Most AI evals measure the wrong thing
Benchmarks tell you how a model does on someone else's problem. Your users only care about theirs.
Editorials
Long reads, short notes, the odd video and photographs from the road. Published whenever there's something worth saying.
Benchmarks tell you how a model does on someone else's problem. Your users only care about theirs.
Fifteen minutes I come back to every November. No commentary needed.
Two scooters, one spreadsheet and the fastest feedback loop I've ever had.
Working backwards works. Working backwards and then editing works better.
Every tooltip tour I've shipped was an apology for a default we didn't have the nerve to choose. Pick the default. Let people change it later.
If a chart doesn't change what someone does on Tuesday, it's decoration. Most dashboards are a gallery.
The clever version needs data you'll only get by shipping the boring one.
Four days, one motorcycle, a lot of fog. Photographs from the road.
Benchmarks tell you how a model does on someone else's problem. Your users only care about theirs.
Two scooters, one spreadsheet and the fastest feedback loop I've ever had.
Working backwards works. Working backwards and then editing works better.
Fifteen minutes I come back to every November. No commentary needed.
Shared because I watched it instead of finishing a spec.
Four days, one motorcycle, a lot of fog. Photographs from the road.
Every tooltip tour I've shipped was an apology for a default we didn't have the nerve to choose. Pick the default. Let people change it later.
If a chart doesn't change what someone does on Tuesday, it's decoration. Most dashboards are a gallery.
The clever version needs data you'll only get by shipping the boring one.