yes i know you can feel it too

Most of this site was made together with a language model: Claude, by Anthropic, working in the site's repository and in my notes. The division of labour was this. The site's concept is mine: a page with no job title, a wave where the title would stand, two faces for one name, the look with its two colours and one typeface, and the rule that nothing on it needs updating to stay true. I drew the mockups the pages were built from. Mine too are the projects, the exhibits, the talks and the research the articles are about, the study behind the touch table, the reading behind the essay, and the decisions in every piece: what goes in, what stays out, what a figure shows. The model wrote most of the code to my brief and my corrections, drew the diagrams from my descriptions and measurements, and wrote the first drafts of the articles from my talks, notes and documentation. I edited those drafts, several times over, cut what was wrong, rewrote what I kept, and sent them back when they still read like a machine. The opinions are mine, and so are the mistakes that are left.¹

What it's bad at

Writing is where I trust it least. Its prose drifts toward the average of what it has read: fluent, confident and a little the same everywhere. That's more than a matter of taste. In an experiment by Anil Doshi and Oliver Hauser, short stories written with a model's ideas were rated better one at a time and came out more alike as a group.² The fluency also covers for errors. Asked for short literature reviews, one model produced references of which more than half didn't exist, and its successor still invented almost a fifth.³ So no reference on this site stands on the model's word. Each one was checked against the work before it went in: that it exists, that the details are right, and that the note says what the work says. The model's habits of sentence are a problem of their own: a contrast where none is needed, a remark on its own honesty, a quotable line to close a paragraph. I try my best to edit away this tendency to spare your and my sanity while reading this.

What it's fairly good at

Code is different, and the evidence there is mixed. In one controlled experiment, developers with GitHub's Copilot finished a small, well-specified programming task 56 percent faster than developers without it.⁴ In another, experienced developers working on large projects they had maintained for years were 19 percent slower with AI tools, and believed afterwards that the tools had sped them up.⁵ My experience sits between the two. It helps most where I can check the result myself: a page renders or it doesn't, a drawing matches the photo or it doesn't. It costs time wherever I can't.

What it makes is mid

There's a more general point, and it's the one I'd make to a client. Ask a model for a user interface, an illustration, a layout, a poster, and you get something. It's mid: not bad, not good, OK. OK is the problem, because OK is what the thing was built to produce, the average of everything it has seen.⁶ A designer sees that at once, and sees what to stay away from. Someone without the training doesn't, and that isn't a failing. Nobody can tell mid from good in a field they haven't practised; I can't in law or in medicine, and I'd be a fool to trust my own eye there. So the work falls to two sides. Designers have to say, to clients and to each other, what these tools are for and what they aren't, and say it before the mid version is in production. And clients have to listen to the people who've done the training, the way the rest of us listen to the plumber. That's the deal in every profession. The models only made the mid version cheap.

Why use it anyway

So why use it at all? Because the alternative to this site was no site. I build exhibits and teach. Writing up projects, making a place for them and keeping it current is work I wouldn't have found the time for, and the projects here would have stayed slides and phone photos in a folder. That's a trade, and this page says which one I made.

What it costs

It also costs electricity, water and hardware, and the total is growing. The International Energy Agency estimates that data centres used about 1.5 percent of the world's electricity in 2024 and expects that to roughly double by 2030, with AI as the main reason.⁷ Per request, the figures are small and argued over. Google reports 0.24 watt-hours for a median text prompt to its own model,⁸ and independent measurements find general-purpose generative models using far more energy than smaller models built for one task, for the same job.⁹ Both can be true at once. What reaches a power plant is the total, and the total is the agency's number.

Where I think the effort belongs

I don't think the answer is for each of us to ration our prompts. That move has been made before. Geoffrey Supran and Naomi Oreskes traced how ExxonMobil's public communication shifted the responsibility for emissions from the companies producing fossil fuels to the people using them, as the tobacco industry had done with smoking.¹⁰ Plastic recycling was promoted for decades by an industry whose own documents, according to a report from 2024, doubted that it could work at scale.¹¹ And where electricity did get cleaner, it was largely because policy created markets for new technology, Germany's feed-in tariff for solar among them, and not because people used less of it.¹²

For AI the levers exist, and they're young. The EU's AI Act requires providers of general-purpose models to document their models' known or estimated energy consumption,¹³ and the Energy Efficiency Directive obliges data centres above 500 kilowatts to publish their energy and water use.¹⁴ Enforcing these, extending them from the training of models to their use, and voting for people who will, is where I think the effort belongs. Using a model less where it adds little is still sensible. It doesn't add up to a policy.

What this note isn't

This is a note about one site and one person's use of these tools, not a review of the field, and its numbers will date quickly; some may have by the time you read this. It also leaves the larger questions, how these models are made, the data they're trained on and the work that goes into them, to people who have studied them. If something here is wrong or out of date, write to me.