Soon anyone can make anything. That’s when it starts.

Four small status lights on a modem. That was my whole day.

I am building my own internet provider. Somewhere on the site there is a block showing a modem: power, DSL, internet, WiFi. When the fixed line drops, one light turns amber and the label underneath changes. It is a tiny piece of interface nobody will ever thank me for. And it took a full working day.

The AI produced dozens of variants in minutes. The animation was too fast, then too slow. The amber was wrong. The spacing between the lights made it look like a Christmas decoration instead of hardware. The pulse was symmetrical, which real LEDs never are. Every single correction that turned it from “technically working” into “actually right” came from me. Not one of them came from the model.

So here is the thing that keeps me up at night, and it is not the code. It is what happens when everyone can generate forty variants of anything in ten seconds, and almost nobody can tell which one is the good one.

The crisis is real. Pretending otherwise is not a strategy.

Let me be blunt about the part most people skip. A severe economic disruption is coming, and it will not look like the previous ones.

Earlier waves of automation hit manual labour. Machines took the arm, not the argument. Knowledge work was the safe harbour: go to school, learn to think, sell your judgement. That harbour is now being dredged. Drafting, summarising, first-pass analysis, boilerplate code, standard proposals, entry-level research – this is exactly the work that models do at near-zero marginal cost.

Therefore the pain lands squarely on the people who were told they had done everything right. Junior lawyers. Analysts. Copywriters. First-line developers. The bottom rung of the knowledge ladder is being sawn off, which is also the rung where you used to learn the craft.

That is genuinely disruptive and I refuse to sugar-coat it. But it is not the end of the story. Because when the cost of execution collapses, something else becomes valuable, and it becomes valuable fast.

The limiting factor is no longer money, time or capacity. The limiting factor is whether you know what good looks like, and whether you can keep saying no to the thirty-nine variants that are almost right. Kill your darlings.

What the modem taught me about scarcity

The interesting part of that day was not the output. It was the sequence.

The model gave me options. I gave it direction. I looked at a variant and thought: this is close, and close is the enemy. Then I said no, again, and described what was wrong in words specific enough to act on. “Too fast” is useless. “The amber should ramp up over 400 milliseconds and hold, because a real LED has thermal inertia” is a brief.

Naming what is wrong is a skill. It is not a technical skill. It is the skill you build by looking at thousands of things made by other people and slowly developing an opinion about why some of them work.

Furthermore, the reference material mattered. I knew what a modem looks like on a shelf at dusk. I had seen enough interfaces to know that four evenly spaced dots read as decoration and four unevenly spaced dots read as a device. None of that came from a prompt. It came from having paid attention to the world.

As a result, the economics have flipped in a way that should make every founder sit up. Building something beautiful used to be expensive. You needed a studio, a design team, a front-end developer, a copywriter, three rounds of feedback and six weeks. Now a one-person company can ship work with the polish of a large brand, provided that one person has taste.

That is the renaissance part. Not “AI makes everyone creative.” The opposite: AI makes creation cheap, which means the differentiator moves entirely to judgement. And judgement scales beautifully for the person who has it and not at all for the person who does not.

Why this is an argument for arts education,
not against it

Here is where I lose half the room. Precisely now, in the middle of an AI transition, schools and universities should be investing in arts and culture. Not cutting them.

Every budget cycle, the same subjects go first. Music. Drama. Art history. Studio practice. They are framed as luxuries, and technology is framed as the serious investment. That framing is now economically wrong.

Because what do you actually learn in a decent arts programme?

You learn to look

You learn to sit with something for longer than three seconds and notice why it does not work. This is the exact skill that separates a useful prompt from a useless one. Someone who cannot see the flaw cannot describe the fix.

You learn to finish

Studio work is brutal about this. There is a difference between a sketch and a finished piece, and everyone in the room can see it. The AI era is drowning in sketches presented as finished work. Knowing where the line is has become a commercial advantage.

You learn to take a critique

Work goes on the wall. People argue about it out loud, with reasons. Then you revise. That loop – make, defend, hear it, revise – is the single most transferable habit I know of. Most companies have replaced it with a Slack thumbs-up.

You learn that history exists

Art history teaches you that everything references something. You build on what came before because you know what came before. Models are trained on the whole corpus; if you cannot recognise what they are echoing, you cannot steer it.

So my argument is not “add more craft hours.” That misses the point entirely. My argument is that the working method of the studio – look hard, name the flaw, defend the choice, revise, finish – belongs inside every subject. In maths. In economics. In engineering. In sales training, for that matter.

What this means on Monday morning

This is not only an education debate. It shows up in your company right now, in the quality of what your team ships.

I see it constantly. A team adopts AI, output volume triples, and quality quietly drops through the floor. Nobody notices because there is more of everything. That pattern is the same one I described in AI is making your salespeople sloppier, not sharper: the tool amplifies whatever discipline was already there. No discipline, no amplification worth having.

The uncomfortable question is not “do we use AI enough.” It is “does anyone here still know what good looks like, and do they have the authority to say no.”

• When you look at what your organisation shipped with AI last month, you can honestly say the weak spots are technical rather than a failure of judgement
• Your team knows where the line sits on a task: at “it works” or at “it’s good”, and that line is explicit rather than assumed
• Work gets critiqued out loud, with arguments, before it goes out the door
• There is a scheduled revision round on meaningful work instead of shipping the first acceptable version
• Someone in the room has the standing to reject thirty-nine near-right variants without being told they are slowing things down
• Your next hire or training budget puts technology and culture side by side rather than treating them as competing line items
• You know which subjects get cut first in your local schools, and whether those happen to be the ones that train judgement

Two or more unchecked and you do not have an AI problem. You have a taste problem wearing an AI costume, and adding tools will make it worse.

The ODB Way

Most AI projects I walk into are not failing on capability. They are failing because nobody defined what good looks like before switching on the machine that makes more of it.

So we start there. In FRAME we look at actual output, not at the tool list. Where does quality break down, and is that a model problem, a process problem or a judgement problem? In MAP we make the standard explicit: what “done” means, who decides, and where the revision round sits in the flow. Only then do we BUILD, because automating an undefined standard just industrialises mediocrity. In PROVE the loop stays open: make, critique, revise, ship. The studio method, wired into AI processes.

That is also why I keep arguing that tooling is the easy part. The harder work is designing the process around it, which is exactly the point of a bad sales process will beat a good salesperson every time, and why the AI stack behind ODB Growth is deliberately small and opinionated rather than long and impressive.

The amber light on that modem block is right now. It ramps up over 400 milliseconds, holds, and the label underneath changes to something a human would actually say. The AI wrote every line of it. I made every decision that mattered.

Soon anyone can make anything. That is not the end of the work. That is when the work finally starts.