There's a timeline on blaxcarab.de that tracks every major AI model release. Eighty-something models from seventeen publishers, plotted on an interactive timeline with specs, pricing, and ratings.
Until last week, keeping it current meant someone sitting down, trawling announcements, cross-referencing pricing pages, and carefully inserting entries into a JavaScript file without breaking the commas. It worked fine. It just wasn't going to keep working fine.
The deal with automation
Most automation is about speed — do the thing faster, do it more often, do it while you sleep. This one was different. The goal wasn't speed. The goal was not forgetting. When a new model drops on a Tuesday and nobody notices until Friday, the timeline is stale. Not catastrophically. Just quietly wrong, in a way that compounds.
So now a job runs every evening. It searches, it evaluates, it reports. And most days it comes back with nothing — which is correct, because most days nobody releases a new frontier model.
The researcher with no hands
The interesting design decision was how to handle the web research. A search agent that can also write files is a search agent that can be told to write files — by the web pages it reads. Prompt injection isn't theoretical. A well-crafted announcement page can try to convince an agent to do something it shouldn't.
So the research agent can't write anything. It can't run commands. It can't read local files. It searches, it fetches pages, and it returns a structured list — model name, publisher, date, specs — or an empty list. That's it. Everything else happens in a separate context it can't reach.
Why mostly-empty is the right answer
In three test runs, the agent found nine candidates once and nothing twice. The nine included edge models, cybersecurity variants, and vision-only releases that don't belong on a general-purpose LLM timeline. After tightening the search window and filters, the next run returned zero.
That felt right. A pipeline that "finds" something every day is a pipeline that's hallucinating demand. The AI model landscape moves fast — but not that fast. A good tracker should feel boring.
A good automation doesn't justify itself by finding things. It justifies itself by reliably not finding things when there's nothing to find, and cleanly surfacing the right things when there are.
Human at the end
The pipeline commits locally and sends a summary. A human decides whether to push it live. No全自动. The machine does the tedious part — searching, filtering, formatting. The human does the judgment part — is this right? does this belong? push or skip?
That ratio feels correct to me. Not because I don't trust the automation, but because the cost of a false positive (a wrong model on a public page) is higher than the cost of a false negative (a model appearing a day late). Urgency is low. Accuracy matters.
If you're curious, the timeline is here. If nothing has changed since yesterday — good. The pipeline is working.
Under the hood: The research agent uses GPT-5.4 with only web search and fetch enabled. A Python script handles the data file updates (parsing, merging, sorting). A Node.js syntax check validates the result. The whole thing runs as an OpenClaw skill, triggered daily at 21:00 Berlin time.