Show HN: HomeBrew HN – Generate personal context for content ranking (hackernews.coffee)

129 points by azath92 ↗ HN
TLDR: Build a quick HN profile to see how little context LLMs need to personalise your feed. Rate 30 posts once, get a permanent ranked homepage you can return to.

Our goal was to build a tool that allowed us to test a range of "personal contexts" on a very focused everyday use case for us, reading HN!

We are exploring use of personal context with LLMs, specifically what kind of data, how much, and with how much additional effort on the user’s part was needed to get decent results. The test tool was a bit of fun on its own so we re-skinned it and decided to post it here.

First time posting anything on HN but folks at work encouraged me to drop a link. Keen on feedback or other interesting projects thinking about bootstrapping personal context for LLM workflows!

23 comments

[ 3.2 ms ] story [ 44.2 ms ] thread
Checked it out and it's offering 2 skim, and the rest are skip?

I had an expectation that it'd go through posts and give me stuff i'd be interested in. Like here's 25 posts that would be interesting?

Only frontpage? no second page? No sort by new, which is my preferred.

I've been wanting this for a while! As an alpha prototype its great. Skip/Skim/Deep dive feels like the right breakdown to me. Having a different color but same UI feels right too.

Other than quality of life stuff (multiple pages for example), I'd like to see it continually learn.

A few things got miscategorized and I'd love for it to naturally correct that with additional input from me.

Really cool!

My results were

- 3 Dive 11 skim, 16 skip.

- 1 dive would have been a skim, 2 skim would have been a skip, 2 skim would have been a dive, 4 skips would have been a skim, 1 skip would have been a dive.

- Total mislabelled: 10, ~30% across and within each category.

Some thoughts:

- it's hard to get it right without a taxonomy of content. I wouldn't skim an essay (Conversations with a hit man). I either dive or skip.

- Skimming is also often a first check before diving, so the categorisation can be a bit artificial.

- I think "Would Read Later" could be an interesting signal/label, like bookmarks on Instagram.

This looks great! At first glance the dive/skim/skip suggestions it offered for me are well judged (I'm now actually diving into the dive ones).
Somewhat tangential but a oddly under-explored side of the LLMs+HN data projects - HN search that does a good job finding HN submission/comment results for the type of things often asked in Ask HN that can be answered by HN searches. 'How do I learn about [some nice thing or another]', etc. People asking this sort of thing often don't know the right keywords so pure keyword search often doesn't work well but more latent-spacey things, in theory, could. Another related one is "can you LLMgenerate something akin to dang's 'Related' posts".
i found the "personal profile" that it created almost more interesting than the actual feed itself. from quite a small sample of posts it had mapped and summarised my interests really well.

i think the bit that needs the most work is classifying each post on the home page; quite a lot of posts that i would mark as "Dive" given its own classification of me ended up as "Skim".

As far as rating posts: user favorites are public, and you could ask for a copy+paste of a few pages of upvoted stories if someone is not using the favorites feature. The stories that have been commented on are also a pretty strong public signal.
I like the idea, but for me the displayed rankings were not particularly good, perhaps it needs a bit more data

Also I know that depending on the days / weeks / mood I will want to read different content from HN, so I guess there should still be like 30% of "random articles" in each category just to create some noise

I felt similarly.

The generated page was really off for me — I had read most of the posts it ranked, at least a little, and most recommended as skips were some of my favorite recent submissions, and vice versa with dives.

On the other hand I'm not sure I'd want to use something like this much as something I like about HN are the pleasant surprises. Maybe as a side page or something if I were really in a rush?

(comment deleted)
It put this post at the top of my feed, which is cool, because it's incredibly relevant to my interests. I used to work on something similar, but way before LLMs were a thing.

Would you be willing to share some more of the architecture/tech stack?

Very interesting, but like others suggested I'd like for it to use my upvoted submissions and comments to build a profile about me.
Funny, I did the swipe thing and then the first result was this post with a [dive] tag. No idea if cheeky or if it actually got that from my choices, but I had a laugh anyway. Neat PoC!

edit: ooh, I see what the swiping did:

## Analysis of user's tech interest: The user demonstrates a strong interest in advanced technical topics, particularly in the realm of artificial intelligence, machine learning, and low-level systems programming/security (e.g., kernel exploitation). They are drawn to articles that involve practical application, model creation, and deep dives into complex technical architectures. Their interest in "Show HN" articles suggests an appreciation for new, innovative projects, especially those with a technical or AI focus. They show less interest in general hardware announcements (like new microcontrollers), historical tech accounts, or very niche, non-AI/ML/security-related programming topics.

Yeah, that's pretty much spot on. Wonder if there's a way to match that against the topics I actually commented on, but at a glance it's pretty cool!

It tried to figure out my interests based off my answers. Little does it know that I'm actually just interested in anything that has a catchy/funny title.
Having to rate the 30 examples made me realise just how much HN is dominated by LLM content these days. Kinda sad.
(comment deleted)
As it is currently written, this is less useful the more niche your interests are. I think for such users, looking at their comment history or upvoted history might be useful instead of or in addition to just sampling recent, popular articles.

In my case, none of the topics I most like to read about and discuss on HN (package management, software freedom, next-gen CLI tools, next-gen shells, philosophy, desktop Linux, functional programming, hacker history, literate programming, Emacs, bitching about common development practices, programming language design, configuration languages) managed to appear in the 30-post sample I used. The profile it wrote for me was pretty good considering that, but definitely not great.

The assessment was also mistaken about my degree of interest in "low level" technical details like binary file formats (in fact it's rather low, although it has gradually increased over time), and my degree of interest in theoretical computer science issues (in fact it's high, but all of the theoretical papers in the sample were about machine learning, which was not an area of academic focus for me).

I do really like the simplicity and customizability of this (exposing the profile as Markdown and making it editable is awesome), and the quality of the results is very good given the tiny input size. But if your primary interests are not super aligned with the mainstream on HN, you won't get a chance to demonstrate that you like them. If users could type a few terms to say what their biggest interests are before running through the samples, this could work even better for people like me.

It would also be interesting if this could work based on article contents and not just headlines. Sometimes I open something and close it immediately, or I open it undecided as to whether I will skim or read closely.

Did you consider using more “traditional” recommendation systems? (and maybe using LLMs to create synthetic preferences…)
I love this. I’ve always considered doing something similar with a traditional recsys model.

The only feature I’d love to see, is there are many posts where I’m more interested in the HN comments, rather than the articles themselves. It would be great to see this incorporated somehow.

Awesome work though. Will bookmark!

During the rating part, am I supposed to be clicking on the link to decide how to rate it? Or should I be basing myself only on the title?
Nifty! Is there a github or source for this?
Personalizing feeds can be tricky without the right data. You might want to try using Mails ai for crafting messages that connect better with your audience. It helped me get more relevant responses and made managing campaigns easier.