AI systems / 18 Jun 2026

Building factories, not features

Naval Ravikant ran a conversation recently with a few people who build real things: the software infrastructure a lot of the internet runs on, supersonic jets, medical devices. They called what AI is doing to their work an industrial revolution. I think that is the right word, and not for the reason most people reach for. It is not that the robots are coming. It is that the unit of work has changed. We have started building factories instead of features.

The unit of work changed

You used to build the thing. Now you build the thing that builds the thing.

I notice it in my own week. I do not write a customer reply, I build the system that writes thousands of them. I do not write a blog post for a store, I built the system that writes fifteen a week and waits for a human to approve them. The reply and the post are output. The artefact I actually made is the factory behind them.

Guillermo Rauch, who runs Vercel, put it cleanly on that call. The job of the engineer is now to produce the factory that produces the output. Once you see work that way, a lot of old habits invert.

Waste tokens, save time

Here is the habit that inverts first. Naval's line is "waste tokens, save time." Stop treating the model's running cost as the thing to optimise. The scarce resource is your time and your attention, not the tokens.

It feels wrong at first, because we were all trained to be careful with compute. But the sums are not close. A few dollars of model usage to save an hour of an expert's afternoon is a trade you take every single time. Throw the problem at the model three different ways. Run it twice. Let it explore routes you would never have hand-written. The waste is the point, because the expensive thing in the room is you.

You do not get stuck anymore

If I had to name the single biggest change, it is not that AI writes good code. It is this. You do not get stuck anymore.

Every builder knows the old failure mode. You hit a wall at two in the afternoon and you are still there at midnight, one error message deep in a problem nobody else can see. That used to cost days. It is mostly gone. When you get stuck now, the model tries another route, and another, until the wall moves. The point is not that the work became perfect. The point is that the floor came up. You rarely lose a whole day to a single wall anymore, and the psychological difference in that is enormous.

The model became a colleague, not a tool

The models grew up while we were watching. A year ago they answered like a keen junior. Now they answer more like a principal engineer. You ask a question and they come back with three routes and the trade-offs between them, not one nervous guess.

That changes your job more than it changes theirs. When the model proposes the options, your work is to choose well. To verify. To say which of the three is right for this business, at this moment. I have written before that agents need a human at the gate, and this is the same truth from the other side. The machine does the producing. The human does the deciding. As the producing gets cheaper, the deciding is most of what is left, and it was always the hard part.

It is not only code

The reason "industrial revolution" is the honest phrase is that this is escaping software.

On that call, a supersonic-jet founder described turbine blade design going from one engineer for one day a blade to two engineers designing a whole engine. Compliance documents that took months now take minutes, because the rules can be fed in as searchable text and checked the way you would run a test suite. The leverage is leaking out of the screen, into metal, and into the paperwork that used to gate the metal.

You do not need a jet to feel it. In the business I run, the same move keeps showing up. Someone who is not an engineer now builds the small tool they always needed by describing it to an agent, instead of waiting in a queue for someone like me.

The autonomous company, in miniature

This is where it gets real for ordinary companies. The line from that conversation I keep repeating: your job is not to work on the thing, it is to train the agent that works on the thing.

That is already how the business I run operates. The team does not answer every comment, draft every campaign or pull every report by hand. They supervise the agents that do, and they step in where judgement is needed. One person now runs the output of what used to be a department. The future people describe as mass unemployment looks, up close, more like a much larger number of much smaller teams.

What stays human

It is fair to ask what is left for us, if the machine produces and the producing is cheap.

The answer the builders kept landing on was creativity, taste and surprise. The model is brilliant inside the distribution of everything it has already seen. It is weak at the genuinely new, the idea that is off the map, the decision about what is even worth making at all. Intelligence multiplies, but only when a person with agency and taste points it at something. A generalist with judgement and a hundred agents now beats a narrow specialist, because the specialism is the first thing the machine learns to copy.

That is a good outcome for builders and makers. The bottleneck moves from "can you make it" to "do you know what is worth making." The second question was always the more interesting one.

The rule

Stop building features. Build the factory, then point it at something that matters.

The work was never the scarce thing. Deciding what is worth making was. That part is still yours.