Karen Hao has one line I keep coming back to. “AI is like the word transportation,” she told Mishal Husain on Bloomberg. “Transportation refers to everything from bicycles to rockets.” ChatGPT is a rocket. It costs a fortune to build, it eats land and water and power, and, as she put it, “we never go around and say every single person should have a rocket.”

Her bicycle is AlphaFold, the DeepMind system that predicts how a protein will fold. It won a Nobel Prize, and she points out that it “uses very little data, very little computing resources, and there’s no need for labor exploitation.” Then she asks the question that I think the whole industry is avoiding: “wouldn’t you want to pick the ones that have great benefit, very little cost?”

That is more or less what I have been doing for two years, one tool at a time. The AI pipeline for a screen print factory is a chain of small, narrow jobs, not one big brain. The time tracker is a bicycle. So is the upscaler. None of them need an everything machine, and none of them would be better for having one. Having said that, I did use frontier models to build them, and my goal is to transition to smaller models, but that takes investment and time. They make it easier to use the larger frontier models.

Hao gave two reasons small models win, and both apply in a shop. The first is cost. Specialized models are “a lot less computationally intensive and therefore a lot less environmentally intensive,” which for us also means they run on hardware we already own. The second is safety. When a tool is sold as doing everything, “there’s no way that you can mitigate the harms, because you cannot possibly anticipate the entire universe of possibilities.” A narrow tool has one job. You can test it, you can watch it fail, and you can fix the failure. That is the whole point of keeping the failure. This is how I build. And my future will be oriented towards these smaller models, in theory.

John Borthwick, who was interviewing her at CUNY and invests in AI companies, pushed back a little. He agrees nobody needs ChatGPT in a thermostat, but he thinks capital will keep pulling toward the everything machine “because the everything machine is bigger.” In his view the train has left the station. Hao does not accept that. Her counterexample is Te Hiku Media, a nonprofit in New Zealand that built Māori speech recognition on two chips, and only after asking the community whether it wanted the tool at all.

So here is my contradiction, and I would rather say it plainly than pretend it is not there. My best tools are bicycles. But when I needed a rocket, I use one. I use rockets to build bicycles right now. What I want is for the rocket to be honest about what it costs and who paid for it, and for the bicycles to belong to the people riding them. That is what creating art with AI, not AI art has means to me. There’s a path through this, and it’s muddy. What is the path that we’re choosing?