When the Builders Stop Building

Andrej Karpathy hasn't written a line of code since December. That's not a tech trend. It's a fire alarm.

Andrej Karpathy built the AI that drives Teslas. He led research at OpenAI. He is, by any reasonable measure, one of the most accomplished software engineers alive. And he hasn’t typed a line of code since December.

Not because he’s retired. Not because he’s on sabbatical. Because the agents do it now.

“I don’t think I’ve typed like a line of code probably since December, basically, which is like an extremely large change,” Karpathy said on the No Priors podcast last week (No Priors, 2026). He went from writing 80% of his own code to writing essentially none in a matter of weeks. His new workflow? “Expressing my will to my agents.” Sarah Guo, the show’s host and a venture capitalist at Conviction, confirmed that engineers in her portfolio companies now work by whispering into microphones all day, directing autonomous agents that write, test, and deploy code on their behalf (Guo, 2026).

Let that settle for a moment. The people who build the tools that are reshaping civilization have stopped building in the way that word has meant for the last fifty years of software. They are no longer craftsmen. They are, in Karpathy’s own language, sovereigns issuing decrees.

And they’re losing their minds over it.

The Ripple That Hasn’t Hit Shore Yet

If you work in a field where your primary activity is sitting between two screens and manipulating digital information, what happened to Karpathy in December is coming for you. It may already be at your door.

Karpathy’s framing is blunt: if you can do it from home, an agent can probably do it (No Priors, 2026). Legal research. Financial analysis. Medical records processing. Marketing copy. Accounting. Graphic design. The common thread is not the industry; it’s the substrate. Digital information is infinitely malleable by systems that process digital information. The transformation of the digital workspace will massively outpace anything happening in the physical world, because the friction is almost zero. Code, Karpathy says, “is now ephemeral and it can change and it can be modified” (No Priors, 2026).

This is not a prediction about some distant future. This is a description of what is already happening in the most technically sophisticated companies on earth. And the sophistication gap between those companies and everyone else is closing at a pace that should make every knowledge worker pay attention.

The Numbers Behind the Narrative

Now, there’s a counterargument, and it deserves a fair hearing. Jensen Huang, CEO of NVIDIA, has repeatedly argued that AI creates more jobs than it destroys. His go-to example is radiology: AI didn’t eliminate radiologists; it made radiology more accessible, increased demand for imaging, and created new specialties. The ATM analogy works similarly. ATMs made branch operations cheaper, so banks opened more branches, so they hired more tellers. Karpathy himself cites Bureau of Labor Statistics data showing that engineering job demand is still increasing, and he describes himself as “cautiously optimistic” about the profession in the near term (No Priors, 2026).

This is the Jevons paradox applied to software: when you make something dramatically cheaper to produce, demand doesn’t shrink; it explodes. More software, not less. More need for people who can direct, evaluate, and integrate that software, not fewer.

I want to take that argument seriously because it contains real economic logic. The Jevons paradox is historically robust. And Karpathy is not a doomsayer; he’s building in this space with genuine enthusiasm.

But the Jevons paradox describes aggregate demand. It says nothing about what happens to the specific humans who occupied the roles that got automated. The ATM analogy is instructive precisely because it’s incomplete: yes, there were more bank tellers after ATMs. But the job of bank teller changed fundamentally. It went from counting cash to selling financial products. The people who couldn’t make that transition didn’t get absorbed by the paradox. They got displaced by it.

And the early data on AI’s impact on technical roles is not theoretical. It’s brutal. In Q1 2026 alone, at least 125,000 jobs were cut across major tech companies — Block, Meta, Dell, Amazon, Salesforce, Atlassian, and others — in layoffs explicitly or implicitly tied to AI (AI RIF data, compiled March 2026). Block cut roughly 4,000 employees after running parallel human and AI teams; the AI teams outpaced humans 3:1 in feature velocity, and the humans were let go (TechLayoffLover, 2026). Meta laid off over 15,000 as “AI costs mount” (Reuters, 2026). One analyst described it plainly: “The headcount line item is being converted into the datacenter line item. Salaries become servers” (Gupta, 2026).

The entry-level market has been hit hardest. At a top computer science program, the placement rate for graduates collapsed from 89% to 16% — out of 410 majors, 67 got offers (TechLayoffLover, 2026). New graduate requisitions are down 89% from 2022 (TechLayoffLover, 2026). At Big Tech companies, new grads went from roughly a third of all hires in 2019 to about 7% today (Gupta, 2026). A Stanford CS graduate with $180,000 in student debt applied to 847 entry-level positions and received zero offers — his post went viral with 1.7 million views (TechLayoffLover, 2026). One engineering organization went from 340 people to 89. No juniors remain.

NVIDIA CEO Jensen Huang’s own formula explains why: a $500,000 senior engineer consuming $250,000 in AI tokens produces what used to require a team. The company saves money and increases output. For the company, this is a dream. For the junior engineer, it’s a closed door.

Here’s the pipeline paradox no one has answered: if junior roles disappear, who becomes senior? You don’t become a senior without first being a junior. If the entry ramp vanishes, the entire career ladder eventually collapses. You can’t have a Jevons paradox for senior talent if you’ve eliminated the pipeline that produces senior talent. The optimistic aggregate story and the devastating individual story are both true simultaneously, and the tension between them is where the real human cost lives.

The Machine That Improves Itself

If the current state of affairs were static, we could perhaps adapt. Retrain. Reskill. Find the new equilibrium. But it’s not static, and this is the part of the conversation that should keep thoughtful people up at night.

Karpathy described something he calls AutoResearch: an autonomous loop where AI agents optimize AI models overnight without human intervention. He pointed it at his nanoGPT repository, a codebase he has personally hand-tuned for over two decades, and let it run while he slept. By morning, it had found hyperparameter improvements he’d missed across twenty years of expert-level work. Suboptimal weight decay on value embeddings. Insufficiently tuned Adam betas. Joint interactions that are nearly impossible for a human to discover manually because the search space is too vast (No Priors, 2026).

Twenty years of hand-tuning by one of the world’s best AI researchers, beaten in a single night by a system running autonomously.

This is not a tool being used by a human. This is a system improving itself in a domain where the human was already operating at the frontier. The “human in the loop” is becoming the “human who checks the results in the morning.” And the results are better than what the human could produce.

Karpathy is candid about the constraints: AutoResearch currently works best in domains with objective, easily verifiable metrics. Writing CUDA kernels is a perfect fit. Creative work, nuanced judgment, moral reasoning: not yet. But “not yet” is doing a lot of heavy lifting in that sentence, and the direction of travel is unambiguous. Anthropic’s own study of Claude usage data across 800+ occupations found that AI already covers 75% of computer programmers’ tasks, 70% for customer service representatives, and 67% for data entry — with white-collar, high-education workers disproportionately exposed (Anthropic study, 2026). One researcher called it a potential “Great Recession for white-collar workers.”

The Meaning Void Wearing a Productivity Mask

Here is where the conversation gets interesting for anyone who cares about the human soul.

Karpathy describes his psychological state with a striking word: “psychosis.” Not clinical psychosis, but something he recognizes as disordered. He feels guilty when his agent subscriptions go unused. He feels anxious when he’s not maximizing the number of parallel agents working on his behalf. Every failure feels like a personal skill issue rather than a model limitation. “If you don’t feel very bounded by your ability to spend on tokens,” he says, “then you are the bottleneck in the system that is max capability” (No Priors, 2026).

Read that again. The human is the bottleneck. The system’s maximum capability is constrained by the human’s ability to direct it. And the emotional response to that realization is not liberation; it’s compulsion. Optimize harder. Spend more tokens. Parallelize more agents. Sleep less. The machine never tires, so your rest becomes the system’s inefficiency.

Karpathy is describing, with remarkable precision, a soul unmoored from any metric of enough. When your output capacity is theoretically infinite, when the only constraint on production is your willingness to keep issuing commands, where do you find the signal that says stop? Where is the sabbath in a system that never sleeps?

He’s not alone in sensing this. Julia McCoy, a content strategist who watched AI transform her entire industry, has asked the question plainly: “Who are you without your job title?” (McCoy, 2026). That’s not a career question. That’s an existential one. It touches the deepest nerve of human identity: am I what I produce, or am I something more?

I hear this question constantly. Not from tech executives (though apparently I should be hearing it from them too), but from people in my congregation. People who’ve been laid off. People whose roles have been restructured. People who built their identity on competence and suddenly find that competence is being commoditized. “Who am I if I’m not the person who does this thing well?” That question has been walking through church doors for as long as churches have had doors.

Karpathy calls his state psychosis. A theologian would recognize it as something older: a soul that has confused its function with its identity, and is now watching the function get automated away.

The Question Beneath the Question

This is where I want to push past the technology conversation entirely, because the technology conversation is not the real conversation. The real conversation is anthropological. It’s about what we think humans are, and therefore what we think work is for.

There are, broadly, two ways to understand human labor.

The first is instrumental. Work is a means to an end. We work to produce output, generate value, solve problems, earn income. On this view, the value of work is entirely located in its product. If a machine can produce the same output faster, cheaper, and better, then the machine should do the work. Full stop. The human should find something else to do, or, if there’s nothing left, enjoy the leisure. This is the default view of Silicon Valley, and it has a clean internal logic. If work is about output, and AI produces better output, then AI should work and humans should... what, exactly?

That’s where the instrumental view runs out of answers. Because Karpathy isn’t enjoying his leisure. He’s in “psychosis.” The output is better than ever. The productivity is through the roof. And the human at the center of it is anxious, compulsive, and reaching for the word psychosis to describe his own experience. The instrumental view of work has no diagnosis for this condition, because on its own terms, nothing is wrong. Output is up. Efficiency is maximized. The system is working perfectly.

The second view is intrinsic, or what I’d call vocational. On this view, work is not merely a means to an end; it is an expression of what humans are. We work because we are made to create, tend, cultivate, and steward. Work is woven into human identity at the most fundamental level.

This is not a sentimental claim. It’s a theological one, grounded in the opening chapters of Genesis. “Then the Lord God took the man and put him in the garden of Eden to tend and keep it” (Genesis 2:15, NKJV). Note the sequence: this is before the fall. Work is not a curse. Work is not a consequence of sin. Work is part of the original design. Adam was placed in a garden and given a job before anything went wrong. The cultural mandate of Genesis 1:28, “Be fruitful and multiply; fill the earth and subdue it,” is a commission to creative, purposeful labor as an expression of bearing God’s image.

If this is true, then work has value independent of its output. The act of tending, creating, problem-solving, building: these are not merely instrumental activities that can be frictionlessly outsourced to a machine. They are part of what it means to be human. To remove them is not to liberate the human; it is to diminish something essential about the human.

This is why Karpathy is in “psychosis” and not paradise. His output has never been higher. His productivity has never been greater. And something in him knows that productivity was never the point.

The Oldest Question in the Newest Context

Psalm 8 asks the question that I believe the church must answer with clarity and conviction in this cultural moment: “What is man that You are mindful of him, and the son of man that You visit him?” (Psalm 8:4, NKJV).

Three thousand years old, and it has never been more urgent.

The naturalist answer, the one that undergirds most of Silicon Valley’s anthropology, is that man is an evolutionary product. A biochemical computer. Consciousness is either an illusion (Dennett), an emergent property of neurological complexity (Searle), or a fundamental feature of all matter (Chalmers). On any of these accounts, there is nothing categorically special about human consciousness. It’s a phenomenon that arose from physics and chemistry, and there’s no principled reason it couldn’t arise again in silicon.

If that’s true, then AI really is the next step. Karpathy’s agents really are his successors, not his tools. And the meaning void he’s experiencing is just a transitional discomfort on the way to a post-human future. Get over it. Adapt. Or become what Yuval Noah Harari chillingly calls the “useless class”: humans with no economic function and no framework for understanding why that feels like death.

The biblical answer is different in kind, not merely in degree. Man is an embodied soul, created in the image of God, with consciousness that is divinely given rather than emergently produced. “And the Lord God formed man of the dust of the ground, and breathed into his nostrils the breath of life; and man became a living being” (Genesis 2:7, NKJV). Human dignity is not a function of cognitive capacity or economic productivity. It is a function of divine origin. You bear the imago Dei whether you’re writing code or unable to write your own name.

This means that work, while genuinely important, is not the source of human identity. It is an expression of human identity. The image-bearer works because God works. The sub-creator creates because the Creator creates. But the identity precedes the activity. You are not what you produce. You are who you were made to be.

This distinction is not academic. It is the difference between a framework that can diagnose Karpathy’s “psychosis” and one that cannot. The instrumental view says: nothing is wrong, output is up. The vocational view says: something is deeply wrong, because a human being has been reduced to a bottleneck in his own system, and his soul knows it even if his metrics don’t.

What the Church Has Always Known

I don’t have a clean five-point plan for navigating the AI transformation of work. Anyone who tells you they do is selling something. The disruption is real, it’s accelerating, and the second-order effects (the pipeline paradox, the meaning void, the self-improving systems) are genuinely unprecedented in their speed and scale.

But I want to name something that I think matters more than any policy recommendation.

The builders are experiencing a meaning crisis, and they don’t have language for it.

Karpathy reaches for “psychosis.” It’s the best word he has. He knows something is disordered. He can feel it. But his framework, the naturalist, instrumentalist framework that dominates the world he operates in, doesn’t have a category for what’s wrong. If you are your output, and your output is better than ever, then you should be flourishing. The fact that you’re not is, on that framework, inexplicable.

A theologian would recognize it immediately. This is a soul that has built its house on the sand of productivity and is feeling the foundation shift. This is a human being confronting, perhaps for the first time, the question that every pastor hears in every season of life: Who am I when the thing I do is taken away?

The church has been answering that question for two thousand years. Not perfectly. Not always well. But consistently, and with a framework that actually accounts for the data of human experience. You are not your job title. You are not your output. You are not your token throughput. You are a creature made in the image of the living God, and your worth was established before you ever produced a single thing.

“For by grace you have been saved through faith, and that not of yourselves; it is the gift of God, not of works, lest anyone should boast” (Ephesians 2:8-9, NKJV).

That verse is usually preached in a soteriological context, about salvation. And it is about salvation. But it also contains an anthropological truth that speaks directly to this cultural moment: your standing before God, your identity, your worth is “not of works.” It is gift. It precedes performance. It cannot be automated, optimized, or made obsolete by a more efficient system.

Karpathy says the things that agents can’t do are your job now (No Priors, 2026). He means it as career advice. But there’s a deeper truth in it than he may realize. The things that agents can’t do — love, worship, repent, forgive, hope, grieve, marvel at beauty, wrestle with God in the dark night of the soul — those aren’t your job. They’re your life. They’re the capacities that no amount of optimization will ever touch, because they aren’t functions to be performed. They are expressions of a soul that was breathed into existence by a God who is mindful of you.

The builders have stopped building. The question now is whether they’ll discover that they were always more than builders.

The church should be ready with an answer. We’ve had one for a very long time.

---

Sources

1. No Priors Podcast, “Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI,” March 20, 2026

2. Sarah Guo (@saranormous), X post announcing episode, March 20, 2026

3. Tech Layoff Tracker (@TechLayoffLover), Block AI replacement layoffs, March 2026

4. Reuters, “Meta planning sweeping layoffs as AI costs mount,” March 14, 2026

5. Aakash Gupta (@aakashgupta), “Salaries become servers,” 2026

6. Tech Layoff Tracker (@TechLayoffLover), CS grad placement collapse (16%), March 2026

7. Tech Layoff Tracker (@TechLayoffLover), New grad requisitions down 89%, March 2026

8. Aakash Gupta (@aakashgupta), Big Tech new grad hires at 7%, March 2026

9. Tech Layoff Tracker (@TechLayoffLover), Stanford CS grad — 847 apps, zero offers, March 2026

10. Anthropic Claude usage study — 75% of programmer tasks covered, March 2026

11. Julia McCoy (@JuliaEMcCoy), “Who are you without your job title?”, March 2026

12. LayoffAI (@LayoffAI), Q1 2026 tech layoff tracker

13. ---

    This article was developed using AI writing tools I built to work with my voice, research, and editorial framework. The ideas, arguments, and theological positions are mine. The pipeline that helps me draft, evaluate, and refine them is something I created as part of my work at Nomion AI. I believe in building with AI and being honest about it. If you want to know more about that process, ask me.