The weekly AI digest experiment: 6 months in

Six months ago I got tired of my own AI news habit.

Open Twitter. Skim a thread. Open Hacker News. Skim another. Open three Substacks I genuinely respect. Realize I've now read 40 minutes of opinions about a model I haven't actually tried. Close laptop. Feel slightly worse.

So I built a thing. It's called ai-tldr.dev and the pitch is boring on purpose: one page, four filters (models, tools, repos, papers), updated automatically, no email signup, no "subscribe to unlock", no thought-leader takes. It's been running for six months. Here's what I learned.

What I actually built

The architecture is unsexy and I'm fine with that.

• A scheduled job hits a list of sources (arXiv, GitHub trending, a handful of release feeds, a few curated newsletters).
• Each item gets normalized into the same shape: title, one-paragraph summary, source link, category.
• An LLM does the summarization pass. It's prompted to be terse and skip marketing language. If the source is already short, we just clean it up.
• Items get tagged into one of five buckets: model, tool, repo, paper, ecosystem.
• The frontend is a single static page with filter chips. Loads in under a second on a phone. No JS framework, no tracking, no cookie banner because there's nothing to track.

That's it. There's no "AI agent" doing anything clever. It's a cron job with taste.

What I got wrong

I overestimated how much people want personalization. My first version had a "follow topics" feature. Nobody used it. Turns out if your digest is short enough, people would rather just scan the whole thing than configure preferences. Configuration is a tax.

I underestimated how much of "AI news" is actually noise. When you summarize the same announcement from five sources, you realize four of them are paraphrasing the fifth. I added dedup by canonical URL + fuzzy title match. That alone cut the volume by ~40%.

I tried to be funny in summaries. It aged badly. A summary written six months ago that was clever in the moment now reads like a tweet you'd delete. I rewrote the prompt to be flat and factual. Bonus: it's also faster to read.

What actually moved the needle

A few things that, in hindsight, mattered way more than I thought:

1. No email gate. The number of "newsletter" sites that block the actual content behind a subscribe wall is wild. Removing that one barrier was the single biggest driver of return visits.
2. Categorical filtering, not search. People don't know what they're looking for. They know whether they want models or tools today. Four chips beat a search box.
3. A canonical URL per item. I link out to the original source, always. The digest is a router, not a destination. Counterintuitively this makes people come back more, not less.
4. Rebuilding instead of regenerating. Early on I'd regenerate summaries when the LLM got better. Bad idea — broke shareable links and felt dishonest. Now summaries are immutable once published.

The boring truth about volume

Here's the chart I wish someone had shown me before I started: AI release volume is not increasing linearly. It's spiky. Some weeks there are three things worth reading. Some weeks there are thirty. If you commit to a fixed cadence ("daily!" "weekly!") you will pad. You will pad with garbage.

I now publish whenever there's enough signal. Some weeks the digest is six items. Some weeks it's twenty-two. Nobody has ever complained. People who say they want consistency actually want quality.

What I'd tell anyone building something similar

• Pick a tight scope and defend it. The site does not cover AI ethics debates, hiring drama, or model benchmarks pissing contests. Not because they don't matter, but because if I tried I'd do all of them badly.
• Make the site work without JS. I'm not religious about it, but a content site that white-screens for two seconds while React boots is a content site I'm going to bounce from.
• Write the summary you'd want to read at 11pm. Not at 9am with coffee. The bar for "is this worth my attention" is much higher when you're tired.

Where I'm taking it next

The next thing I'm probably going to ship is a tiny "what changed this week" view — basically a diff of which categories had the most movement — because that's the question I keep asking the data manually.

Six months in, the takeaway is unglamorous: most "AI products" are over-engineered. A boring page that loads fast and respects your time still works.