Of Gods and LLMs

Everyone's scared of AI right now. It's going to take your job, hollow out the economy, make us all obsolete. Our work is bound up in our identity. We don't give it up easily.

As the saying goes, history doesn't repeat itself, but it rhymes, and the man-versus-machine story isn't new. In the late 1800s, the folk hero John Henry raced a steam-powered drill to see who could lay more rail. He won, and dropped dead with the hammer still in his hand. Garry Kasparov beat IBM's Deep Blue in their first match in 1996, only to lose the rematch a year later. Now a chess engine running on the phone in your pocket can defeat the strongest human players alive.

But here's what's missing. Like the classical Greek gods, there's something about machine perfection that's fundamentally inhuman, and something about being flawed and mortal that turns out to be the whole point. The Japanese have a practice called Kintsugi: they repair broken ceramics with gold, and the piece comes out more beautiful for having been broken. That's not a coincidence. It's a clue about where human work stays valuable in an AI-maximalist future, and it's the opposite of what everyone's afraid of.

Part 1: Neural Networks

In order to understand LLMs you have to understand how they work. Put extremely simply, they're able to guess the next word in a sentence based on your prompt and the training data fed into the model. I'll explain using a more relatable example.

Picture deciding whether to go surfing. You weigh a few inputs: wave height, wind, water temperature, your willingness to put on a cold wetsuit. But you don't weigh them equally. Wave height matters a lot; water temp matters a little; how much you had to drink the night before matters a lot. In your head, you multiply each input by how much it matters, add up the results, and if the total clears some threshold, you go.

Your lifetime history of decisions to go surfing or not is the training data, and given enough of it, you can build a model that's statistically likely to predict whether you'd go surfing in a given set of conditions. This would be your Large Surfing Decision Model.

If this were a real model: the inputs are the conditions (waves, wind, temp, hangover). The nodes are the little units each weighing those conditions toward some outcome. The layers are how those judgments stack: raw conditions feeding into concepts like “is the surf good” and “am I up for it,” feeding into the final call. And the parameters are the weights themselves: the thousands of “how much does this matter” numbers that your training data settled into.

A modern LLM is this exact structure, just with billions of parameters across dozens of layers, trained to predict the next word instead of your next surf session. To give you a sense of scale, here are the estimated parameter counts across OpenAI's GPT generations:

• GPT-3 (2020): 175 billion
• GPT-4 (2023): ~1.7 trillion estimated
• GPT-5 (2025): ~2–5 trillion estimated

Instead of wave height and water temperature, the models are trained on the entire corpus of the written human record. An amazing feat in and of itself. All of that compute and inference to guess what the next word in a sentence should be.

And therein lies the weakness.

Part 2: The Midwit Problem

LLMs are the midwit. And they have to be. Structurally.

In order to predict the next word with a high degree of probability, the model chooses the most likely response. Which means it chooses the safe, consensus, statistically probable response to your prompt. And how do we, as humans, tend to judge things that are safe, average, and predictable?

Boring as Fuck.

Statistically probable businesses are: McDonald's. Denny's. Holiday Inn. The Ford Taurus. PF Chang's. The Gap. Supercuts.

There's nothing wrong with these businesses. They're predictable. They deliver exactly what you'd expect, you'll never write home to your parents about them, and they're so inoffensive you might have forgotten some of them exist. These are the statistically probable businesses that serve a simple purpose and, when run perfectly, deliver their value in exactly that way. That is the problem of average, and it is the problem of midwittery.

You can test this yourself. Clear the context of your favorite LLM, or log into a fresh session. Ask it something you have an opinion on. What's the best book? What's the best movie? It will give you the most plain-jane, read-it-in-a-textbook answer: The Great Gatsby. Gone with the Wind.

LLMs are statistics machines. They have no sense of value.

Which raises the harder question: what is value, actually? What makes one thing matter more than another? Pirsig spent a whole book on this and gave it a name — Quality.

Part 3: Imperfection Is the Point

In Zen and the Art of Motorcycle Maintenance, Pirsig uses the characters in the story to investigate the nature of Quality and arrives at the unsettling conclusion that Quality can't be defined, only recognized. Crucially, it requires caring. Quality shows up only when someone with stake in the outcome is paying attention. You can't notice the quality of a motorcycle's tuning if you don't care about how it runs.

LLMs, with their immaculate ability to predict the next word, are functionally correct, but they don't care what the outcome is beyond interpreting your prompt. Claude Opus 4.7 doesn't have an opinion on what a good exhaust note sounds like. ChatGPT 5.5 would never design the Porsche 911 with a rear-mounted engine: a layout that's technically wrong, harder to drive at the limit, and the entire reason a 911 feels like a 911.

LLMs have competence without commitment, skill without stake, productivity without caring.

Another example of lacking stakes is the myth of the Greek gods. Why would the immortal gods of Olympus envy a bunch of puny mortals? They had everything: supernatural power, beauty, endless time. And yet the myths keep circling the same strange admission: the gods envied us. While their endlessness was the source of their affliction, our finiteness gave weight to every action of our lives.

Was this the last pomegranate you'd ever eat? It might be, and it will taste sweeter for it. The gods had an eternity of ambrosia and couldn't be bothered to taste any of it. Nothing was ever at risk for them, and nothing at risk means nothing at stake, and without stake, there's nothing to care about. The gods couldn't care, so the gods couldn't taste. An LLM has the same affliction. It has limitless information: all the ambrosia in the world, every word ever written by a human being. And no reason to prefer Dostoevsky to Tolstoy.

Part 4: Logical and Psycho-Logical

So in the world of intelligence, Denny's and Supercuts are going to go to the LLMs. If you need a boring, textbook-correct answer, the friendly neighborhood AI will solve that problem quickly, cheaply, and efficiently. The free market will drive the cost of commodity intelligence down as automation takes off: any repetitive, mechanical task you've ever done at a white-collar desk job will get done by a robot instead. Boring will stay boring, but at least it'll be cheap. I'm personally of the opinion that's a good thing. Robots are better at being robots than people are. Who'd have thought?

So what are humans better at?

Being human.

In a world of cheap, commodity intelligence, human imperfection and finiteness is what will drive value. Maybe your dad had an old Eames chair in his study and you have a fondness for bent plywood. Maybe you had too many tequila shots on that trip to Baja one summer and now you drink gin instead. Taste is being imperfect in a particular direction: enjoying brutalist architecture or mid-century modern furniture is caring about one thing rather than another, tied to our finite, imperfect experience of being human. That personal bias is exactly what gives us preferences, values, and taste.

An LLM will never be a stand-up comedian. ChatGPT has never had to sit in therapy unpacking a three-month breakup with a toxic ex.

The work of being human needs humans: to know which problems are worth solving, and to bring a human perspective to them. Rory Sutherland of Ogilvy is the king of solving problems by addressing the human component, not the technical optimization. One of his often-quoted examples in Transport for Humans is the Eurostar train. Why spend billions making the train faster when you could put Wi-Fi and comfortable seats on it and make the time on the train more pleasant for the passengers? To take the example to its logical extreme: add champagne and supermodels, and many would ask for the ride to take longer. A highly inefficient train ride, but certainly an enjoyable and memorable one. What Rory understands is what humans value. If you ask an LLM “I have billions to spend, make me the best train system,” it will generate something efficient, cost-effective, and safe to the point of forgettable: solving the problem in the most straightforward way possible.

None of this is unprecedented. We've already seen what industrialization, the computer, and the internet did to work. So this leads to the logical conclusion: robots will do the things that robots are best at — cold, rational efficiency — and humans will do what we're best at: being human. Handwritten letters are rare now precisely because emails are free. As AI makes everything-that-can-be-cheap cheap, the things that stay expensive — human attention, human care, human time — are what will signal value.

Yes, AI will take your job. In many ways, it already has.

That doesn't mean there's nothing left for you to do — it means the thing left for you to do is the part only you can.

In an AI-maximalist world, the race to the bottom on cost efficiency has already been won by someone. Value. Judgment. Opinion. Perspective. These are all expressions of Pirsig's Quality: someone caring for flawed, irrational, fundamentally human reasons. And since we're in the business of serving other humans, showing humanity is everything.

Part 5: Counterintuitively, AI Will Make Humans More Human

Being human is a crazy experience. None of it seems very predictable. No one is perfect. No one is exactly the same, and that's what's interesting about us. There's a reason nearly every iteration of science fiction has some kind of anti-emotional or synthetic character, Mr. Spock, for the very purpose of holding up a blank canvas against which to explore the human experience.

The more competent the machines get, the less reason there is for a human to compete on competence. No one needs a John Henry to drive steel faster than a steam-powered machine. What's left, and what may be the only thing left, is the part that was never about being correct: the specific, the flawed, the felt, the chosen. AI doesn't make us less necessary. It's the opposite: AI needs humans to decide what's valuable, and value is created by the finite, unique experience of being human.

So the future of work isn't that AI takes your job. It takes the part of your job that any sufficiently good machine could do, and leaves you the part only you can do.

An LLM can explain Kintsugi. It will never understand what makes a broken bowl more beautiful.