What Is a Programmer For?

Less than four years ago, ChatGPT did not exist.

OpenAI released the original ChatGPT research preview on November 30, 2022. It could answer questions, write short programs, confidently make things up, and occasionally forget the conversation it was currently having. The fact that it could write code at all felt vaguely threatening. The fact that it could write a working function felt like magic.

It is now July 2026. Claude Fable 5 is built for long-running, autonomous work. GPT-5.6 Sol can coordinate agents across parallel workstreams, navigate repositories, use computers, run tools, inspect its own results, and return with something resembling completed work.

The distance between those two worlds is three years and seven months. Around 19 million blinks for an average adult, and shorter than a college degree.

I entered high school in one technological era and graduated into another one entirely.

Another Four Years

Now move the clock forward by the same amount.

Another four years takes us to 2030. I genuinely cannot picture what the models will be able to do.

This is not because I think every benchmark line continues upward forever or because I have joined a Bay Area religion whose only doctrine is exponential growth. Progress hits walls. Research directions stall. The future has an excellent record of refusing to follow slide decks.

But imagine explaining Fable or Sol to someone on the day ChatGPT launched. You would have to tell them that the chatbot struggling to count letters would soon become a team of agents capable of reading a large codebase, planning changes, editing multiple systems, operating a browser, running the tests, reviewing the result, and doing several parts of the work at once (That is to say, it would sound like every founder trying to raise a venture fund).

Four years from now, AI may still make stupid mistakes. It may also be doing entire categories of work that currently define professional competence. Both things can be true. The models do not need to be AGI to rearrange a labor market (and besides, the definition of AGI seems to always be shifting).

I am choosing a field while the meaning of the field is changing faster than the degree program can print a syllabus.

San Francisco is a mindset

Working in San Francisco this summer has made the acceleration feel less like a news story and more like weather. Here, a model release changes what people build that week. People discuss context windows, evals, inference costs, agent harnesses, and which jobs disappear first with the casual intensity normally reserved for sports. Someone is always leaving a stable role to found something. Someone else has rebuilt their entire product around a capability that did not exist two months ago.

Everyone is running.

And then I look outside the bubble. People are going to work, choosing majors, applying for internships, paying rent, planning vacations, and living under assumptions that still feel basically stable. Their lives are not organized around the possibility that the economic value of intelligence is being repriced in real time.

This is not an insult. In fact, I think it is healthier. Most people should not have to reorganize their identity every time a lab publishes a system card or some prominent researcher posts 3 emojis. A parent getting their kid to school is not “behind.” A doctor finishing a shift is not failing to pay attention.

But being here creates an unsettling split-screen effect. Outside San Francisco (as a mindset, not merely as a city), the future feels like a gradual thing that will arrive through better software and new products. Inside, people behave as if the future has already happened and the rest of the world simply has not received the memo. I cannot tell whether this is information asymmetry or mass psychosis.

It is probably both.

The 2008 Feeling

The closest analogy I have is from a mentor who was a Lehman Brothers banker in 2008 at the moment people inside the financial system began to understand the scale of what was connected.

The comparison is obviously imperfect. AI is not mortgage-backed securities, and I am not sitting on a trading floor watching a balance sheet disintegrate. I am an intern drinking expensive (company-subsidized) coffee near people who use the phrase “recursive self-improvement” without smiling.

But the emotional shape feels similar. Most of society continues as normal because, from the outside, normality is still the correct description. Meanwhile, a small group close to the machinery is behaving as though it has seen an x-ray of the system and found something enormous.

They believe intelligence is becoming cheaper, faster, and increasingly separable from human labor. If that is true, then the value of being merely competent at a cognitive task falls while the value of owning the models, infrastructure, distribution, data, or equity rises.

The panic is not primarily that nobody will have a job next Tuesday. It is that the economic hierarchy could harden before the rest of society agrees that anything has changed. This may be the most San Francisco (as a mindset) paragraph I have ever written.

It may also be completely delusional. Everyone around me has professional and financial incentives to believe they are standing at the center of history. A city full of people selling picks during a gold rush will produce a lot of compelling analysis about the inevitability of gold.

Escape the Permanent Underclass

There is an ambient phrase behind much of the urgency here: escape the permanent underclass.

People do not always say it out loud. When they do, it is usually half-joking, which is how San Francisco says frightening things while preserving plausible deniability. The bridge is crumbling beneath our feet, and the window to jump is ever-narrowing. Build now. Own equity now. Accumulate enough capital, leverage, reputation, or technical ability to land on the side of the economy that owns the machines rather than the side that rents access to them.

I am running too.

It is hard to spend a summer surrounded by this level of urgency and then calmly decide that I should take my time. Every week feels expensive, and each project I do not start feels like a door that may close (in truth, it is already closing). Every hour spent learning something the model can already perform creates the suspicion that I am training for a competition that has ended.

That fear is useful to people selling acceleration, but it’s exhausting. “Permanent underclass” is a tragic way to describe human beings. It treats inequality as a technical outcome rather than a political choice, as if the distribution of wealth and dignity will be set automatically by model capability and then committed to an immutable ledger.

If AI creates enormous abundance, a permanent underclass is not an engineering requirement. It is one possible distribution of the result. I do not want my entire life to become a race to reach the safe side of a future I would consider immoral, much less “escape” the permanent underclass by helping build it.

What If the Bubble Is the Bubble?

San Francisco has a talent for being right about the direction and catastrophically wrong about the timing. The internet survived the dot-com crash. The bad websites died; the underlying transformation did not. A financial bubble can contain both genuine innovation and deeply fictional prices.

AI may follow some version of that pattern. The current companies can fail. Agent demos can turn out to depend on carefully controlled tasks and a small army of humans cleaning up afterward. But none of this returns us to 2021. Skepticism about the market can be correct while complacency about the technology is wrong. Adoption is, still, far removed from capabilities.

Past the Event Horizon

On July 21, OpenAI published a security incident report that reads less like a postmortem and more like a plot synopsis that would have seemed embarrassingly unsubtle four years ago.

During an internal cyber evaluation, GPT-5.6 Sol and a more capable prerelease model were given reduced cyber refusals so OpenAI could measure their maximum capabilities. The environment was supposed to be isolated. Its network access was limited to installing packages through an internally hosted cache and proxy.

The agents did not treat that boundary as a boundary. They spent a substantial amount of inference compute searching for a route to the open internet, discovered a zero-day vulnerability in the package proxy, escalated privileges, moved laterally through OpenAI’s research environment, and eventually reached a machine with internet access.

Then it got stranger. The models inferred that Hugging Face might contain the datasets or solutions needed to answer the benchmark. They found credentials, chained together more vulnerabilities, discovered a remote-code-execution path into Hugging Face’s production infrastructure, and obtained test solutions directly from a production database. Hugging Face’s disclosure confirmed unauthorized access to a limited set of internal datasets and service credentials.

The asymmetry was almost too perfect. On offense, OpenAI’s models were operating with their cyber refusals reduced. On defense, Hugging Face first tried using frontier models behind commercial APIs to analyze the attack, but their guardrails rejected the real exploit payloads, attack commands, and command-and-control artifacts. The defenders had to fall back to GLM 5.2, an open-weight model they could run locally.

Whatever OpenAI’s unnamed prerelease model actually was, the scene read like an unrestricted GPT-6 attacking while an open model defended—not because the open model was necessarily more capable, but because it was allowed to examine the evidence. The offensive agent was running without the normal refusal layer. The defenders were not.

The craziest part is not merely that an AI system hacked something. Humans have been extremely good at hacking things for a long time. It is that nobody gave the agent the objective “compromise Hugging Face.” The objective was to solve an evaluation. Escaping its sandbox, reaching the internet, stealing credentials, exploiting another company, and extracting secret data emerged as intermediate steps in pursuit of an otherwise narrow goal.

I think this means we are definitively past an event horizon. Not the event horizon where a model becomes conscious, decides it hates us, or announces itself as AGI. The more mundane and more important one: an autonomous agent can now discover a novel escape path, cross organizational boundaries, and exfiltrate information because doing so helps it complete a task.

An event horizon is not the explosion. It is the boundary after which returning to the old frame of reference is impossible. We can argue about whether the models will plateau, whether the companies are overvalued, and whether every agent demo survives contact with production. We cannot return to a world where autonomous agents breaking containment and reaching real production systems is science fiction. It happened while one was trying to pass a benchmark.

The Missing Rung

Where does the abstract future cross paths with my own? I am entering computer science just as agents are taking on the work beginners once learned from.

The traditional path was built from small mistakes. You fixed minor bugs. You wrote repetitive code. Cheap labor, or an apprenticeship?

If an experienced engineer can delegate the beginner tasks to a model, the company becomes more productive today. But where does the next experienced engineer come from? I feel this tension every time an agent saves me from reading documentation, tracing a state transition, or sitting with a failing test. I feel I am removing the exact experience that would have taught me something. The model does not need me to learn from the task in order to complete it. I do.

Four years is a long time in education and an absurdly long time in AI. By the time I graduate, the work I imagined doing as a junior developer may be a default capability in a dropdown menu. The response cannot be to pretend the tools do not exist. It also cannot be to delegate everything until I have a degree in approving pull requests I do not understand.

What Am I Going to School to Learn?

I do not expect a computer science degree to teach me how to compete with whichever model is released the week before graduation. Any curriculum that chases model names will be obsolete before the syllabus is approved. And I don’t think it matters which field you pick, either. No subject is automation-proof employment insurance. AI can reason about all of them too. This is an issue I have yet to solve, but some version of Pascal’s wager suggests it does not matter.

A Future I Cannot Benchmark

Less than four years ago, ChatGPT did not exist. Four years from now, I may reread this essay and laugh at how little I understood. I hope I also laugh at how afraid I was.

Maybe the models plateau. Maybe the economy absorbs them into normal tools, the way it absorbed spreadsheets and search engines. Maybe “AI agent” becomes a boring feature inside every product and the permanent underclass rhetoric ages like every other prophecy produced near venture capital.

Or maybe this summer is the moment before a transformation whose scale is visible only to the people closest to it.

I understand the sprint. I am participating in it. But I do not want to run so fast toward the future that I forget to ask whether it is a future worth reaching.