Signals

We just shipped a long-context eval harness that runs ~40× cheaper than the open ones. The trick wasn't a smarter judge — it was caching the rubric tokens. Writeup tomorrow.

Omnea is hiring our second founding engineer in London. TypeScript + a real product surface to own. If you've ever wanted to write the schema AND the marketing site in the same week, DM me.

Promoted to Staff at Anysphere this week. Grateful to the editor team — building Cursor has been the most fun I've had writing software. Onward.

Hot take: most agent failures aren't reasoning failures, they're memory failures. If your agent can't remember what it just decided, it's not going to get better with a bigger model.

Speaking at LocalFirst Conf in Berlin next month on building a CRDT-backed dev environment that actually feels native offline. Slides will be public. Come say hi if you're around.

After 3 months of interviewing eng candidates, the strongest signal isn't the system design round. It's whether they can describe a previous decision they'd reverse — and why.

Modal raised our Series C. We're going to spend it on the unglamorous stuff: better cold-starts, better GPU scheduling, and a much cleaner billing surface. Building for the next 10 years.

Reminder that for most teams 'we need vector search' really means 'we need better metadata filters'. Start there. You can always add embeddings later.

Edge runtime gotcha of the week: setTimeout doesn't always survive between requests on certain cold paths. If you're using it for debouncing — don't. Queue it.