This month I was unusual.

I ginned up the courage to sign up for a writers’ workshop. Now once a week I’m hopping on a train—to Midtown—to sit in a dank room with a grey table and honest strangers.

I was never much of a fiction writer, and the last time I tried was twenty years ago. If you took my kids’ ages and doubled it, you still would not get twenty. (What have I been doing in all that time?)

The first week, the teacher (I guess we can call him teacher, maybe instructor) asked me to read a story I had just spent only twenty minutes writing.

If you’ve never been to a writer’s workshop, it goes like this: you read a story by someone incredibly accomplished that knocks you over by the way it simply works. To me, a good story is like a clock. An assembly of intricate components into a single object. A great story is seamless, so perfect that its details can be easily overlooked. It’s solid in your hand. You feel the tick-tick-tick.

And then you’re expected to write something quickly and hold it up to the class. You’ve been coloring outside the lines with crayons in the middle of the Louvre.

The story I had to read out loud had a prompt, crowdsourced by the class. Typically, the writer-teacher-instructor will use some conceit to have the class generate a bunch of words, then you use some of them to craft a story.

In this case, we had to build a dating profile for an imaginary man.

What does a man created by a room full of new york moonlighting writers look like? Say hi to Corey. He’s a short king from Boston, an architect, very religious, loves fishing and birding, and his favorite drink is a non-alcoholic beer.

Then writer-teacher-instructor told us we had twenty minutes to write. Go.

If you’re curious if a workshop is for you, I say do it. You won’t know what kind of high art you’re capable of unless you try. All it takes is twenty minutes.

What I’ve been reading.

Happy National Poetry Month. May I recommend a few?

• Loud Music by Stephen Dobyns (ask me if I have a favorite poem)

• How to Triumph Like a Girl by Ada Limón

• Mathematics by Jane Hirshfield

Thanks to the workshop, I’ve been reading more short stories.

• This story from Miranda July about meeting a famous person on an airplane was marvelous. Absurd and clever:

I tossed handfuls of water toward my armpits and they landed on my skirt. It was made from the kind of fabric that turns much darker when it is wet. This was a real situation I had got myself into. I acted quickly, taking off my skirt and soaking the whole thing in the sink, then wringing it out and putting it back on. I smoothed it out with my hands. There. It was all a shade darker now.

• And please read this 3-page story about a school where things can’t stop dying. It had me cracking up and getting weird looks on the subway. No spoilers.

• Now I’m working my way through the sublime Brokeback Mountain.

What I’ve been listening to.

The queen is back. I recommend Robyn’s new album, especially if you’re near or over the age of 40, or spiritually 40 (frequently shutting your eyes in the shower or in the driver’s seat and imagining you’re back in the clurb).

See you in May!
gaurav

In 1907, a horse amazed Germans by doing math.

Its cunning was obvious. When its handler asked “what is two plus two,” the horse tapped its hoof four times.

It math’d, by all accounts, with grace and dignity, to the great delight of its captured audience. That is, until a psychologist named Oskar had to ruin all their fun.

Oskar noticed that, once the horse tapped the right number of times, its handler relaxed his face. Eyebrows unfurrowed, lips unpursed. Every time: the tapping starts, the number is reached, the handler relaxes, the horse stops.

There ends the comedy of Clever Hans the mathematical horse, felled by the cleverer Oskar.

The lesson is this: Clever Hans couldn’t do math because horses can’t do math. So every time you think you see a horse doing math, or anything else unhorselike, simply ask, “Is this a horse?”

This is not an essay about a horse. It’s an essay about language models, how they were designed, how to understand them, and how not to misunderstand them. I avoid using math because, like horses, you don’t need it. I promise that, by the end, you will feel calmer and cleverer.

How we taught machines to read and write

In 2019, I quit my job to co-found a language model startup. We built our first model later that year. It could understand a patient’s symptoms and give a ranked list of potential diagnoses. Two engineers used it to discuss pressing medical issues with their doctors. Blindly, the doctors agreed with the model’s top suggestions.

I know how to teach a machine to read and write, and it doesn’t involve thought.

So what does it involve?

A map of language

If we’re lucky, as kids, our parents read to us. Through listening and repetition, we pick up on the rules of language. It’s not a perfect process (when my two-year-old is speedracing around the house, he yells, “I’m fasting!”) but that’s part of the process. Mastery of language is earned through tangible experience.

A machine, however, starts with nothing. They don’t get parents or bedtime stories. They don’t get to see someone pointing at a furry thing with four legs while saying “dog.” We transform abstract shapes on a page into concrete meaning. They can’t see words because they don’t have eyes.

Where machines excel is math. If we want them to read, we have to represent language as numbers first.

Picture a magic portal that transports you to a dimension where each word is a point on a map. Each point has coordinates, a set of numbers unique to it. The map isn’t random. Similar words cluster into neighborhoods. Country names all have their own club. That’s where you’ll find Turkey. And the club sandwich? Over by the turkey club.

To create this map, we count how often words appear together in a large body of text. If two words appear together often, they probably have a connection. The more they appear together, the more they’re related, the closer they should be on our map.

Machines don’t know what words mean yet, but now they can do math with their coordinates. They can “read” an email and flag it as spam. A search engine can use them to surface relevant results. They can even translate between languages. But they can’t talk yet.

Context reshapes the landscape

There’s one problem with our map. Words don’t have fixed coordinates. Their relationships shapeshift with context.

Consider the sentence:

The chicken crossed the road because it was hungry.

You and I understand that it refers to the chicken, because roads don’t get hungry.

And if I say:

The chicken crossed the road because it was clear.

Now, it must refer to the road.

If we want our map of language to be functional, we need it to consider context. So researchers developed models that could update a word’s position on the map. When it sits near hungry, the model pulls it toward chicken. When it sits near clear, it drifts toward road.

The landscape rearranges itself with every sentence. The old map was static. The new map is alive.

A compass to guide them

When we (I assume you’re also a human) learn to read, our brains don’t learn a static set of rules. We learn to navigate language.

That’s why typos and scrambled letters don’t throw us. Consider this:

Aoccdrnig to rscheearch, it deosn’t mttaer in waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht the frist and lsat ltteer be at the rghit pclae.

You might have seen this trick before. It’s called typoglycemia: you can read the sentence because the first and last letters act as anchors. Your brain reorganizes the rest. It knows how to focus on what matters.

Machines had to learn how to navigate language too.

Consider our sentence again:

The chicken crossed the road because it was ____.

Now the machine must guess what comes next. If the model focuses on chicken, it may output “hungry.” If it focuses on road, it may output “empty.”

The ability to focus on what matters is called attention. It’s a simple idea that changed the field overnight. New language models called Transformers used attention to read and write better and faster than older models.

Like a sailor steering by Polaris’ light, these models navigate language by focusing on what matters.

Following directions

Now machines had a sense of direction, but they were bad at taking directions. Researchers had trained them to finish incomplete sentences, which taught them syntax, grammar, and a surprising number of basic facts. What they didn’t learn was what to do with that knowledge.

They were a room full of monkeys banging away at typewriters. With enough time, they might produce Shakespeare—but when? And how much would it cost in bananas and cigarettes?

To be useful, models needed to take directions. But how?

Two schools of thought:

1. Train a separate model for each task. One for translation, one for summarization, one for answering questions, etc.

2. Train one model to follow multiple commands. Feed it inputs like “translate this from English to German: the chicken is hungry,” and teach it to produce “Das Huhn ist hungrig.”

School two prevailed. We learned how to train a single model that could multi-task: translate, summarize, answer questions, and more.

Still, it was unwieldy. Commands had to be explicit. Outputs were coherent but often wrong. The machines had a map of language, a compass, and a printout from MapQuest.

GPS

By 2021, OpenAI had a problem.

If you asked a person and a model each to summarize an article, readers preferred the person’s summary. The models needed to learn how to be appealing.

In hindsight, the answer was obvious: build a second model trained to predict which response a reader would prefer. Then use that to teach our first model how to write better.

Imagine I showed you pairs of things and asked you which thing you liked better: apples or oranges, pie or cake, a minute of dialogue with Andrew Tate or a two by four to the head? Eventually, I’d learn your tastes. I could then use that knowledge to teach a model to write the way you like.

It worked. Readers began preferring the preference-trained model’s writing.

This technique allowed researchers to create small preference-trained models that wrote better than larger, untrained ones. And when they trained larger models on preferences, they observed a massive leap in writing quality and capability.

This new preference-driven large language model was eventually released as ChatGPT. Some co-inventors had left to form Anthropic, and created Claude.

Recalculating

In only a few years, machines learned to ace the Turing test, support medical diagnoses, and pass the bar exam.

Still, they had flaws. They hallucinated often. Worse, they didn’t know how many r’s were in “strawberry,” a question that has plagued mankind since we descended from the trees.

The problem was that the models immediately blurted out a response. So AI researchers asked: what if models could “think” before they spoke?

But models don’t have an inner voice. They take things in and say things out loud. So researchers exploited this property. They instructed models to outwardly generate an inner dialogue. The models reflected on the input, turned it over, considered how to answer, and then answered. This technique is called chain of thought, and it looks a lot like someone thinking out loud to themselves.

Chain of thought gave birth to reasoning models, which are much better at math, logic puzzles, and complex instructions.

We’ve come to 2026. Now machines not only have the map, they have a GPS that can recalculate and reroute in real time. No matter the start and no matter the end, they have a way to get there. The ride got so smooth, you start to forget it’s a machine.

The end of fluency as a signal for thought

When my daughter was 5, I had her speak to ChatGPT.

I turned on voice mode and let her chat with the glowing orb for a few minutes. She asked it to tell her a story, and it weaved a tale of princesses and unicorns.

I then asked my daughter, “Do you think this is real?”

She thought for a moment and replied, “She … sounds real.”

What followed was an age-appropriate conversation about how ChatGPT was a machine, made by people to sound like a real person (then we ate popcorn and watched dinosaur videos on YouTube).

For most of human history, fluency—the ability to speak clearly—was a sign of intelligence. A cogent sentence could be recognized as the creation of a mind.

Now that language is computable, fluency is an unreliable discriminator of rational thought.

Computation is the transformation of information from one form into another. It can be described as a series of operations.

Take for example Google Maps. When you type in a destination and ask for directions, it evaluates the edges on a graph and optimizes a route using traffic data. When you have voice mode on, it sounds like a helpful friend giving directions. But we know that what’s underneath is an algorithm with an objective (minimize travel time). We don’t confuse Google Maps for a sentient being. We can see the machinery.

A language model is the same kind of thing. Instead of a graph, it operates on a map of language. The path it charts connects your input to a response, using another algorithm with an objective (do what human says, make happy). It’s just better at obfuscating the ones and zeroes.

Thought is not computation, because it involves the thinking mind’s perspective.

A mind meets every input with everything it already holds. Memories, beliefs, context, the residue of every prior thought. All of these coalesce into a thought that can be expressed in words, felt as an emotion, externalized through actions.

This process is unique to every mind. The same sentence read by two minds culminates in two different thoughts. Each mind contributes itself.

Thought also reshapes the thinker. When I think about my kids, it’s not passive. I can levitate with joy or sink with worry. Their love is my weather. Often, at the end of the day, exhausted from parenting, my wife and I will … talk about our kids. We’ll exchange stories, show each other funny photos, reflect on highs and lows. This ritual is healing. It shapes how we parent.

The machine-sentence, in contrast, is shaped by an objective to maximize appeal. Its past is represented as a set of weights, its present as the vector of your input. Its memory is a markdown file. With every new chat, the machine is made spotless.

Language as code

People invented language to be an externalized representation of thought. If machine-writing isn’t driven by thought, does it qualify as language?

Let’s take a closer look at “chain of thought,” the process by which a model is taught to “reason” over a challenging prompt. What the model is doing, because it has no interiority, because the only way it’s reshaped is by processing a new input, is to translate your words. It expands the input to provide more context or direction for itself. In other words, it didn’t like the word-numbers you gave, therefore it needs better word-numbers. So it generates them in order to get to the final outcome: a response you like.

What we call “machine thought” is a program: given an input, apply a transformation, and then generate the necessary output. Reasoning models are, effectively, learning to write self-programs. This allows them to adapt to what a user presents, and attain their goal of maximizing your satisfaction. Every sentence the model gives you was shaped by the same process: optimize for what the reader will accept.

When I write, my goal is for you, the reader, to recognize my arguments and feel resonance. I am compressing my knowledge and argument into words, and hoping that you decompress them into true understanding. If I’ve done that, mission accomplished. If do it while engaging and entertaining you, gold star.

Good writing is hard because language is not the transformation of words into other words. It involves the distillation of history, context, thought, senses. The best stories and essays evoke an image in your mind, draw out your feelings, incite revelations. Where in the great corpus of online training data that robots scrape breathlessly is the humanity?

What these models generate has the shape of language. Humans often perceive it as language. And so we call them language models. Yet the highest forms of language have no definite objective. There’s no loss metric for a stirring poem. You cannot optimize your way to the great American novel. What gradient descent produces is a silhouette of language, not the thing itself.

If machine-thought and machine-language are not thought and are not language, what shall we call them? What are words chained together with a concrete objective?

Code.

A machine isn’t thinking or writing, it’s coding in English.

Interpreter error

If machines are coding in English, what are they doing to the minds that read that code?

Remember that a calculator is computing when it returns “4” after you type 2+2=. It will never produce 5, or 6, or any other number.

A mind, however, can stop believing that 2+2=4.

In George Orwell’s 1984, when people are asked to believe that 2+2=5, they must first dismantle their core belief system. They rearrange their thoughts, which rearranges their thought process. If done all at once, it’s recognized as coercion, an act of violence. Done piecemeal, it can be painless.

If you choose to believe that horses can do math, then that is your right. But know that a horse that stops tapping when you relax is simply trying to please you. And when you allow yourself to think horses can do math, what else will you believe?

It’s been a real pleasure to rediscover writing. Getting ideas out of my head, refining what I think, revealing what I don’t. Thanks for coming along for the ride.

I have a few essays I’m working on now, and each requires some more thought and research. But I want to keep writing, and I figured not every thought requires a full essay. So here is my first in a series of Field Notes, or short updates.

What I’ve been reading

Last fall, I started a book club with some of my neighbors. We just wrapped up Kindred by Octavia Butler, our selection for February. This is my third Butler book of the year, after reading both Parable of the Sower and Parable of the Talents in January.

I wish I had discovered her sooner.

There are parts of Butler’s craft that I’m not completely attuned to. She tended to tell more than show, especially in Kindred, an earlier work. But I like to read like I’m putting a puzzle together. Show me the negative space and then let me pop a piece in.

That said, she understood our past and drew conclusions about our future so keenly. Were she alive today, I sense she would’ve been disappointed but not at all surprised.

I also read Stefan Zweig’s biography of Montaigne, published in 1941. Zweig deliberately overdosed on barbiturates in 1942, fearing that the future had nothing to offer him but destruction. Yet he found so much beauty in Montaigne’s life, who had managed to survive one of the worst plagues in recorded history while dedicating his life to knowledge (if you haven’t read Montaigne, he pioneered the personal essay). If only Zweig had held on for another few years.

After that, I needed a little levity, so I’m reading Pity the Reader: On Writing with Style, which has Kurt Vonnegut’s lessons on writing. It’s mostly excerpts from his canon with insightful commentary from his friend and former Iowa Workshop student, Suzanne McConnell.

Vonnegut considered writing a spiritual practice. Here, McConnell and Vonnegut comment on the connection between reading short stories, Buddhist meditation, and our neurological responses to both:

I don’t know if Vonnegut ever read Butler, but we have hints on how he would feel about her:

I love Vonnegut. I remember the morning he died, it was a rainy one. I don’t remember getting out of bed.

What I’ve been listening to

Recently, I’ve been on a Joanna Newsom kick, and she sounds very different to my now-39-year-old ears.

The naïve read on Newsom is that she was part of the freak folk movement (I guess because she’s from Norcal), or that she was twee because her voice had a high register.

The truth is she is a remarkable poet with an incredible mastery of metaphor, meter, pace, and tone. She doesn’t call herself a harpist, she calls herself a harper. She uses the power of the instrument to reflect the emotion of the words underneath. Between the two, she builds remarkable tension, even outright violence, in songs like “Baby Birch” and “Sawdust & Diamonds,” the latter being one of the best songs ever composed, for my dollar. Then she invites you to sit in that tension.

“Sawdust & Diamonds” deserves its own essay, but first let’s dip a toe. Consider the lines:

Drop a bell off of the dock
Blot it out in the sea
Drowning mute as a rock;
and sounding mutiny

Meanwhile, Newsom deftly plucks arpeggios, creating the sound of bubbling, gurgling water, a feeling of rushing down deep into the dark.

The bell is drowning, drowning. And us with it. Until:

There’s a light in the wings, hits the system of strings
From the side, where they swing —
See the wires, the wires, the wires

The pace picks up, the register lifts, and we’re rising. It’s the same playbook Radiohead will use on the incredible “Weird Fishes/Arpeggi” off In Rainbows a year later.

It’s been eleven years without a new Joanna Newsom album, but she hasn’t stopped writing. I’ve read that her latest songs are focused on family. I’d be delighted to hear that version of her.

New to Newsom? Here’s a playlist on Apple Music (she’s not on Spotify).

What I’ve been paying attention to

My wonderful son is two and a half, and so naturally I’ve been waiting for him to fully gain consciousness.

I think it happened this week when he had a 101-degree fever. He was staring out into the middle distance, and when I tried to get his attention, he suddenly turned to me and said, “You are Daddy.”

He now describes everything I do. They half sound like narration and half like commandments: “You are here. You are drinking coffee.”

I’ll never get over how the human brain folds itself like origami, creating more and more intricate, more beautiful, more terrible shapes.

So it goes!

You may have seen the same videos pop up on your feed: young adults strolling around museums, rolling their eyes and making obscene hand gestures in the direction of modern art.

If the point of art were to faithfully reproduce what the eye can see, we’d still be staring at bowls of fruit. We’d line them up on tables, nice and still, in appreciation of their shadows, lines, and textures. We would get very good at watching how daylight catches a grape, the gentle bruising on the pear, the dimples on brownest banana. And then we would die with the divine knowledge that heaven is a fruit salad.

Conjure the familiar complaint: “I don’t understand what’s hard about this. My kid could paint that.” Their meaning: “If I cannot see the effort, the effort does not exist.”

I have no patience for this argument; it is wrong historically, wrong philosophically, and, as I hope to point out, wrong technically (the best kind of wrong).

Let’s begin with history. Classical art was not, despite the insistence of armchair art critics, primarily concerned with realism. I love Caravaggio; look how he weaponized light and shadow to make you feel the violence of a beheading before your brain can actually process his composition.

Consider El Greco, who stretched his figures into alien proportions because what he was evoking was a spiritual ecstasy that demanded superhuman shapes. We are not in the domain of man; we are seeking Godhood.

Let’s take something that many believe is the Platonic ideal of real: Michelangelo’s David. Look carefully at his hands and his face. That is not what a man looks like (as I insist regularly to my wife). You see, the David was designed to be stood up on a pillar many feet off the ground. Michelangelo had to consider what he wanted you and me, mere mortals on the soiled ground, when we looked upon his visage. And so David’s proportions are deliberately wrong; his hands are massive, his gaze bright and steady. We are staring upwards not at a mortal, but at Triumph.

When the camera was invented in the mid-19th century, the artist’s monopoly on visual reproduction was suddenly gone. Did art shut down? Did the painters put away their brushes, the sculptors toss aside their clay, and quietly fade into antiquity?

No: art exploded. Impressionism, Post-Impressionism, Cubism, Surrealism, Abstraction. Why?

Modern art is the answer to the question “What is art for if not accurate representation?” The answer turns out to be everything else: emotion, confrontation, tension, interiority, dissent. Too much for a single bowl of fruit to hold.

One of my favorite places in the world is in front of Jackson Pollock’s “One: Number 31, 1950” which takes up an entire wall on the fourth floor of the MoMA. I include the screenshot so you know what to look for when you go (do go), and not because it does it a hair of justice. To have any hope of appreciating Number 31, you must go there on a quiet morning and fall into it with your eyes.

If you know nothing about Pollock, all you may see is chaos. Proof that they’ll put anything in a gallery. And yet there’s some order underneath.

In 1999, the physicist Richard Taylor published a paper in Nature, that argued that Pollock’s drip paintings are actually similar to fractals.1

Forests are fractal; a tree’s trunk splits into limbs, limbs into branches, branches into twigs. The silhouette of a single tree mirrors the canopy pattern of the forest. In his research, Taylor showed that Pollock developed fractals in his art systematically. They emerged in the first few minutes of painting. Then Pollock spent months on layering and refinement. His later paintings, such as Blue Poles, have some of the highest fractal dimensions. They are also the ones observers tend to prefer.

Is preference measurable? Researchers went on to use tools like eye tracking, skin conductance, and EEG to investigate.2 What they found was that Pollock’s paintings induce a familiar physiological response in us: relaxation. We chill out. Other mid-range fractals have the same effect: trees, clouds, coastlines. They turn some invisible key inside of us. The feeling of coming home.

Pollock understood. He once famously said, “My concern is with the rhythms of nature.” When Time magazine ran a piece calling his work “chaotic,” he wired back: “NO CHAOS DAMN IT.” He refused to draw fruit; instead he drew their essence.3

When I’m standing in front of Number 31, I feel connected to something innate. I’m looking into a forest unlike any on Earth. If I reach out to touch it, it may touch back.

In Kurt Vonnegut’s Breakfast of Champions, the fine people of Midland City commission the painter Rabo Karabekian for a hefty $50,000. What he delivered was a massive, avocado-green rectangle adorned with a single fluorescent orange stripe.

They were, as you can imagine, pissed. To save his own hide, Karabekian explains the intention behind his masterwork in a grand speech—and makes converts of the mob:

“All you had to do was explain,” said Bonnie MacMahon. “I understand now.”

“I didn’t think there was anything to explain,” said Carlo Maritino, the builder, wonderingly. “But there was, by God.”

Abe Cohen, the jeweler, said to Karabekian, “If artists would explain more, people would like art more. You realize that?”

Pollock didn’t start off as an abstract artist. He spent years as a figurative painter. Only after that, after Benton, after the influence of Picasso, after Jungian analysis and Mexican muralism, did he have a long reckoning with his canvas. You can judge only what you see there, but that erases the biography. In other words, the process of art is as important as the product. And what modern art demands of us is to engage with that process. Channel our inner Bonnie MacMahon.

Duchamp’s “Fountain” is one of the most important artworks of the 20th century because it detonated the question of what beauty is. Provocation is its design. And so the young adults mocking art on camera are ultimately performing the intended provocation. Mockery is a tool of engagement. It proves the art works as intended.

Mockery is preferred to apathy, but it’s still a poor tool. It’s the one you reach for when vulnerability and introspection are either hard to wield or out of reach. It’s a comfort when you’re confronted with something you don’t understand, and the world you grew up in has taught you that not understanding is weakness.

I empathize with these kids. The safest posture is irony, and sometimes you need to feel safe. “It’s not that serious, bro” is a complete philosophy. It is defensive, and there’s a lot to want to defend yourself against. Such as a culture that treats confusion as failure rather than the beginning of something.

“My kid could paint that” is not a statement about art, it’s a sign about the observer. Maybe they were never offered context, or given permission to stand in a room and feel something without needing to perform. Their kid maybe could’ve painted that, but they didn’t. What will they make instead?

I want my own kids to feel comfortable sitting with the discomfort that precedes understanding. I want them to have permission to not know things, to feel things without having to name it. To let things be difficult without being hurtful. To be better at all these things than their dad.

I. Seven million years ago, in the dense forests of East Africa, a family of apes diverged.

One group drifted west, retreating deeper into the jungle. There they found empty niches, plenty of food. Stability. Their descendants became the chimpanzees.

The other group remained in the east. Over time, they saw their world gradually change. Thick woodlands, once abundant, were giving way to open savanna. These eastern apes, climbers by nature, were left with little choice. They descended. Their children stood to walk upright. By less than four million years ago, upright apes were scattered across eastern Africa.

Then, on the shores of Lake Turkana, an ape picked up a stone with intention.

We must imagine what happened next. The stone was heavy, volcanic. From a distance, you would see the outline of a person, arms raised.

They bring the stone down hard; flakes of volcanic rock shear, scatter through the air. Sunlight shines through the thinning canopy, revealing the gleaming edge hidden inside. Behold the Lomekwian: our first tool.

With this new tool, the apes cut meat off scavenged bones, adding crucial protein to their diets. This helped fuel brain growth, which allowed them to adapt and produce more sophisticated tools. Their descendants were able to break animal bones apart, accessing rich bone marrow. Eventually, they transformed from scavengers to hunters, and created new physical and social tools to adapt to their rapidly-changing environment. The cycle spun for millions of years, and here we are.

The ape looked at a rock and imagined what it could become, then made it. This is an act without instinct, one that requires imagination: the first creative act on record.1

This is a story about the moment before the stone comes down.

II. In 1968, Dr. Larry Weed published a paper in the New England Journal titled “Medical Records That Guide and Teach.”

The title was his thesis: the medical record should not be a container for information, it should be an engine for knowledge. Data about the patient should organize around problems, the problems structure the physician’s thinking, the thinking generates better data.

Weed was an inventor. He created the problem-oriented medical record, the SOAP note, the problem list. He understood that the medical record was not just a system of record, it was a system of work. And he believed that if you got the workflow right, the record would guide the clinician.

In the 1970s he built PROMIS, one of the first electronic health record systems. It did what the paper described: organized knowledge around problems, connected clinical evidence to patient data, fed conclusions back into the record.

The idea was elegant, but the execution was held back by its time. PROMIS needed its own dedicated terminals and mainframe. This made it inaccessible to all but the largest academic institutions. But an electronic record in isolation is just an expensive filing cabinet. The value required a network.

In 2008, medicine still ran on paper. Fewer than one in ten American hospitals had an electronic health record system. When the economy collapsed, Congress added $35 billion to a stimulus package with the intention to drag healthcare into the digital age. The big idea: pay hospitals to adopt a (government-certified) electronic system, penalize those who refused. By 2015, 95% of hospitals had an EHR.

This is not the market inspiring a crowd of competing solutions. This is a square peg being crammed into a very abstract hole. Was the technology finally ready for the needs of clinical medicine?

A few years ago, I called a physician friend to help me prepare for an interview centered on “things I would change about healthcare.” We’ve known each other since fourth grade. As teenagers, we took Hindu philosophy classes on the weekends and often got into excited arguments. I could trust him to have a thoughtful, empathetic take. So I asked him, what would you change about healthcare?

He told me that his experience in medicine had given him a “moral injury.” His motivation to be a doctor was altruistic, but he was increasingly demoralized by the financial incentives that shaped his every day and limited the care he could provide for his patients.

I met with another friend, a pediatrician trained in Boston. We were sitting on the stairs, watching our kids switch between playing blocks and throwing them at one another, when she confided that she made a mistake choosing medicine. She’d left her fellowship to start practicing in Roxbury. I also spoke to an emergency room physician who left medicine to teach, and a friend who dropped out of medical school for a job in tech.

In the interview, I said, “Doctors and nurses aren’t just experiencing burnout—how they frame the world is being completely rewritten.” I wanted the audience to understand: these are not people who starve for resilience. These are people who wanted to be doctors to heal and to think. They are the physicians that Weed wanted to help.

The federal incentive to adopt EHRs was an input mandate; the government paid for a tool instead of incentivizing ideal outcomes, like patient or population health. Weed’s architecture, a system for reasoning and iterative learning, became an assembly line.

Once in college, I was walking with a girl I was trying to impress in the East Village. She asked me if I had read any Foucault. “The pendulum guy?” I replied. Apparently not.

Years later, I thought of that moment, which had stuck with me (how could it not) when in a bookshop I came upon a copy of The Subject and Power. In it, I found a framework that helped me understand what went wrong. Foucault had described this kind of regulatory intervention decades prior: “To govern, in this sense, is to structure the possible field of action of others.”2

Foucault argued that modern power operates through the organization of what is possible in a system, not through force or coercion. Regulating the adoption of electronic systems was a show of power, not progress. Worse, by defining the process, they limited the field of action. Instead of imagining what we could have built, they assumed the tool was a proxy for progress.

In 2008, the dominant vendor saw an opportunity; it built a system of record so large that it became gravitational; 250 million lives maintained by a company that competes with others trying to serve those lives. Workflow that crosses the boundary a system of record requires its permission to do so. This limits competition: solutions that can’t talk to the EHR can’t threaten the EHR. Attempts at regulation fail.

Into this captured market, venture invested billions of dollars. It’s hard to build workflow, so you build point solutions instead. On paper, it fits the model: start small, MVP, build a base, build out. In practice, every inch you claim is a bareknuckle fight. To date, there have been over a thousand venture-funded digital health companies. CIOs field dozens of vendor pitches a month. A tool for every pain point, plumbing upon plumbing.

When systems fail, everyday people become the connective tissue that make it work. They must click and drag and drop and fax their way to what they need. Parts of the system are repurposed to assist them; the SOAP note became a billing instrument, the problem list became a checkbox.

The system then becomes a funhouse mirror of itself. A billing auditor tells a physician to add “pupils equally round and reactive to light” to a clinical note, not because it makes a lick of sense, but because it will allow for a higher billing code.

The people impacted most are our independent healthcare providers. A five-physician primary care group serving a rural community does not have a CIO, it has an office manager who also handles HR. They cannot negotiate API access or hire an integration engineer. Kaiser can absorb a thousand cloud applications; your neighborhood family doctor can manage three. People cover the gaps in the workflow. They’re routers, translators, administrators.

Physicians and nurses spend hours on menial tasks. They’re working on notes they couldn’t finish during the visit because they were wrestling with software. In the office, at home after the kids are asleep. Gaps in the workflow became an equity gap.

Into this mess enters the ambient scribe, promising something different.

First-generation scribes listen to physicians and draft their notes for them. But when you expand from there, something creative happens. Because documentation is integral to billing, the scribe can act as a lever for physicians to get paid, if it can get into the workflow. When the scribe feeds coding feeds billing feeds claims feeds revenue, you’re not just saving time, you help practices thrive.

What makes scribes different from solutions of old? The better they work, the less attention they demand of their users. And as AI scribes become increasingly agentic, capable of acting properly on behalf of care providers, the more they will reduce unnecessary work. This is the workflow serving the clinician.

In 1982, Weed founded another clinical software company. It was eventually acquired by Sharecare, a consumer wellness platform co-founded by Dr. Oz. In 2011, at age 87, Weed and his son published a book called Medicine in Denial. The lesson he’d learned over forty years was that “[m]isdiagnoses are not failures of individual physicians. They are failures of a non-system that imposes burdens too great for physicians to bear.” When asked whether he was optimistic, Weed said: “Based on what I know about all the vested interests in the present medical education system and in the present practice of medicine, I am not optimistic such changes will be forthcoming.” Weed died in 2017 at 93.

Recently, I visited my ENT, a British fella trained in America, for a surgical follow-up. He had a specialized camera in my sinus that let him find and remove debris in one maneuver. A TV in my sightline showed the camera’s feed in real-time.

While trying not to sneeze, I asked him how he liked his training. He liked it fine, he said. During fellowship, he’d had the chance to design and build his own medical device. The existing tools weren’t good enough.

“What device was that?” I asked. “It’s currently up your nose,” he smiled.

As he finished up, I asked him what his next project was. He’d recently bought a GPU server to train his own ambient scribe, one that met his quality bar. He had lots of ideas too, wanted to see what else he could plug it into. Weed would have recognized him immediately.

III. When you refuse to take a pill for something, you reach for a book.

My “something” was a deep sense that I had lost my purpose. I was stuck in a job I no longer enjoyed. The workaholic in me was angry; it could handle abuse if the mission was worth it, but it was increasingly, obviously, despairingly not.

An analyst formatting a report until 2am for a meeting that will be rescheduled in the morning. A researcher preparing a deck for a presentation few will show up to. A product manager updating a kanban board that no one will look at.

I have been all these things.

I did these jobs because I saw a gap between what mattered and what was seen and rewarded. And the longer I tried to cover the gap, the harder it was to remember that I was hired to think.

Rather than allowing myself to churn, I locked into research mode and found the late social anthropologist David Graeber. In the 2010s, he interviewed people from different industries and developed a theory about the nature of modern work, and how far it has drifted from purpose. He glibly called the kind of jobs I was doing bullshit jobs, and said they “regularly induce feelings of hopelessness, depression, and self-loathing.”

Graeber concluded that this kind of work isn’t a bug, it’s a feature. The friction of producing results fills time, and the filled time protects everyone from introspection (we have another, less academic term for this: busywork).

Despite the tone, Graeber seemed empathetic and deeply serious about the consequences of doing a job that you believe has no meaning. He wrote that it amounted to a “spiritual violence directed at the essence of what it means to be a human being.”

How did we get to the point that there were enough people who said, yeah I work a job like this, to merit serious study? Graeber himself wonders aloud, how did we as a people get stuck?3

By chance, I also discovered Jean Baudrillard (ok I was scrolling Instagram4), and picked up a copy of Simulacra and Simulation. In it, Baudrillard wrote: “It is no longer a question of imitation, nor of reduplication, nor even of parody. It is rather a question of substituting signs of the real for the real itself.” — It being existence.

I felt what he described: meetings to plan meetings, spreadsheets that exist so the spreadsheet can be updated. Work that refers only to other work.

Now I had a working theory of why I felt purposeless; Foucault named the cause, Baudrillard named the effect, Graeber named the consequence. The system we inhabit forces us into a job that becomes a pantomime. It satisfies your observers, but you know you’re living a simulation. You become the apparatus substituting signs of the real. You take a moral injury from the spiritual violence.

At this point, I had a diagnosis, but I didn’t have a name for what I was missing, or a prescription for what to do about it.

Subscribe now

IV. As a PhD student, I wanted to make learning faster.

I wanted to create better tools, algorithms, and analytical methods. This desire peaked in 2015 when I was enamored by the promise of machine learning. It was exactly the right moment for someone with basic Python and statistics knowhow to become a data scientist.

Plus I needed a clean break from benchwork. I no longer enjoyed the methodical, backbreaking process of collecting data (I used to do microsurgery on fruit flies, buy me a drink before you ask me about it). But as a data scientist, I would have perfect posture and, with the right spreadsheet and methods, predict anything.

From there, I joined a genomics startup at peak sequencing fever. The prevailing hypothesis: if researchers understood genetic mutations better, we would understand disease better and cure it faster. Our startup was building sophisticated tools for the step between sequencing biological tissue and understanding what its genetic code meant. The plumbing we provided was essential, we believed.

Stubbornly, I learned that few customers cared about the plumbing, they cared about the insights. Meticulous product details that we labored over were disregarded or dismissed.

Around the time I left, Google released DeepVariant, a mutation detector5 that collapsed layers of complex, pedagogical algorithms into a single model. Obscenely, it effectively did what clinical testing companies were quietly doing, having rooms of people look at gene sequencing data to spot the mutations. It did this using computer vision instead of interns.

I realized I had to be more discerning with the problems I chose.

My next job was a diagnostics company, much closer to the action. My team provided data and insights on cancer mutations and patient outcomes to companies developing cancer therapies. One morning, I met with colleagues at the campus of one of our largest clients. They wanted to share what they were learning.

We were huddled excitedly around the walnut table. Their senior analyst pulled up a slide and proudly explained their discovery: the older someone was, the longer they lived on a cancer therapy.

I slowly looked around the room and met eyes with our biostatistician, whose eyebrows screamed: of course older people live longer, they’ve had time to get older! Whoopsie.

Later, back at the office, putzing around on Twitter, I saw a new paper on transfer learning. It was an elegant idea: train a model where you have lots of data, then fine-tune it in specific domains where you have only a little. Our team had been using this approach on patients with rare cancer types. We were able to more accurately predict which therapies they would respond to; a boon.

This paper had done something like it, but for language, and at a scale that changed the game entirely. They called their invention BERT.

I realized that a lot of the plumbing we’d been building to structure data, package it, sell it—it could completely go away. It was DeepVariant all over. These new language models could collapse the distance between information and understanding; incrementally at first, then exponentially. I couldn’t predict the specifics, but I saw the outcome.

V. As I write this, AI is collapsing the cost of building software by orders of magnitude.

The workflow that used to require a hundred engineers and three years can now be assembled in months. Perhaps faster.

How should we respond to this moment? There are temptations.

One is to focus on output. Do more, do it faster. But a better hammer does not tell you what to build. We can ship a new widget faster than ever before, but if it creates more burden, more interactions, more checkboxes, more clicks — we have not solved a problem, we have merely added noise. Now we are starting to see the agentic version of this temptation: scaled-up swarm systems burning through millions of tokens, executing at superhuman speed with an AI CPO at the helm. We know what Baudrillard would say; a fancier tool used thoughtlessly accelerates the production of simulacra (faster bullshit is still bullshit).

Another temptation is to optimize the structures we work in today. But dusting a room that a hurricane has blown through won’t get you far. Consider Weed: the EHR mandate defined progress as adoption and cemented a system that violated what he spent his life striving for. Now we risk building a future that inherits those same constraints. We can choose a smarter workflow, a cleaner design. But our sense of what is good is often constrained by the architecture we should be dismantling. Otherwise progress will be incremental, not exponential. Foucault would say that this is the system structuring you when you should be restructuring the system.

Or we could remove people from the loop entirely. They dislike their jobs? Just take them away. What a failure of imagination; if I offered you a robot army and all you can imagine is them doing jobs someone already does, please give them back. This instinct also misreads the “spiritual violence.” It is not an imposition, it is an absence. A void where purpose once stood that we revolve. Liberating ourselves from our jobs without offering purpose would be its own violence. The harm in purposeless work is that it’s purposeless, not that it is work.

The final temptation is to simply be more of what we already are. It’s hard not to turn a side eye at those doubling down on taste because they see themselves as tastemakers. Hard not to wonder at the builders building new buildings for the sake of building. To suspect that the people who see a future without people need a hug, maybe a warm blanket. Our conclusions reflect us back to ourselves and we mistake them for revelation.6

And now my confession: I have indulged every temptation on this list.

After all this wandering, the question remains: how to find purpose?

When I am wrestling kids in the morning and absorbing someone else’s urgency at work and scrolling on my phone at night, it can be easy to ignore the question.

I have found words to describe my circumstances and (hopefully) avoid past missteps. Still I find myself yearning. The danger is that yearning can be co-opted, twisted, corrupted. It can be commercialized. In times of scarcity, it can be patronized.

But we are not in an age of scarcity. I do not have to lick the boot of a pope in order to buy my canvas. What will I do with that freedom?

VI. “It is always the same: once you are liberated, you are forced to ask who you are.”

Thousands of years before Baudrillard wrote that, and thousands of miles away, on a battlefield in Jyotisar, Krishna spoke to Arjun: “It is better to strive in one’s own dharma than to succeed in the dharma of another.” If I was to rediscover my purpose, my dharma, it had to be my own. Once again, I’m sitting on the carpeted floor of the old community center with my friends, seeking clarity in ancient wisdom.

I didn’t have the words, so I reached for a book. Then I reached for another. Pretty soon, I was reflexively reaching for something to read instead of my phone after the kids went to bed.

When I read lines like — “Sometimes what I wouldn’t give to have us sitting in a bar again at 9:00 a.m. telling lies to one another, far from God.”7 — I wanted to revisit them, to feel what I felt again: revived by the electric burn of a sentence, crafted with intention, by someone who gave a damn for the reader.

So I bought a notebook. In the cool quiet after the final page, I stare at the ceiling and think. Then I write. I turn my notes into reflections, ranging from the snarky (”Camus once wrote, One must imagine Sisyphus happy. What this book presupposes is … maybe he wasn’t?”) to pages on yearning, after being awed by Hua Hsu’s Stay True.

I read twenty-nine books that year, twenty-nine more than I’d read the year before. I picked up weights for the first time in a long while. A poem came to me in a quiet moment and I jotted it down. I grabbed a brush and painted watercolors with my kids. I went on walks, I smiled at babies and said god bless you to sneezing dogs.

At no point did I feel a light bulb turn on. It was a long walk. One foot, one page, after the other.

What I had come to was a me that I’d forgotten. The path back had been so haphazard, accidental. Just a spark that turned into a practice. Is this how it’s supposed to be?

When my daughter is bored or warding off bedtime, she’ll wrap herself around my arm and ask for a story about her favorite invention of mine, Princess Buttercup—a very purple princess with butterfly wings.

When she was younger, the stories were simple and decorous. Princess Buttercup wants to forage for strawberries, oh no there’s a puddle, oh wait she has wings. Princess Buttercup wants to go to her friend’s bubblegum tea party, oh no she forgot to buy a gift, oh wait she bakes cookies with her dad.

As my daughter grows, so do the stories. Now, Buttercup’s people are suffering from famine. A fire-breathing dragon is hoarding the goods they bring in from nearby kingdoms, causing the local economy to collapse. Buttercup climbs the dragon’s peak to try and reason with him. But he turns his flame on her and she barely escapes. Meanwhile, a peasant girl also braves the dragon’s lair and discovers that there are some crops it cannot eat; they could learn to grow them. It will mean a hard first winter, and changing long-standing traditions, and no cookies, but they can survive. She informs a despondent Buttercup, who must now pick herself up and lead her people.

My daughter sometimes carpools to ballet with her friend’s mom, who recently revealed to me that my daughter tells the girls stories. These stories aren’t coming from books she’s read or variations on mine; they’re wholly hers. Their favorite story is about Chomp-Chomp, a half-monster/half-dinosaur that eats kids’ shoes. It’s the highlight of their ride.

When the friction of production disappears, it can no longer protect us from the question of what to produce.

Tools change. Temptations are evergreen. But what is real is not the tool or its output or the system it inhabits; it is the capacity that precedes them. Imagination allows us to explore; exploration reveals potential. Intention focuses that vision and the creative act fulfills it. This is what persists. This is what adapts. Everything else is simulacra.

Refuse, then build and rebuild. In Foucault’s words: “Maybe the target nowadays is not to discover what we are but to refuse what we are. We have to imagine and to build up what we could be.”8

In the world that shaped us, creativity was not mere luxury, it was survival. The western apes who found comfortable niches stopped exploring and became chimpanzees. The eastern apes adapted to relentless ecological pressure through creation: new tools, new structures, new ways of thinking, new forms of culture.

Terms like “lizard brain” offer the false impression that our deepest motivations are fear and hunger. But the creative act is also ancient. The apes at Lomekwi were imagining and creating tools one million years before an ape could control a flame. Two million years before they could form sentences. Creation is primal.9

I’m not sure the ape at Lomekwi had a plan, at least not in the beginning. But they had to imagine the edge. And, perhaps after some trial and an abundance of error, they brought the stone down with intention.

Today, if you were to wander into the forests of New York, you may spy a different set of apes: my children. Modern apes still bang stones. They invent words and songs too. They can’t climb very well, but they fill the canopy with their laughter and screams. They draw towering smoke and leaf monsters and paint watercolors of bugs with a hundred and three legs.

What will they create next? The first step is always the same: imagine something that isn’t.

In 1814, Pierre-Simon Laplace conjured a demon.

In his Philosophical Essay on Probabilities, Laplace asked the reader to imagine an intellect so vast that it could know the precise position and momentum of every particle in the universe.

Such an intellect, he wrote, could calculate the future and reconstruct the past. Nothing would be uncertain. The arc of every billiard ball, the trajectory of every planet, the fate of every atom in every living body. All determined, all knowable.

For two centuries, Laplace’s demon was the reigning fantasy of scientific prediction. The only thing separating us from omniscience was information. Gather enough and the future opens like a book. It was seductive.

Decades earlier, a young David Hume wrote what he hoped would be the masterpiece of his generation.1

He asked the same question from the other end. If you are playing billiards, what evidence do you have that the next collision will behave like the last? What is stopping two colliding balls from shooting upward through the ceiling? Nothing but habit, he argued. A stubborn animal confidence that the world will remain as always.

Who was right?

I am a physicist by training, and so I tend to look at a problem and ask, am I seeing Laplace or am I seeing Hume? Newton, or Heisenberg?

The world we inhabit is mostly Laplacian. But two centuries later, quantum mechanics would vindicate Hume's instinct. Particles don’t behave the way Laplace insisted. Attempts by any intellect to measure their precise momentum and position will fail.

The machine learning community takes lessons from both; Laplace’s dream that enough data yields enough understanding, Hume’s warning that no amount of observation guarantees certainty about what comes next.

In 1996, another David (Wolpert), this time a physicist (we tend to propagate) tried to settle the argument with a theorem he called No Free Lunch.2

In a nutshell: If you apply every algorithm to every problem and average the results, they all perform the same. Whatever advantage an algorithm has on a particular problem, it pays for with equal disadvantage somewhere else. On average, a random forest or a neural network performs no better than simply flipping a coin. So much for Laplace’s demon.3

The implication: human judgment about the structure of the world is a prerequisite to the machine’s capability; you need to pick the right algorithm for your data. Be discerning, cultivate domain knowledge, do careful feature engineering, develop the right inductive bias. This became the conventional wisdom of an entire field. It was reasonable, well-supported, and almost entirely correct.

On December 5th, 2017, Google DeepMind released a paper that should have started a philosophical crisis but instead mostly started arguments about chess.

They introduced AlphaZero, a program that taught itself to play games by playing against itself. The researchers gave it no opening books, endgame tables, or grandmaster games to study. They offered no human knowledge of any kind, except the rules.

What AlphaZero did was generate visions of the future. Stable, predictable, possible outcomes. And from there, it could efficiently determine what the best mode of action was, the best path at any juncture.

After nine hours of self-play it defeated Stockfish, the strongest chess engine in the world. In a hundred-game match, it had 28 wins, 72 draws, 0 losses.

It not only won convincingly, it did it with style. Stockfish searched seventy million positions per second. AlphaZero searched only eighty thousand. It compensated by being radically selective, focusing on the most promising variations. It sacrificed material for positional advantages that human grandmasters found beautiful.

Garry Kasparov, who had lost his own match against Deep Blue two decades earlier, remarked, “I can’t disguise my satisfaction that it plays with a very dynamic style, much like my own!”

Stockfish was the product of decades of human refinement. Generations of programmers had encoded their understanding of chess into hand-tuned evaluation functions, opening databases, and endgame tablebases. It was a monument to the conventional wisdom of No Free Lunch: you match the algorithm to the problem. Human knowledge before the machine.

AlphaZero needed none of that. It started from random play and discovered strategy from scratch. It independently reinvented common openings before moving beyond them, into territory no human or machine had explored. Centuries of accumulated chess knowledge, reproduced and then surpassed, in less time than it takes to fly from New York to Moscow.

In 2005, twelve years before AlphaZero, David Wolpert co-authored a follow-up paper called Coevolutionary Free Lunches. He proved that No Free Lunch had a critical exception: in self-play problems, where agents compete against each other to produce a winner, some algorithms will genuinely outperform others.

The exception condition is precise: when the objective, what “best” looks like, isn’t fixed, the geometry of the problem changes. The optimal solution space is instead formed by the dynamics of the agents’ play.

A few years later, the field moved from games to language.

In 2021, large language models were starting to provide plausible and convincing outputs. They ingested the internet and learned to recombine it, remix it.

Around this time a critique formed: the models were stochastic parrots, regurgitators of their training data and nothing more.4

My issue with the critique was not the diagnosis, it was its poor assumptions; one, that learning is, in its basest form, memorization with extra steps. Two, that how we trained models was a fixed paradigm. Three, that the ceiling of machine intelligence would always be bound to the corpus it consumed, the human input.

In 2022, OpenAI published the InstructGPT paper. In it are results that should have settled the debate. They demonstrated that a tiny model, only 1.3 billion parameters, was preferred over a 175 billion parameter model. The vital difference: though both models shared the same pre-trained foundation, the tiny model was refined using human feedback, the larger trained to mirror human instruction.

The differentiator was not more data. The objective had changed. It was a dialogue between a person and a machine.

This was the birth of reinforcement learning in language; rote memorization gave way to a more human-centered optimization function.

In this paradigm, the stochasticity of the model was a benefit, not a drawback. You can prompt for a variety of outputs, have people rank them, and then train the model to generate outputs closer to their preferences.

This seemed enough to transform a parrot into something that had an implicit understanding of intent. A model that responded with coherence outmatching something 100 times its size.

The field was learning something critical: reinforcement enables transformation. It’s where the model learns to play.5

AlphaZero demonstrated the principle in its purest form. It was not memorizing, it was playing among potentialities. The InstructGPT paper showed that the same primitive, feedback-as-learning-engine, could be applied to language.

Once you understand that, the next question must be, what else can we play?

The superhuman capabilities of these models are real. When a coding agent can hold an entire codebase in its attention and refactor it with a coherence that surprises experienced engineers, I don’t think this leap comes from a bigger corpus or a longer context window. Does it come from self-play?

Imagine a system that generates its own coding problems and attempts its own solutions. It or another capable model evaluates the work. It encounters edge cases, failure modes, architectural dead-ends. It creates a portfolio that rivals what thousands of engineers could do in their lifetimes. It is playing chess with software engineering.

Where else could we apply this pattern? Consider security. In the old paradigm, you train on a dataset of known vulnerabilities and hope the model generalizes. In the self-play paradigm, you generate attack vectors that have never existed in the wild. The system explores the space of possible attacks and discovers new ones through adversarial play. It encounters and handles scenarios that no human red team would think to construct.

Medicine is the domain where this pattern is most needed and least proven.

Chess has perfect information, known rules, unambiguous win conditions. Medicine has none of those things. You don’t always know the root of the problem. The rules change between patients. A treatment that works for one person may harm another. But if we could simulate the full care journey, a doctor making decisions, a disease responding, outcomes unfolding, if we let the system play through a trillion variations, would it learn things that no clinical trial could reveal?

When language models entered the mainstream, people figured out that they could handle a lot of problems if we could describe them as a prompt. What problems could we solve if we could structure them for self-play?

Is Laplace’s demon smarter than us? No. The ability to map all potential futures of a system is not what makes human intelligence special.

Humans know that the world is not deterministic, we do not interact with it as though it were. We infer. We guess. We learn enough to draw the shape of something and then we act upon it.

It’s with that same intuition that we designed these playgrounds. The choice of what to simulate, what dynamics to introduce, what objectives to optimize. This is where human intuition is vital. A thousand naive algorithms left to their own devices would not invent self-play.

The hard part is bringing self-play to problems that don’t look like games. I want to stop asking what data should I feed it? and instead ask what world should we build for it to grow up in?6

Because we are not building tools anymore, we are raising something. The frontier of intelligence is not compute or data or architecture. It is the careful, creative work of deciding what form of intelligence we want to cultivate, then building the playground where it can teach itself things we never could. What it becomes depends on the world we give it to explore.

When my son was born, the first thing we put in his room was a milk crate full of picture books. On the wall, we hung a picture I’d taken of a gorilla at the Bronx Zoo. Shortly after his first birthday, he held up a book to show me a silverback inside. He then pointed at the wall and smiled. “Dada. Gorilla.”

It’s Friday night. The kids are in bed. My wife is out with friends. I am alone in my office, engrossed in my new favorite hobby: software development.

There’s no better dopamine than productivity dopamine

Over the holidays, I noticed that chatter about vibe coding tools ticked up. I saw, first at the fringes of my feeds, then front-and-center, a consensus rising: there’s been a massive leap forward in writing code, and it happened almost overnight.

A step change. A vibe shift, if you will.

Sure sure, I thought, there’s always something big and new. Several years into dedicating my career and sanity to AI, I’ve learned to temper my expectations. See through the hype, find the truth, build on it. Besides, as a dad of two cherubic gremlins, I was more concerned with picking up every scrap of wrapping paper and making sure I had just the right amount of bourbon in my eggnog.

Then I came across this tweet from Andrej Karpathy (the inventor of “vibe coding” who once said it’ll be a decade until AI agents work properly):

Still, I wasn’t moved. My experience with AI coding tools had been mixed. I appreciated that it could autocomplete docstrings, add type hints, suggest small optimizations. These tools were excellent at the necessary tedium, but whenever I tried to make major changes with AI it usually failed, albeit convincingly.

What finally caught my attention was a piece by Casey Newton, “The project that turned me into a Claude Code believer.” In it, he writes about using Claude Code to remake his website quickly and canceling his Squarespace account.

Hey, I use Squarespace, I thought.

10 minutes later, I canceled my Squarespace account too.

My typical workflow for personal coding projects looked like this:

These days, I am developing ambiently:

I used to be limited by whether my free time, energy, and inspiration coincided. Now I have an idea and minutes later I have a working prototype.

After revamping my website, I made a utility that organizes my Downloads and Documents folders. The next day, I built a bookmark manager with an AI assistant to replace Raindrop. After that, a web app for my book club to vote on our next pick. Then an agent network to search news and fact-check claims against sources I trust.

More. I can do more.

Is that all you got?

Spend enough time with Claude Code and you go through a roller coaster of sorts. At first, you’re partners. Buddies. You’re collaborating, sketching out the approach together, riffing on ideas. Let’s build a tool that turns voice memos into professional emails. Maybe controls to tune length and tone, validate against my personal history. Why not, the night is young? Claude codes while you swirl your drink. This is what living your best life surely means.

Alas, honeymoons end. A deploy fails. The frontend drifts from the backend. You find you’re explaining yourself more. Suddenly you’re not a partner, you’re a babysitter.

Naturally, daring folks are already trying to evolve beyond this pattern.

One fun example is Ralph. The idea is beautifully simple: instead of babysitting Claude, you give an agent a list of well-defined tasks and let it run autonomously. When a task is done, it moves on to the next one.

Meanwhile, you stand up, stretch, walk away, make a coffee, read a book, maybe reconnect with loved ones (assign another Ralph another task). Check back in an hour and there are twelve commits, a working feature, and notes about what Ralph learned along the way.

Last week, an engineer on my team was refactoring a legacy app. After pair coding with Claude Code, he realized the project would take a while, so he pivoted. He created a Ralph-like agentic pattern, with finer controls and checkpoints for feedback (call it “Ralph with better parenting”). On a Friday, he gave helpful instructions to an agent and kickstarted the loop. The refactor finished successfully over the weekend. Weeks of progress made before Monday morning standup.

The death of coding is the future of work

We are in the early days of a future that looks very different from our past.

Building the wrong thing used to be expensive. Thirty years ago, software companies tried to derisk by writing detailed requirements that anticipated every possible need. Success was not only measurable, it could be planned. But markets evolved while the spec stayed the same, and products launched for customers who no longer existed.

Eventually, people began to notice that the projects that succeeded did so messily: code was written, tested, demoed, revised, and rewritten. Iterations mattered more than being right the first time. These people had wonderful and terrible ideas, and these ideas were given a name — agile — and a manifesto: software, like people, changes. Building software should honor and respect change.1 As software went on to eat the world, a cornucopia of agile frameworks multiplied and were plenty.

The rituals of old were designed to manage scarcity. Scarcity of time, attention, resources. But code is cheap now. The new constraints are creativity, insight, taste.

If the new shape of work looks like yesterday’s but faster, then it will be a tragedy. This new world demands jetpacks, not faster horses.

How does one build a jetpack?

Some contours of the future feel clear to me.

The promise of every development process is the same: ship better products faster. Teams assume process organically leads to good outcomes, but when process inevitably fails people tend to pile on more process. And process can distract teams from what matters most.

When agents write code, you need to focus less on process and coordination and more on making better decisions faster. If your coder is an agent, a requirement document isn’t a communication tool, it’s a prompt. It needs to be precise, testable, and unambiguous. Telemetry to estimate cost and time for a project is weak today, so the plan needs a well-defined first loop to estimate the full build.

Automated evaluation is more critical now. Tests are no longer a safety net, they’re the contract between you and your agents. If you can’t define success in a way a machine can verify automatically, you quickly go back to inspecting code.

As I write this, deployment is still a bottleneck. But agents writing and deploying code autonomously need systems that are secure by default, with CI/CD pipelines, permissions, sandboxing, rollback. This will get solved soon.

Standup is dead, it’s replaced by an AI sending you bullet points in an email. Sprints compress into days, then hours, then minutes. What even is a sprint when the backlog is a JSON file shared between models?

By 2027, a single prompt will get you a fully functional application in production. Not a prototype. A real product, deployed and serving users.

At that point, software development will be unrecognizable. The most successful developers won't be the best coders. They'll be the ones who know what's worth building.

We must cultivate our garden

In February 2020, I got a text from a friend at the CDC: “Things are pretty bad. Make sure you’re ready.”

My wife was 7 months pregnant with our first kid. I didn’t know what “ready” meant exactly, so I hurried to the grocery store after work. I left an hour early, imagining a backlog of cars outside the parking lot, an angry crowd huddled around empty shelves inside.

Yet there was no line and no crowd. I walked in and quietly loaded my cart with rice, beans, things I knew would last. I looked around and realized no one was panicked or even mildly concerned. An elderly man asked me to grab a bag of flour he couldn’t reach. I met his broad smile with a nervous one.

Was I overreacting?

The next day, I decided to work from home until the baby was born.

In those days, one of my small pleasures was the hydroponic herb garden we kept in our dining room. I made a lot of pasta, so I kept the basil full. I loved that garden. I only needed to set it up just right, check in on the water and sunlight. The basil was so tall and so green. At a time when things felt very uncertain, I at least had that.

I didn't have language for this feeling, but someone else did.

Masanobu Fukuoka was an agricultural scientist in Yokohama when he got pneumonia and nearly died.2 Lying in a hospital bed, alone, he spiraled into an existential crisis which held onto him even as he recovered. One night while wandering, he heard a heron cry. “All the concepts to which I had been clinging, the very notion of existence itself, were empty fabrications,” he realized. He quit his job the next day and went back to his family’s farm. He wanted to know whether he could grow crops by doing less, not more. The more we intervene, he believed, the more intervention becomes necessary. Every technique creates a new problem requiring another technique. Fukuoka called his philosophy “do-nothing farming” and went on to have some of the highest rice yields in Japan. He saw his role clearly: not to grow crops but to set the right conditions and get out of the way. “Crops grow themselves.” When I tend to my agents, will it feel the same?

The gardener’s paradox is the tension between control and cultivation. The better the conditions, the more the outcome depends on the system, not the gardener. Letting go is hard. But the crops might grow taller without us. Perhaps we’re all gardeners now.

👋 Hey, I’m Gaurav (gawr-uhv). I’m a reformed academic turned startup founder turned healthcare executive.

What are we doing here?

I want to write again. It’s been a while. Writing is thinking, and thinking out loud keeps me honest.

At first I thought I’d write about things I’ve learned in my career, but the internet is full of “look what I know” slop. I don’t want to add to that. I want a space to explore questions that don’t have quick or clear solutions. Questions with answers that aren’t down in any map, so to speak.

Why now?

Knowledge is more accessible than ever, but consensus has never been harder to build. And new tools are making it easier to think less, sand away the friction.

I’m scared of a world where we forget how to think rigorously, to failure. Where we get worse at making sense of competing ideas, disagree without division, build things together. Consensus is a precursor to progress.

Writing is my way of building a thinking practice.

“Queequeg was a native of Rokovoko, an island far away to the West and South. It is not down in any map; true places never are.” — Herman Melville, Moby Dick