Human obsolescence.

Last Tuesday I shipped a brief that used to take me an afternoon. The model did the research, the outline, the first three drafts, and a cleaner fourth. I spent eleven minutes on it — two of those minutes were me staring at the screen deciding whether I still needed to be in the loop.

The client could not tell.

That is the whole essay. Not "AI will take jobs." Not "humans will always be needed." The more precise thing: a specific layer of the work is already gone, and most of the public argument is happening one or two layers above the one that actually moved.

The wrong conversation

The obsolescence discourse has two official positions, and they are both slightly dishonest.

The first says nothing important is changing — that models are autocomplete with a marketing budget, that "real work" will always need a person, that last year's job description is a law of nature. This is usually said by people whose week has already been rewritten and who have not updated the story they tell about themselves.

The second says everyone is about to be replaced, so the only rational move is panic, regulation, or a personal brand about being "irreplaceably human." This is usually said by people who have not yet sat next to a production agent on a Tuesday and watched which parts of the job actually fail.

Both positions treat "the human" as a single object. That is the mistake.

A job is not a soul. A job is a bag of verbs. Some of those verbs are already cheaper, faster, and more consistent when a model does them. Some are not. The interesting work is naming the layers — and then deciding, on purpose, which layer you still want to occupy.

Layer 01 — a verb is gone

Task obsolescence is not a prediction. It is a changelog.

In the last twenty-four months, inside this lab and in the systems around it, a set of verbs quietly transferred:

• first drafts of almost anything with a known shape
• research that used to mean twenty tabs and a half-finished Notion page
• first-pass code, then second-pass code, then the refactor you were avoiding
• the outbound call that used to wait for an SDR to finish lunch
• the meeting summary, the changelog, the "can you just write this up"

None of these were the job. They were the calories the job burned. When they leave, the job looks the same on LinkedIn and completely different at 9:14 on a Tuesday.

The tell is not that the model is "better than you." The tell is that you stop reaching for the verb. You open the agent instead. After a few weeks the old motion feels like writing a letter when you could send a text — technically still possible, emotionally ridiculous.

Layer 02 — the bag of verbs thins out

Roles die slower than tasks, which is why people feel gaslit.

A "writer" is still a writer after drafting leaves. A "researcher" is still a researcher after the tab-work leaves. An "SDR" is still an SDR after the first call is a voice agent. The title survives because titles are social objects. The Tuesday does not survive, because Tuesdays are made of verbs.

What actually happens is compression. The role that used to be 80% production and 20% judgment becomes 20% production and 80% judgment — or, more often, 20% production, 40% judgment, and 40% system work nobody put in the original job description: evals, retries, schema drift, the dashboard, the 3am page.

If you add those layers without weighting, you get a comforting story ("humans still do the important parts"). If you weight them by hours, you get the RemainMeter: most of the week is already a system. The remaining human is not a smaller version of last year's human. It is a different job wearing the same name.

This is also why "just learn to prompt" is a weak strategy. Prompting is a task. Tasks transfer. The people who stay in the bag are the ones who own the outer loop — what the system is for, when it is allowed to act, and who is awake when it is wrong.

Layer 03 — you thought you were the verb

Identity obsolescence is the one that hurts, and the one the internet cannot stop performing.

If your self-image was "I am the person who writes the brief" or "I am the person who takes the first call" or "I am the person who can hold the whole codebase in my head," then a model that does those verbs for eleven minutes is not a tool. It is an insult.

A lot of the loudest takes are grief with a thesis attached. That is allowed. Grief is data. It is not a strategy.

The useful move is to retire the verb as identity before the market retires it as a role. Keep the taste. Keep the accountability. Let the calories go.

A working axiom: if the part of the job you would put on a conference badge is the part a model can now do unattended, you are already late to the next layer.

I have had to do this twice in two years. The first time I was "the person who ships the agent demo." The demo became a weekend. The remaining job was the boring middle — evals, retries, the Tuesday. The second time I was "the person who writes the long thing." The long thing still needs a person. The blank page does not.

What the remaining human is for

Not poetry in the loop. Not a ceremonial "human review" checkbox that rubber-stamps whatever the model emitted. Not "bringing the soul."

In practice, the remaining 38% looks like this:

1. Taste under constraint. Deciding what to build, what to ignore, and what "good" means when the generator can emit a hundred plausible versions before lunch. Taste without a constraint is vibes. Taste with a constraint is the job.

2. Accountability. Someone still gets the 3am page. Someone still signs the thing that ships. Someone still sits across from the client when the brief is wrong. Models do not take pages. If your system has no named human, it is not autonomous. It is unowned.

3. The boring middle. Evals that don't lie. Observability. A dumb deterministic outer loop. The work I already wrote about, because it is the only work that turns a demo into a system. This is not glamorous, which is why it is still available.

4. Distribution and trust. The model can write the post. It cannot be the person the market remembers as having been inside the window. It cannot take the call after a bad one. Trust is still a human residue.

The 13 free hours are the fork. You can spend them producing more of the work that just became free — more drafts, more calls, more first-pass code — and become a very fast version of the person you were in 2024. Or you can spend them on the layer that did not transfer: choosing the window, owning the failure, publishing from inside the work before it is obvious.

One of those is a volume strategy. The other is a position.

How to occupy the remaining layer on purpose

Three things I now treat as non-negotiable, because the alternative is waking up inside a role that has already thinned out and pretending it hasn't.

Name the verbs that have left. Write them down. If you cannot list the parts of your week a model now does without you, you are arguing about identity while the changelog has already shipped.

Move your name onto the outer loop. Version the prompt. Own the eval suite. Be the person who can answer "why did it do that on May 14 at 09:02?" That is a job. "I write the first draft" is increasingly not.

Spend the free hours on a window, not on more calories. The 12-month-before-it's-obvious window is still the highest-leverage place an independent technical person can sit. The model makes it cheaper to operate inside that window. It does not tell you which window to pick. That refusal is the remaining human.

Human obsolescence is real. It is also narrower, earlier, and more useful than the keynote version.

A layer of the work is gone. The title has not noticed. The identity is grieving. The remaining human — if you take the job — is the person who decides what the system is for, and who is awake when it is wrong.

Pick the layer. Let the verb go. Keep the page.