Lots of positive feedback here, but after a quick listen, I'm not a fan. To summarize my thoughts...
1. The experience just feels too sluggish. For example, I opened the HN homepage, skimmed all of the headlines, read the comments on the top post, and it probably took about 30 seconds. I was done, feeling like I got all the information from HN that I needed, with the intention of checking for new posts later in the day. This tool took an entire minute to brief me about a single post.
2. It's not very practical. Usually 1 or 2 posts catch my attention on the HN homepage each day. This tool is most likely going to give me information about the wrong posts. You could improve this with some type of algorithm that learns what information I listen to and what I skip, but it's not ideal. Or, perhaps I could click the headlines I'm interested in, and a custom audio summary is generated.
3. Lastly, I think it removes the human experience of HN. I like to read exactly what people are posting. Everyone is unique, and it's interesting to see how people interact, along with their choice of words and tone. Erasing all of that and listening to a robotic summary just sucks the soul straight out of the community. It reduces the connection to the people here, which I think is the best aspect of the site.
Thanks for sharing though, it's interesting to see this idea brought to fruition.
I think it’s a case of this not fitting your use case.
It wouldn’t replace reading the front page for me but I could really see it replacing a podcast on my morning walk. Especially given the absence of news spin or adverts. I’ll definitely be giving it a try tomorrow.
4. Skimming HN is useful in text-format because you can go deeper on something that catches your eye by following the link. The podcast format removes this because you can't click on something you hear and dive deeper into it. Skimming without the ability to dive into something is not useful, see next point.
5. High-level summaries are not my use case for HN, or any social media. They don't provide real value, just the illusion of value. What I want to dive deep into stories I am curious about. That can be done with TTS, but I would need to curate my feed first, and then use the podcast format to dive deeper into my curated stories.
I almost feel like this product would be more useful if you remove the LLM aspect and instead let me paste a list of HN threads into it and just TTS all of them, including the full comments. Then I could listen to this long-form content while driving or doing something else.
You're mostly hinting at a very obvious shortcoming of synthesized speech: Its sequential. The phenomenon is most obvious if you look at screen readers using speech synthesis. Its a fundamental problem of the medium, which some devs will discover independently, now that tts has a new surge.
https://gist.github.com/SMUsamaShah/e7c9ed3936ba69e522f8cb38...
This userscript (can use as a bookmarklet) lets you catchup to all unseen stories quickly. Open HN, click bookmarklet, any new stories will be appended with a "(NEW)". Old stories, will show much their rank was changed. Can set a flag to hide seen stories, or set comment and points threshold to hide/show old stories.
To me it makes it very quick to catchup. I just go over a few pages and can quickly see every new story that I have not seen yet.
Wow this is awesome. I've actually been working on something very similar. In my case though I focus on individual stories. Like the commenter above I feel audio doesn't work so well for general overview.
I haven't done a showHN yet because it's not quite ready but you can see where it's at here:
Brilliant idea. I think this has real value as well: as I get older, I find that I have less energy for reading, but I also notice I often miss really cool stuff that was briefly on the front of HN.
I like that it summarizes the comments too. There are often real gems buried in there. (I assume you're only taking a few top ones?)
I think a great improvement could be made with personalization. Most of what's on the front page isn't personally relevant to me, and there's a lot of cool stuff on the new submissions page that never catches on. So it would be nice if a system could learn what kind of stories I personally respond to, and show me (a summary of?) those -- even if they aren't currently trending.
Last.fm came out 20 years ago (proving you don't even need AI for amazing recommender systems), but it seems personalizing your experience never really caught on. (Yeah, the YouTube algorithm kind of does this, but you unfortunately have no real control over it.)
> as I get older, I find that I have less energy for reading,
As I get older (50 now), I have 0 patience for anything else than reading; it's much faster than listening and rereading stuff I didn't get the first time is easier and faster. I listen to stuff (including zoom/calls) to stuff that doesn't hold much importance/value; maybe it triggers something, then I go read about it instead.
Thanks for the comment, I'm really enjoying the discussion it has sparked.
Yes, I'm just taking the top comments, along with a few child comments, in order to not exceed the context window of the model.
Regarding personalization, there's definitely lots of potential. HN can be so random though, sometimes you find things that you didn't even know you needed (intellectually). I guess as with most recommender systems, it's about a balance between exploration and exploitation.
Maybe an MVP could filter for specific keywords and add those posts to the model input.
> Last.fm came out 20 years ago (proving you don't even need AI for amazing recommender systems)
What makes that recommender system not AI? Wikipedia say it uses "collaborative filtering" which Gemini says is a form of AI. AI started nearly 70 years ago.
agreed on energy for reading. do you think it’s that we’re getting older or that the friction associated with consuming information is just getting lower and lower over time?
Would be cool to create embeddings for historical HN posts, and then use a users favorite posts to personalize the post selection by averaging the embedding vectors for a users favorite posts then doing a cosine similarity search to select stories most likely to be of interest to a user.
Although it would be even better to use a users like history, but I’m not sure if/how those can be accessed.
Speaking of, I’m curious how other folks use embeddings. I know you can average multiple embeddings together, but is anyone else doing other translations and having success? Thinking of King - Man + Women = Queen, It seems a lot of the time I see questions being directly used as inputs for semantic search/RAG. I wonder if it might make sense to create a large set of question-answer pairs and embed them and then determine the average translation to move from “question space” to “answer space”, then when you embed questions you apply the translation on the embedding to move it into “answer space” before performing RAG, or maybe this would just add too much noise?
This assumes that favorites have some kind of meaning. Favorites are public (IIRC) and I basically use them to track and share interesting/funny comments. Upvotes on the other hand, are private and are more inline with the things I care about.
hmmm, I can't speak to people using word2vec in conjunction with RAG, but the other use case is actually pretty common. (you don't need to generate answers though in my experience).
For each document intended for ingestion into a vector database:
- Use an LLM to generate a list of possible questions that the document is capable of answering (essentially equivalent to generating a quiz)
- Map these question embeddings back to the original documents
- Store document, document chunks, question 1, question 2, etc. into the vector database
So now when a person queries your RAG, you have the direct link from user query -> doc chunks, but additionally the transitionary link from user query -> similar query -> doc chunk.
Hah! Here's what it had to say about itself this hour:
> In a more self-referential turn, we examine the launch of HN Update itself. This initiative aims to provide hourly news broadcasts, summarizing top stories from Hacker News. While listeners appreciate the concept of curated news, there are concerns about the accuracy of the summaries and the potential for bias in representation. Community discussion suggests that while the value of such a service is evident, trust in the accuracy of the content remains a critical issue.
Pretty neat but it seems to make stuff up. It took a meta comment from this post[1] about the website formatting and suggested the community was worried the C++ memory safety proposal would make code hard to read on mobile. It is hard to trust the other summaries after hearing that.
Apparently you can add "don't make stuff up" to the prompt and it helps. I'm not sure the exact phrasing but probably something like only using what's in the text given.
> This formatting and font mixing is difficult to read on mobile.
I guess for us it's obvious that it's a meta comment, but I can understand the confusion. Still, it could have figured out that "readability on mobile" doesn't really apply to C++, a programming language.
I'll add a section to the prompt reminding it that comments can be meta or even non-factual (gasp), so that it doesn't try to shoehorn meaning into comments like this one.
Pretty interesting project. I have a special, yet weird, emotional attachment to these HN-based projects because they act like extensions or add-ons that could improve HN in one dimension.
While I don't feel like using this tool when I sit in front of my computer or when starting doom scrolling on my phone, I certainly would like to use it when I need to quickly check HN while being busy.
Some quick feedback:
1. First, regarding the UI/UX:
- The title and subtitle don't tell much about the app.
- In the audio track, while you can see the progress of the playing audio, it doesn't show the total time nor the current time.
- It would be useful to highlight the audio track into segments, each one representing a story.
- It would make sense to mention how many stories are being summarized (seems to be 5 right now).
2. For a better use case, I believe it would make more sense to have this tool as a mobile app or a PWA to easily access it even from your car's infotainment system while driving.
3. Tightly linked to (2), having this tool as a mobile app and making it available as a widget, with a Play button, would reduce the number taps needed to play the broadcast.
Next step- create an HNN news network with chyrons and AI news anchors that you can have running on the TVs in your office 24/7 the way banks have CNBC and Bloomberg News running
I would actually use that, especially if it had hourly summaries of the days news like some real news networks do, so you don’t have to watch all day but could just chuck it on when convenient.
The best part about these AI-reads-content things is the potential to let users do things like choose the voice, the talking speed, etc. As a non-American I hate listening to those American talking-head news channels because it’s jarring for me when I’m used to my own accent, even if I do want to catch up on some current events or something.
I started listening to it after this very submission became #1 on HN, so it was very meta to listen to it talk about how it might be making stuff up...
Thanks! An archive shouldn't be too hard to implement.
Would a morning briefing include more of a curated selection of news, like the most interesting/most discussed news of the day? Currently I just take the top 5 posts from the main page.
Damn, looks like you beat NotebookLM by a year and a half!
Is the code for this available? I'm particularly curious how you did the multiple speakers and voices.
NotebookLM has the issue that they keep switching sides, like one will be the student and the other the teacher on a subject, but then they'll suddenly switch in a way that makes no sense.
Interesting, I actually did a search before submitting mine, but I narrowed it down to the last year only. Yours being 2y/o didn't show up. You were ahead of your time!
There used to be a commercial that would play on Canadian TV stations. It showed the view of a river, once with calm and tranquil sounds and then played again with ominous sounds like from a horror movie. When you are presented the river with tranquil sounds, you are happy and calm. When you are presented the river with eerie sounds, you dislike it. The message at the end of the video was to inform Canadians to think more critically about the media they consume, and to question if they are being led down a forced narrative rather than being given facts to question for themselves.
When I see modern day Reddit, Instagram, Facebook, etc, almost all of the videos are backed by music which is trying to bias the user. It makes me uncomfortable.
No, you are not the only one, background music is pretty common though.
Often the background music is just too loud compared to the voice. Professional productions use a technique called ducking, where the music is dynamically made quiet in relation to the volume of the voice.
Personally, I prefer no background music at all, and I wonder if there is a case to made for the professionally balanced case?
No, it's pretty annoying really; I don't know why people do that. I don't like it on youtube/podcasts either. It adds nothing at all at best and distracts at worst.
Not at all, I feel the same. In fact, last time I had that reaction was yesterday, listening to an audiobook which was employing the same dubious technique.
Same boat re: audio speed. I actually speed up the voice in the backend by 1.16x . Above that I was getting too many artifacts in the audio. The nice thing about doing it at that point is that I can handle the music and the voice separately, i.e. the speed of the music stays unchanged.
Speeding it up in the player will also speed up the music, which is not very zen. But I guess I'll just add it to the player and let people decide how fast they want to go.
This strikes a balance with one's use of time and "fear of missing out" (FOMO). Thank you, yunusabd, for creating this service. As a side note, I've been following HN headlines and comments for a very long time, but only just now created an account, just so that I could write this comment to express my thanks to you. On the subject of wishing for additional features: I agree with someone else who commented about how it would be nice to have some form of a daily summary, or maybe broken into three portions of summaries per day: morning, afternoon and evening. Additionally, those three portions would also (ideally) remain available and not be removed, so that a person might go back and review a particular range of times/days. This would be much more practical to enable a person to be able to keep up with most of the developments, but without needing to check in once every hour.
This is really great stuff, love the added "news-station" type music! If something like this we're to be monetized would there be any issues regarding copyright?
Thank you! And that's a really good question. Since it's summarizing the articles, I would assume that there shouldn't be any issues regarding copyright. Regarding the comments that it's using, I think HN generally has some rights to them, although again it's more of a summarization. Generally HN seems cool about these things.
Now, if you were to scan the homepage of the BBC and create a news broadcast from it, they probably wouldn't be too happy about that, understandably. I have no idea how this would fare in court though.
This is so good! I could use that mixed in together with a spotify station for modern Radio experience. I miss the days where you could just do something else and listen to the radio host. Streaming partially supplanted it but not to the same extent, sadly.
Enjoyed this a lot! Particularly enjoyed the meta update about this thread. There are some neat suggestions here, but if this could come up on my podcast feed as-is I would listen to it.
Yeah, if you look at the page source, there's actually no text in the document body. At one point I will have to use a JS-capable browser, to capture pages like this one.
Last week I had a similar idea as you. I created a webpage called https://zeli.app that automatically helps you parse the content of posts from the HN homepage, rewrite the titles, and generate abstracts.
So, I ended up with these post fragments, and I tried feeding these title + abstract lists to Claude 3.5 Sonnet to string them together into a 5-minute English podcast, emphasizing on making these fragments flow cohesively.
Honestly, I wasn't completely satisfied with the results because the topics switched too quickly, and each post only had 2-3 sentence introductions, making it sound less interesting.
Anyway, it was exciting to find someone with a similar idea as me, but I still think the lack of interesting content is the main issue.
"Tech enthusiasts, welcome to 'Claude's Crazy Tech Emporium'! I'm your old friend Claude. Today, we're taking you on a fantastic tech journey, from the maze of subscription services all the way to nuclear-powered data centers. Fasten your seatbelts, we're taking off!
First, let's talk about those love-hate subscription services. Ever tried to unsubscribe and couldn't find the exit? Like being stuck in a maze, looking everywhere but finding no way out. Don't worry, the U.S. Federal Trade Commission has heard our cries! They've recently introduced the 'Click to Cancel' rule. Imagine canceling a subscription as easily as deleting an ex's contact. Ah, the sweet taste of freedom! But don't get too excited, because...
Amazon is tempting us again! They've just launched a series of new Kindles, including their first color Kindle and the all-new Kindle Scribe. Now you can not only read color e-books but also write and draw on them. It's the perfect gift for those who love doodling in paper books! But book lovers, beware, this isn't a license to scribble in library books. Speaking of which...
Have you heard about the 'Transition Year' in the Irish education system? It's like a 'dream school year' for students! During this year, students can try various courses, from aeronautics to art, from programming to car maintenance. Isn't this like a real-life 'Hogwarts'? I even wonder if they have a course on 'How to Create Magic Books on Kindle'. But if you really want to experience magical technology, then...
You must check out the Apple Vision Pro! Someone recently used this device that looks like it came straight out of a sci-fi movie on a plane. Imagine wearing this high-tech headset, watching 3D movies at 30,000 feet - fellow passengers might think you're a time traveler from 2050. Just a reminder, don't scream out loud if you're watching a horror movie, or the flight attendant might think you've spotted a UFO! Speaking of unidentified objects...
Recently, someone benchmarked so-called 'AI PCs'. The results show that these computers might not be as intelligent as we imagined. It seems AI still prefers to roam in the cloud, reluctant to move into our computers permanently. Maybe AI thinks our computers are too cramped? But don't worry, because...
Amazon is finding a new home for AI! They've recently quietly invested in a nuclear power developer. Looks like they're planning a 'nuclear' upgrade for their data centers! Imagine, every tweet you send might have a hint of nuclear energy. Don't worry, this won't turn your phone into a mini nuclear reactor. Although, if it did, we'd never have to worry about low battery again, right?
Finally, let's look at the 'nuclear' war in the WordPress community. The dispute between Matt Mullenweg and WP Engine is more dramatic than 'Game of Thrones'. It seems even in the open-source world, court intrigue is unavoidable. Developer friends, besides coding, learn some workplace politics too. But remember, in the programming world, the one with the most beautiful code usually wins, not the one best at playing politics.
Well, that's the end of our 'Crazy Tech Emporium' tour. From the maze of subscription services to nuclear-powered data centers, to power games in the open-source world, we've experienced quite a lot, haven't we? I hope you've had a good laugh and learned something on this crazy journey. I'm Claude, see you next time! Remember, in this crazy tech world, curiosity and a sense of humor are the best survival tools. Bye-bye!"
Fantastic idea, good implementation. Very valuable and useful!
Do not listen to the chattering nabobs of negativity in the comments!
Your impl is valuable; gives a quick summary of top-n stories. Useful for somebody who dips in/out of HN now and then. It may not be for everyone (indeed, as various commenters have suggested), but its useful for me. And it;s a great example of marrying AI-generated summary with podcast style quick 5-min summary.
Just like listening to BBC news summary.
A few suggestions:
- perhaps make it personalizable (based on HN user?) or atleast provide an option to summarize top 15 (or 10 or 20?) stories
- offer a way to go beyond the first page of HN
- offer a way to summarize the most-commented stories
I'd be interested in a podcast app that allows me to subscribe to written RSS content and uses AI to (transparently) convert them to audio podcasts.
The time/attention it takes to engage with written content can be a barrier for me. The hands-free experience of listening to something is so convenient, I find I engage with audio far more often than written text.
The only major downside of audio is the challenges around notetaking, and recording the snippets of info I want to hold onto.
Neat! Personally hourly feels a bit much but a daily briefing that can fit inside a commute or on a short walk would be perfect. Might be the first ai podcast I'd subscribe to.
I added a basic speed setting, have a look! Also looking into the archive idea, which will be a bit more involved, since my current infra is kind of minimalist.
I would like this except for it to be a daily podcast of the unique set of links and things that made it to the front page with more than 5 comments ordered by points and then comments and then have that read out to me with a summary of the comments. Now that. That would be amazing.
Is there a way to filter out certain type of stories/submissions?
Or is there any HN filter or web-reader (if any) that can help me do that? I know HN doesn't do submission tags but still if something allows some kind of "type/kind" hiding/removal.
Can the speed slider support .1 increments? I'd prefer 1.2x :)
Also, it would be useful to have an index of Stories on the side for each segment and possibly locators within the horizontal audio segment where index/story items are located.
i think this is cool, i'm happy you did it... i really listen to podcasts for the hosts, though. i think the banter and the personality and the perspective are things that i really enjoy. getting to know people through what they choose to talk about or how they frame things is part of the whole charm.
if i want a fast rundown of HN, i'll read HN. you did make me kind of crave a 15 minute weekdaily "morning news" style podcast with some tech journalists that runs down the most talked about HN posts, though.
Interesting project, I think it would be nice to also have a volume slider as well as a speed slider. I manually set the volume via wavesurfer.media.volume = 0.10; to keep from hurting my ears.
Frankly this is super fantastic. Thank you. Any possibility to make this longer, and split it into sections with updates on previous stories at the end? I love this concept a lot.
cool. and, a neat side effect of catching this late is that the HN Update broadcast I heard (labeled "10/20/2024, 9:59:47 PM", assuming Pacific time) included & described itself in the audio. at the time of writing, it's a top story currently on the front page!
This is so neat. I been wanting to build something for HN and can't believe I didn't think of this. It was also cool too hear the update with a mention of HN Update as well considering it's treanding! :)
1) The website is slightly wide on my phone, so it wiggles left and right when I scroll. I usually fix this by pinch-zooming out the tiny amount necessary to align it to my viewport width, but your website apparently suppresses zooming.
2) The AI announces itself as having the name “Jane Doe”. What's the point in this? I know that American news with real anchors do this (having the anchor announce their own name), but this is not universal so it feels foreign to me. Since AIs don't have names, it feels like this was shoehorned in just to sound like American commercial news.
Fun stuff. It does feel like NotebookLM (and others) are hurtling us toward a future that seems inevitable: all content is public domain, and people consume it in many transformed ways.
Those that get ahead of the curve and make their content publicly available and semantically well structured will see their ideas thrive.
Lots of positive feedback here, but after a quick listen, I'm not a fan. To summarize my thoughts...
1. The experience just feels too sluggish. For example, I opened the HN homepage, skimmed all of the headlines, read the comments on the top post, and it probably took about 30 seconds. I was done, feeling like I got all the information from HN that I needed, with the intention of checking for new posts later in the day. This tool took an entire minute to brief me about a single post.
2. It's not very practical. Usually 1 or 2 posts catch my attention on the HN homepage each day. This tool is most likely going to give me information about the wrong posts. You could improve this with some type of algorithm that learns what information I listen to and what I skip, but it's not ideal. Or, perhaps I could click the headlines I'm interested in, and a custom audio summary is generated.
3. Lastly, I think it removes the human experience of HN. I like to read exactly what people are posting. Everyone is unique, and it's interesting to see how people interact, along with their choice of words and tone. Erasing all of that and listening to a robotic summary just sucks the soul straight out of the community. It reduces the connection to the people here, which I think is the best aspect of the site.
Thanks for sharing though, it's interesting to see this idea brought to fruition.
I think it’s a case of this not fitting your use case.
It wouldn’t replace reading the front page for me but I could really see it replacing a podcast on my morning walk. Especially given the absence of news spin or adverts. I’ll definitely be giving it a try tomorrow.
Agree and I would add,
4. Skimming HN is useful in text-format because you can go deeper on something that catches your eye by following the link. The podcast format removes this because you can't click on something you hear and dive deeper into it. Skimming without the ability to dive into something is not useful, see next point.
5. High-level summaries are not my use case for HN, or any social media. They don't provide real value, just the illusion of value. What I want to dive deep into stories I am curious about. That can be done with TTS, but I would need to curate my feed first, and then use the podcast format to dive deeper into my curated stories.
I almost feel like this product would be more useful if you remove the LLM aspect and instead let me paste a list of HN threads into it and just TTS all of them, including the full comments. Then I could listen to this long-form content while driving or doing something else.
You're mostly hinting at a very obvious shortcoming of synthesized speech: Its sequential. The phenomenon is most obvious if you look at screen readers using speech synthesis. Its a fundamental problem of the medium, which some devs will discover independently, now that tts has a new surge.
https://gist.github.com/SMUsamaShah/e7c9ed3936ba69e522f8cb38... This userscript (can use as a bookmarklet) lets you catchup to all unseen stories quickly. Open HN, click bookmarklet, any new stories will be appended with a "(NEW)". Old stories, will show much their rank was changed. Can set a flag to hide seen stories, or set comment and points threshold to hide/show old stories.
To me it makes it very quick to catchup. I just go over a few pages and can quickly see every new story that I have not seen yet.
Wow this is awesome. I've actually been working on something very similar. In my case though I focus on individual stories. Like the commenter above I feel audio doesn't work so well for general overview.
I haven't done a showHN yet because it's not quite ready but you can see where it's at here:
https://news.gipety.com/hn/10842381/k/218/s/three-years-as-a...
I think it would be quite a nice way to give a second life to classic episodes of past showHNs and askHNs.
Re: 3. The problem with reading HN on your phone is that it's generally frowned upon to do it while driving a car.
[dead]
[flagged]
Brilliant idea. I think this has real value as well: as I get older, I find that I have less energy for reading, but I also notice I often miss really cool stuff that was briefly on the front of HN.
I like that it summarizes the comments too. There are often real gems buried in there. (I assume you're only taking a few top ones?)
I think a great improvement could be made with personalization. Most of what's on the front page isn't personally relevant to me, and there's a lot of cool stuff on the new submissions page that never catches on. So it would be nice if a system could learn what kind of stories I personally respond to, and show me (a summary of?) those -- even if they aren't currently trending.
Last.fm came out 20 years ago (proving you don't even need AI for amazing recommender systems), but it seems personalizing your experience never really caught on. (Yeah, the YouTube algorithm kind of does this, but you unfortunately have no real control over it.)
> as I get older, I find that I have less energy for reading,
As I get older (50 now), I have 0 patience for anything else than reading; it's much faster than listening and rereading stuff I didn't get the first time is easier and faster. I listen to stuff (including zoom/calls) to stuff that doesn't hold much importance/value; maybe it triggers something, then I go read about it instead.
Thanks for the comment, I'm really enjoying the discussion it has sparked.
Yes, I'm just taking the top comments, along with a few child comments, in order to not exceed the context window of the model.
Regarding personalization, there's definitely lots of potential. HN can be so random though, sometimes you find things that you didn't even know you needed (intellectually). I guess as with most recommender systems, it's about a balance between exploration and exploitation. Maybe an MVP could filter for specific keywords and add those posts to the model input.
> Last.fm came out 20 years ago (proving you don't even need AI for amazing recommender systems)
What makes that recommender system not AI? Wikipedia say it uses "collaborative filtering" which Gemini says is a form of AI. AI started nearly 70 years ago.
agreed on energy for reading. do you think it’s that we’re getting older or that the friction associated with consuming information is just getting lower and lower over time?
> Yeah, the YouTube algorithm kind of does this, but you unfortunately have no real control over it.)
You do, indirectly. Just need to curate your last watched videos.
Sometimes I feel like I got put into a certain genre or bubble or if things autopplay a when I sleep I'll not ice my front page being taken over.
I just go delete some of those videos from my recent list I can see visible improvement.
Or just start a few video on the topic you want to see and then it's all you'll be recommended.
> Yeah, the YouTube algorithm kind of does this, but you unfortunately have no real control over it.
Wouldn't that argument hold true even if it was implemented here?
> as I get older, I find that I have less energy for
Side comment: When a person says something like that, they might be speaking of only themself, but there's a different parsing that many will hear.
Ageism is a real problem in our field, and one thing we can do is to not accidentally feed it.
Would be cool to create embeddings for historical HN posts, and then use a users favorite posts to personalize the post selection by averaging the embedding vectors for a users favorite posts then doing a cosine similarity search to select stories most likely to be of interest to a user.
Although it would be even better to use a users like history, but I’m not sure if/how those can be accessed.
Speaking of, I’m curious how other folks use embeddings. I know you can average multiple embeddings together, but is anyone else doing other translations and having success? Thinking of King - Man + Women = Queen, It seems a lot of the time I see questions being directly used as inputs for semantic search/RAG. I wonder if it might make sense to create a large set of question-answer pairs and embed them and then determine the average translation to move from “question space” to “answer space”, then when you embed questions you apply the translation on the embedding to move it into “answer space” before performing RAG, or maybe this would just add too much noise?
This assumes that favorites have some kind of meaning. Favorites are public (IIRC) and I basically use them to track and share interesting/funny comments. Upvotes on the other hand, are private and are more inline with the things I care about.
hmmm, I can't speak to people using word2vec in conjunction with RAG, but the other use case is actually pretty common. (you don't need to generate answers though in my experience).
For each document intended for ingestion into a vector database:
- Use an LLM to generate a list of possible questions that the document is capable of answering (essentially equivalent to generating a quiz)
- Map these question embeddings back to the original documents
- Store document, document chunks, question 1, question 2, etc. into the vector database
So now when a person queries your RAG, you have the direct link from user query -> doc chunks, but additionally the transitionary link from user query -> similar query -> doc chunk.
I wonder if dang uses embeddings, or it just comes naturally.
Hah! Here's what it had to say about itself this hour:
> In a more self-referential turn, we examine the launch of HN Update itself. This initiative aims to provide hourly news broadcasts, summarizing top stories from Hacker News. While listeners appreciate the concept of curated news, there are concerns about the accuracy of the summaries and the potential for bias in representation. Community discussion suggests that while the value of such a service is evident, trust in the accuracy of the content remains a critical issue.
My version added "and yes, to our listeners... hi mom!"
Very cute and kind of impressive.
> in a more self-referential turn
:o wow, the AI is becoming self-aware!
Actually makes me think if its possible to embed prompts into story headlines and/or content to do nefarious things...
Pretty neat but it seems to make stuff up. It took a meta comment from this post[1] about the website formatting and suggested the community was worried the C++ memory safety proposal would make code hard to read on mobile. It is hard to trust the other summaries after hearing that.
1: https://news.ycombinator.com/item?id=41899828
AI making stuff up? Unbelievable.
Apparently you can add "don't make stuff up" to the prompt and it helps. I'm not sure the exact phrasing but probably something like only using what's in the text given.
True, that is not a good look.
> This formatting and font mixing is difficult to read on mobile.
I guess for us it's obvious that it's a meta comment, but I can understand the confusion. Still, it could have figured out that "readability on mobile" doesn't really apply to C++, a programming language.
I'll add a section to the prompt reminding it that comments can be meta or even non-factual (gasp), so that it doesn't try to shoehorn meaning into comments like this one.
Pretty interesting project. I have a special, yet weird, emotional attachment to these HN-based projects because they act like extensions or add-ons that could improve HN in one dimension.
While I don't feel like using this tool when I sit in front of my computer or when starting doom scrolling on my phone, I certainly would like to use it when I need to quickly check HN while being busy.
Some quick feedback:
1. First, regarding the UI/UX:
2. For a better use case, I believe it would make more sense to have this tool as a mobile app or a PWA to easily access it even from your car's infotainment system while driving.3. Tightly linked to (2), having this tool as a mobile app and making it available as a widget, with a Play button, would reduce the number taps needed to play the broadcast.
I hope this serves as a constructive critic.
Thanks a lot, all very good points!
Next step- create an HNN news network with chyrons and AI news anchors that you can have running on the TVs in your office 24/7 the way banks have CNBC and Bloomberg News running
I would actually use that, especially if it had hourly summaries of the days news like some real news networks do, so you don’t have to watch all day but could just chuck it on when convenient.
The best part about these AI-reads-content things is the potential to let users do things like choose the voice, the talking speed, etc. As a non-American I hate listening to those American talking-head news channels because it’s jarring for me when I’m used to my own accent, even if I do want to catch up on some current events or something.
I hate it but I also love it.
Im reminded of the talking heads in Batman Beyond that were presumably AIs.
I started listening to it after this very submission became #1 on HN, so it was very meta to listen to it talk about how it might be making stuff up...
Great project!
Wow... Really great.
Can you add some historical archive to listen n hours ago (or morning briefing like) summary as a feature?
Thanks! An archive shouldn't be too hard to implement.
Would a morning briefing include more of a curated selection of news, like the most interesting/most discussed news of the day? Currently I just take the top 5 posts from the main page.
I did https://radio-hn.pages.dev/ last year in the same idea, just once per day. Posted it somewhere here at the time.
Damn, looks like you beat NotebookLM by a year and a half!
Is the code for this available? I'm particularly curious how you did the multiple speakers and voices.
NotebookLM has the issue that they keep switching sides, like one will be the student and the other the teacher on a subject, but then they'll suddenly switch in a way that makes no sense.
Interesting, I actually did a search before submitting mine, but I narrowed it down to the last year only. Yours being 2y/o didn't show up. You were ahead of your time!
Why is it that some people can accept - even like - music behind a voice that tells something?
I strongly dislike it. It makes it harder to listen to the voice.
Am I the only one?
Totally agree.
There used to be a commercial that would play on Canadian TV stations. It showed the view of a river, once with calm and tranquil sounds and then played again with ominous sounds like from a horror movie. When you are presented the river with tranquil sounds, you are happy and calm. When you are presented the river with eerie sounds, you dislike it. The message at the end of the video was to inform Canadians to think more critically about the media they consume, and to question if they are being led down a forced narrative rather than being given facts to question for themselves.
When I see modern day Reddit, Instagram, Facebook, etc, almost all of the videos are backed by music which is trying to bias the user. It makes me uncomfortable.
No, you are not the only one, background music is pretty common though.
Often the background music is just too loud compared to the voice. Professional productions use a technique called ducking, where the music is dynamically made quiet in relation to the volume of the voice.
Personally, I prefer no background music at all, and I wonder if there is a case to made for the professionally balanced case?
Thanks for the feedback, looks like you're not the only one. Personally I enjoy it, but I can understand that some might find it distracting.
I just need to figure out a way to implement a toggle for the music, while also playing nicely with the playback speed etc.
No, it's pretty annoying really; I don't know why people do that. I don't like it on youtube/podcasts either. It adds nothing at all at best and distracts at worst.
Not at all, I feel the same. In fact, last time I had that reaction was yesterday, listening to an audiobook which was employing the same dubious technique.
should add option to turn off background music
It causes sensory overload for me, I abhor it.
+1
Love it. Reminds me of the also useful Hacker News Recap from wondercraft but it looks like that stopped updating as of October 1st (https://www.wondercraft.ai/our-podcasts/hacker-news).
Would be great to have a playback speed button as well. (I can't sit through any audio at 1x.)
Same boat re: audio speed. I actually speed up the voice in the backend by 1.16x . Above that I was getting too many artifacts in the audio. The nice thing about doing it at that point is that I can handle the music and the voice separately, i.e. the speed of the music stays unchanged.
Speeding it up in the player will also speed up the music, which is not very zen. But I guess I'll just add it to the player and let people decide how fast they want to go.
I added a basic speed setting, hope 3x is enough ;)
Did anyone notice it reporting about itself now? This will definitely go into the archive :)
Haha indeed, got it as a last story
Haha yes!!
This strikes a balance with one's use of time and "fear of missing out" (FOMO). Thank you, yunusabd, for creating this service. As a side note, I've been following HN headlines and comments for a very long time, but only just now created an account, just so that I could write this comment to express my thanks to you. On the subject of wishing for additional features: I agree with someone else who commented about how it would be nice to have some form of a daily summary, or maybe broken into three portions of summaries per day: morning, afternoon and evening. Additionally, those three portions would also (ideally) remain available and not be removed, so that a person might go back and review a particular range of times/days. This would be much more practical to enable a person to be able to keep up with most of the developments, but without needing to check in once every hour.
This is really great stuff, love the added "news-station" type music! If something like this we're to be monetized would there be any issues regarding copyright?
Thank you! And that's a really good question. Since it's summarizing the articles, I would assume that there shouldn't be any issues regarding copyright. Regarding the comments that it's using, I think HN generally has some rights to them, although again it's more of a summarization. Generally HN seems cool about these things.
Now, if you were to scan the homepage of the BBC and create a news broadcast from it, they probably wouldn't be too happy about that, understandably. I have no idea how this would fare in court though.
Unrelated but reminds me of GTA V's radio channels in cars. The news talking about what you just did.
This is so good! I could use that mixed in together with a spotify station for modern Radio experience. I miss the days where you could just do something else and listen to the radio host. Streaming partially supplanted it but not to the same extent, sadly.
I'd love to see a list of the articles covered so as I'm listening I could jump to an article and get more data.
Enjoyed this a lot! Particularly enjoyed the meta update about this thread. There are some neat suggestions here, but if this could come up on my podcast feed as-is I would listen to it.
"First up a webgpu wifi simulator has emerged, although details remain sparse due to trouble extracting data from the linked resource"
Sounds about right
Yeah, if you look at the page source, there's actually no text in the document body. At one point I will have to use a JS-capable browser, to capture pages like this one.
It was pretty cool to hear the broadcast reference this. 100% think it should be more concise and have a greater depth of personality.
Last week I had a similar idea as you. I created a webpage called https://zeli.app that automatically helps you parse the content of posts from the HN homepage, rewrite the titles, and generate abstracts.
So, I ended up with these post fragments, and I tried feeding these title + abstract lists to Claude 3.5 Sonnet to string them together into a 5-minute English podcast, emphasizing on making these fragments flow cohesively.
Honestly, I wasn't completely satisfied with the results because the topics switched too quickly, and each post only had 2-3 sentence introductions, making it sound less interesting.
Anyway, it was exciting to find someone with a similar idea as me, but I still think the lack of interesting content is the main issue.
I like what you've done with zeli!!
It is a shame it did not get so much traction when you did your showHn.
One thing that I think would be useful alongside your summaries would be tags for each item. Have you considered that?
Lack of interesting content?
Sample here:
"Tech enthusiasts, welcome to 'Claude's Crazy Tech Emporium'! I'm your old friend Claude. Today, we're taking you on a fantastic tech journey, from the maze of subscription services all the way to nuclear-powered data centers. Fasten your seatbelts, we're taking off!
First, let's talk about those love-hate subscription services. Ever tried to unsubscribe and couldn't find the exit? Like being stuck in a maze, looking everywhere but finding no way out. Don't worry, the U.S. Federal Trade Commission has heard our cries! They've recently introduced the 'Click to Cancel' rule. Imagine canceling a subscription as easily as deleting an ex's contact. Ah, the sweet taste of freedom! But don't get too excited, because...
Amazon is tempting us again! They've just launched a series of new Kindles, including their first color Kindle and the all-new Kindle Scribe. Now you can not only read color e-books but also write and draw on them. It's the perfect gift for those who love doodling in paper books! But book lovers, beware, this isn't a license to scribble in library books. Speaking of which...
Have you heard about the 'Transition Year' in the Irish education system? It's like a 'dream school year' for students! During this year, students can try various courses, from aeronautics to art, from programming to car maintenance. Isn't this like a real-life 'Hogwarts'? I even wonder if they have a course on 'How to Create Magic Books on Kindle'. But if you really want to experience magical technology, then...
You must check out the Apple Vision Pro! Someone recently used this device that looks like it came straight out of a sci-fi movie on a plane. Imagine wearing this high-tech headset, watching 3D movies at 30,000 feet - fellow passengers might think you're a time traveler from 2050. Just a reminder, don't scream out loud if you're watching a horror movie, or the flight attendant might think you've spotted a UFO! Speaking of unidentified objects...
Recently, someone benchmarked so-called 'AI PCs'. The results show that these computers might not be as intelligent as we imagined. It seems AI still prefers to roam in the cloud, reluctant to move into our computers permanently. Maybe AI thinks our computers are too cramped? But don't worry, because...
Amazon is finding a new home for AI! They've recently quietly invested in a nuclear power developer. Looks like they're planning a 'nuclear' upgrade for their data centers! Imagine, every tweet you send might have a hint of nuclear energy. Don't worry, this won't turn your phone into a mini nuclear reactor. Although, if it did, we'd never have to worry about low battery again, right?
Finally, let's look at the 'nuclear' war in the WordPress community. The dispute between Matt Mullenweg and WP Engine is more dramatic than 'Game of Thrones'. It seems even in the open-source world, court intrigue is unavoidable. Developer friends, besides coding, learn some workplace politics too. But remember, in the programming world, the one with the most beautiful code usually wins, not the one best at playing politics.
Well, that's the end of our 'Crazy Tech Emporium' tour. From the maze of subscription services to nuclear-powered data centers, to power games in the open-source world, we've experienced quite a lot, haven't we? I hope you've had a good laugh and learned something on this crazy journey. I'm Claude, see you next time! Remember, in this crazy tech world, curiosity and a sense of humor are the best survival tools. Bye-bye!"
Fantastic idea, good implementation. Very valuable and useful!
Do not listen to the chattering nabobs of negativity in the comments!
Your impl is valuable; gives a quick summary of top-n stories. Useful for somebody who dips in/out of HN now and then. It may not be for everyone (indeed, as various commenters have suggested), but its useful for me. And it;s a great example of marrying AI-generated summary with podcast style quick 5-min summary. Just like listening to BBC news summary.
A few suggestions: - perhaps make it personalizable (based on HN user?) or atleast provide an option to summarize top 15 (or 10 or 20?) stories - offer a way to go beyond the first page of HN - offer a way to summarize the most-commented stories
Good job! Happy to contribute if its on Github
Nice. Would it be possible to add a link to the actual submission as it is being discussed?
Definitely! I was thinking to just show a list of all the links under the player, would that work for you?
I'd be interested in a podcast app that allows me to subscribe to written RSS content and uses AI to (transparently) convert them to audio podcasts.
The time/attention it takes to engage with written content can be a barrier for me. The hands-free experience of listening to something is so convenient, I find I engage with audio far more often than written text.
The only major downside of audio is the challenges around notetaking, and recording the snippets of info I want to hold onto.
Neat! Personally hourly feels a bit much but a daily briefing that can fit inside a commute or on a short walk would be perfect. Might be the first ai podcast I'd subscribe to.
That seems to be the general consensus, I think it makes sense to move in that direction and beef it up a little.
On a related note, what are some cheaper alternative to Open AI's text to speech? Last I checked it came out to ~$9/day for their text-to-speech.
Really neat! How was the waveform visualization achieved?
Looks like https://wavesurfer.xyz/
I like this idea a lot. Archived entries would be nice as others have suggested. It would also be nice to be able to control playback speed.
I added a basic speed setting, have a look! Also looking into the archive idea, which will be a bit more involved, since my current infra is kind of minimalist.
Great idea! I've always wanted to catch up with tech news while walking/driving, particularly from HN.
I would like this except for it to be a daily podcast of the unique set of links and things that made it to the front page with more than 5 comments ordered by points and then comments and then have that read out to me with a summary of the comments. Now that. That would be amazing.
Next step heygen avatar's doing desk presentations with small videos overlayed to show the websites etc.
Amazing, I would love to do a version of this for more niches. Any pointers on how you built this?
Is there a way to filter out certain type of stories/submissions?
Or is there any HN filter or web-reader (if any) that can help me do that? I know HN doesn't do submission tags but still if something allows some kind of "type/kind" hiding/removal.
Can the speed slider support .1 increments? I'd prefer 1.2x :)
Also, it would be useful to have an index of Stories on the side for each segment and possibly locators within the horizontal audio segment where index/story items are located.
i think this is cool, i'm happy you did it... i really listen to podcasts for the hosts, though. i think the banter and the personality and the perspective are things that i really enjoy. getting to know people through what they choose to talk about or how they frame things is part of the whole charm.
if i want a fast rundown of HN, i'll read HN. you did make me kind of crave a 15 minute weekdaily "morning news" style podcast with some tech journalists that runs down the most talked about HN posts, though.
Interesting project, I think it would be nice to also have a volume slider as well as a speed slider. I manually set the volume via wavesurfer.media.volume = 0.10; to keep from hurting my ears.
its actually been done, but in slightly different manner
https://t.me/hninsights/836?comment=441
Sounds good. Listened to some of it -- good.
Suggestion, have better audio:
(1) Enunciate the words more clearly.
(2) To help with (1), slow down and speak fewer words per minute.
(3) At the end of a sentence, don't drop voice volume and enunciation clarity.
Frankly this is super fantastic. Thank you. Any possibility to make this longer, and split it into sections with updates on previous stories at the end? I love this concept a lot.
"although details remain sparse due to issues extracting content from the linked resource"
Unfortunate that it adds stuff like this, which doesn't seem helpful to the listener.
This is probably the coolest use case I have seen of HN API
Not sure this is something I'd regularly use just because I prefer reading the news to listening to it, but I still found it pretty cool.
I see lots of generated content with AI, would be interested how much of this is actually listened by a real human after the first days of excitement.
The AI voice is good. What are you using to generate it?
Wow—this is indistinguishable from an NPR piece.
cool. and, a neat side effect of catching this late is that the HN Update broadcast I heard (labeled "10/20/2024, 9:59:47 PM", assuming Pacific time) included & described itself in the audio. at the time of writing, it's a top story currently on the front page!
This is awesome, I wish it did a rolling 8 hour instead that is updated once a hour if that makes sense.
Fantastic idea! I really liked extracting meta information from Comments as well. Good luck!
vaguely related someone also made a Hacker News bot on Bluesky
https://bsky.app/profile/newsycombinator.bsky.social
Can you make one of the playback speeds 1.25x (it jumps from 1.0x to 1.5x)?
This is awesome. I like it
I wonder what it will do when it encounters a (year) postfix.
I would love to have a transcript of the generated podcast.
This is so neat. I been wanting to build something for HN and can't believe I didn't think of this. It was also cool too hear the update with a mention of HN Update as well considering it's treanding! :)
Excellent Idea
I have two super nitpicky things to say.
1) The website is slightly wide on my phone, so it wiggles left and right when I scroll. I usually fix this by pinch-zooming out the tiny amount necessary to align it to my viewport width, but your website apparently suppresses zooming.
2) The AI announces itself as having the name “Jane Doe”. What's the point in this? I know that American news with real anchors do this (having the anchor announce their own name), but this is not universal so it feels foreign to me. Since AIs don't have names, it feels like this was shoehorned in just to sound like American commercial news.
AI slop generator.
Fun stuff. It does feel like NotebookLM (and others) are hurtling us toward a future that seems inevitable: all content is public domain, and people consume it in many transformed ways.
Those that get ahead of the curve and make their content publicly available and semantically well structured will see their ideas thrive.
Here's my user test: https://news.pub/?try=https://www.loom.com/embed/9e8b8d454ee...