waywo.eamag.me
Here's my feedback..
1. Please, consider to always include the latest What Are You Working On posts. For instance, this latest one posted 16 hours ago is not included yet (https://news.ycombinator.com/item?id=45869146).
2. Also, it may be better to include ONLY top-level comments and not replies. For instance, the second row appearing on the list says "Thanks for the the kind words.." by @rubansk is a reply and not a root comment which may be of a little value here.
3. For the UI, when I sort by columns "AUTHOR" or "POSTED", the table would take a different size?! It would be nice to keep the size of the table fixed.
Just my two cents..
It would be more interesting if you could tag the projects based on their subject matter: What problem are they working on, what subject area, what change are they hoping to bring to the world?
I just finished working on consulting project that involved tagging corrective and preventative actions (CAPAs) for a lab to help them organize some of their QA efforts. The use of LLMs for tagging free-form text is a common task and I thought it would be fun to experiment with different strategies for improving the consistency of tags. The article above presents a good approach because it's a streaming solution, but does come with drawbacks (more overhead to set up and treats older data differently than new data). Commenters recommend using a batch approach by collecting all the text up front and then using various strategies to cluster and generate tags, then use an LLM--giving it the predefined tags in its prompt. Once you have enough good tags you could train your own smaller model to generate tags. The batch methods have lower overhead, but take more time for tweaking and experimenting for your specific dataset.
For generating embeddings, I used Cohere's v4 embedder. I found that using HDBSCAN for clustering the embeddings of the tags was much more helpful than using K-means. I also learned that training a Pytorch MLP to predict multiple tags was superior in every aspect to training a tree-based model and gives very good precision, but just OK recall due to the difficulty of nailing all the tags right. I also compared gpt-5-mini and claude-haiku-4.5 for generating tags. Gpt-5-mini was much slower, but cheaper and better at generating good tags. Claude-haiku-4.5 was not far behind and was much faster due to the absence of thinking tokens, but much more expensive. The metric I used to compare the LMMs on their raw tagging ability was Scikit-learns's homogeneity_score.
1. Does it include comments from ALL previous WAYWO posts?
2. Add navigation by page too. Currently I can press next, but it does not reflect in the URL, which means I can't jump to middle or at the end.
Looks like that paging is artificial and data is loaded once on first load. You could also add how many posts to show on single page.
1. Optionally allow all pages to be shown at once (no pagination).
2. Provide a "download all as JSON" button.
3. Allow the use of an LLM to filter through these comments.
Great idea! I'd love to see some stats on the high and low level tags.
For quickly skimming the list, I think it would be useful to add a short, standardized summary for each project. These summaries should be easy to generate automatically with any LLM of your choice.
I have a similar tool for organizing the “Who Is Hiring” threads [1] that uses GPT to provide a quick overview. It’s still running on an old model version from 2023, yet it has been working surprisingly well for over two years now with basically zero maintenance.
- I wanted to publish it asap and get feedback, that's why it's a static website with some ui bugs, you can see a json with comments in a static folder so updates have to be with something like GitHub actions or I have to connect a db
- I included replies on purpose because I saw people saying "I'm also working on this, here's my project" and I wanted to index them
- tags are created with an llm and then there is another deduplication pass to combine them. You can imagine some pros and cons of this approach. I also didn't use embeddings with clustering because it didn't really work for another project of mine https://eamag.me/2024/Automated-Paper-Classification#icml-em...
- it flew under radar first time I posted, but shot to a front page after a second chance pool. If you want to contact me, collaborate, delete data or help to improve this website - there are contact options on my blog and an anonymous feedback form in the footer!
- When you select some tags, list can overflow horizontally and first column with tags is only partially visible. Happened to me when selecting "hardware" tag.
- When you select prev/next or tag, list could scroll to top
- It also often picks up irrelevant comments: https://news.ycombinator.com/item?id=45421459
Still, a few things I like. I used old.reddit.com, but even that seems to be a bit better than the default UI on hacker news. I don't mind the UI, but I found old.reddit.com easier and faster to use (the new reddit UI is just garbage though). It would be nice if hackernews could add more features that are SIMPLE and also simple improves to the layout - again, just very small and careful changes; people dislike any change to their workflows, so hackernews should make these conservatively and only little; and perhaps with a limit per year or every 5 years or so.
Content-wise I have no huge issue, although I'd like more grouping, as some news are interesting, others not so much (to me). Reading just the title is often not enough; some good articles have horrible titles and vice versa.
I've thought long and hard about how a new comer could get around the issues but never came up with anything rock solid. SMS / Email is maximum sign up resistance but could help. Also being able to browse before making an account is key for new users but then you're open to bots.
Anyone got some good idears?
Aside from that, the voting system X uses for community notes is pretty neat and something similar could work as well.
I've spent a bit of time on this as well... wanting to spin up a modern BBS and thinking maybe just a TUI over SSH might be enough to get something interesting without opening up to the heavy spam bots.
I just posted recently in the thread so would have expected to see my post somewhere on this site near the front page. Granted, I guess HN already has a new/hot algorithm, so perhaps you didn't want to reinvent the wheel and instead focus on search.
Despite being a software dev and more invested than the average person, I don't feel like there are that many websites or applications I use really. But there are the best part of a thousand webapps being developed there. I was mulling over starting a language learning platform, but there seem to be 133 being worked on just by people who've posted to HN to say they're working on one.
No wonder most of us seem to struggle with getting traction.
On mobile I’d prefer to be able to just keep scrolling instead of having to tap ‘next page’, scroll to the top, then scroll down again
s.replace("thousands of comments","HN")
And HN is a website than can be loaded on a smartphone even with superslow connectivitiy :-))
Couple suggestions:
- Randomize the home page results
- Link to random site