibrating gauges, reading pressures, making sure temperatures are where they need to be. Now he sits in a hospital room mentally timing IV drips. This one runs twenty minutes. That one runs one-twenty. He knows when each bag should finish and sometimes calls the nurse before it does, a minute early, the way I imagine he has called control rooms his whole life: not because something has gone wrong but because he is the kind of man who pays attention to when things are about to.
Between drips, he calls back to the office to check if a bill got approved. Then he walks down to the hospital billing section. Then he comes back and asks my mother if she’s eaten. She says the chai has too much milk and not enough ginger. He says he’ll send someone to the gumti for a kadak chai. She says nahin, rehne do. He asks again about the food.
Theseus replaces a plank. The ship doesn’t notice.
My parents met once before they married. One meeting. Chai going cold. Her parents listening from the kitchen. A yes that would echo for thirty years. Now the chai has too much milk and not enough ginger and they are arguing about it in a hospital room and this, I think, is what the philosophers missed. They were asking whether the ship is the same ship. They should have been watching who keeps sailing it.
My cofounder got married last week. I introduced them years ago, him and my childhood friend. I watched them walk around fire and I thought: another ship launches. The mandap was up by morning and down by nightfall. Mango leaves and marigolds. The fire burned for an hour. But my grandmother used to say the mandap doesn’t hold the marriage. The marriage holds the mandap. She would say this the way she said everything, like it was obvious and you were slow for needing to hear it.
I have said yes to things that changed completely. A company that became a different company. A plan that survived nothing. You say yes and then everything you said yes to transforms and you are still there, holding the shape of a promise the world has rearranged around you. My father knows this. He said yes to a girl over cold chai and thirty years later he is timing her IV drips in a hospital in Guna. You keep showing up. The instrument changes. The attention doesn’t.
My father comes back from the billing section. My mother is asleep. He checks the drip. He sits down. He waits.
Neti neti. Not this, not this. Strip away what changes and something remains. My father can’t pronounce the name of the drug but he knows when the next blood draw is scheduled. My mother asks if I’ve eaten and I say no and she clicks her tongue and this is the thread.
She wakes up and turns to me. So, that girl. Have you called her?
Some things even chemotherapy cannot pause.
Theseus reaches Athens. The philosophers want to know if it’s the same ship. I think Theseus looked at them the way my mother looks at me when I’m being too philosophical. Beta, the chai is getting cold. Why are you making simple things complicated?
My mother is recovering. My father is timing the next drip. My cofounder is married now to my childhood friend. Somewhere two strangers are walking around fire, and the mandap will come down by nightfall, and the wood is already beginning to change.
The ship was never the point. The sailing was.
]]>And yet.
I was debugging a neural network at 2 AM. Loss function stuck, gradients vanishing, the usual. I gave up and put on 3 Idiots for the twentieth time. Comfort viewing. Rancho was explaining how fear makes you stupid, and I was half-watching, half-thinking about why my model wouldn’t learn, when he said something about the heart being easily fooled, tell it “aal izz well.”
Then it hit me. The model learns backward. Rancho lives backward. And somewhere in a Feynman diagram, an electron is doing exactly the same thing.
What if the direction we call “forward” is the problem?
Richard Feynman had this idea that sounded insane. A positron, the antimatter twin of an electron, is just a regular electron traveling backward through time. When matter and antimatter annihilate, Feynman saw one electron bouncing off the present like a wall, reversing direction, heading back the way it came.
The math checks out. The equations of physics work the same in both directions. Play any fundamental interaction in reverse and it still holds. The arrow of time we feel, yesterday gone, tomorrow coming, that’s just statistics. Entropy. The improbability of spilled milk jumping back into the glass.
At the particle level? Freedom. No obligation to move forward.
I think about Ranchoddas Chanchad. The name itself is a joke, a pseudonym stolen from a servant’s son, worn by a kid who snuck into engineering college because he loved machines. Everything about him runs backward. The system says memorize first, understand later (maybe never). Rancho understands first. Joy first. The system says get the degree, then you have permission to learn. Rancho learns without permission and the degree chases him.
When his friend Farhan finally calls his father and says “I want to be a wildlife photographer,” that’s a reversal. Twenty-two years of moving forward through someone else’s expectations, and one phone call traces it all back to the source. What did you actually want, before anyone told you what to want?
Backpropagation works the same way.
Data flows forward: input to hidden layers to prediction. The prediction is wrong. It’s always wrong at first. The gap between prediction and truth is called loss.
Here’s the magic. The loss flows backward. Through every layer, every connection, every weight. Each parameter asks one question: how much of this error is mine? Then it adjusts. Millimeter by millimeter, the network rewires itself toward truth.
The forward pass is arrogant. It makes claims. The backward pass is honest. It traces error to its source.
When Raju stands on the ledge, Rancho reaches backward. He channels Raju’s paralyzed father, speaks as him, goes to the root. The suicide attempt is just the loss signal. The real error started years ago, in shame, in fear, compounding silently.
You fix a network by backpropagating. You trace the error home.
The Yoga Sutras call it pratyahara. Withdrawal. Tracing attention backward before it attached to objects, before it became craving. Following the gradient of suffering to find which ancient weight keeps producing the same pain.
Moksha, liberation, comes from muc, to release. Release from forward momentum. From the assumption that more accumulation will eventually equal peace.
It won’t. The function itself needs questioning.
Feynman’s electron bounces off the present and heads back. Rancho looks at a system designed to move students forward through fear and asks, what if we start with joy instead? A neural network makes a wrong prediction and the error travels home, finds every weight that contributed, asks each one to change.
We accumulate so much. Degrees, identities, fears we’ve learned to call wisdom. Layer after layer until the original signal is buried. The child who was curious before anyone told them what curiosity was for. The kid who loved cameras and engines before learning they belonged to separate careers.
When Farhan’s father finally says “live your life,” we cry because a gradient reached all the way back. Twenty-two years of accumulated error, finally acknowledged. By reversing.
Rancho disappeared at the end of the film. They found him years later, in Ladakh, teaching children, having built a school, having won the Padma Shri without ever chasing it. He was living at the source. He’d been there all along.
The positron knew all along: you can always turn around.
Too much to unlearn. Only one life. The backward path is shorter than it seems.
]]>My mum was diagnosed with HER2+ breast cancer recently. I’ve spent the last month sitting in hospitals, talking to oncologists, and learning more about cancer than I ever wanted to know.
One thing kept coming up in every conversation, every paper, every video I watched: we’re optimizing for the wrong thing.
When someone says “cancer research,” the image that comes to mind is scientists in labs developing the next breakthrough therapy. Treatment advances like immunotherapy, CAR-T cells, and precision medicine get the headlines and the funding, and for good reason. My mum is on TCHP, a regimen that includes Herceptin and Perjeta, which are targeted antibodies that attack HER2-overexpressing cells specifically. Twenty years ago, HER2+ breast cancer was among the deadliest subtypes, and now it’s among the most treatable because of exactly this kind of research.
But here’s what the survival data actually shows:
Localized cancer, caught early while still in one place, has an 80-99% survival rate. Regional cancer that has spread to nearby tissue drops to 20-25%. Metastatic cancer that has spread throughout the body lands in single digits.
Same disease, completely different outcomes based on when you find it.
Detection timing, by a massive margin.
Greg Simon ran Biden’s Cancer Moonshot and was in charge of coordinating the entire US cancer effort. He also has leukemia himself. When asked why we don’t prioritize early detection more, his answer was blunt: “Money. Lack of focus. And there’s not a lot of glory in creating an early detection program compared to developing a new drug.”
New treatments are exciting and screening programs are boring, so we keep pouring resources into making late-stage cancer slightly more survivable while people die of early-stage cancers that a basic blood test could have caught.
You’re probably not at high cancer risk yourself, but your parents are entering the age range where cancer becomes common. Half of all men and a third of all women will be diagnosed with some form of cancer in their lifetime, and most of those diagnoses happen after 50.
I realized I had no idea what tests my parents should be getting, how often, or where to get them done in India. The information exists but it’s scattered across medical guidelines, diagnostic chain websites, and oncologist recommendations that assume you already know the basics.
For your mother, the big three are breast cancer, cervical cancer, and colorectal cancer.
Breast cancer is the most common cancer in Indian women, and mammography every 1-2 years starting at age 40 is the standard recommendation. There are also newer AI-powered thermal imaging options that work particularly well for women with dense breast tissue where mammograms can miss things.
Cervical cancer is almost entirely preventable with screening, yet only 1.9% of eligible Indian women have ever been screened. A Pap smear or HPV test every 3 years from age 21-65 catches precancerous changes before they become cancer.
Colorectal cancer is rising in India and often ignored because the screening involves either a colonoscopy or talking about stool samples. The recommendation is colonoscopy every 10 years starting at 45-50, or an annual stool test which is simpler and can be done at home.
For your father, prostate cancer screening through PSA testing is worth discussing with a urologist starting at age 50, though it’s controversial because of false-positive rates. Oral cancer is where India has genuinely high rates due to tobacco and paan usage, and if your father uses tobacco or areca nut, annual visual oral examination catches precancerous lesions early. The Kerala Oral Cancer Screening Trial showed 81% mortality reduction through simple visual screening. If your father is or was a heavy smoker, annual low-dose CT for lung cancer is the only screening proven to reduce lung cancer deaths in smokers.
Most diagnostic chains offer bundled cancer screening packages that include tumor markers, basic imaging, and consultations. For parents who haven’t been screened in years, these packages cover the essentials without requiring you to coordinate individual tests.
Full-body MRI scans are available in India and screen for tumors across brain, spine, chest, abdomen, and pelvis in a single session. Worth knowing that MRI can’t evaluate lungs well and has high false-positive rates for incidental findings, so it’s an addition to standard screening rather than a replacement.
Multi-cancer blood tests are newer and genuinely interesting. Some Indian companies now offer tests that screen for 30+ cancer types from a single blood draw, with clinical validation showing around 90% sensitivity. These complement standard screening rather than replacing it, but if you want comprehensive coverage, they’re worth looking into.
This is where screening recommendations shift significantly.
The general rule is to start screening 5-10 years before the earliest age of diagnosis in your family. If your mother was diagnosed with breast cancer at 45, you should start mammograms at 35-40 rather than waiting until 50.
Genetic testing becomes relevant when patterns suggest hereditary mutations. Consider it if a parent was diagnosed with breast or ovarian cancer before 50, if multiple family members have the same cancer type, if there’s a family history of multiple different cancers, or if male breast cancer appears anywhere in the family.
One finding that surprised me: 30% of Indian breast and ovarian cancer patients carry hereditary mutations, compared to about 12% in Western populations. If there’s family history, testing is more likely to find something actionable here than the international statistics might suggest. Tata Memorial Hospital offers free genetic testing for eligible patients, and several commercial labs offer BRCA testing with genetic counseling included.
Since my mum has HER2+ breast cancer, I looked into this specifically. HER2 status is not hereditary because it’s a somatic mutation that happens after conception rather than something present in the germline that gets passed down.
However, there’s a documented association between HER2+ breast cancer and TP53 mutations (Li-Fraumeni syndrome), particularly when diagnosis happens before age 41. If there’s a pattern of multiple cancers at young ages in your family, or rare cancers like childhood sarcomas, TP53 testing is worth discussing with a genetic counselor.
For most people whose parent has HER2+ breast cancer without other red flags, standard population screening guidelines apply.
For your parents over 50, the starting point is booking a comprehensive screening package if they haven’t been screened recently. Make sure mammography is happening for your mother every 1-2 years, get colorectal screening started, and if there’s tobacco use, add oral cancer screening to the list.
For yourself, start by documenting your family’s cancer history including cancer type, age of diagnosis, and every blood relative you can gather information on. If that history shows patterns like multiple cases, young ages, or the same cancer types recurring, pursue genetic counseling. And start age-appropriate baseline screening, which means Pap smears from 21 for women and self-exams monthly.
The conversation to have with your parents is simple: ask when their last cancer screening was. Most won’t have a clear answer. Book it for them.
Early detection has better survival statistics than any drug we’ve developed, and the constraint isn’t technology. It’s getting people to actually use what already exists.
]]>I spent six months watching two founders build nearly identical products in the same market with the same resources, and one of them collapsed while the other became unrecognizable in their growth. The difference had nothing to do with intelligence, funding, or market timing. The founder who failed got a small rush of satisfaction every time someone told him his idea was brilliant. The founder who succeeded got that same rush when someone told him his idea was broken and explained exactly why. They were optimizing for completely different rewards, and those rewards were quietly determining every decision they made.
Your brain releases dopamine not when you experience pleasure, but when you anticipate it. This matters more than almost anything else about human psychology because wherever you have trained your brain to anticipate reward becomes the invisible gravitational center of your entire behavioral universe. You will, without conscious awareness, orient your days around the dopamine sources you have cultivated, and those sources will compound into either capability or dependency over the years. The question of where you get your dopamine is not philosophical abstraction; it is the most practical predictor of what your life will look like in ten years.
Consider the person whose dopamine fires when they receive validation for their ideas. They will learn to present ideas only when those ideas are polished enough to receive praise, which means they will stop sharing rough thoughts that could benefit from early feedback. They will unconsciously filter their social environment toward people who agree with them because agreement feels good and disagreement feels like pain. They will experience criticism as identity threat rather than useful signal, and they will optimize their work for being right rather than for getting it right. These are not character flaws; they are the logical behavioral consequences of a reward system that was trained to fire on external approval.
Now consider the person whose dopamine fires when their ideas improve, regardless of who caused the improvement or whether it required admitting they were wrong. They will seek out disagreement the way a miner seeks ore because every substantive objection is free labor toward a better outcome. They will share incomplete thoughts eagerly because early-stage feedback has the highest leverage. They will become genuinely indifferent to the question of who was originally right because that question is simply not connected to their reward system. Same raw intelligence, same opportunities, completely different trajectories.
Some dopamine sources leave behind residue that makes your life better. The hit comes and then you have something to show for it. Making things happen is perhaps the purest form of this because each successful act of agency trains you toward more agency, and agency compounds. You learn that you can move reality, and this knowledge makes you more likely to attempt moving reality again. The chess player who gets their dopamine from understanding why they lost will improve faster than the player who gets their dopamine from winning because the former can extract value from every game while the latter can only extract value from half of them.
Physical capability works the same way. The runner who gets their hit from the process of building endurance accumulates cardiovascular health, metabolic flexibility, and the kind of discipline that transfers to other domains. The gym becomes a place where effort converts directly into measurable progress, and this trains a mental model that effort leads to progress generally. Creating from nothing operates similarly because every completed project leaves behind both the artifact itself and the expanded capability that made the artifact possible.
Understanding something deeply is another compounding source, and it differs from collecting information in the same way that building muscle differs from buying gym equipment. The moment when a confusing domain suddenly has structure, when you can see how the pieces fit together, leaves you permanently more capable in that domain and in adjacent domains. This kind of understanding cannot be faked because it requires the actual work of confusion and resolution, and that work is precisely what makes it compound.
Other dopamine sources borrow from your future wellbeing. Validation is perhaps the most insidious because it feels exactly like useful feedback while training you toward performance instead of growth. The like count, the praise from respected people, the nod of approval when you share something clever all register as reward, but they create dependency on external sources that you cannot control and they orient your work toward what will be validated rather than what is true or useful. You can build an entire career on validation-seeking and appear successful from the outside while becoming increasingly fragile and increasingly unable to take the kinds of risks that matter.
Being right is different from understanding in a way that matters enormously. You can be right about something and learn nothing from the experience because the reward already fired when you received confirmation and there was no remaining incentive to examine why you were right or what you might still be missing. Arguments become particularly dangerous territory because the dopamine hit from winning an argument is disconnected from the hit of actually learning something, and most people cannot tell the difference from the inside. They feel good after winning an argument and interpret that feeling as evidence of growth when in fact they have just practiced rhetorical dominance.
Consumption without creation is another extractive pattern because it provides novelty, stimulation, and even the feeling of learning without leaving any residue behind. The person who watches five hours of YouTube videos about woodworking feels like they have engaged with woodworking, but they have built nothing and developed no capability. The information passes through without transforming anything, and the time is simply gone. Outrage has been industrialized by social media platforms that discovered righteous anger is among the most reliable dopamine triggers, and they have built systems that deliver it in optimized doses throughout the day. The person who checks Twitter for outrage develops a tolerance that requires increasingly extreme content while their baseline capacity for nuanced engagement with complex topics atrophies.
Evolution designed certain reward pathways to solve specific survival problems, and modernity has created superstimulus versions of the original triggers that exploit these pathways without providing any of the original benefits. Food engineering has produced combinations of salt, fat, and sugar that trigger reward responses far beyond what any natural food could provide, and these responses are untethered from actual hunger. The person reaching for snacks at 10 PM is almost never hungry; they are bored, anxious, or avoiding something uncomfortable, and the snack is functioning as a micro-escape rather than as nutrition. This matters because every use of food as mood regulation weakens the capacity for internal mood regulation and strengthens the association between discomfort and reaching for something external.
Pornography operates on similar principles but with higher stakes because it involves the reproductive reward system, which evolution made particularly powerful. The pattern of infinite novelty with zero effort and escalating intensity trains the reward system toward passive consumption while atrophying the capacity for real intimacy, which requires effort, vulnerability, and the acceptance of someone who will not provide infinite novelty or frictionless access. The deeper pattern is that pornography often has nothing to do with arousal and everything to do with numbing, avoiding, or the ritualized escape from whatever the person does not want to face. This makes it structurally identical to other avoidance behaviors while being more difficult to discuss openly.
Social media scrolling exploits the ancestral need for social information because knowing the status hierarchies and alliance structures of your group was genuinely survival-critical for most of human history. The platforms have built infinite scroll because stopping the feed would reduce engagement, and they have optimized for content that triggers strong emotional responses because strong emotions drive more engagement. The result is a simulacrum of social connection that feels like connection while providing none of the benefits of actual connection, and the person who meets their social needs through scrolling will underinvest in real relationships while not noticing the substitution.
Every extractive dopamine source shares certain structural features that become recognizable once you know what to look for. They all involve low or zero effort for high reward, which is the signature of an exploit because your reward system did not evolve for a world where reward comes without effort. They all involve escalation over time as tolerance builds and the same dose stops working, requiring either more intensity or more frequency to achieve the same effect. They all have invisible opportunity costs because the snacking session or the scrolling hour or the pornography use does not feel like it cost anything, but it cost the time, the energy, the momentum, and often the self-image that would have been available for something else.
Most importantly, they all function as mood regulation substitutes that prevent the development of internal capacity. The person who regulates anxiety by scrolling never develops the ability to sit with anxiety and process it, and anxiety therefore remains just as threatening as it always was while the behavior required to escape it becomes more entrenched. The person who regulates boredom by snacking never develops the ability to use boredom as a signal that something in their life needs attention, and boredom therefore remains something to flee rather than something to listen to. The person who regulates loneliness through pornography or parasocial relationships never develops the vulnerability and social skills required for real intimacy, and loneliness therefore remains a chronic condition that can only be temporarily numbed.
You cannot simply remove an extractive dopamine source because the void demands filling and willpower is a finite resource that depletes under sustained resistance. The question is always what will replace the source you are trying to eliminate, and the answer has to be something that either scratches a similar itch through compounding means or addresses the underlying void that the behavior was masking. Bored scrolling can become bored walking or bored building because these provide novelty and stimulation while leaving residue behind, but this only works if the boredom was the actual driver rather than a symptom of something deeper. Anxious snacking can become anxious exercise or anxious conversation because these address the physiological activation of anxiety while providing genuine benefits, but this too only works if anxiety was the actual driver.
Lonely pornography is the hardest case because the replacement requires building actual connection capacity, which is slower and less reliable than any substitute and requires the kind of vulnerability that the person has likely been avoiding. There is no clean answer here, only the recognition that the longer the avoidance continues, the more atrophied the connection capacity becomes and the harder the eventual rebuilding will be. The person who recognizes this pattern at twenty has dramatically better odds than the person who recognizes it at forty, not because of any moral difference but because of the simple mathematics of habit formation and neuroplasticity.
The honest questions are not complicated, but they require a willingness to look that many people do not have. What do you reach for when you are bored, and what would it mean if that answer changed? What do you reach for when you are anxious, and what does that behavior do to your underlying anxiety tolerance? What do you reach for when you are lonely, and is that thing building or substituting for real connection? What would you be embarrassed to have someone see the complete logs of, and what does that embarrassment reveal about your own assessment of the behavior? What do you consistently tell yourself you will moderate, and what does your failure to moderate reveal about the strength of the pull?
The deepest version of this inquiry is simply: what do you do when no one is watching and no metric will be affected? When there is no external reward structure, no audience, no accountability, what do you gravitate toward? That answer reveals your true dopamine sources with uncomfortable clarity because everything else is performance and self-narrative. The people who build things that matter tend to have an unusual relationship with this question because the work itself generates the reward and the external success is almost incidental, a byproduct rather than a target.
The twenties are the critical window because this is when dopamine patterns calcify into the infrastructure of a life. The neuroplasticity that makes habit change relatively easy at twenty-two makes habit change increasingly difficult at thirty-two and genuinely hard at forty-two. The person who spends their twenties regulating mood through consumption, receiving sexual reward from screens, receiving social reward from metrics, and receiving food reward from engineered snacks will enter their thirties with atrophied capacity for delayed gratification, for real intimacy, for genuine connection, and for the kind of sustained discomfort that meaningful work requires. This is not moral failing; it is the predictable consequence of training a reward system in certain ways over sufficient time.
The inverse is equally true and equally important. The person who wires their reward system toward agency, toward improvement over validation, toward creation over consumption, toward embracing difficulty as the price of growth will enter their thirties with tailwinds that continue compounding. Each year of good dopamine training makes the next year easier because the capacity for delayed gratification expands, the tolerance for productive discomfort increases, and the behaviors that generate compounding returns become more automatic. The choice is not really a choice at all once you see the structure clearly; it is simply a question of whether you will make the investment while the investment is still cheap or whether you will pay the much higher price of rewiring later.
The question of where you get your dopamine is the question of what kind of person you are becoming, and you are always becoming something whether you choose it consciously or not. The honest audit is uncomfortable because it reveals the gap between who you think you are and what your actual behavior patterns suggest, but that gap is precisely where the leverage lives. The person who can look clearly at their own dopamine sources and make deliberate choices about which sources to cultivate has discovered one of the few genuine cheat codes for a good life, and the person who cannot look will continue to be shaped by forces they do not see and cannot therefore influence.
]]>Neural networks spend most of their time multiplying matrices. Smaller number formats (FP16, bfloat16, FP8) mean less memory, less bandwidth, and faster math units. The challenge: gradients can be 10^6x smaller than weights, so narrow formats cause underflow. Loss scaling fixes this by multiplying gradients before they shrink to zero. bfloat16 works without scaling because it keeps FP32’s range. FP8 needs per tensor scaling and two formats (E4M3 for weights, E5M2 for gradients). For inference, INT8 and FP4 quantization map trained weights to smaller formats using calibration data to find optimal scale factors. The trend: lower precision with smarter scaling. Neural networks need dynamic range more than precision.
I spent years training models without thinking about how numbers are stored. FP32, FP16, bfloat16, FP8. These felt like implementation details that PyTorch or TensorFlow handled for me. Then I started working on inference optimization and realized: the way we represent numbers is often the actual bottleneck. Not the algorithm. Not the model architecture. The numbers themselves.
A neural network is just matrix multiplications, additions, and nonlinearities applied billions of times. Every single one of those operations uses numbers. If your number format is 32 bits, you need 32 bits of memory per value. You need 32 bits of bandwidth to move it. You need 32 bit math units to compute with it.
Cut that to 16 bits and you halve memory. Halve bandwidth. And modern GPUs have dedicated 16 bit math units (Tensor Cores) that run 2x faster than their 32 bit counterparts. Cut to 8 bits and you halve again.
The catch: smaller numbers mean less precision and less range. Use too small a format and your gradients disappear. Your activations overflow. Training diverges.
The history of deep learning numerics is the history of figuring out how small we can go without breaking things.
A 32 bit float (FP32) has three parts:
The value equals: (−1)^sign × mantissa × 2^exponent
More exponent bits give you wider range. More mantissa bits give you finer precision. With only so many bits to work with, every format makes a tradeoff.
FP32 has enormous range (roughly 10^−38 to 10^38) and high precision. Overkill for most neural network operations. The question became: where can we cut?
IEEE FP16 has 5 exponent bits and 10 mantissa bits. Its range spans roughly 6×10^−5 to 65,504. That range is too narrow.
Here’s the problem: neural network gradients vary wildly in magnitude. Early in training, loss values might be in the hundreds. Gradients flowing back through dozens of layers can shrink by factors of 10^−6 or more. Values smaller than 6×10^−5 become zero in FP16. This is underflow.
When gradients underflow to zero, weights stop updating. Training stalls. The network learns nothing.
Early attempts at FP16 training failed for exactly this reason. Gradients vanished.
The fix is elegant. Before backpropagation, multiply the loss by a large number (say 1024 or 65536). This scales up all the gradients by that same factor. Small gradients that would have underflowed now stay in the representable range.
After backpropagation, divide the gradients by the same scale factor before updating weights. The weight update sees the correct gradient values.
This is loss scaling. The 2017 Mixed Precision Training paper by Micikevicius et al. showed this works across CNNs, RNNs, GANs, and language models.
But picking the right scale factor is tricky. Too small and gradients still underflow. Too large and gradients overflow (become infinity). The solution: dynamic loss scaling.
Dynamic loss scaling starts with a large scale (like 2^24). Every iteration, it checks for infinities or NaNs in the gradients. If found, it skips the weight update and reduces the scale. If not found, it periodically tries increasing the scale. The algorithm automatically finds the largest usable scale.
There’s another problem with FP16 training. Weight updates are tiny. A typical learning rate is 0.001 or smaller. Multiply that by a gradient and you get a very small number to add to each weight.
If your weight is 1.0 and your update is 0.00001, adding them in FP16 might give you 1.0 again. The update is too small to change the FP16 representation. This is the weight update problem.
The solution: keep a master copy of weights in FP32. Every iteration:
The FP32 master weights can accumulate tiny updates over many iterations. The FP16 copy is just for fast computation.
Google introduced bfloat16 for TPUs. It has 8 exponent bits (same as FP32) and 7 mantissa bits.
The key insight: range matters more than precision for neural networks. With 8 exponent bits, bfloat16 has the same range as FP32 (roughly 10^−38 to 10^38). Gradients that would underflow in FP16 stay representable in bfloat16.
The tradeoff is precision. 7 mantissa bits versus FP16’s 10 means coarser values. But neural networks are remarkably tolerant of noise. The slight precision loss rarely hurts accuracy.
bfloat16 often works without loss scaling. The range is wide enough that most gradients stay representable. This simplifies the training loop.
FP8 cuts to 8 bits total. NVIDIA, ARM, and Intel jointly proposed two FP8 formats in 2022:
E4M3: 4 exponent bits, 3 mantissa bits. Range up to ±448.
E5M2: 5 exponent bits, 2 mantissa bits. Range up to ±57,344.
Why two formats? Different tensors have different needs.
Weights and activations in the forward pass benefit from precision. Their values cluster in a relatively narrow range. E4M3’s extra mantissa bit helps here.
Gradients in the backward pass have extreme dynamic range. Some gradients are 10^6 times larger than others. E5M2’s wider range prevents underflow.
The recommended approach: use E4M3 for weights and activations, E5M2 for gradients. Some networks can use E4M3 everywhere. Some require both formats.
FP8’s narrow range means most tensors need scaling. The general approach:
This per tensor scaling is more involved than FP16’s single loss scale. Each tensor might need its own scale factor. The scaling factors themselves are stored in higher precision (typically FP32).
Modern hardware and libraries like NVIDIA’s Transformer Engine handle this automatically. But understanding it matters for debugging and optimization.
INT8 (8 bit integers) has been popular for inference quantization. Why not use it for training?
Integers have no exponent. All 8 bits go to the mantissa (plus sign). This gives high precision within a fixed range but zero flexibility in that range.
Neural network gradients vary by factors of 10^6 or more. INT8 cannot represent this variation with a single scale factor. You would need per layer or even per channel scaling, dramatically increasing complexity.
Floating point formats handle dynamic range naturally. Each value carries its own scale in the exponent bits. This is why FP8 works for training while INT8 remains primarily an inference format.
NVIDIA Tensor Cores accelerate matrix multiplications in reduced precision. To trigger them, your tensor dimensions must meet alignment requirements.
For Hopper architecture with FP8:
If your tensors don’t meet these requirements, the computation falls back to slower CUDA cores. A 768×768 matrix multiplication runs on Tensor Cores. A 765×765 might not.
This is why padding to nice dimensions often improves performance dramatically. The math is the same but the hardware utilization is completely different.
When multiplying two FP8 numbers, the intermediate results are computed in higher precision. On Hopper GPUs, FP8 matrix multiplications accumulate into FP32.
This matters because matrix multiplication involves many additions. Adding two FP8 numbers would lose precision rapidly. By accumulating in FP32, we maintain accuracy even while inputs and outputs are FP8.
The outputs are then cast back to FP8 (or FP16 or bfloat16) for storage. This pattern (low precision for compute and storage, high precision for accumulation) is central to mixed precision training.
On NVIDIA H100 GPUs:
Memory bandwidth follows similar ratios. Moving FP8 tensors requires half the bandwidth of FP16, quarter the bandwidth of FP32.
For memory bound operations (which dominate inference), precision reduction translates almost directly to speedup. For compute bound operations (like large matrix multiplications in training), the Tensor Core throughput increase is the limiting factor.
Microscaling (MX) formats take scaling further. Instead of one scale per tensor, MXFP8 uses one scale per block of 32 values.
Each block of 32 FP8 values shares a single FP8 E8M0 scale factor. This allows finer grained adaptation to local value distributions within a tensor.
The benefit: MXFP8 can use E4M3 everywhere, including for gradients. The per block scaling captures dynamic range that would otherwise require E5M2.
The cost: more scale factors to store and manage. But the overhead is modest (one 8 bit scale per 32 values) and the hardware support on Blackwell architecture makes it efficient.
In PyTorch with AMP (Automatic Mixed Precision):
from torch.cuda.amp import autocast, GradScaler
scaler = GradScaler()
for input, target in data:
optimizer.zero_grad()
with autocast(device_type='cuda', dtype=torch.float16):
output = model(input)
loss = loss_fn(output, target)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
autocast automatically runs operations in FP16 where safe and FP32 where necessary. GradScaler handles dynamic loss scaling. The master weights in FP32 are managed automatically by the optimizer.
For FP8 training, NVIDIA’s Transformer Engine provides similar abstractions:
import transformer_engine.pytorch as te
from transformer_engine.common.recipe import DelayedScaling
fp8_recipe = DelayedScaling(fp8_format=Format.HYBRID)
with te.fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
output = model(input)
Some operations are numerically sensitive:
AMP and Transformer Engine maintain lists of which operations run in which precision. You can customize these lists if your model has unusual numerical requirements.
The trend is clear: lower precision with smarter scaling. FP4 (4 bit floating point) is already appearing for inference. NVIDIA’s Blackwell architecture supports NVFP4 with per block scaling.
For training, the constraint is gradient representation. As scaling techniques improve (microscaling, per layer adaptive scaling), we may see FP4 training become viable.
The fundamental insight remains: neural networks need dynamic range more than precision. Every generation of hardware and algorithms exploits this more aggressively.
Understanding numerics lets you make informed tradeoffs. When your training diverges, you’ll know to check for gradient underflow. When your inference is slow, you’ll know which precision reductions are safe. The numbers matter.
Training happens in FP32 or FP16. Inference deployment often needs INT8 or FP4. The conversion process is called quantization. There are two main approaches.
Post Training Quantization (PTQ)
PTQ converts a trained model without retraining. The process:
For INT8, you need a scale factor and zero point per tensor (or per channel). The scale maps your observed range to the 256 integer values. Values outside the range get clipped.
The calibration dataset matters. It should represent real inference inputs. If calibration sees values from 0 to 100 but inference sees 0 to 200, half your values clip to the maximum and accuracy collapses.
Percentile calibration helps. Instead of using the absolute min/max, use the 99.9th percentile. This clips rare outliers but preserves the bulk of the distribution.
Quantization Aware Training (QAT)
Some models resist PTQ. Accuracy drops too much. QAT fixes this by simulating quantization during training.
The forward pass inserts fake quantization operations. Weights are quantized and dequantized before each layer. Activations get the same treatment. The network sees quantized values during training.
The backward pass uses the straight through estimator. Gradients flow through the fake quantization as if it were an identity function. The network learns weights that are robust to quantization noise.
QAT typically recovers accuracy that PTQ loses. The cost is retraining (or fine tuning) the model.
Why FP8 simplifies inference
If you train in FP8, inference is straightforward. The model already uses FP8 weights and activations. No conversion needed. No calibration needed. Deploy directly.
This is a major advantage of FP8 training over FP32/FP16 training with INT8 inference. The formats match. No accuracy surprises at deployment.
FP4 quantization specifics
FP4 has only 16 possible values (4 bits). Representing weights accurately requires block scaling. NVIDIA’s NVFP4 uses one FP8 scale factor per block of 16 values.
The process:
Per block scaling captures local variation. One block might have values near zero, another might have values near one. Each gets its own scale.
For activations, the scaling must happen at runtime. The calibration approach records typical activation ranges, then applies appropriate scales during inference.
When does quantization fail?
Some layers are sensitive:
The fix: mixed precision quantization. Keep sensitive layers in FP16. Quantize the rest to INT8 or FP4. Tools like TensorRT and vLLM do this automatically based on sensitivity analysis.
]]>