Wondering who is doing the most interesting work related to personal agents that learn from you, your interests, interactions, and then do horizon scanning for information related to what you care about and provide alerts/notifications?<p>Anthropic apparently has a private beta of something called "Syft - Early access Al assistant that promises to proactively send personalized daily briefings on topics you care about. Could be very good for alerts."
My Intelligent RSS reader and agent YOShInOn ingests about 110 RSS feeds. In the last cycle it ingested about 7500 articles and chose 270 it thought I would like and mixed in another 30 randomly chosen for variety, validation, calibration, etc. It groups articles into 20 clusters and shows me the top articles of each cluster rather than top articles overall, otherwise I might see nothing but arXiv papers about recommendation systems. The main UI looks like TikTok, I have to "like" or "dislike" items and that balanced positive and negative samples is good for the model.<p>I might look at another 100 articles on screens that show articles likely to get on the HN front page, articles likely to get a lot of comments, etc.<p>It uses embeddings from sbert.net and uses classical methods from scikit-learn such as the probability calibrated SVM and k-means for classification and clustering. It takes maybe three minutes to build a new prediction model so there is no problem doing it every cycle.<p>It picks an article, then I favorite it, then I pick some things to post to Mastodon or Hacker News, those articles get queued and posted at a controlled rate.<p>It's highly successful at what it does and I am not afraid to demo it. Back end is python with aiohttp, front end is HTMX. My favorite client is an iPad. Runs on a "Gaming PC" with all the RGB bling<p><a href="https://mastodon.social/@UP8/110568695732952713" rel="nofollow">https://mastodon.social/@UP8/110568695732952713</a><p>but it would not need such a powerful machine. What it does need is some commitment, that is, you would have to judge 1000 or so articles for the system to start making good predictions.<p>If I were to try to make a product of it I'd be very concerned about that "cold start" and if it was a SaaS play like Stumbleupon I'd probably lean on the "collaborative filtering" approach because it pools your data with other people's data and also gives popularity signals that would make better recommendations quicker -- however I like the YOShInOn feed better than commercial feeds and I like that is "clean" and based on my preferences and my preferences alone.<p>Personally I find it a hassle to look at an article and decide what exactly to link. There is a popular article about a scientific paper and the choice of which one to post depends on if the paper is open access, how obnoxious the ads on the site are, which one has better illustrations, etc. I'd like to get it so it can give me a few choices of what link and title to post, particularly when I am on mobile.<p>It has search but I don't use it much. It wouldn't be hard to make it a bookmark manager ingesting single web pages.<p>I also envision a "pro" version of the product which maintains multiple classifications at a time aimed at professional "searchers" who could be patent searchers or recruiters or salespeople or researchers who do meta-analysis, open-source intelligence, etc.