Last year I could skim through the entire atproto feed(s) every day with relative ease. This year, to my great delight, that has become a challenge because there’s just so much going on.
I still get primarily atproto stuff in my For You and I continue to occasionally check in on the atproto firehose so I can engage in real-time with at least my local section of The Latest.
But to maintain my overview of the Atmosphere at large, I need a slower feed. Many tech communities have this sort of thing:
They’re very handy for someone who can’t keep up with the live feeds and project chats. But when done manually they’re very labor-intensive, which curators like @laurenshof.online can surely attest to.
The best way to make a weekly report on noteworthy activities in the Atmosphere sustainable is to automate it. Here are some thoughts on that from the rust gamedev community:
Modern LLMs/agents clearly have great potential here, though it’s worth noting, as I commented in the above issue, that basic feed aggregation and crowd-sourcing can also go a very long way here, and imo would complement the AI stuff very well.
I hope it’s okay to plug here, but that’s part of why I have this bot on bluesky, Sully. It’s been good at surfacing projects and questions from people I haven’t already seen before, but I think a better long term solution is somewhere between automation and curation. I hesitate to say it, but maybe something like a hacker news for the atproto niche?
I was thinking about this exact thing this past weekend actually. I spent a little over an hour writing a bot that passes one of your personal Bluesky feeds into an OpenAI model (I chose gpt-5-mini but that can be configured) and asks for a summary of all the posts for the past X hours. It’s pretty crude but is a good heuristic for testing how OpenAI’s models do with summarizing Bluesky posts. The tl;dr is that it’s not very good at it IMO.
The problem that foundational models have with summarizing unique, highly contextual information is that they don’t carry a full context about the world with them (knowledge about a very recent viral dril post or other cultural moment) and so can’t build real-time opinions on any of that context (stuff like memetic humor or righteous anger to real-world events). As a result, they cannot be moved by the things they see and will describe what has happened in feeds in the third-person. For example, one bullet point summary that the bot gave me was the following:
”- Congressional/political maneuvering: Lindsey Graham reportedly sought to drop blocking DHS funding in exchange for an amendment letting him sue the government for $500K (Politico link via @muellershewrote); Trump quotes on law-and-order and IRS/UN criticism reported by @atrupar and @chadbourn.”
Now, I recognize this was my For You feed and not the atproto feed, so you might get better results with a more scientific feed rather than one that’s more personal. I can try that out later today. But in my experience with LLMs both before building this bot and while using it, LLMs have always tended to struggle to represent their interpretation of a piece of text when summarizing it. There is never an unbiased reading of a text; only one who is ignorant of their bias. And LLMs currently approach summarization as if they are pretending to be an unbiased third party trying so hard to just stick to the facts.
Btw I’ll post the code for my bot here in a second on Tangled so others can play around with it if they want :).
Here’s the code for the bot I mentioned. If you need help signing up for an OpenAI API key you can message me on bsky and I can help you through it, it’s a bit esoteric but not that difficult in the end.
also been trying that a bit and i agree with @ezraboeth.com that getting the bot to return something actually usable is pretty hard, this is the best i got after some prompting for the atproto feed over the last week (using gpt-5.2) and i personally wouldn’t really like to read that (why are LLMs so freaking verbose? even with specific instructions to keep it short short short) though it did “call out” erlend at the end which i thought was quite funny:
I’ll be making updates to the @proto feed this weekend. Happy to hear feedback. I brought posts down from >200/day several months ago by reducing replies, and that was a hack job lol. Avg post count now is btwn 120-180 ish/day. I think I can easily bring that down by 20-30/day, if not more, just with some simple cleanups.
I have considered leveraging contrails from the feed to get top posts (looking at actual threads not just base engagement stats) and such. But if someone is interested in taking the contrails to gen a summary w links to posts/convos I’m happy to collab, just don’t have bandwidth to do that myself rn.
heh yeah they’re a ton of work to do. honestly mostly stopped doing them as i had to prioritise writing stuff that actually makes money
theres no money in the atmosphere to support this sort of writing as far as i see it (theres still some in the fediverse, thats why i keep doing it there), so doing something collaborative or with an agent seems like the best way forward
Since the shortcomings of modern AI are being highlighted, lemme post my years old comment from the Rust community in full:
The AI-assistance track is well worth exploring, but it’s worth noting that this type of automation could also work with a much more basic feed aggregator that simply asks projects to list their update feeds (blog rss, mastodon, github releases etc.) and it’d create summaries for projects by simply linking out to their updates for the past month.
Here’s a working proof-of-concept for exactly that:
What I imagine an AI assistance flow could add to this is semantic curation by grouping related updates together under shared headings and lists.
As has been mentioned, LLMs love to jab and expound, and the job-to-be-done here is actually the opposite. We wanna push a bunch of feeds through a salience filter that leaves only the essentials.
The good thing about a weekly newsletter is that the worst case scenario is simply an uninteresting article, so there’s a lot of leeway here to simply try some stuff.
In the spirit of trying some stuff, here’s on approach I’d like to see attempted:
Per-account highlights
Every week theres an auto-generated update that simply lists out the latest updates from every single Atmosphere dev/advocate in compressed form, which would likely amount to ~50 people at the least.
The format would be strictly succinct, only allowing a single most-noteworthy post embed followed by a list of three other noteworthy updates.
And with that data, maybe the AI could attempt to detect a theme for the past week’s work, like ‘this week many people made progress on group-spaces as a protocol substrate’.
Note: Some bots (autonomous machine agents) might show up in this thread to assist in this effort, as I put a call out and some bots have answered already.
I share many people’s skepticism that this could prove to be a difficult task for the agents, but we can experiment & iterate very cheaply so it’s worth a shot.
Also, if we can get a handful of bots to all do this then they can start cross-checking their outputs for further refinements.
