The Frontier
of Intelligence

Engineering write-ups from work we have actually shipped — evaluation, agent architecture, RL post-training and the infrastructure underneath. It is not a news feed and it is not thought leadership.

The delivery and research team. Posts carry named authors, and the series posts are written by the people who built the thing being described.

When there is something worth writing up, which works out at roughly monthly. We would rather publish six useful posts a year than fifty filler ones.

Yes — the newsletter goes out weekly with the latest write-ups. You can sign up from any post without creating an account.

Quote freely with attribution and a link back to the source post. For full republication, ask us first.

No. Everything here is written by people who did the work described, which is the only reason it is worth reading.

The AI systems guides are the structured entry point; the blog assumes you already know roughly what an evaluation suite is. Start there and come back.

Yes, and where a figure comes from a specific engagement we say so. Benchmarks are reproducible from the public evaluation harness with the same seeds and hardware.

Usually. Book a working session and mention the post — we will put the right person on the call.

Sometimes. If several people ask the same question it tends to become a post, so asking is worth doing.