How We Turn Three Chat Groups Into a Weekly Newsletter

Cairo runs three private AI user groups: one for general AI discussion at every skill level, one for people actively building with AI and sharing what’s working, and one for people using AI to help with investment research.

They’re good groups, curated almost entirely by hand by me. The signal-to-noise is high, mostly because the people in them are doing the work rather than commenting on it. Which means that in any given week the members collectively surface more genuinely useful AI news than I could find on my own — and it all promptly scrolls away, because that’s what chat does.

The Rosetta exists to stop it scrolling away. It’s a curated digest of the week’s most useful news, tips, and lessons, generated in-house on our own pipeline.

The pipeline

Ingest. A scheduled job reads the groups and lands new messages in a database — sender, timestamp, source group, content, links.

Categorize. A local model sorts each message: news, articles, tools, tutorials, community chatter, junk. This is a well-shaped job for a small model — bounded input, fixed output set, no creativity required. It runs on our own hardware for free.

Blurb. For anything carrying a link, a second pass reads the linked page itself and writes a short summary from what it actually says.

Curate and score. Candidates get ranked, and the strongest items make the issue.

Publish. The issue renders to HTML and PDF, with the archive on the way to Cairoglyphics along with a signup, so you can get it by email instead of remembering to look.

Why not just point a model at the internet

Because that’s a different product, and a worse one.

There is no shortage of AI newsletters generated by a model reading the same twelve sources everyone else reads. They’re fine. They’re also interchangeable, and you can tell.

What makes ours worth reading is the filter: fifty people who build things chose to share this. A human with relevant expertise decided an item was worth their group’s attention before any of my machinery touched it. That’s a quality signal you cannot synthesize, and it arrives free.

The pipeline isn’t finding the news. The members are. The pipeline is making sure their find doesn’t disappear under the next forty messages.

The design lessons

Curation is the product. The generation is the easy part. Everything valuable happens in choosing what goes in — and most of that choosing had already been done by people, before the software started.

Categorize before you summarize. Sorting into buckets is cheap and easy to check. Summarizing is expensive and hard to check. Doing the cheap step first means the expensive step only ever runs on things worth spending on.

Keep the boring work local. Categorization and blurb-writing are high-volume, low-judgment jobs. Every one of them that runs against a frontier model is money spent on a task a model in my lab does perfectly well.

Store the raw messages. Not just the output. Thousands of messages sitting in a database is a genuinely valuable archive — searchable, re-processable, and useful for things the original pipeline was never designed to do. I’ve since mined it for material I’d completely forgotten writing, including a few of the posts on this site.

That last one is the tip I’d press hardest. Whatever you’re processing, keep the input. You will want it later for a reason you can’t currently imagine.

The projects, experience and opinions here are mine. AI helped me turn my notes and build records into this piece and polished it for Cairoglyphics.ai.