Backscroll is an agent that scrolls your feeds. It scrapes my twitter and instagrams feeds and imports as much historical likes and bookmarks to send to a llm to compile a feed that you really want to see.
Its been fetching my Twitter and Instragram home timelines just as I would if I opened up the app several times an hour for the past few months. Even if you were to scroll your feed for .15 mph 24/7 or effectively the entire day, the total tokens you'd read would be under 200k costing a fraction of a dollar to send through a frontier LLM to process and a fraction of a cent if you were to run through a decision model thousands of times a second.
If Michelangelo and Da Vinci had access to short form video they wouldn't have been nearly as infuential, apart from GLP-1, AI, short-form video and feeds seem to be letting a growing portion of society detach from the subtlties of reality. A human being fed ai generated slop to produce ai generated slop doesnt sound like a productive society meant to engage, challenge, and enable to create depth of art, science, and stories worth sharing.
The feed was made in several iterations, due to lack of time I'm going to quickly summarize how the initial version went and how it works now that decision models are fast, cheap, and powerful.
The initial version would periodically have an "inbox" where a lot of tweets we'd collect directly from my feed via a logged in browser/legal tricks(i promise) and then sent to a LLM to grade against a rubric. This rubric would evolve as I spent a couple hours a week labeling tweets I liked or didn't like, I could leave comments and the next time the rubric generator would comb through, it would update the rubric and shift how future tweets would be scored.
Using X-Algorithm's Thunder and Phoenix it would mix and match the tweets based on several heuristics with low time decay such that even good tweets could show up more later in the day or even a day later. Some tweets would be recycled and shown throughout the week in case I "dwelled" on them for longer than normal or if I saved them through Backscroll.
A simple way of checking and evaluating if the quality of the feed and eval loops were good was to take all of my bookmarmed tweets(which I personally think are really good) randomly pit them against tweets and then evaluate if the rubric/other scorers could highly rank my bookmarks.
My end goal was to have a feed that would be automagically ranked with a LLM via a RLM. You can see a variatn of this when I cloned TikTok 1-to-1 with functional comments, messages, and a feed generated by a RLM.
The rise of decision models made this project even more reasonable and feasible since now you could just throw all the tweets in parallel against a decision model, the deicsion model would take in the same rubric and then output a distribution which you'd mix with time decacy and other signals(seen tweet, etc) to rerank and serve a feed.
This also enables-- given a high enough scores -- to immediately ping me via Telegram. I don't really need real-time info but I guess I have it.
Some more iterations on the initial design of the feed, it felt sort of modern but effectively if you asked claude generate me a UI that looked like Vercel x Bootstrap x Apple and make it fully black. It had the structure to waht led to the Backscroll today, but didn't have the depth necessary to make it feel like it was made by Surya Dantuluri.
More feeds
What should the way you interact and manipulate your feed feel? I was inspired by the Apple color gamut panel you get when trying to change the color of a pixel.
What if we changed the settings button with a more dynamic one? This felt too much like eye candy and eventually scraped from the design in production.
A bunch of older design experiments, one was just fully cloning Twitter interface 1-to-1.
