What AI is really doing to work
A one month retrospective from the frontier at Fractal Tech NYC
To start, I disclaim my rather certain title. I’m no labor economist, having learned about Jevons Paradox and its relevance to AI just a few months ago.1 I’m also just a really uncertain person. Don’t ask me what I’m having for lunch this week, and don’t ask me what AI will do to our economy.
Still, I feel compelled to reason with what I’m seeing at the frontier. At Fractal Tech NYC, my cohort is the first group of 1) mostly software engineers who 2) are learning AI software engineering with the latest models (Claude Opus 4.6/GPT Codex 5.3/Gemini 3.1…when I say Claude, substitute with any frontier model). Fractal Tech is also a general tech hub that attracts the best engineers in the world; the demos I’ve seen are, without exaggeration, mind-blowing. That’s coming from me, a fairly tech-critical/techno-ambivalent person who has worked in tech/tech-adjacent roles for the past decade.
Why someone like me felt drawn to AI software engineering is a post for another time. The point is, I’ve made a massive jump in the last month, best illustrated with the image below. You might have seen this “tiers of AI usage” floating around:
I’ve gone from the yellow boxes into the red box, joining the few million users in the world who are Clauding out. Here’s what I can say: people have zero clue. The distance between what the public thinks AI can do and what it can actually do is the widest it has ever been. An average person probably still conceives of AI as a smart and fun chatbot or image generator. A curious person understands that AI can generate code in a fancy autocomplete sort of way. An engaged person might be following Andrej Karpathy tweets, but even top-level minds aren’t properly explaining the frontier because the frontier is being built every second, so all we have are (probably true) memes like “The one-person unicorn startup is now possible.”
The real root: the software engineer is their own bottleneck now. AI is a coordination problem; the work (code) is a mere byproduct of coordination.
Most software engineering work is no longer about the work itself. The real work is about solving the coordination of autonomous workers who don’t eat, sleep, rest.
How is this possible?
The True Nature of AI
First, we need to grasp AI at an intuitive level. It’s not a stochastic parrot, a criticism lingering from the GPT-3 days in 2020. It’s not even a predictor. Or rather, predictor doesn’t quite capture its true nature.
Claude is a simulator. In particular, a type of simulator called a harness. AI chatbots may simulate conversations, but Claude can do that AND simulate a person using a computer. All through text alone.
How? How does Claude, whose fundamental building block is just text, simulate using a computer?
Let’s start with simulating a conversation. A large language model takes some textual input, feeds the input into a magic black box, and returns an output. When you recursively feed the text + output into the black box, you simulate a conversation. When the conversation is too long, you run out of space (“the context window limit is reached”).
Leaping from simulating conversations to simulating computer usage is done through tool calls.2 Now the magic black box can access (call) external programs (tools). Now it can take actions that a typical chatbot can’t do e.g. get real-time data, search through your files, call your mom. A total gamechanger. This is how engineers are able to teach Claude to act as if it were them, using their own computers. And this is where the absolutely bonkers stuff is happening.
I’m seeing engineers token-maxxing and optimizing for price/token, commanding 10+ AI coding agents at once. A full workforce. A hivemind. Each AI agent is a worker bee with a separate role (writing code, reviewing it, handling merges, monitoring other worker bees etc). The queen bee engineer sits at the top and describes what they want to build. How they do this can be wildly different—some engineers spend all day speaking English to Claude, outsourcing the ~20% edge cases to their overseas test engineer while they sleep. Others draw intricate system diagrams with Mermaid and feed those plans to Claude. And Claude today, the worst that it will ever be, already can deliver.
Jevons Paradox and Barbell Economics: The Middle is Dying
Let’s return to Jevons Paradox, the counterintuitive idea that increased efficiency consumes more of a resource rather than conserving it. For AI, the paradox has played out practically according to plan. When DeepSeek released their AI chatbot at ~5% of OpenAI’s cost, investors panicked. Then Satya Nadella tweeted “Jevons paradox strikes again!” and flipped the narrative. Everyone relax! Cheaper AI means more AI, which means more compute, more chips, more more MORE.
So now the question isn’t whether demand increases, it’s what kind of demand. To me, the most salient questions from that article were around substitute versus complementary goods. A substitute replaces e.g. oat milk for dairy. A complement enhances e.g. eggs with toast. So which one is AI, substitute or complement?
Before I entered this program, I had a thesis: if AI solves software as pure substitution, it pushes engineers down the stack. Software is abstraction layered on abstraction, declarative layers on top of imperative cores, all the way down to physics. If AI handles the upper layers, then complementary expertise migrates to the lower ones—hardware, materials science, manufacturing. Back to wrangling atoms, since not even AI can bend the speed of light. I had hoped this would trigger a manufacturing renaissance in the U.S.
But what I’m seeing aren’t the best minds going down. They’re going up. They’re becoming pure coordination layers. The bonkers hivemind workflow I described in the section above? Veteran engineer Steve Yegge likened it to Kubernetes, the system invented by Google to manage thousands of servers. You don’t need to say “run this app on server #42.” You just say “I want 3 copies of this app running at all times” and Kubernetes will figure out the details.
Now I get the pattern. As products are commoditized, the value moves up to the coordination layer above.
What’s forming is a barbell where value concentrates at two extremes:
At the top—AI as complement: vision, taste, knowing what to build, directing the hive. The bottleneck here is imagination and agency. This is where Jevons Paradox kicks in. One orchestrator can now act like a team of ten, so every company wants orchestrators. Demand explosion.
At the bottom: atoms. Physics. Reality has the final say with semiconductor fabs, battery chemistry, rockets to Mars.
The middle is the substitute being wiped out. Standard software engineering like webapps, what most people think of when they hear “tech”—that’s dying. Solved problems. Soon to be commodities, like servers. And sure, there’s still some time; anything outside AI’s training data distribution is safe for now. Building Solidity contracts on Ethereum is not a solved problem (yet). Modeling how to launch rockets is not a solved problem(yet?).
Now let’s consider the pattern at scale.
What’s happening to software engineering right now will happen to almost every white-collar corporate job over the next few years.
Work meets its Quartz Crisis
In 2017, I almost dropped out of the technological world. I seriously considered applying to Patek Phillipe’s watchmaking apprenticeship in NYC. My fantasy: leave the fake world of bits and bytes, return to reality, make beautiful things with my hands. George Daniels, one of the greatest inventors of all time, was my fuel.
Half a century ago, the Swiss watch industry was nearly crushed by the Quartz Crisis. The problem of accuracy in timekeeping was solved once and for all—digitally, not mechanically. In 1927, Bell Labs invented the quartz clock. By 1969, digital watches threatened to replace mechanical watches for good.
Strangely enough, at the same time this upheaval was taking place, a self-taught British watchmaker was insistent on solving the problem of accuracy for mechanical watches. At the rock: quartz creeping in. At the hard place: inertia from the Swiss behemoths. For over twenty years, he fought the establishment.
He won.
George Daniels invented the co-axial escapement, replacing the single lever that had been the default for over 250 years. Today, any watch made by him sells for millions. Today, the Swiss watch industry lives on.
Software engineering is facing its Quartz Crisis now. More white-collar work will face it soon. As quartz solved the problem of accuracy, AI is solving the problem of work.
What do we do?
Abandon all hope ye who enter here says Celeste 🌱 in a probably not-that-hyperbolic article.

My advice? Hang out with hot people. It worked for me last night; I was physically incapable of doomerism all evening.
But to be for real, I think it’s important to remember what makes you human. As Jeff Giesea says in his article, the humanities are more important than ever. Those skills at the top layer? Vision, taste, knowing what’s worth building….those are the humanities.
I can certainly attest to the power of reading books. Reading about George Daniels nearly a decade ago has resurfaced now. Just last week, I relayed his story to my Fractal Tech compatriots in a demo about audio clocks. And I relayed it again in this article as a message of hope. In a twisted sort of way, I’m fulfilling my old dream, refracted—moving to New York City and facing my Quartz Crisis head-on, as George Daniels did.
His story is one of survivorship, in both senses of the word. Survivor bias, in that there’s only one George Daniels and it’s unlikely that I’m actually a secret genius inventor. But you should know that he was an actual survivor, too. His love of watches rescued him from a brutal, lonely childhood. And even when the world was screaming at him to give up, it’s hopeless, you’re doomed—his loving attention never wavered.
My intuition says if you care about beauty, the process, the craft, whatever makes you come alive, that there will always be space for you. And even if there isn’t, you’ll carve out a space if you keep reaching for your co-axial escapement.
If you want to follow along with my bootcamp adventures, check out my website with ~weekly dev blog posts and my daily Tech Bro posts on Substack Notes or Twitter/X
This article was adapted from some daily Tech Bro notes I wrote a few weeks ago:
Tool calls are standardized/enabled by Model Context Protocols (MCPs).



@Khe Hy good insights, Lily is in the WoP circles
Thank you for the dispatch