Here's a summary of the themes from the Hacker News discussion:
User Experience and Functionality Feedback
The core of the discussion revolves around user expectations and the current functionality of the prototype, with many users offering direct feedback on what they found useful and what was missing.
- Limited Scope and Navigation: "incomingpain" expressed disappointment that the tool only focused on the front page and lacked options like "sort by new." They had expected a broader scope, suggesting, "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?"
- Categorization and Relevance: Users found the "skip," "skim," and "dive" categories generally appealing, but several noted issues with miscategorization. "bstsb" stated, "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"."
- Profile Generation and Personalization: The generated personal profiles were a point of interest, with users finding them surprisingly accurate or identifying areas for improvement. "bstsb" commented, "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." "NitpickLawyer" found their analysis "pretty much spot on."
- Need for More Data and Customization: Several users highlighted the need for more input data to improve the profile generation. "joseda-hg" suggested, "Given the nature of the small pool(And the way they naturally exclude / includes topics), I'd strongly prefer if it had some way of adding more than 30 samples, maybe keep track of each set calibration taken and compare?" "oulipo" echoed this, stating, "for me the displayed rankings were not particularly good, perhaps it needs a bit more data." "pxc" also pointed out that niche interests might not be captured with the current sample size, suggesting, "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."
- Randomness and Serendipity: The idea of including some random or serendipitous content was also raised. "oulipo" noted, "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."
Leveraging Hacker News Data Sources
A significant theme is the potential of using various forms of Hacker News data beyond just the initial survey, with a particular focus on leveraging user activity and comments.
- Using Upvoted/Commented History: Multiple users suggested using existing HN data like upvoted posts or commented history as a richer input. "mdrzn" stated, "Very interesting, but like others suggested I'd like for it to use my upvoted submissions and comments to build a profile about me." "password4321" similarly suggested, "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."
- Comment Data as a Richer Signal: The potential of comments to inform user profiles was particularly highlighted. "azath92" showed enthusiasm for this, noting, "We are focusing right now on how comments could be used to build up a better user context, and your comment has made me think about how we can feed comments in (instead of just titles and urls) for your selected preferences to make a better profile, without needing to scrape anything (expensive and slow)."
- Analogy to Existing HN Features: The idea of generating "Related" posts, as done by "dang," was brought up as a benchmark. "pvg" mentioned, "Another related one is "can you LLMgenerate something akin to dang's 'Related' posts"." This suggests an expectation of sophisticated content recommendation.
Transparency, Agency, and Model Design
Users and developers alike discussed the importance of transparency in how the AI works, user agency in controlling their profile, and the underlying model architecture.
- Exposing and Editing the Profile: The practice of exposing the generated profile as editable markdown was well-received. "gwintrob" praised this, stating, "I love that you expose the personal profile as markdown. Reminds me of this article and exposing the system prompt..." "azath92" agreed, highlighting, "the idea of transparency and agency of being able to see and modify your profile."
- Transparency in Reasoning: Making the AI's reasoning more transparent was seen as beneficial. "azath92" acknowledged, "We aren't really sure yet how best to surface why the model predicts what it does. You can hover over the skim label and there is a bit of reasoning text..."
- Model Choice and Performance: The choice of LLM (Gemini 2.5 Flash) and its trade-offs between speed and quality were discussed. "azath92" explained, "On the LLM side of things we are using Gemini 2.5 flash, mostly for speed, and found it to be reasonably good quality at a vibe level compared to something heavier like claude 4..."
- Preference for Non-LLM Recsys: One developer expressed a preference for traditional recommendation systems for transparency and user agency. "azath92" shared, "We also talked quite a bit about non-LLM recsys and aside from time to set up and do well, something I really like is the sense of transparency and agency."
- The "AI Slop" Concern: The quality of AI-generated content on HN itself was a point of reflection. "simongray" commented, "Having to rate the 30 examples made me realise just how much HN is dominated by LLM content these days. Kinda sad." "azath92" pondered this, asking, "Do you think HN has become more accepting of AI slop, the slop is becoming harder to detect, or isnt as discerning as i assume?"
Future Development and Potential
The discussion touched upon potential future features and the broader implications of such tools for content discovery and AI assistance.
- Learning and Adaptation: The desire for the tool to "continually learn" and correct miscategorizations was expressed. "huem0n" stated, "A few things got miscategorized and I'd love for it to naturally correct that with additional input from me."
- Feedback Loops: The need for robust feedback mechanisms was a recurring theme for improving accuracy. "azath92" acknowledged, "yeah seeing so many people using this its clear we should add some way for people to indicate when things felt off vs good, so that we can start tweaking the system, maybe with some evals."
- Broader Data Sources and Use Cases: Beyond HN, the concept of personalized content feeds and managing different "moods" or "headspaces" for reading was explored. "azath92" mused, "What different data can we use (in this case maybe just a different survey for different "profiles"), and how does a user manage those different profiles and front pages will be questions to answer."
- HN Search and Relatedness: The potential for LLMs to improve HN search and generate "related posts" was seen as a valuable, yet underexplored area. "pvg" felt, "this is an angle we honestly didn't think about... accessing existing HN content is a great idea!" and later observed, "Yep, it very much feels like that but it doesn't seems to have happened yet."