It's so refreshing browsing this compared to YouTube, it feels like I don't have to constantly be on high alert at being baited into some trap. Makes you wonder what kind of internet could be possible if all these tech giants weren't just optimizing for engagement and addiction.
YouTube mobile app has a "your custom feed" where you supply English language prompts and get a relevant custom feed. My go-to has been "new and unseen science & technology videos predominantly from people I follow but occasionally interspersed with others' you think I might like in a similar vein"
Genuinely curious, is this sarcasm? I don't use any YouTube mobile app, but doesn't "new and unseen [...] videos predominantly from people I follow but occasionally interspersed with others' you think I might like in a similar vein" just sound analogous to "recommend me videos I'd like to see".
It is also available in the PWA. Yes, it's a real feature with a Gemini-powered prompt. Yes, GP describes a prompt that approximates the actual algorithm.
I just wrote "nothing but Sesame Street parodies" and the feed does what it says on the tin.
Now this project gives me some ideas of things to implement:
- pick the thumbnail from the middle of the video
- allow curated subscriptions, i.e. I subscribe to some curated list made by someone (in this case it could be the database of videos Karpathy has gathered from HN, but there could be others)
At some point I was almost implementing DeArrow support, but using it for a while I don't think it is too great, often the alternative titles people give are unreliable, and honestly I don't know how they exist without being spammed or trolled to death. I do have SponsorBlock support though (I feel like SponsorBlock would work better if it allowed high-quality AI bots to submit blocks though).
Can I search with your app for ONLY the latest uploaded videos and filter for content length?
Will it display whatever bullshit recommendations YouTube feeds?
Really looking for an option to only have "current stuff" - YT constantly recommends me stuff from 3 years ago which I already watched, even with my highly curated account.
Interesting. I consciously, actively smack YT with a rolled-up newspaper in the form of 'never recommend this channel' specifically to make it only recommend me stuff that I've already watched, and not try to pipeline me in some algorithmic direction such has happened at some periods. I don't think I have to draw people a picture here.
I've never imagined actually wanting an algorithm that would find things I'd actually like, having been burned so badly by 'algorithm that finds things it wants to get me to like'. Are you sure you want an algorithm to do that?
Global search is provided by YouTube, so I can't guarantee anything.
I could add some local filters if necessary, but doesn't YouTube already provide its own filters based on when the video was published? Cathode exposes these.
YT search sucks, and this by intention to keep the people scrolling. The sorting & co is all completely broken, YT recommends continuously stuff from years ago etc.
It makes me yearn for what we lost. I remember when the Internet was all like this page, and it was bloody wonderful. I hope we get to a point where hosting becomes functionally free, and the brain rotting advertisement spam becomes a thing of the past.
But we didn't lose it - it's right here, and if this one goes down you can make your own with 2 hours of an Opus 5.5 subscription. Honestly, I whipped up something similar to this the other day for Arxiv papers.
The same Firefox that, on android at least, asks you if you want to open the native YouTube app (or be asked again next time) for every link to YouTube or every time you type youtube.com in the url bar? But that's even more annoying than the ads!
Same feeling. For me the trap was mostly titles and thumbnails, A/B tested to trap my time.
I made myself a thing that tries to answer the title + thumbnail from the transcript, and shows me the answer if there is a good one. Never launched it, but it's at youtldr.com if you want to try.
My plan was to try to find the "answer" for videos in comments. Never really got around doing it, but LLMs make the possibilities more interesting and easier.
Edit: this is also a good opportunity to ship something like this: https://hn.youtldr.com, like Karpathy's HackerTube, but with the new videos from "HN-vetted" channels.
It’s true, we are just optimizing for engagement, which is a first-order (and our best current) approximation of happiness. I’m hopeful that as the attention economy evolves, we will evolve objective metrics that more accurately reflect the long-term health and happiness of users.
One of the positive effects of the recent AI wave is that it shows more folks are warming up to the idea of paying a subscription for tech. Perhaps, the next generation will grow up with tech that doesn't rely on addiction to earn ads money. The downside is that the gap between "folks who can afford tech" and "folks who cannot" might grow even more.
Let's be realistic. Most of the videos on hackertube.org wouldn't exist if it weren't for the enormous sums of money YouTube pays their creators because of...ads.
I would also be more impressed with it if they created their own video processing and serving stack.
The vast majority of all yt content makers do NOT pay the bills with ad rev but sponsors.
These platforms are sucking every penny they can from everyone at every step because number must go up. That is what they are suppose to do being public.
A private company like steam with a sub pay model where creators still can do sponsorship videos throwing out the attention algos would be a better product and still be plenty profitable.
Not all companies should be public and often the best are not.
I (and millions of other people) literally had a netflix streaming subscription for 15 years (up until I cancelled it last year after 4 or 5 price hikes). The desire and ability to pay for stuff has always been there - but at the same time, there's a reason youtube has youtube content and netflix has netflix content.
> The desire and ability to pay for stuff has always been there - but at the same time, there's a reason youtube has youtube content and netflix has netflix content.
Funny thing is: I can subscribe to Netflix for one month a year, and watch everything that I want to watch within a week or two ... which is to say there isn't much there that interests me. With YouTube, I can typically find an hour or two of content to watch per week.
I'm not claiming that YouTube is better. It's simply a case of different people having different interests. While YouTube is almost certainly a poor option for higher budget productions, it seems to work out for lower budget productions where the producers are willing to seek out their own revenue streams.
It includes YouTube Music and it removes ads. Apparently it also marks where sponsorships end in videos (though I'm not sure that is Premium only). Seems like a great deal to me.
yt by default is basically free with ads so the public has viewed it as “that” kind of product, mass content of vastly dif quality level.
Netflix was pay to get high quality well known product along with new high quality content.
For well over a decade they were totally different. You cannot just flip a flag and change that. It isnt people refusing to pay, it is because yt is not netflix and never will be for as long as yt is in the name.
Businesses will seek out more than one revenue source for the same product. For example: newspapers and magazines had advertising, even though the customer paid for the product. Cable television had advertising, even though people paid for cable (and cable companies paid for the channels).
It isn't going to stop data collection either. If anything, it will improve the quality of the data collected.
I'll definitely be taking a look, but it seems that the videos are still hosted on Youtube?
The main issue with Youtube are the ad rolls. The video content itself can be great. But I wouldn't wish those ads on my worst enemy. It's mind poison. Straight poison into your consciousness. AI generated ads for every conceivable image and status vulnerability. And they now do this weird thing where the ad is delivered as a pop song. Like a Taylor Swift engineered pop song for a Korean face mask. The song is 10 mins long about how this woman was happily married then divorced then saw her husband with a woman half his age and how that made her feel and how her self worth was tied to her marriage and her looks but also she's not just that but also somehow the solution is a Korean face mask. It was really traumatizing for me. You can see how much it infected my mind. It's poison.
Thank you for giving me a glimpse. Reading half through your comment made me realize I never saw YouTube ads since the AI cancer, naturally, I wondered what it feels like.
I've been a bit fascinated by my lack of interest in the feeds that are curated for me. They are all things I'm supposedly interested in but I just am not when I see them in that format.
I realised after a while that one of the main reasons I go online is to expand my horizons. I'm actively looking for something new or interesting or surprising that I might not have encountered or thought about before. I am also looking for perspective - give me a broad picture of how the world is so I can calibrate and introspect where my own perspective differs and could be updated. I want to see the things that don't fully align with my ideas.
So it turns out personalised feeds are actively curating against what I'm looking for. I'm curious if others are like me or is this just something weird about me. I think it is very much why I like Hacker News format. It's in the broad areas of my interests but it is anything but curated just for me.
I think many experiencing the same thing but we're way too small drop of a drop in the bucket that it doesn't even resolve in all of their measurements and A/B tests. The vast majority of the "gradient signal" is from normies and their patterns, desires and tastes are different. Basically I think we see a massive deterioration of quality because we are a rounding error.
I can’t reject this line of thought strongly enough.
The vast majority of signal is from a plurality of opinions, and the outcome is an average of those inputs, usually tweaked a bit to make them more addictive.
Your thought that “the algorithm doesn’t serve me personally because I have exceptional taste” is not a unique observation. Everyone who consumes these systems has the same thoughts. Everyone is a rounding error, whether you’re in a knitting community, or you’re a worlds leading expert in deep learning.
I have the same experience, and I don't think it's just us.
I suspect what's going on is we're experiencing the long-term effects of AI recommendation algorithms getting plugged into everything. I'm not anti-AI. I think neural nets are incredibly powerful tools and I really like using them as an aide for my coding work. But they clearly do only a first or second level of pattern matching(asterisk), and I think that has a fundamental impact on their overall usefulness and desirability.
I can't explain this any better than Melanie Mitchell in her "AI for Thinking Humans," so I'd generally recommend that book to really understand the issues. But one way of thinking about it is that the AI keys in on patterns that are indeed there in its training data, but very often don't pick up on the same patterns as a human or treat patterns "not in the same way" as a person would. So you get the uncanny valley effect where the recommendations share some features of things you like and find interesting, but often miss the "soul" of the thing.
A good example of this is I recently watched the movie "X-Men" (2000). I then got a bunch of superhero movie recommendations. On the surface, this makes total sense. I watched a superhero movie, I get more superhero movies. But the reason I like that movie is it portrays racism/nationalism in a way that isn't totally stereotyped while still being PG-rated and easy to watch, without the seriousness of something like "American History X" (it was also a Patrick Stewart tribute). It's very unlikely that an AI can make that level of inference. I actually would enjoy watching more movies like that (as opposed to your point about wanting variety and something more exploratory), but what are the odds it's going to recommend some vague theme like that instead of "more superheroes"?
I also agree with your point that the underlying assumption we all want "more of the same" is a big thing that makes the modern internet feel more narrow. So you put those 2 together and it feels like we're being put into a narrow Skinner box that approximates aspects of things we like, but is really more of a confined environment designed to extract our time and attention while only fulfilling the shallowest, barest minimum of our wants and desires.
I just want to be the one in control of that feed that is curated towards myself. An AI that is local and learns about my likes and dislikes and does not share that info with third parties for their benefit. There is a vast sea of content and a small fraction of it is of interests to me and/or might be helpful to me even if I am not consciously seeking it at any given moment. I should be able to look at its rules, see how it evolves and tinker / reason with it the way I would with a personal assistant or life coach. And I would like to use that to filter out content that is a waste of brain cycles. And to find people, ideas, experiences that I might click with based on its global knowledge of the mindscape.
Youtube feeds work the other way around it appears. Content seeks you. Things promoters think you may click on haunt you. There are videos I see are there in the feed and I don't click on because I am not interested and they linger there, trying to dupe me to click on. Absolutely disgusting.
It was built by nerds but optimized on the data of everyone, so it ended up appealing mostly to the average person.
Maybe that's also the solution to a question I've had for years. The ads I see on diverse social platforms are terrible, but if they are matched with algorithms that try to fit me in as an average person, then it makes sense that, as a huge nerd, they miss the mark.
I find it amazing how close the front page of this is to what I actually watch/want to watch on YouTube. I guess I shouldn't be surprised, but it feels like an unpersonalised feed (besides being based on Hacker News) shouldn't have a significantly better hitrate for me than YouTube's "personalised" feed.
Dumb question maybe, but aren't these two metrics supposed to be aligned? How can they keep me watching if they don't show me stuff I want to click on?
One common trick in Shorts is for the topic of the video to not be revealed until the end, while leaving enough clues that you think the topic will be worthwhile. "So anyways, I was reverse engineering this project with Claude and I came across this little detail that I think some of you might be interested to know. This IoT product that I've got in this box right here is a smart doorbell that quacks when you step on it. But actually by holding this button for five seconds you can make it meow instead."
If you swipe away when it shows it's about an IoT doorbell you watched half, and if you swipe away after seeing the button makes it meow you watched the whole thing. In either case the algorithm thinks you liked it more than average.
In long-form videos this doesn't really work but you can use vague titles to achieve higher click rates. Taken from my actual homepage: "DON'T GIVE ME HOPE, DON'T LET ME HOPE THIS MUCH" "Did you know about these...?" "Calcium Carbide Trick" "The WORST skill to train on OSRS" "They Were Busy With Their Big Plans" "If you were trying to destroy the economy..." "What happens when we don't need each other any more?"
You know which one I'm not going to click on? "FRACTRAN: 14 fractions that produce infinite primes". That's because I already know FRACTRAN is an esoteric programming language written in fractions.
On the one hand there's stuff you think is worthwhile to learn about, and on the other there's the dopamine brain candy that youtube throws at you along the way. They just want to keep you clicking, they don't care whether you feel good about it afterwards.
Do a good job training it and this isn't much of an issue, but give it an inch...
While the project is very interesting in and of itself, the process section that has been edited a bit later to the post is even more revaling:
> A little bit more on my process. I use Claude Code (desktop app for Mac). I use Opus 5.5 on Medium effort as my default. I don't actually do anything resembling traditional engineering anymore (read/write code, run commands in a terminal, or etc.), I only talk to Claude and have it do everything. In this case, the only "manual work" I did was: 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box and 3) I bought the domain name. I worked on this in between Claude's actual work so it still feels a little bit like team work :) Once the box was up I gave Claude the IP and told it update, deploy, secure, harden and optimize the code, and it did the whole thing.
It really does sound like you do not need to run models like Opus 5.5 on high reasoning effort for most tasks.
Interesting project, but have you considered publishing the curated videos as automatically maintained playlists (organized by topic, popularity, or recency) on a YouTube channel?
To me the most valuable part of this is the curation pipeline. I'd love to have access to those recommendations without needing to adopt an entirely separate viewing experience.
I like the idea of playlists, might even work to ask Claude to make a few. Did you have in mind the ability to create and publish them on the site somehow?
I was actually thinking of publishing them directly to YouTube, under a HackerTube channel, with playlists automatically generated and maintained by your existing curation pipeline.
Yes please, not sure why this does not exist yet. I would even think that almost all YouTube video links from HN would make a better experience compared to the algo feed.
> Everything behind this site is yours to reuse. The easiest way: give the zip to your AI agent (Claude Code or similar) and tell it what you want, such as a different interface, a weekly email of talks on your topics, a terminal browser, or an analysis. The zip includes a DATA.md written for agents: how the files fit together, the gotchas, example records, and how to walk you through getting YouTube metadata if you want it.
This is what my dad (87) needs, filtered YouTube. He sits a lot on YouTube and often he is recommended a lot of negative World War 2 content. He is often trapped in that and although he also sees positive videos about other interesting things, depressing content gets the most attention. He has been to a geriatric doctor who also said he needs to stop watching that negative world war 2 content, since he is impacted by it.
I looked in YouTube if there is a way to filter out on key words or specific content like war, but could not find anything. YouTube Kids is a thing, but that is too restrictive.
YouTube is great for him if we could filter out stuff. This hackertube comes close. If anyone could make an app for the iPad that could filter content out by keywords that would be great.
Even with a fresh dump, it would lag behind YouTube, since a video only gets in once someone links it on HN.
I took all the channels HN keeps recommending (3+ videos in Karpathy's data) and made a page with their fresh uploads as they come in: https://hn.youtldr.com
(I make YouTLDR, which writes the one-line answer under each video.)
Yeah, this is the first thing I looked for and when I couldn’t find it I easily decided it’s something I would never come back to. Add this feature and it’s something I would come to frequently
plus one when in programming mode; i have playlists on YT; but this helps to focus... and mostly I find HN has very interesting things that spawn a ton of thoughts...
They should have have a section only for non entertainment stuff. I don't want to see memes, funny videos . unless im in the mood for it .. I just want things I can learn from most of the time.
Youtube sometimes displays an 'Ask AI' button under the video. I recently searched 'Pride of America' (the cruise ship) and the top result was interesting but alarmist/clickbaity, along the lines of 'Don't make these 7 mistakes'. 'Ask AI' let me ask "what are the 7 mistakes": it told me, and in 10 seconds I could deem that video a waste of time. Very efficient.
Overall this 'Ask AI' button is going to smash content creators who rely on clickbait. As a user, it greatly improves my ability to skip bad/clickbaity content, which I really appreciate.
This weirdly feels a lot like my own "subscription" tab.
It almost seems like we need a system where people you trust recommend interesting content, as opposed to algos thzt shove content which will produced anger / engagement / more ads revenue.
Great idea. You know, it's weird, this looks very much like my feed in 2010, in terms of quality.
I recently logged into an account that hadn't been used since then, it was like a time capsule! But within a few days it figured out who I am and started showing me manmade horrors beyond comprehension.
So it sounds to me like there's a niche for a 2010tube as well.
But is this compliant with Google / YouTube terms?
I have wanted to make custom YouTube wrappers for specific interests that avoid ads and distractions. There are communities like say fans of a certain music genre etc that would find the curation etc useful. But I always thought posting publicly with your name attached is a bit grey/risky.
Some time ago i build defcon.videos (down now) which basically was a similar project just full of videos from defcon conferences. Google came at me quite fast, because by their terms of service its forbidden to store a page like this if using API or automated grabbing of videos from the index (which i did logically not manually porting 100s of videos and metadata). They call in an alternate search index which is forbidden by them. Also they claim if you pull video data via their api your only to store it for limited time.
Understandable: creators often update their video metadata (title, description etc.) or make public videos back to private or unlisted, and it is in their own right to do this.
Third parties that copy data will cache stale versions even when the creator doesn't want to share them anymore. Seems fair that YouTube doesn't condone this.
Great idea and works really well. Super easy to find interesting stuff!
If you ever add a link section to other sane video platforms, you might want to add media.ccc.de which has talks of nearly 500 hacker conferences. Also without spam.
I like this a lot. I was just thinking about a site where you could share your YouTube cookies and then use anybody's curated algorithm because you can curate it quite fast in signed out mode.
I’m always surprised how bad YouTube’s recommendation algorithm seems to be.
Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic.
I rarely use TikTok, only to view links people send me (similar to YouTube), but the recommended content is miles ahead, making it much more difficult to close compared to YT.
My YouTube feed is pretty good; it's full of technical discourse on topics that I enjoy learning about, ranging from electronics hacking to civil engineering. That's probably the result of a long number of years of continuous training:
Good video, with a positive combination of presentation quality and a topic I enjoy? Watch it -- maybe now or later, but it gets watched. Thumbs up. Maybe even subscribe. That's what I'm there for.
Mediocre video, whether due to content or presentation? Watch parts of it, but probably not with any sense of completeness. No rating.
Bad video, with content I don't enjoy or with off-putting presentation? Stop watching. Thumbs down, and/or back out to use three-dot menu and "Not interested"
Bad ragebait scaremongering video with emotionally-distressing title? No watching at all. Three-dot menu and "Don't recommend this channel."
---
I get a lot of good stuff. Some mediocre stuff. Very little bad stuff. Almost zero ragebait.
It works well enough that an unfiltered, logged-out view of YouTube is completely and utterly appalling by comparison. That place is fucking awful without training.
(And no ads, because ads have never been a thing for me on YouTube and I experience no compulsion to change that.)
I think that most people make little use of the negative feedback options, mainly because they are more clicks away (dark patterns). A addon that gives them a button next to the thums up could help.
Yeah, perhaps I'm more-willing to put effort into it than some others are and making it easier would help us all out.
I do wonder if the relative inaccessibility of negative feedback paths is due to malfeasance/dark patterns/whatever or just stupidity, though.
If the purpose of YouTube is to show ads to people (it is), then that means maximizing viewership. In this way, it doesn't really matter what people are watching as long as they're actually watching it. More hours of streaming consumption = more hours of ads viewed, and the clock doesn't care if it's positive and enjoyable content or rage-inducing clickbait.
So with meaningful feedback, the good stuff should be able to float up to the top. That would gain viewers, which gains ad placements, which gains money. Everyone wins.
---
Unless, of course, I'm wrong about society. Maybe the majority of potential viewers really do enjoy sitting around and being angry and/or fearful, and find some kind of emotional reward by being incensed.
And clearly, some are: Switching on an AM radio in Anywhere, USA demonstrates this well (the vitriol is off the charts).
I'm just not willing to accept that this represents most people, at least not just yet.
(But if that is an accurate portrayal of the state of things, then: Ragebait means more viewers and more dollars. And thereby, there can be no hope for the platform ever recovering at all. If it pays more to get worse, then it shall simply become worse.)
From that list 731 are blocked in my rss feed. I do read and cache rest.
I sometimes check through website what new has entered the system. It is surprising to see something there new. It is so exiting to see something outside of my bubble.
YouTube algorithm has a lot to comb through, but I am too surprised how suggestion are bland, and unattractive. I am also surprised how difficult is to find new channels through YouTube search.
>Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic.
I've got years of history and it does the exact same thing for me. I watched one news video about the pope and now half of my feed is Catholic news channels.
I buy a mattress on Amazon and I get recommendations for... ten other mattresses? Like what's the logic here, I have an entire garage where I just keep stacking them up to the ceiling? Just show me something else. Is there a human being out there whose like "I bought a French press last week, what could I need this week, five more French Presses?"
I do not understand how these dystopian surveillance companies with their infinite data points are so bad even at this
Built (and still am) something like this, but hand-curated (watched end-to-end) rather than source-filtered. v0 test run: https://cabin.watch
no distribution means I'm having a bit of a hard time finding good videos fast enough (rejection rate through the roof rn), so send any if you know some.
In my opinion this is the part that's holding back _a lot_ more people from publishing. It's a barrier that's surely about to fall. Still interested what steps you took (and manually, or Claude did it all?)
Curious about the hardening and optimizing - how much extra prompting is spent to ensure a ‘clean’ software supply chain for this, versus having Claude at its own discretion use third-party packages/libraries/etc (or not)?
would love to see a write up on this too. my own tech experience has been all back end stuff and doing everything myself, so tying into "the way things are done online" aint so easy. and my patience for reading 3-4 books on it is close to zero.
> Buy a VPS, publish to it, buy a domain and link it, then keep paying for both in perpetuity.
I've always been put off by the continuing financial obligations personally. Like I'll invest a considerable amount for one time things easily, but year after year with constantly increasing costs? No fuckin way.
I've listened to plenty of podcast episodes with CGP Grey and he wasn't secretive about the fact that he does not really enjoy work that much. He also maintained that once he quits he won't make any announcements, but rather just silently disappear.
I assume he has decided that he has enough money and retired.
He (or someone else) is still changing the old video titles and thumbnails to optimise views, though. This particular video used to be called Humans need not apply.
The more popular this gets, the more incentives there are for bad actors to flood hackernews with bullshit. Though I think it's safe to assume it won't be too popular and only a few people will use it for a while.
People want stuff recommended to them from similar people without a profit incentive, that's quite hard to come by... Myanimelist and other longer-form content lists work better because of the nature of the content.
Would love to see something similar with other forums, 4chan perhaps? That's also a forum where the incentive structure doesn't lead to advertising.
I attempted something similar a few months ago, but specifically for documentaries after I found two amazing documentaries in the comments sections (the Disney living characters one [1] and the Mars rover engineer one [2]).
I mostly used some basic string matching to pull out documentaries and put them in a spreadsheet [3]. I sorted them by random and was going through one a day, but the results weren’t great so I stopped.
I should give it another go with a more hands-off fully vibe-coded approach.
One feature I really like about this is how you've included a (presumably LLM generated?) summary of what Hacker News recommended a video, and then the actual comments recommending it. Nice to be able to see why a video is there and what people like about it.
Out of curiosity - was this created using claude code? Cli or gui? I am not even remotely getting that kind of thoroughness from claude with opus 5.5 or anything really. It seems to want to stop as soon as it thinks it figured things out and then the rest of the conversation is just itself saying that its previous claims were wrong.
I'm not sure I follow, what is it exactly that your bots can't do? This is a relatively simple project implementation wise, since it only has to serve static content. LLMs are also OK at manually sifting through text passages to classify them semantically, if you ask them to.
> Opus downloaded all comments of HackerNews, filtered them, ran Sonnet over them in a Workflow (Claude Code feature), got rid of spam, and narrowed everything down to 40K comments recommending a YouTube video.
This part. And Andrej even mentioned he uses medium! I wonder what went into the prompt that prompted the agent to dig deep. Mine seems to stop at very mediocre results fairly consistently and then I just reprompt it a couple of times and it does some back and forth contradicting itself.
I have projects and I made real apps using only LLMs, but it feels tedious. Definitely not a 2 hour thing. And definitely not magical and smooth sadly :(
Well, not sure what I can add here, because this is exactly the type of semantic classification that I expect LLMs to do well, and can't really say why your agents failed to do it.
I've done similar work simply by giving a model a database of text, and asking it to do a unit of work for every passage. "Checking for spam" would be the model checking surrounding messages and user's history, likely an opinionated verdict, but would work. Difficulty comes in when those units of work become complex or have to work in tandem, which isn't the case here.
> HackerTube was built by Claude Opus 5.5, Anthropic's AI model, working in Claude Code: the data pipeline, the search and recommendations, and this site.
I haven't used CC in a while, but I would bet this is sol that made this.
however I still get the same issues with sol getting more apologetic and is more likely to stop working when i tell it is wrong, instead of fixing course and trudging ahead.
I love the idea of different ways to access YouTube. We should be able to make a website like this that is essentially a meta channel around a topic. Great work
Ah, I like this idea. I also had a different thought of accessing YouTube content in the sense I wanted to know all the tech live streams happening today on a particular topic.
This is such a good use of tech! Although I have heard in the past that youtube is quick to ban any use of its content on a third party site. I hope I am wrong and this stays.
Everyone has praised it already, so I won't be original, but I have to say that I've been looking for something like this for quite a while now. I consume a lot of long form videos on yt, similar to what you can find in your feed, but somehow yt keeps boxing me into very narrow space after a while with same videos over and over again. I've been trying to escape this trap by creating new accounts but eventually I keep finding myself in the same box.
I was thinking earlier how the comments or annotations could be implemented in the right way, to fix the problems with the Youtube's comments system. It is certainly not trivial, and one can argue whether this is even desirable. But it would be cool to be able to reference parts of different videos and other resources more fluently.
For example, Kurzgesagt channel prepares their mini-lecture videos very carefully, and they often have a document on their site, referenced in the description under the video, and which annotates the video, and justifies the choices they made in how to present things. The document gives quotes and references to academic sources. If one could navigate and update narratives like that, sort of like a wiki, with one-click references to the relevant snippets of the video, maybe that would work. But even for their videos, the comments section in Youtube is largely useless.
World wide web was unleashed by putting together in a seamless way a few things which were already common before it. One click hypertext navigation was already a standard thing for the technical manuals locally on PC. But making it work on-line, instead of typing several strings in command line made all the difference. Maybe today we could tweak the commenting system to make the comments much more convenient to navigate than they are today. Keep similar comments from multiplying unnecessarily. Allow finding and linking relevant information. I do not have a thought out solution, but there is definitely some room for innovation here, if one approached it imaginatively.
You may want to change the logo, it is a different color but close enough to cause confusion. Not sure what a trademark court would say but it would definitely be enough to draw intimidation from Google if they wanted to.
Interesting to see this person, who really spent effort to understand HOW things work, to brag (it is publicity for his employer, of course, but still) how he does not do any "real engineering" any more, making some clone of youtube (yes yes, "a much better clone, with quality content").
It is perfectly well engineered content for the target audience ahead of the IPO. Gotta do something to push your stock options...
The funny part is that you can do such a youtube clone with something outdated like Qwen 3.5 35B, no need for the latest SOTA frontier trillion weight model.
Nitpick, maybe adding the youtube dislike counter estimation for videos might be a nice touch (yes it's an estimate but it's pretty decent). Though on second thought, perhaps these videos don't need them.
thanks for putting this site together, it's awesome! Even more so is might be your process. Can add some color about your prompts and how you direct the work. I find myself taking much longer for different but comparable types of work and would love to learn a bit about how you direct the model. I use Astra but hope there is some transferable bits here.
My prompts are getting shorter and dumber because the model "gets" what I'm doing even with fewer words. It's almost to the point where if you give it too much, you're constraining the model with too many requirements, when the thing it would have done otherwise was better/smarter. So I feel like the prompting meta is moving more towards fewer bits administered just right. I go back and forth on whether I should make a video on it because on the one hand people have these questions but on the other my best practices look so dumb it's almost embarassing.
don't think of your video as the "best way to do it", we'd love a video of how you specifically go through these things rather than a hyper-optimized way to go about it. it may be true that the next guy in here has a much more advanced process, but it surely doesn't hurt seeing someone with a 'bit of history' on the matter jump-n-run their way through small projects like this. I believe it would bring some comfort to those who feel like they're a thousand years behind already.
I created a new YouTube account that turn of history and that's my default. If I want to watch something I have to type and search, there is no recommendation and it's been great so far. Sharing this as I think someone might not know about this and they might find it interesting.
I also found turning off history and heavily pruning subcribed channels was the only way to fix youtube sloppification; particularly avoiding shorts being heavily forced by google
Andrej this is pretty awesome. Encompasses the need for all of us devs (junior through leads) to continue learning, while embodying what our current reality looks like in the development of the site itself.
Will say tho, your videos themselves have been the biggest help so far, so I would rather see more of those in the future. But I'll take this as a consolation :)
Its so refreshing to see the GUI. especially, since most of these videos are handpicked(?!) by users as suggestions and not originally intended to swim on the ratrace! AdBlockers helps you w/ UI distractions but the continious sloppy videos equipped q/ faulty thumbnails was/is my biggest distraction so far and thats what picked my attention on this fresh take.
I believe YT clients is inevitable (not just UI but something similar to this) similar to RES[1] happened for reddit!
I'm on subscription but I had Claude add it up: 504M input tokens (90% cached), 7.6M output tokens. 90+% of that was Sonnet calls to label the comments, all of the actual work, the app and deploy etc was ~1M tokens of Opus 5.5, I compacted context once. i.e. very small project by modern standards.
I wish someone would make a decision model that’s as good as sonnet but like 10-100x cheaper. Because for this project which is a pretty small dataset you’re spending ~$300 on sonnet usage. like jev can probably do this task but it’s not smart enough to reliably be used
GLM 5.3 Flash is incredibly cheap and imho as good as or better than Sonnet. I used it recently to fix an old server that my family had running serving some websites, and by fix it I mean I got it to log in over ssh, figure out what's running there, copy it out, log in to a brand new droplet on DO, set up a fresh instance of apache with everything patched up to latest, copy that ancient content over, make sure it runs with latest PHP which required several rounds of fixes, optimize the system so it runs on the smallest 512MB droplet, harden the system so that ideally no one ever needs to look at it for another 10 years - it took about an hour, and cost a grand total of 4 american.... _cents_
I use Opus at work since we have a nearly unlimited AI budget, but for personal projects I just use GLM 5.3 and I don't see much difference between that and Opus.
I think most people could consider the cost as being whatever percentage of the subscription quota was used multiplied by the cost of the subscription.
YouTube is at high risk at the moment, because both sides of the marketplace hate it. The best creators are just as frustrated as users like Karpathy by the current state of things. I predict a subscription-based, properly moderated alternative may gain critical mass soon.
I pay for YT premium but noticed the quality of recommendations is subpar yet I still end up wasting time watching it. So I decided I need to limit my feed to my subscribed videos only, which YT app makes annoyingly difficult.
How does HN handle this? I suppose downloading videos for consuming outside YT own app works, but what's the best solution for this, especially with Apple TV as a player?
Personnaly I disabled recommendations/watch history/etc... so the youtube homepage is blank, there is absolutely nothing on it.
I first started by moving all my subscriptions to grayjay(where I also disabled the home page), that way I just click the subscriptions tab and there is everything there, you can even create subscription groups if needed.
Then as I had some issues with grayjay desktop(not the mobile app which works great) I then simply moved all my subscriptions to miniflux and newsrafts(which I now rarely use). I use feedflow on android and reminiflux on desktop as I don't like the miniflux frontend.
So now if there is a new channel I want to follow I just add the channel rss feed to miniflux/newsraft.
So for apple tv I guess you could use a RSS app but I am not sure the user experience would be great, so using grayjay could be an option as you can stream the video to the TV's but that would require another device (phone/computer) as it's not available on ios/apple tv.
I guess a good starting point would be to simply disable everything in your youtube settings to start off (watch history, recommandations, etc...).
I have an extension that blocks all recommendations and forces me to subscribe to the creators I like and search for content myself. Recommendations aren’t optimized for your benefit. Cutting out the recommendations really helps to not get sucked in.
You can make the ultimate anti-short attention span player by intercepting all keyboard/mouse inputs to prevent navigating away and requiring you to watch the whole video ◡.
Is it right that there haven't been any interesting or good videos for the past 5 months? I filtered by “upload date,” and the most recent videos are from 5 months ago. Still, I like it. Is it open source? I'd like to know what kind of tech stack was used here.
What makes this so good is not the curation it’s the age in which confirms our current views after years of bad information. The truth is setting us free and this confirms much
Not sure how but adding a live up to date version of https://news.ycombinator.com/from?site=youtube.com would be nice to see in channel form, I don't think this matches recently recommended (unless based on # of upvotes) - still thank you for putting this curation together.
I've been wanting to make that work for a long time! Assuming I understand you correctly: for HN to segment youtube domains so they show as youtube/channel the way github/username, twitter/username, etc., do.
I think I got stuck on needing some account to access the API that will tell you the channel name.
I thought they meant turning that page into a YouTube-like view (being able to filter by channel would also be very cool, though), I was going to do something similar in my own HN extension, but after trying an initial lazy embedding approach [1], asking for declarativeNetRequestWithHostAccess to be able to bypass HN's Content-Security-Policy feels icky.
I recently built my own feed. On an Android phone, you can have an AI agent SSH in and make changes to it if you're rooted. I replaced the feed on my home screen with links my AI thinks I'd find interesting - it's pretty good!
Really cool idea, you might consider bundling taps when you click a video so it auto-plays instead of opening up the native web browser for youtube. That would make it feel more like the official client. Could easily hide it behind a “buffering” screen.
that's really cool,
I don't know if it's possible to add `Return YouTube Dislike` integration to shows the dislike counts. API (OSS): https://returnyoutubedislike.com/docs
and this one might be tricky but SponsorBlock integration also? to auto-skips sponsor segments and introsusing crowdsourced timestamps. API: https://wiki.sponsor.ajay.app/w/API_Docs
This is a great concept, but I want to see more recent videos. Not videos uploaded 6-9 months ago. It would be great to be able to filter for videos uploaded in the last month.
I’d like to add one more thing, 3Blue1Brown has created a well-organized archive of YouTube channels launched during all the 3Blue1Brown Math Exposition Contests from 2021 to 2026. There are a lot of great YouTube channels worth checking out!
I’ve had the UnTrap for YouTube extension installed for a while to remove all the shorts, feeds and algorithmic content and my YouTube consumption has gone down by like 90%. Feels so much better not consuming all that mental junk
Very interesting. I find that my youtube feed is nicely curated to my interests. Adapting my interests with the algo well. Discovery is mostly en pointe. It is only when I use an unlogged-in browser that I see the world outside the Matrix. That is a slopfest indeed.
What I would really like, now that you have opened up this amazing concept, is an ability to feed my own algo this entire collection. Inject all those videos into my own saved playlist. That would be anotherr cool use of this collection. Is there a way of doing that? Like some sort of algo plant food.
What do you usually do to fine tune your algorithm? Mine always seems to corrupt after some time even if I like and subscribe to specific channels. I have to delete all my history once a year or so.
This is a nick interest site, a video version of harkernews. I am surprised to see videos from many years ago which surely never be seen in YouTube due to too many contents rendered by algorithms
This is cool and all, but what I see is an Anthropic model harvesting (again) human output and curation. That is the part providing almost the entire value of this aggregation wrapper. What I would have preferred to see the model undertake: is somehow imbibe the median Hacker News user persona (if there is one, the reccs did serve as a proxy for that in this case) and work with YouTube API to surface a list of videos with low viewership or traction, but serve the custom taste of this audience at high SNR. I know talk is cheap, but the latter to me is a bicycle and sewing machine, while the former a neatly raked sandpile.
Sorting would certainly be great. I'd especially welcome sorting by some kind of quality metric. The "thumbs down" count is not available via API anymore, but you could use thumbs up per views as a proxy, although that might require some reweighing, since videos with many views usually have fewer thumbs up.
You could also add another layer of voting on hackertube.org, so you'd get a more hackernews-focused voting signal.
What a great idea. Does anyone know of a way of 'getting' / downloading all those URLs (maybe with the video title)..? I'm afraid of the site disappearing.
The site won't disappear (imminently, as I'd like to use it personally!) but head over to the about/ section and download the .zip with the underlying data, which I put up for this purpose.
One thing worth exploring would be how the recommendations compare to YouTube's algorithm. Hacker News upvotes and discussion history provide a different signal for what people find interesting, but I wonder how well that translates into video recommendations.
I don’t live alone, but I mostly watch YouTube on the TV. As I don’t eat at my desk (where I spend a lot of time), I usually go on youtube and try to find a nice recommendation or a new video from my interests. But my ipad and my ereader are always near in case I don’t find anything interesting (often these days). I’ve plenty of things in flight to read and HN is always a backup option.
Would you be willing to share the code base of this site (sans videos list)? I know a person with a list of approx. 4500 curated youtube video links they've been maintaining for an entirely non-HN related purpose that would like to build a similar site and populate it with the video URLs. Right now it's in a much more clunky, awkward format.
I currently use downloaded and curated youtube videos on my local server ( mostly for my kid so that I don't have to deal with youtube autoplay or her clicking something kid would ), but this is just a reminder of much google has changed for the worse.
HackerTube Plus is only $5.99/mo. Among other amazing features it includes our new AI Amanda! You can connect Amanda to your Google account and get personalized recommendations.
It makes me want to login in, to share the history and recommendations between screens. I think a neat way to do that would be to authenticate with some other service that can just store and sync an encrypted blob that is the Local Storage of the site. That could be fairly private and doesn't put any load on the first site, it wouldn't need accounts. This could be a browser feature but then it doesn't work between browsers.
It might be neat to integrate DeArrow to fetch thumbnails instead of just arbitrarily picking picking a 50% screenshot. It could also be used to replace the titles on the server side.
How do the categories work? I'd download the video transcripts and embed the text (maybe summarized first). This way you could support custom categories via embedding similarity.
The average thumbnail on YouTube these days is some guy looking like HOLY SHIT, tearing his hair in amazement looking at something with an arrow pointing at some detail. "TOP 10 XYZ. YOU'LL NEVER BELIEVE #6". Like and subscribe. Utterly repulsive.
i'd prefer to see the original thumbnail though so that i can recognize the video if i have seen it on youtube already. i also expect that the kind of videos that tend to be recommended on HN are less prone to have clickbait thumbnails.
Hey Andrej, thanks for the effort. It's highly appreciated!
In your video regarding your own use of LLMs[1] you mention
https://arena.ai/leaderboard and https://labs.scale.com/leaderboard to
keep track of the status quo. That video is from February 2025 and
https://artificialanalysis.ai has gained quite a lot of popularity but
also earned criticism repeatedly due to updating the benchmarks right
before a new model release.
Would you still recommend the arena and scale leaderboards?
And if you don't mind, what's your opinion wrt artificialanalysis?
No i don't recommend those anymore. Currently I recommend Epoch's ECI. AA was kind of bad but recent changes are trending a little bit better. I think benchmarks are in a really bad state right now otherwise, I basically don't trust them, good related reading https://epoch.ai/data/benchmark-reviews-documentation/includ...
Tired of low-effort YouTube clickbait slop? Just click on the Subscriptions heading to go to https://www.youtube.com/feed/subscriptions. You'll see only videos from the creators you have subscribed to. This one-click trick boosts the signal-to-noise ratio dramatically.
HN's Firebase-based API (https://github.com/hackernews/api) makes that easy, until you want to reconstruct the threads (i.e. who replied to who) and then it's annoying.
For years* I've wanted to replace that API with one we operate ourselves, which will simply return a JSON version of any HN URL.
I've finally started working on this! Perhaps my favorite thing about LLMs is how they make projects possible which would otherwise not have gotten started.
Edit: perhaps I should explain what I mean by that last bit.
The way I want this new API to work is (1) HN should use it internally to generate HTML pages, and (2) there can't be any regressions - nothing should change in the web version. This is a pain to get right, as it's easy subtly alter the details, and it's annoying to regression-test everything. Also, (3) with a design like this, it is easy—and critical not to—leak data that ought to remain private (e.g. people's IP addresses)
This is the sort of pain and annoyance that LLMs are amazing at taking care of. I get to focus on the parts I want to think about—the design and (yes) the code—and have a tireless assistant do the gruntwork and review. Without this I don't think I'd have gotten around to working on it.
* (edit: over a decade in fact - see collapsed reply)
I went into the blog, and saw only two entries in 2026. I dont know how to feel.
I dont know if its because the new state of things is that all knowledge is now secret and cant be shared.
Or if its because Opus 5.5 has turned even Karpathy into a product owner/manager so now the only interesting things are products (hackertube in this case).
Either of them makes me sad. Although its probably both.
I guess I learn something new about the HN crowd every day. When I loaded the page, half of the first screen were videos by IDW folks.
Pro tip for YouTube proper - turn off history and only view the “subscribed” page, and search for content that you really want to see. Oh and of course turn off autoplay and limit shorts to 0 views per day.
Yes. It appears to be random but the first time I loaded the page, on mobile, half of the videos featured well known IDW and populist types like Joe Rogan. So hardly what I would expect from culling recommended videos by a technical crowd.
I think you may have just gotten a bad roll since I have been scrolling on it for some time and haven’t seen any videos in that category. But I agree that that isn’t what I’d expect from the HN crowd
> I'm sure this algorithm works well for the median person on the internet, but it definitely does not work for me
I see this sentiment expressed a lot regarding a variety of ways the internet has been enshittified. Are we really sure the median person would prefer this to better content if they were presented with an alternative?
And even if they would - what are the ethics of knowingly feeding slop to the masses?
Controversial opinion: The current state of the internet is not bad cause big tech has focused on increasing engagement. It has deteriorated cause normies have saturated the space. Projects like this are good because they have gone through the normie filter of hackernews.
I think if the big tech had focused on engagement while the users had remained mainly nerds and geeks, the result would've been extraordinary.
Appreciate the effort, yet I think there is a place this is needed even more:
> It was Friday night at 11pm, I was tired, I went to YouTube to watch something and realized that I was looking at an enshittified wall of slop.
Please do the same thing on Social Media: Facebook, 'X', reddit, instagram, etc . Maybe find a 'hacker signal' that identifies quality content across those platforms and have them aggregated for the public so we don't have to spend time in the fascist network prisons.
The greatest trick - if it even can be pulled - would be setting it up in a way that it can't be sued out of existence by the angry platform overlords - if you can do this, then you are truly a genius, sir.
And lol ... Can you imagine how savage this is, with Rich's disgust with ORM and entangled concepts and like the audience sits there with their ActiveRecord and 10 levels deep inheritance and rails 'easy way' of doing stuff. Probably was a bigger shocker than dHH's keynote this year. lol
Unfortunately HN is just another algorithm. It roughly aligns with what I already see on my youtube but includes some terribly old videos and a mix of videos from creators I'm currently subscribed to and channels I know I never want to watch again like Joe Rogan, Sabine Hossenfelder, Lex Fridman etc. This is indeed a HN video site and I hate it lol.
The youtube video recommendation algorithm isn't that bad if you're careful which videos you watch. Sometimes you'll take a risk on a video and it'll punish you with a week of similar videos though.
I didn't publish the server because it's a vibe coded mess, but if you point your Opus 5.5 at the zip and ask for YouTube-like site it will be basically instant. This part was like 3 min.
EDIT: one caveat is that i couldn't distribute the YouTube data so you'd have to create a (free) API key for that part. Luckily it's only ~1 min of hassle.
- Filters on search and topic pages: upload date, duration, type (talk, lecture, documentary, ...), and sort by upload date, view count, etc. (ty jahala, kiuiras, mixtureoftakes)
- Filters are now on the home page too, no need to search first. Picking one shows all matching videos, e.g. every lecture over an hour. (ty kkotak, macote, Aboutplants)
- Sort by "Recently recommended": orders by the latest HN recommendation instead of upload date, so a 2012 talk that someone linked last month floats to the top. Also a chip on the home page. (ty beaconflint, mixtureoftakes)
- Sort by like ratio as a quality signal: likes/views, damped so small videos can't win on a handful of likes. (ty gpugreg)
- "Play all on YouTube": turns the videos on a page (up to 50) into a temporary YouTube playlist, so it works in the YouTube app on phones and TVs. A start toward real HackerTube playlists. (ty twalichiewicz, waschl, victor9000)
- Removed the YouTube-looking logo. (ty doginasuit)
- HackerTube Longs: every video over an hour, right where YouTube puts Shorts. (ty gosen)
- Bonus: "Hidden gems" filter (loved on HN, under 100k views on YouTube), and "Evergreen only" to hide news-of-the-week videos.
It is true and it is worth pointing out. Still, you have to admit that what he did here shows the distance of how easy it is to get from prompt to a working prototype if you have a clear idea of what you want to do (edit: his blog claims 3h.. which seems like a neat outcome if it includes gathering hn info from scratch ).
