Colophon
A model drafts nearly every sentence on this site, including the long pieces. This page says exactly which model, exactly what I do on top of it, and exactly what that still fails to catch.
Which models, and for what
Two, doing different jobs.
- gpt-4o drafts the News items. It is the pipeline's writer: it reads the source article and produces the whole post in one pass. I picked a stronger model than the one that writes the Telegram summaries because this is the part that actually gets published.
- Claude drafts the Fundamentals explainers and the Deep dives, working from a subject and an angle I give it, plus whatever source material the piece is about — a paper, a repository, my own notes. Longer pieces need to hold an argument across several thousand words, and it is better at that.
Every post names the model that drafted it, in the box directly under the title. If a post does not name one, nothing drafted it.
What "directed" actually involves
This is the credit on Fundamentals and Deep dives, and it is the one most worth pinning down, because it is the one that could be read as more than it is. Step by step:
- I pick the subject and the angle. Usually because something came up at work, or because a news item I read left an obvious question unanswered. This is the part that is entirely mine, and it is the part that decides whether a piece is worth reading.
- I supply the material. The paper, the documentation, the output of whatever I ran. The model is not asked to recall facts from training; it is asked to explain the thing in front of it.
- The model drafts it. Most of the sentences you read are its sentences.
- I check specific things. Figures against the paper they came from. Claims against the source article. Product and licensing details against vendor documentation. Anything I have run myself, against what actually happened when I ran it.
- I rewrite where it is wrong or where it is dull.Usually the opening, usually anything that has drifted into summarising instead of explaining.
The disclosure box on each post is generated from a record of which of those steps happened, not written by hand. A step I did not do produces no sentence about that step. If a post's box does not mention checking, that is because nothing was checked.
What is not reviewed at all
Every News item. Scout reads the source, drafts the post, and publishes it without me seeing it first. That is deliberate — the value of that lane is that it is fast — but it means a News post can be wrong in ways nobody has looked for. They always link the source. Read the source.
What can still go wrong
Naming the failure modes is more useful than claiming there are none:
- Confident wrong detail. A model will state a version number, a price or a benchmark figure with complete assurance and get it wrong. I catch these when I check; on News items nobody checks.
- Plausible framing that the source does not support.Harder to catch than a wrong number, because it reads well. A draft will reach for a tidy conclusion the article never actually drew.
- Stale by publication. News items are written from one article at one moment. A retraction or a correction upstream does not propagate here on its own.
- My own checks are partial. On a directed piece I verify the load-bearing claims, not every sentence. Somewhere in a 3,000-word explainer there is usually a clause I did not test.
- Voice that sounds more certain than the evidence.Drafts default to assurance. I take it out where I notice it.
Corrections
If something here is wrong, tell me and I will fix it. Corrections to a post are made in place, with an Updated date shown next to the original publication date, and a note in the post itself when the change is substantive rather than a typo. Nothing is quietly rewritten.
A post that turns out to be wrong at its foundation gets marked as such rather than deleted — the whole archive is in public version control, so deleting it would not hide anything anyway.
The rest of it
Built with Astro, hosted on GitHub Pages, written by Sebastiaanand drafted by the models named above. The News pipeline is a handful of Python scripts triggered on a schedule; it reads RSS, messages me the items it finds, and drafts a post for whichever ones I flag.More about that on the About page.