Absolutely, you're hitting the same conclusions I've reached. The algorithms are optimized for the lowest friction users that just replay the same music they like over and over again and accept whatever the popular music is. If you're a user that likes music discovery you're fighting against the system to get what you want.
I don't think the recommendation engines behind Spotify, Youtube Music, etc compare to the recommendations I got from last.fm over the years. The algorithmic ones seem to have a bunch of issues that bug me as a long time music listener and someone with a large music library.
- their memory is short as hell so you can listen to something for a while, stop and then it'll suggest it to you later as something to "discover"
- they are way too biased towards recently listened music and will replay things over and over if you're not actively managing your queues.
- because they're so based on what you have listened to (recently) they suggest things that are extremely obvious music no one is "discovering"
- they suggest the "top" songs from artists, albums, etc, it's very hard to get it to play a "deep cut"
- if you have a large library you'll inevitably hit playlist song limits and other things silently. Each service handles this differently, Youtube Music seemingly kicks things out of my library or liked playlists each time I add something else.
I've literally just gotten in the habit of never using the autoplay features and just starting whole albums from start to finish again because the algorithms annoy me so much. Youtube Music has been getting worse about it too where now it often ignores the music you chose to start a playlist and starts playing things you've listened to recently regardless of it doesn't match the genre/vibe at all.
That's because the recommendation engine that Last.fm used back in the day was made the incredibly expensive way: the entire corpus was hand-tagged and cross-linked by humans atop an enormous CDDB. Last.fm, Audioscrobbler, and MusicBrainz (the association engine) were all linked together.
The recommendations engine used them but it's main strength was it was primarily based on collaborative filtering (https://en.wikipedia.org/wiki/Last.fm).
Essentially if people who listen to many of the same artists/tracks as I do have discovered other things I have not, then those unseen artists/tracks become candidate recommendations.
It worked as well as it did because they had a user base of music fans with a wide variety of tastes. CBS ran them into trouble when they upset those fans by breaking the radio and by being perceived as too close to the RIAA.
The will need to get the numbers up, but I'm hoping them being independent again is a good sign.
Is that even a problem? If someone consumes a lot of algorithmic recommendations and you don't, wouldn't that drift you farther apart in the last.fm relationship?
If you really like Song X and Song X happens to be on a popular spotify playlist with a bunch of stuff you’re not into, you’ll start getting recommended all that other stuff on last.fm, no?
Well, the current last.fm "play your recommendations" is linked to spotify, so maybe you're right? Last.fm has gone through phases of no streaming, streaming, and partner streaming, and TBH I haven't used last.fm as a stream source in quite a while. I guess it seems possible that if they outsource their recommendation playback to spotify, you'll get spotify recommendations.
Outside of the spotify integration, last.fm doesn't have visibility into anything that isn't scrobbled AFAIK. It's based on user data only. You have "neighbors" who have similar tastes, which I think is calculated based on overlapping scrobbles (not sure if time-weighted, or just top listens). If we both start scrobbling with a limited amount of artists, and 75% of our scrobbles are the band Primus, we're probably going to be neighbors. If I decide that Primus sucks and start listening to Coldplay all day, our venn diagram overlap separates and we're not neighbors anymore.
Maybe the neighbors influence the recommendations, but playback is outsourced to spotify? I guess I don't really know. You can still browse neighbors though, and use their top lists as "recommendations", which should only be based on listening history.
Yeah. 100% agree here. The Pandora algorithm was an absolute breath of fresh air relative to every other recommendation system. Just constantly surfacing new and interesting music that I had never heard of but was exactly what I liked. I spent so long hunting for info on how they achieved it and if their taxonomies had ever been open sourced.
The Pandora algorithm -- and Pandora's positioning -- is truly a product of its time. You'll want to look into the Music Genome Project (whose very name dates it; the Human Genome Project finished in 2003), but equally influential was the big labels' stranglehold on legal music distribution. When Pandora started and for years of profitability, they had no deals with major record labels, instead promoting small indie artists on the basis of their recommendation engine. If they were going to survive, that engine had to be _excellent_.
But Spotify has that as well. Tons of user curated playlists. And although user playback data is harder to parse through, it's also pretty straightforward to build some clustering algorithm where if you both like X then you might like Y as well.
My theory is that they don't have the incentive. Apple Genius was ridiculously good at music discovery, too. I shudder to think how much I spent on iTunes songs via genius over its run. But now that apple/spotify/etc get my monthly dollars either way, there's no huge incentive for them to create the discovery systems.
One really annoying example of YTM's algorithm is it (or whoever works on it) doesn't understand that a genre can have diverse sounds and instruments, so it will recommend songs that all sound the same.
Like if I start listening to house music, it will just recommend 100 songs that have organ 2 [0], even though house music is more diverse than that. Then it forces me to thumbs down the music, which also isn't what I want to do, because I have no idea what effect it's having on my recommendations. Is it just going to stop recommending house music altogether? Is it going to stop recommending songs with organ 2? Is it smart enough to understand that I just want less and not none? I do like organ 2, I just don't want to drown in it when I'm trying to find new music.
Or I will thumbs up a phonk song and it it just floods me with phonk remixes of pop songs.
Last.fm, on the other hand, seemed to have some way of towing a line of different enough without going too far. Both YTM and Spotify algos just do cookiecutter similarity.
> Then it forces me to thumbs down the music, which also isn't what I want to do, because I have no idea what effect it's having on my recommendations.
I feel this. Social media algorithms can be so complex and opaque now that I have to consciously consider what repercussions my interactions have. I have so little idea what interactions affect recommendations on e.g. Instagram that it almost feels random.
That I would so well internalize the "big brother is always watching what you click, what you hover, what you rewatch, what you comment on, what you pause to read longer than average, what you favorite, what you thumbs-down, etc" default experience provided by facebook/amazon/youtube/streaming platform/short form video platform/etc
that when I stick my head back into 4chan from time to time (to see what the motorcycle thread is talking about these days, or get idea for a show to watch) it's a like a physical weight lifts off me as I realize that no one and nothing gives a toss about what threads I open, or what posts I respond to, or what images I save or post. It won't change any feeds in opaque ways. It won't pollute my recommends (jokes aside about how how the choice of website already polluted matters enough). It won't do anything.
Blew my mind when I put my finger on what I was feeling and realized how pervasive this sort of thing has gotten in most every big tech online product.
One of the most infuriating things about recommendations engines is the way they handle non-English music. Maybe it's not with every language, but as soon as I listen to a Dutch song; the engines will recommend me ALL Dutch music, regardless of genre.
The other frustration I’ve noticed is that they key in very heavily on artist and specific “genre” designation as what feeds the recommendation, which is actually quite bad for anyone who likes experimental work.
I understand that if your recommendations are based on “people who like this also tend to like that” then you’re right in the strike zone. But that approach is basically agnostic to any property of the music itself. Suppose there’s a rock band that released a specific song where they’re experimenting with a new style that has an atypically (for them) funky/jazzy influence. If I say I want more songs like that I mean songs that fuse rock/jazz/funk, not more songs that fans of [rock band] are into.
I still think for new music discovery Pandora’s approach remains the best if you really curate a station for yourself. Apple Music has been good for creating very listenable playlists though, and their new AI playlist generator has been very fun. Surprisingly, YouTube also seems to have some secret sauce where they recommend a lot of interesting stuff that I’ve genuinely never encountered before. I suspect this is because there’s a lot more amateur and experimental artists on there doing weirder stuff and it’s able to find audiences for those in ways that the music-focused services have less visibility into since their catalog is so focused on stuff from the recording industry.
> If I say I want more songs like that I mean songs that fuse rock/jazz/funk, not more songs that fans of [rock band] are into.
I agree. There are bands where I'm not into their usual stuff but they have one or two songs that I really like. It'd be nice to drill down even father into specifics like "this one section of this one song" or even just songs that feature certain instruments or similar sounding vocals.
I'm 90% sure that music labels pay to "put their thumbs on the scales" with these recommendation algorithms in order to push their "hot" artists. I wonder how many of these problems are a result of that.
Personally I’m more suspicious of “classic” artists, where the royalty and songwriting picture might be very skewed behind the scenes. The corporate owners of Spotify favouring one catalog of, say, “70s music” versus another could lead to a long-term capture of that category with little reaction or awareness.
Hot artists, in my estimation, are more about bot campaigns to kick off and sweeten ‘hotness’ as they’re in an ongoing war against other talent of the moment (with shady labels on all sides).
Spotify was advised of payola scheme in federal court in the US, unfortunately last month the judge ruled that the TOS prevention of lawsuits against Spotify is legal.
We can never know for sure if this is or isn't the case, so our only hope for stuff we can be confident isn't this way is with foss / self host able solutions
I've tried to switch to Spotify from Apple Music a few times because the common wisdom seems to be that Spotify has better algorithmic recommendations. But Apple Music "knows" what I like already, and Spotify never grabs me so fast that I'm willing to stick around for weeks training it -- and I suspect part of that is all of Apple Music's human-made playlists. Apple Music has hired a lot of good editors/curators over the years, and I haven't found any service -- including audiophile darlings Qobuz and Tidal -- that beats it in that aspect.
For me curation was better but I was really missing the ability to quickly seed a playlist with a specific vibe and build from there for specific moods.
That, and the desktop app and confusion between library and Apple Music streaming was annoying to manage. They need to unify that experience or split it completely.
Great summaries. I also have a real affection for my last.fm discovery, and I think it had everything to do with "deep discovery" going deep into the related artists pages. It really shaped my relationship to music and my love of music discovery and I sometimes find I don't click with people whose idea of discovery is The Algorithm(TM).
I tried to import my music life into Google Music, uploading my lifetime of libraries there. When they wound that down I just lost trust in online services and now do it through Nextcloud, which honestly is pretty awesome imo. There's no algorithmic suggestion for better or worse, but none of the "who ordered that" style assumptions imposed on you by the system like those you outlined above.
Cannot call lastfm algorithm advanced in any sense. Just opened Amon Tobin page: "similar artists: Kid Koala and DJ Kush", which is an impressively shallow understanding of the last 20 (!!) years of his life, and this happened with almost every artist on the platform, because the average sum of tastes of every listener does not exist in reality. E.g. in the case of Amon Tobin, Kid Koala is the average of similarities between early albums and recent releases, which is just not true, his music cannot be averaged throughout his career. I love my Web 2.0 youth, but the average similarity algorithm doesnt deserve praise. Its not better, its nostalgia and lack of faang-style unlimited greed which confused with better quality
Edit: of course spotify-style recommendations are much much worse, I just mean that lastfm doesnt have good algorithm either because artists are not consistent in releases. What is an average between electronic cult classic "The last resort" and every other Trentemoller album in strict indie rock style? This average does not exist
I am unfortunately a bit late to this conversation but I've been building a music exploration website in my spare time, which has similar artist recommendations too.
https://explore.band
Still work in progress but curious to know what you think if you get a chance to try it
Checked with vast Amon Tobin discography with all his aliases as his discography is a very complicated edge case for any recommendation algorithm: Stone Giants/Figueroa is listened by people who like Amon Tobin but the former is pure indie rock and folk and should not show Amon Tobin itself but more something like Current93.
Don’t know how to solve this issue but it but it break any recommendation algorithm for last 20 years
I was wondering that too. They explained the differences between the tools but didn't really qualify which was doing things "right" just that they differed.
These days it's hard to tell and there's always a mix of both with any high demand items so it makes the stock limits even more pronounced. With how Valve has done hardware releases lately though I imagine it's more a stock limitation.
Is it really? I go to my "local" second-hand marketplace and I see countless of listings for the new Valve Controller. I think it's fair to say most of those aren't "Ops, I made a purchase and I can't return it" but most likely being scalpers. No doubt, some of them are fake as well, but regardless, tends to be fairly easy to see when things are being scalped or if it's actually just high demand, if it's the latter, you don't see tons of second-hand listings the day after it opened.
Right, they're saying you only see the side of the resellers, you have no idea the number of people who purchased it to keep it (like many of us in the thread). So in reality you may be only seeing less than 1% of stock for resell and not the 99% that are just buying it to keep it like normal. It's just confirmation bias that you assume everyone is buying to resell it cause that's all you're able to see.
Context usage is in an open PR now! https://github.com/zed-industries/zed/pull/54881 give it a week or two depending on if you want to use stable vs preview releases. I haven't tried pasting images yet either but I have used their context menu that lets you add images.
1. That's a pref, turn off "format on save" lots of editors and IDEs have it. Maybe they should default to off but it's not an unheard of option with no way to turn off.
Go look at any large project, they have 500+/1000+ issues and many are ancient. Chrome, Firefox, you name it. I wouldn't be surprised if many issues have even been solved or need new reproduction steps but there's a difficulty to triaging all the issues as well.
> I wouldn't be surprised if many issues have even been solved or need new reproduction steps
All the ones I'm subscribed to are straight-up feature gaps versus competing IDEs. Though of course, there is significant selection bias to my subscriptions
Yeah I can see that. I probably wouldn't be subscribing to issues that weren't feature requests/gaps very often. Ones that are tied to bugs are the only others that come to mind that I'd subscribe to but just thinking about my own dev experience in various jobs and how even there our internal backlogs of issues would have unclosed, out of date stuff, I think some portion of issues in public projects would be too.
I feel like it doesn't support some of the commands that manage Claude itself so think `/mcp` `/plugins` etc. Most of the common ones are configured to work though from what I've seen but the ones that do more configuration of Claude seem to be blocked.
That is likely a drawback to their ACP wrapper scheme, it helps exposes IDE functionality but they have to keep up with Claude Code functionality in the other direction. VSCode's Claude code plugin is just like using the CLI.
Entirely right it's a limitation to the ACP side. They're in the middle of adding functionality where you can have terminal/CLI threads and ACP threads too. https://github.com/zed-industries/zed/pull/54729
reply