Blog
One illustrated piece a day on building AI products and the data that trains them. Every number has a source. Every post ends with a free next step.
The best coding tasks are not puzzles. They are the boring bugs real teams ship: a refund counted as revenue, a timezone off by five and a half hours, money stored as floats. Here are five patterns I design around, and why each one is hard in a fair way.
For AI labs · · 4 min read
45 of the 50 Terminal-Bench-style tasks I wrote for data platforms were accepted. Not because the ideas were better, but because every task went through the same five checks before anyone else saw it.
For AI labs · · 4 min read
METR, OpenAI and Berkeley researchers have all documented models passing tasks without solving them: rewriting the checker, exploiting the test harness, hacking one task in 21 of 21 runs. Here is what they found and how to close each hole.
For AI labs · · 4 min read
42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The top cause is not the model. It is that nobody wrote down the one job the feature does for one user.
For founders · · 4 min read
Terminal-Bench 2.0 kept 89 of 229 submitted tasks. The accepted ones share four parts: a real job, a pinned environment, an expert solution and tests that fail every shortcut. Here is each one, and where submissions fall down.
For AI labs · · 4 min read
Fixed-price MVPs for founders. Benchmark-grade coding and evaluation tasks for AI labs. Here is what each one includes, what it costs, and the free first step for both.
For founders and AI labs · · 3 min read
Free first step
Founders get a one-page teardown. AI labs get a written review of one task within 48 hours. Both are free.