What a strange article, from somebody who should understand the underlying technology (click on the “books” tab - the author is a technologist).
This is not about AI, the author is mostly just pointing out that Spotify was not designed for classical music.
This is a product issue. Spotify DJ is essentially “shuffle with some voice interludes”. There’s probably some non-AI code in there to explicitly prevent it from playing an album end to end.
Besides, AI is not one thing. It’s weird to generalise “This beta spotify feature doesn’t serve me, hence AI is useless”. For example, when the author says “if it can’t do this, how could it compose music?”, that’s a category error.
Honestly the whole post and tone are just baffling. It’s mixing up all sorts of opinions and trying to put them under one umbrella, and about 50% of the text is just name dropping specific classical pieces.
I happen to agree that the Spotify DJ feature is terrible, but I think this is a very ineffective way of presenting the argument.
His 'Annotated Turing' (a reproduction of On computable numbers, with an application to the Entscheidungsproblem, with explanation and walk-through) got me into CS vs. prior interest mainly in EE.
> Microsoft provides two frameworks for developing Windows applications: MFC (Microsoft Foundation Classes) and Win32. MFC (Microsoft Foundation Classes) is a Microsoft framework for developing Windows applications in the C++ programming language. Win32 is a collection of functions and data structures provided by Microsoft for the development of Windows applications. [0]
> For example, when the author says “if it can’t do this, how could it compose music?”, that’s a category error.
Given the author's background I believe it's intentional ragebait. It's as ridiculous as saying LLM can't count the number of Rs so it cannot generate grammatically correct sentences. No way he really thinks the logic is sound.
If you have the slightest knowledge of classical music you would know it should not be mixed like in a dj set, and you would not optimize your dj algorithm for it.
I think you could pick out a movement from this and then a movement from that. I can see somebody wanting to have classical music playing all day without having to pick out specific tracks, like listening to the radio.
I think the auto DJ feature is already well capable of that: having tracks playing the all day.
But if you want to preserve the original composition of classical music, you have to play the track start to finish, preferably with a small pause between tracks as well.
Interesting, i never heard of anything similar before, but i'm quite sure the classical music fans would also hate on him for ruining the original compositions.
> For example, when the author says “if it can’t do this, how could it compose music?”, that’s a category error.
That isn't really a category error. It's more begging the question. It makes the assumption that the ability to DJ music is the same ability as being able to compose music, and uses that assumption to suggest the conclusion that a failure to DJ classical movement would necessarily result in the failure to compose same. A category error would be assigning a property to AI that it cannot have. It would look more like, "if AI can't DJ music, we have no way to know what color it is."
I’m fairly certain Spotify’s core meta data adheres to the US music industry largely set / reinforby Nielsen.
I’m curious why the author would want to happen with the feature if not move from 1 artist to another
> It makes the assumption that the ability to DJ music is the same ability as being able to compose music
And yet an awful lot of musicians are also DJs. It's almost like spending a lot of time playing music and watching how people react to it give you a good sense of how the underlying processes of creating it can work.
When "me" is most classical music and this is a music platform I think the critique is not unwarranted. They could adapt it with special system prompts for classical.
Probably 50% of my listening is classical. If I want to listen to classical I just listen to albums. I’ve never had a problem with this. The concept of a DJ for classical music is just kind of weird.
Spotify is still bad for classical music because you can’t ex. search by composer or label of find alternative recordings of the same piece etc. If you know what album you want already its ok, but if you like classical you should really consider IDAGIO.
Isn’t fundamentally the issue that for any symphony by Beethoven or whoever that there are thousands of recordings of performances? So if I decide I want to listen to a certain one then I also need to pick a particular performance that a particular orchestra did a certain time?
I’m British so I guess the equivalent would be Radio 4. But no, not really tbh. I just find what I want and listen to it. I know some people really like Radio 4 though.
Cynthia Solomon has shared a treasure trove of rare classic videos of Seymour Papert, Marvin and Margaret Minsky, kids programming Logo and playing with turtles, and many other amazing things at the MIT AI Lab, MIT Media Lab, and Atari Cambridge Research:
Sorry but at this point this beta product is over 2 years old. It should be better by now. Unlike your "baffling" comment at least he wrote the article himself.
The article isn’t about the DJ feature at all, despite claiming to be. It is very clearly and openly about Spotify not catering to classical music in general. It starts by calling all people who listen to anything other than classical music “illiterate”!
AI DJs for music feel a bit like AIs writing restaurant reviews. Possible in theory, but fundamentally I don’t really care what a machine thinks, I care about what a human, preferably an expert human, thinks.
I listen to a lot of DJ mixes on YouTube (Hör Berlin is great, for example) and part of the appeal is what this particular DJ picks: what kind of music are they listening to in the country they’re from, how are they interpreting it, what are they mixing it with, etc. For some DJs there’s also kind of a personal visual brand, like musicians themselves.
The idea of an anonymous AI picking an optimized list of music kind of defeats the purpose.
I like using the automatic lists in soundcloud to discover new music. Often its hit or miss but it can surface some great tracks... Its intentional though, gotta have your finger on the skip track and heart...
Right but a good DJ introduces you to new music while fitting the track into the set as a whole. It’s not a random music discovery process, and oftentimes I’ll end up mostly preferring to listen to a song as a part of the set, not individually.
To use the food analogy again: sure, if you just eat random things on the menu, you might find new foods that you enjoy. But it’ll be a much better experience if the chef / restaurant is introducing you to new foods in an intelligent way, not randomly or “We see you like chicken, so try this other chicken dish.”
no idea how spotify ai specifically works (i don't use that service) but:
> fitting the track into the set as a whole. It’s not a random music discovery process
there have been plenty of attempts to analyze music and to automate track matching like the music genome (going back to '99) and while human DJ's definitely have their place (i actually listen to lots of those) it's not inconceivable that a lot of modern music could also be mixed and matched automatically with at least half-decent (to a human) results.
P.S. found the article itself pretty funny - like a nerdy, methodical complaint, just funny to read
I don’t think it’s impossible or anything, I just don’t think it would really result in anything particularly interesting. The best DJs often add such an obscure reference or song, dialogue from a movie, etc. that comes from their own individuality. Music recommendation systems seem to mostly operate on a tagging/descriptive basis, because obviously they don’t have real lives to draw these references from.
If an AI would make interesting DJ mixes that aren’t merely collections of similar music, I think they’d need to be constructed in a totally different way.
Check out Paul Lamere's talk about playlisting that he presented at ISMIR 2010 (The International Society for Music Information Retrieval has conferences about all this stuff, and Paul founded The Echo Nest, which Spotify later bought):
ISMIR: The International Society for Music Information Retrieval
>Tutorial 4: Finding A Path Through The Jukebox -- The Playlist Tutorial. The simple playlist, in its many forms -- from the radio show, to the album, to the mixtape has long been a part of how people discover, listen to and share music. As the world of online music grows, the playlist is once again becoming a central tool to help listeners successfully experience music. Further, the playlist is increasingly a vehicle for recommendation and discovery of new or unknown music. More and more, commercial music services such as Pandora, Last.fm, iTunes and Spotify rely on the playlist to improve the listening experience. In this tutorial we look at the state of the art in playlisting. We present a brief history of the playlist, provide an overview of the different types of playlists and take an in-depth look at the state-of-the-art in automatic playlist generation including commercial and academic systems. We explore methods of evaluating playlists and ways that MIR techniques can be used to improve playlists. Our tutorial concludes with a discussion of what the future may hold for playlists and playlist generation/construction.
Using recommendation engines feels like wading through the sewers. Eventually you find some gems but you have to pass a lot of shit on the way. Listening to actual DJ sets is gem after gem. Only problem is if you are looking for stuff to dj with yourself, most of what they are playing is yet unreleased or private edits that never come out.
I haven't tried AI DJ, so I can't comment on that, but I find it hard to empathize with the author. Not because the criticism lacks merits, but because there is no real attempt to explore the pro/cons of the tech. I see this pattern often with people who complain about AI. They pick a narrow case where it isn't good at and use it to dismiss the whole thing. AI isn't a human, it's going to have its limits.
Same thing I saw in AI-assisted coding. People complaining how AI- enabled some XYZ security risk, it's bad, it's crap. This could be true, but why ignore the fact that you create a full blown native Mac app, with a single sentence? That should be good for at least a few things. Right?
Basically it's because what "AI" can do is extremely different from what "AI evangelists" claim it can do.
I haven't seen a single "AI evangelist" address any concerns and limitations, other by than "throw more AI at it" or "it will get better in 5 years, just in time for cold fusion".
> you create a full blown native Mac app, with a single sentence
Like they created a full blown C compiler that "could compile linux" but in reality didn't pass its own tests?
If you constantly cry wolf, no one's going to believe you when the wolf actually comes.
> I haven't seen a single "AI evangelist" address any concerns and limitations
You see what you choose to focus on. I come across many people who are excited about the possibilities of AI-assisted coding, who are frustrated by its limitations, who share strategies for overcoming or avoiding those limitations, and s on. For a concrete and famous example, I would put Andrej Karpathy in this category. Where are you looking that you're not finding any of these people? linkedin?
In my experience the people who are excited about ai assisted coding are people who aren't good at coding in the first place and don't care about quality, consistency, or understanding what they are having it write, and people who have a vested interest in ai coding tools being used (leadership who want to say "my team uses ai" and "ai experts" who have a personal brand dependent on ai being successful)
The user you're replying to has made many similar posts like this. I previously tried engaging in good faith. I try not to fall into the XKCD 386 trap now, my time is better spent with Claude Code. Hope I can help save you some time too!
> Basically it's because what "AI" can do is extremely different from what "AI evangelists" claim it can do.
You always have people at both sides of the aisle though - people who say it can do much more than it can, and people who say it can do much less.
It's the same with all technologies - robotics, crypto, drug discovery, the internet, digital cameras, quantum computing, 3D Television, self-driving cars - it was probably the same with the steam engine. All of these will have had people who said that the technology would be useless and die (e.g. Napoleon and the steam engine), and others that would have said it was totally transformative.
Pointing to people who hold extreme opinions 'for' a particular technology that are overly-bullish, and then dismissing the technology based on that, isn't a particularly good strategy in my opinion.
> Like they created a full blown C compiler that "could compile linux" but in reality didn't pass its own tests?
Who's "they"?
> If you constantly cry wolf, no one's going to believe you when the wolf actually comes.
Who's "you"?
You seem to believe all AI advocates are of the same hivemind and they somehow think and behave collectively. Have you considered that they might be different people with individual opinions and motivations?
It's easy to address the limitations of AI by simply not using AI for those. No one forces you to use AI for tasks where its capabilities are limited; regardless, there are plenty of tasks where they aren't.
AI is very good at some things and very bad at others. Early on, many thought chess would be one of the last things mastered by computers, but they were wrong. It makes no sense to take the statement "AI is extremely bad at this task compared to humans" and conclude that AI must be useless or a waste of time.
In this case, the AI DJ is bad at picking out classical music. Okay, sure, whatever. But that doesn't automatically mean the AI DJ is bad at everything.
> Like they created a full blown C compiler that "could compile linux" but in reality didn't pass its own tests?
You are strawmanning hard here. Who is "they"? You are putting all "AI evangelists" into the same blob here, and instead of answering the questions at-hand you ignore them and respond in an ad-hominem style by attacking a project that someone else made, completely unrelated to this entire thread. That is not good faith discourse!
So you want to bring every conversation on the topic down to the level of the most idiotic fanboys making the most outlandish claims that are easiest to shoot down?
If this was JUST directly in response to these “AI evangelists”, a group which I’ll ignore that you’re unfairly treating as a monolith, that’d be fine.
However, every post here that says the slightest thing positive about AI’s abilities is always met with “yeah well it can’t do my dishes for me so it’s total garbage!” BS.
You yourself are bringing up “making a compiler” out of nowhere. Nobody but you brought that up here. Yet you’re using it as the be-all end-all yard stick, simultaneously completely ignoring and completely proving the argument that you’re replying to.
It’s amazing how big a % of the developer community has started acting like intentionally unintelligent petulant children the moment they’re faced with an iota of the sort of job security risk they’ve been inflicting on others for decades. Some of you need to grow up.
This appears to be a troll account, that only ever engages in heated discussions. Please, do not engage with it, folks :) On a related note, has anyone noticed actual bots commenting on HN? I sometimes feel discussions are a bit weird here.
Oh, I was not comparing it with out-sourced development and was instead comparing it with developing it oneself.
Sure, outsourcing is similar, but the difference is one uses a process that is inherently probabilistic and will show up in every result, while other just depends on the probability of you getting a good team.
I suspect the unspoken premise was that it was all in context of people who - just like those who hire contractors - don’t have the capacity to do it themselves.
In this context I suspect a SotA LLM could sometimes beat some cost-comparable UpWork professionals in both quality and spec adherence. In other words, if you need an app and can’t do it yourself and have a tight budget, LLMs are quickly becoming a viable option for more and more complex apps (still only simple ones before it produces junk, but progress is pretty appalling)
>beat some cost-comparable UpWork professionals in both quality and spec adherence...
I am not sure I want to keep paying for something that needs some amount of luck on my side, to be useful. Writing elaborate plans for LLMs also feels a bit pointless when there is no hard and fast rule about how much of it will be followed ..
Apparently some people appear to be doing it, but I am afraid it is not something that will have a universal appeal..
> This could be true, but why ignore the fact that you create a full blown native Mac app, with a single sentence?
I would guess it's for the same reasons that you're ignoring all the fixes necessary to get to an actual "full blown native Mac app". It's rarely a single sentence unless your app does something trivial like printing Hello World.
What a strange take - you dismiss valid criticisms of Spotify product, just to venture off into the land of "well you can create a mac app with one sentence" as if that would matter here.
> This could be true, but why ignore the fact that you create a full blown native Mac app, with a single sentence?
Isn't that a bit overblown? I just fired up Copilot in VSCode and typed in "make me a DAW plugin that will inject MIDI control changes into the track output" and it didn't even know where to start.
I've tried using Spotify and similar services that try to track your preferences but they're just, I don't know, boring. I much prefer the challenge of a human-picked DJ set.
I usually listen to dublab (los Angeles, cologne, and Barcelona) and nts1 (usually London) and nts2 (location rotates). They have 1 or 2 hour DJ sessions (live or recorded) and your hear some music that you normally wouldn't be exposed to and sometimes you hate it but usually not.
I've tried using Spotify and similar services that try to track your preferences but they're just, I don't know, boring. I much prefer the challenge of a human-picked DJ set.
The significant problem that AI faces in automatically curating something is that the input data is usually pretty terrible. It's based on either similarity of the thing being curated which doesn't work because people don't want things to be too similar or to dissimilar, or it's randomness which doesn't work because it's too discordant, or it's based on patterns in the data (people who listened to X listened to Y, so recommend Y to people who listen to X) which works but only if the listener's taste aligns with the majority. If you introduce multiple sources of patterns in the data you quickly lose any variation and things stop being interesting.
This is a hard problem. No one has ever really solved it, despite Spotify, Netflix, YouTube, etc investing hundreds of millions into the space. Humans are probably just too fickle to accept that an algorithm can choose for us. It lacks the social proof that a tastemaker like a DJ brings.
> the input data is usually pretty terrible. It's based on either similarity
I found luck just using a LLM to chat about my tastes, what I like, what kinds of songs I want to discover ... it does a good job and is able to also give me background.
I think what you're describing is what people working with recommender systems call serendipity. Maximizing serendipity, while maintaining relatively high relevance/recommendation success rate, is supposedly a pretty difficult problem to solve. I'm not sure if LLMs have changed that.
These are great, thank you so much for sharing the recommendations. I tuned in to NTS and casually just kept on listening for a very long time. If anyone else has good recommendations, I'm all ears. Thank you.
Check out mixes by Blackest Ever Black label (now defunct) from NTS and Berlin Community Radio, listening to them literally feels like a journey. Funny part, sometimes they use a contrasting tune to end a mix, which creates a feeling similar to movie credits roll in the end.
Spotify should have twitch-style, human hosted, radio shows (intentionally no video) with live chat. with full access to the catalog, it would easily bring back communities built around good music.
That would be good experimnent and could actually work.
I would love to try it however they would have to solve "global song availability" and "Sponsored songs only Stations".
But if they did try there is the chance of some niche communities forming.
It wouldn't even need to be live to begin with. A narrated playlist with a DJ and basic control functionality such as fading into songs or a voice over.
Not trivial but doable and I wonder why they never tried that.
It has been tried. I don't remember its name, but I remember that they have changed names at least once. It's a pretty obvious "app" for Spotify's API which they opened up a few years ago.
you have to do your own search and play, but some of the stuff by unknowns and famous artists giving back is profound, they KNOW when they hit it, all live, mostly acoustic and all useing musicians, no tape, no sequencers.
listen to one such performance, and maybe you dont need anything else for a week.
This isn’t really related to the core argument, but I think the author would be better served on just about every count by switching to Apple Music (Classical). The discovery and organization mechanisms are built for classical music first and make the whole project of finding, saving, and enjoying the material way better. They include PDFs of the booklets, for goodness sake! (And let you cross-shop recordings of the same piece by different performers so, so easily.)
> I’m aware that many people are unfamiliar with this musical tradition, but it forms one of the sturdiest pillars of what we casually refer to as “western civilization.” Plus, it’s a whole lot of really enthralling music.
I really had to push to keep reading past this part.
But this piece doesn’t really say anything surprising anyway. Spotify isn’t for classical music. There are other services that are.
> Am I naïve in expecting Artificial Intelligence to be smart? Is my interpretation of the word “intelligence” too literal?
I wish more people would ask themselves those questions.
Sadly Charles himself didn't appear to conclude that yes, it's naïve to expect AI to be "smart" (whatever that means) and yes, he and many other people get hung up on the word "intelligence" in AI, a field that's been called that since the 1950s.
The real problem with Spotify's DJ is that, if you use it a lot, it gets into a feedback loop where it keeps playing the same songs that it serves you up because it thinks you like them. It's pretty bad at finding new music which is ironic because I find Spotify's Discover Weekly algorithm to be quite good (sometimes)
Or someone cueing up well-known pop/dance tunes at a wedding/disco. Last time I was at one they weren't generally firing up symphonies, string quartets then doing a deeper drop of a heavy hitting baroque banger to see the bodies hit the floor.
Yes it is. "Classical" without further context means any part of the tradition of Western art music with written score. Classical-era classical music is a subset of classical as a whole.
The Spotify AI DJ makes some pretty cool sets for me, but I listen to Hyperpop and Outrun type electronic music, stuff like that. A DJ spinning sets of classical music is pretty weird haha. I’d recommend just listening to the Classical New Releases playlist, which is excellent.
I do wonder how people can be satisfied with automatic music playlists. I was entertained by this for maybe a few hours when Pandora was new, but they all seemingly always devolve into either playing weird shit, playing the same 50 songs over and over again, or playing whatever new release shilled crap the record companies are paying to promote. Yet it seems like everybody else these days is a Spotify addict. I guess most people are fine with it.
Pandora is the only one that even remotely came close to something worthwhile, for me. It usually picked stuff that I wanted to hear; and that was a decade ago. Every other selection service regularly fed me garbage.
Pandora was worthless, though, because of their skip limit (even in the paid version). Even with its effectiveness, it would still feed me junk.
This guy is a classical music guy, though, and all the pickers suck, for that. Classical has been treated badly, forever. I am extremely disappointed that Apple segregated classical into its own app, because I have always enjoyed mixing it in with my regular music.
One thing about classical music, is that every performance is a “cover.” Who performs the piece is just as important as who wrote it. None of the selection services seem to understand that.
MP3 tags are pretty much worthless. They are incredibly limited, and I don’t know why they have never been improved.
What would you add to MP3 tags? ID3v2 already has separate fields for section/title/performer/conductor/composer/lyricist, it isn't the spec's fault Spotify doesn't use them.
How would we classify Zappa, or Secret Chiefs 3? Are they jazz, alternative (a worthless category), rock, pop, heavy metal, comedy? Depending on what you listen to, it could be any one of them. Also, each song could be in multiple categories. Boz Skaggs was known for disco-style pop, but he was an outstanding blues performer, and many of his songs reflect that mix.
This is really a music industry problem, and software just reflects that. The bug is really in the Requirements phase.
Well, it's less of a technology problem than it is an industry one. You can have multiple entries in the genre list and they're freeform, for example Ambient;Electronic, in both ID3v2.3 and ID3v2.4. For Vorbis Comments, you have multiple GENRE= tags. Some players support this.
In my interactions with distributors, it seems streaming services tend to support up to two genre classifications; though they're pretty outdated and general (even more general and dated than the Winamp genre list). I don't think they use the metadata presented much in the classification; in fact Spotify does its own estimation of 'energy' and other subjective emotions using various classifier algorithms.
I listen to a lot of old music - 1950s, 1960s. I don't really have peers who listen to it so discoverability is a real issue. Pandora was amazing for me ~20 years ago, it introduced me to songs I never would have heard. Especially in the 50s you had a lot of "one hit wonders" so just listening to a band wasn't a great way to find other songs that I would like.
I don't really use Spotify so I can't compare but Pandora was awesome. I've found Youtube playlists to be the best replacement so far.
For me, it's been 'gone' since 2017, when they shut down in Australia. They used to be available everywhere, but locked off international audiences when laws became inconvenient.
(Obviously you could VPN in, but it's a meaningful hassle.)
Not really. I stopped using it at some point a very, very long time ago, and I don't recall why. I've moved to Youtube so it hasn't occurred to me to revisit.
Spotify's discover weekly was genuinely good when it first came. It was on another level from other recommendation services. Maybe 90% of the music I've bought on Bandcamp, I would never have known about if it wasn't for discover weekly (Bandcamp's own recommendation/discovery features are lousy).
But somehow, probably from a combination of rights owners gaming it and Spotify gaming it, DW is a pale shadow of its former self.
I've been working on and off on something [1] that tries to address this problem through somewhat manual curation. You choose what genres you're interested in, and get auto updated playlists created from your music library.
I have a few other experimental features in the pipeline that will expand the music selection, but they are not there yet.
My observations are that the average person is bothered by the slop of modern playlists full of AI music, but they don’t care enough to do anything about it.
Personally I dropped playlists long ago for YouTube dj sets which are a million times better than Spotify’s AI dj. Some of this is not a tech failing but the DJs have access to unreleased tracks, their own private edits, and are more willing to do more bold things. The AI DJ will never drop a surprise change that makes the crowd scream.
It's very miss-or-miss; you need to be willing to thumb down 95% of what Pandora thinks you'll like. But with enough care, it's a good discovery channel.
> Am I naïve in expecting Artificial Intelligence to be smart? Is my interpretation of the word “intelligence” too literal? And when an AI behaves stupidly, who’s to blame? The programmers or the AI entity itself? Is it even proper to make a distinction between the two? Or does the AI work in so mysterious a way that the programmers need no longer take responsibility?
IMO this is a programming/prompting failure - not a failure in the general capability of 'AI'.
We can prove that an AI can understand this with a basic prompt:
This is a minimal base prompt, with no fine-tuning, with the same user prompt, which shows that an AI will respond correctly by default. Presumably either the AI they are using is a weak model, or their prompt is encouraging the model against this (e.g. maybe the prompt says 'return one song based on the suggestion, and then songs from similar artists after')
> I’ve heard people claim that an AI can compose music. But how can that be when it can’t even grasp basic concepts in music?
Trying to infer the underlying capability of AI to generate music based on a badly-prompted Spotify DJ feature is always going to have it's limits. The proof of 'can AI compose music' will be in the eating of the pudding. AI models have already been able to compose classical music to some extent, and can grasp music theory, so after this point it's just going to be a matter of quality/taste.
Your reference to prompting is pretty disgusting since you try to shift the blame to the user. All the prompts were crystal clear. Trying to shift any blame on user error is non-sensical stupidity or dumb manipulation in this case.
Also, might I recommend looking at the way the world is, not the way the world might be. This is one of the ugly AI tendrils this disgusting industry is putting into everything, bringing ruin to the world. This is the actual reality of it, making the world a dumber, less interesting more stupid place.
> Your reference to prompting is pretty disgusting since you try to shift the blame to the user.
I'm shifting 'blame' to Spotify, rather than the user or the AI model - although blame is probably a pretty strong word anyway for what is probably just supposed to be a fun DJ feature.
> All the prompts were crystal clear.
We don't know what the prompt is, because the FULL prompt will be a combination of the base prompt plus the user prompt. It's trivial to show that a modern model with a minimal base prompt will return correctly (as per my original post), so IMO there is probably something in the base prompt which is encouraging the model to return differently.
I wanted to clarify the first two points, but i'll not respond to the rest of your comment as it's a bit overly-emotive (calling what I say disgusting, rambling about the downfall of society as a whole etc).
> it was a skill issue on the part of the Spotify engineers writing the internal system prompt for their slop DJ
Spotify are currently making a big deal about not writing any code - I attended a webinar this week where one of the slides proudly trumpeted this fact:
“
0 lines of code
Spotify's best engineers have not written a line of code since December.”
> Your reference to prompting is pretty disgusting since you try to shift the blame to the user
Users are often to blame in many varied cases and there should be no taboo around discussing this. I think maybe some people hear that you should never blame rape victims for rape and then go running wild trying to apply that as a general principle of never blaming anybody who is in any way a victim of anything, even when the "victimhood" is simply some piece of trivial software not working well. But we're not talking about rape so your intense rejection ("disgusting") is completely off the mark.
He doesn't really even dig into the quality of Spotify's AI DJ apart from pointing out, in a very roundabout way, that it was designed for popular music.
Classical is a harder (or at least different) problem and it's why specialist apps like Apple Music Classical exist.
Classical isn’t harder. It’s just so niche that leadership at spotify never bothered. It has a whole different taxonomy; it’s composer, not performer based etc.
Spotify isn’t against new taxonomies outside of weatern pop music. The India launch, where ragas were super important, shows it. But the Indian market is vastly larger than the small (albeit loud) number of classical music enthusiasts.
All that is to say, it’s a business decision, not a tech or AI problem.
(I really like classical music, too, btw., so please don’t read this as me not respecting that user base.)
> I don’t listen to pop songs. I prefer music of the 500-year tradition (...)
And who apparently wants to stream music, it is wild he's not subscribed to Apple Music Classical which exactly circumvents all complaints in this article...
Spotify DJ is terrible - just like their Apple TV app (which is perpetually buggy). However, their radio mix feature is so good that it's become the #1 reason I can't switch.
In this case, I see the author's point. The DJ isn't being advertised as "a narrow tool to select some random pop tunes". If an average person is told this is AI, has a full text interface and responds with "sure I'll do what you asked" and appears to understand, then they expect it to do what it is asked.
We're told its better than people at selecting songs (e.g. has the combined wisdom of all music and music experts), basic requests like "play the first movement of Beethoven's 7th" don't sound hard for an average person with limited / no musical expertise. If I said "please play the entire 7th symphony", and the tool responds with "sure, I'll play the whole thing", then proceeds to play the Beatles, I'd say that's a fair thing to point out as a shortcoming.
Its only obvious to tech people that understand that the technology has extreme limits and only works well on areas with abundant high quality data and labels, and can't be expected to reason like a person at all in many cases, that those limits seem as obvious as hammer / screw-driver. And that given how spotify developed these models, they probably didn't really intend classical or test that area -- so it fails despite sounding confident.
But maybe we should stop advertising screwdrivers as universal intelligence? There's a lot of mott and bailey going on. When AI makes mistakes its "just tools, stop expecting intelligence." However, when people question the AI hype its "humans make mistakes too, LLMs are truly reasoning and better most humans already." And "the entire labor economy will be replaced, human DJs will cease to exist.".
In this article we see proof that the words people use to describe a phenomenon influence how they think about that phenomenon: what they expect, what they assume, how they reason about it. https://en.wikipedia.org/wiki/Linguistic_relativity
Every time someone calls an LLM "AI", their brain faults a little more.
This is the profession of marketing's greatest success: inflicting so much damage on the rest of the world.
I briefly tried it when they first launched it, but in less than an hour decided I hated it.
Which I really should have anticipated since I generally dislike music radio "DJ"s too and Spotify's AI DJ is trying to be like one.
In particular it would do things like start playing tracks with no bearing on anything I'd ever listened to, like local South African music which is very far from universally preferred here. I also got the feeling it was pushing "promoted" tracks with little regard to what I would likely like, just like real life radio stations.
I also don't care to have some voice interrupting the music all the time.
I was hoping it would kind of be like their other "radio"s, but it would be more explorative to finding more "similar" tracks to what I have listened to, without seeming to get stuck in a repeating play list.
I suppose it's a cool gimmick for people who are prefer the broadcast radio experience.
I asked this thing to play me some instrumental EDM tracks and it couldn't handle the task. I don't think classical music is even remotely viable. Spotify already really sucks at it. Pouring AI on top definitely won't help the main issue which is gaping holes in relevant content. It just doesn't exist on the platform in most cases.
I'm not surprised. I used the AI DJ twice, on separate occasions, and it played me the same songs, in the same order... Suffice to say I have not used it since.
What’s interesting is this feature (Spotify DJ) really excels when you give it qualitative input “workout music that pairs well with a sunrise” - and can deliver stronger results that hunting for a playlists
> I don't listen to pop songs. I prefer music of the 500-year tradition that encompasses [list of like 50 composers]... one of the sturdiest pillars of what we call "western civilization"
> The use of the word “song” for instrumental music — that is, music that is not sung and hence is not a song — is borderline illiterate.
This guy comes across as incredibly obnoxious. It's shit like this that gives classical music a bad rap as stuffy and unapproachable.
But yes, Spotify and the like are terrible for classical music. Apple Music has a separate app for this, which does a pretty good job and addresses most of these complaints.
But he already explains why it won't work at the beginning. If stuff is cataloged according to a pop paradigm, why would we expect to be able to reassemble it according to a classical one?
Presumably a pop DJ would also mess this up. It's like going to an Indian restaurant and asking what Dim Sum they recommend.
The only reason a human would be able to do this task is that they might be trained in how to find classical music, and they have spent some time learning what is what in that world.
But a Spotify AI is of course going to be trained on the prevailing classification system only.
The whole pitch of AI is that the model is going to be able to make exploit general knowledge outside of the local scope of the problem, just like a human would do. So I would also expect that it would be able to transfer his knowledge of classical music learned in language training, and apply that to the Spotify database.
The classical classification system is equivalent to the classification of cover versions in popular music. Of course, most audio software handles cover versions poorly too, but it's not like it's a completely unknown problem.
As of a few months ago I get AI slop tracks in virtually all YTM automatic playlists, after a few normal tracks, so I abandoned the platform entirely.
Call me when Spotify and YT collaborate with Deezer on labelling AI music as such. Yes, it's a nuanced concept, but the soup YT was serving me was extremely obvious, which was easily confirmed by checking the throughput of the "artists".
As the DJ is an interface to shuffle, and the author specifically wants to listen to unshuffled music the lack of intelligence may not be entirely in the AI.
> I should mention that my perspective might be a little different from most people’s because I don’t listen to pop songs. I prefer music of the 500-year tradition that encompasses (in roughly chronological order) composers such as Tallis, Byrd, Dowland, Gesualdo, Monteverdi, Lully, Blow, Corelli, Purcell, Vivaldi, Rameau, Handel, Bach, Scarlatti, Haydn, Mozart, Beethoven, Rossini, Schubert, Berlioz, Mendelssohn (Fanny and Felix), Schumann (Robert and Clara), Chopin, Liszt, Wagner, Verdi, Brahms, Puccini, Mahler, Debussy, Strauss, Beach, Schoenberg, Ives, Ravel, Stravinsky, Berg, Price, Copland, Shostakovich, Carter, Boulez, Gubaidulina, Pärt, Reich, Glass, Eastman, León, Adams, Saariaho, J. L. Adams, Wolfe, Higdon, Adès, Thorvaldsdottir, Mazzoli, Shaw, Fisher, and many others.
There are apps specifically dedicated to classical music and there are many youtube channels for classical music, with sheet music[1], with visualizations[2], with videos of concerts.
Spotify and it's drop-in competitors were never good for classical music. This article is just another rant on this issue, by someone to whom classical music is so important, a pillar of western civilization, but not important enough to look for other ways to listen.
This is funny. Someone with fine musical tastes uses a feature designed for people who just want background music or radio style picks.
Of course it’s not very smart. It’s probably a shuffle biased by a music classifier RAG that puts Beethoven alongside Trent Reznor’s slower instrumental works.
As a Spotify user, I often wonder how much they're constrained in their choices by their contracts with music publishers. As an example, the fact that you don't have an option to downvote a song - ie, signaling that you don't want to hear it - is such a feature gap that I can't believe it's there by choice.
I wouldn't be surprised if creating a truly great AI DJ was also hindered by this kind of legal shackles.
Spotify is filled with payola, and their claims about it are intentionally extremely misleading while not explicitly fraudulent.
It shows up in all Spotify-generated playlists, so I refuse to listen to them. I assume
their shitty AI recommendations are similarly filled with cancer.
It's amazing: How AI can now generate accurate images from a textual description but storefronts and music platforms can't recommend suggestions actually matching what I like.
Instead they're just thin veils around paid-promotion.
While I sympathize with the issue and have experienced similar problems with classical music, I found the listing of composers and the holier-than-thou attitude (because “pop is bad”) grating and soured the rest of the post.
Ha, despite all that the author exposes themselves as a filthy casual anyway by focusing on the work itself, as if Spotify were looking up a score. Instead “of course” we are looking for a recording, principally keyed by, for example, conductor (orchestra), director of music (choral), and/or a soloist or key ensemble members. Searching by work is like typing in “Hallelujah” to find a version by someone other than Leonard Cohen.
Snobbery sniping aside, I empathize with their sentiment, and their work was worth reading. Spotify’s whole UI is far too complicated and I wish they would go the Facebook route of breaking out the separate products into separate apps. Jumbling podcasts, pop music, and covers — sorry, classical music — is a bit weird.
>the author exposes themselves as a filthy casual anyway by focusing on the work itself, as if Spotify were looking up a score
Isn't it the job of a DJ to pick a good recording? Petzold's test seems reasonable to me. As a classical listener, if I want a specific recording I'll just play that recording. The main function of the DJ is music discovery. Perhaps they know of good recordings I haven't already heard.
I haven't listened to radio for over a decade, but back when I did I listened to BBC Radio 3, where the DJ played classical. "DJ" does not necessarily mean somebody beatmixing dance music in a club. Spotify's "AI DJ" is obviously meant to simulate a radio DJ.
I am indeed a human. The variable quality of my contributions here ought to attest to that!
My grandfather was a typesetter and print designer. My other grandfather was part of Gill’s circle and his bookplate was inscribed by him. My first and only kickstarter in which I participated was Linotype: The Movie. I am currently reading Jury’s Type Designers of the Twentieth Century. I also have Peace’s catalogue of Gill’s inscriptions on my desk. Justin Knopp from Typoretum set my personal card from his digitized collection of rare founts. I’m interested in type and page design and I do like em dashes.
But I also just really like iOS’s automatic replacement of 2x hyphens with a dash.
A lot of this discussion makes more sense if you know the history of The Echo Nest and their acquisition by Spotify.
The Echo Nest was one of the most interesting music-tech companies ever built: a music intelligence platform spun out of MIT that analyzed audio, metadata, web text, artist similarity, genre structure, and playlist construction. Spotify bought them in 2014 specifically to strengthen music discovery and recommendation. At the time, Spotify said the deal would let it use The Echo Nest's "in depth musical understanding and tools for curation", and even said the Echo Nest API would remain "free and open" for developers.
If you ever used the old Echo Nest APIs, Remix SDK, demos, Music Hack Day projects, or Paul Lamere's experiments, that was a golden era. Echo Nest had open APIs for artist similarity, track analysis, playlisting, "taste profiles", ID mapping across services, and beat/segment-level music analysis. Paul Lamere's whole ecosystem of demos came out of that world: Boil the Frog, Sort Your Music, Organize Your Music, playlistminer, and later Smarter Playlists. His GitHub still points to a lot of that lineage, and his blog is still active. In fact, he posted just this month about rebuilding Smarter Playlists after ten years of use.
The sad part is that the open developer platform mostly did not survive the acquisition. By 2016, developers were being told that the Echo Nest API would stop issuing new keys and then stop serving requests, with migration to Spotify’s API instead. Community discussions at the time also noted that some Echo Nest capabilities, especially things like Rosetta-style cross-service mapping, were not really carried over.
That's also why Spotify's current AI DJ is so frustrating. The problem is that "AI DJ" is not the same thing as a system that deeply understands musical structure, discography semantics, performance history, or classical work/movement hierarchy. It's a recommendation + narration layer, not a true MIR-native musical intelligence system.
If you're interested in the research side of this field, the conference is ISMIR: the International Society for Music Information Retrieval, which is literally dedicated to computational tools for processing, searching, organizing, and accessing music-related data. That community is still very active. The ISMIR site describes MIR exactly in those terms, and the 2010 Utrecht conference included Paul Lamere's tutorial, "Finding A Path Through The Jukebox -- The Playlist Tutorial."
>gffrd on June 26, 2023 | parent | context | favorite | on: Show HN: Mofi – Content-aware fill for audio to ch...
>Yes! It was "Infinite Jukebox," created by Paul Lamere ... it was awesome because it would analyse a track, then visualize its "components" and you could watch as the new "infinite" track looped back on itself and jumped from point-to-point in the original track to create an everlasting one.
He created some excellent products from the Rdio API, and later Spotify ... and I believe his analysis engine ended up being the foundation upon which Spotify's _play more tracks like these_ capability is based.
>Looks like he's moved over to publish on Substack -- there's a recent(ish) post reflecting on 10 years of Infinite Jukebox:
>However, that wasn't the end of the Infinite Jukebox. An enterprising developer: Izzy Dahanela made her own hack on top of mine. To make it work without using uploaded content, she matches up the Echo Nest / Spotify music analysis with the corresponding song on YouTube. She hosts this at eternalbox.dev. It runs just as well as it ever did, 10 years later.
>DonHopkins on June 28, 2023 | parent | context | favorite | on: Show HN: Mofi – Content-aware fill for audio to ch...
>I was working on some music retrieval stuff in 2010, so I joined the EchoNest developer program and played around with their web apis that let you upload music and download an analysis that you could use in all kinds of cool ways. They had an SDK with some great demos and example code. I discussed it with Eric Swenson and Paul Lamere, and had the chance to hang out with Paul Lamere and Ben Fields at ISMIR 2010 (the International Society for Music Information Retrieval conference) in Utrecht, where they gave a tutorial about playlisting:
>Tutorial 4: Finding A Path Through The Jukebox -- The Playlist Tutorial. The simple playlist, in its many forms -- from the radio show, to the album, to the mixtape has long been a part of how people discover, listen to and share music. As the world of online music grows, the playlist is once again becoming a central tool to help listeners successfully experience music. Further, the playlist is increasingly a vehicle for recommendation and discovery of new or unknown music. More and more, commercial music services such as Pandora, Last.fm, iTunes and Spotify rely on the playlist to improve the listening experience. In this tutorial we look at the state of the art in playlisting. We present a brief history of the playlist, provide an overview of the different types of playlists and take an in-depth look at the state-of-the-art in automatic playlist generation including commercial and academic systems. We explore methods of evaluating playlists and ways that MIR techniques can be used to improve playlists. Our tutorial concludes with a discussion of what the future may hold for playlists and playlist generation/construction.
>[...]
Some of the most interesting Echo Nest descendants are still around in one form or another. Paul Lamere's current/public projects include Smarter Playlists, and his GitHub still highlights SortYourMusic, OrganizeYourMusic, playlistminer, and BoilTheFrog. Glenn McDonald’s Every Noise at Once is another major descendant of that tradition: an enormous map of music genre space. Glenn's own site still describes it as an "inexorably expanding universe of music-processing experiments", and the public genre pages now explicitly say they're a long-running snapshot based on Spotify data through 2023-11-19. After Spotify's layoffs in 2023, TechCrunch reported that Glenn lost access to the internal data needed to keep Every Noise fully updated, which is why it now feels more archival than alive.
Back in 1998 when I was working on The Sims 1, I proposed in my review of the design document something I called "Moody Music": essentially a soundtrack plus a synchronized semantic/emotional control track that could affect gameplay over time. The idea was that music wouldn't just decorate the simulation; it would change it: influencing mood, motives, relationships, skills, timing, and even triggering events at specific musical moments. I wrote that up in my review of the 1998-08-07 Sims design document, along with the broader idea of letting the game recognize a player's own CDs and fetch associated "moody tracks" from the network.
Don’s review of
The Sims Design Document,
Draft 3 – 8/7/98:
>I have some ideas about how the music could effect the game, that
I will write up more completely later. In a nutshell, the people in
the house could have a cd or record collection to choose from, each
record an object that has the sound (audio wave and/or midi) and a
“moody” track synchronized with the music. Playing the music
also plays the moods into the environment that the people pick up
on. Music can subtly effect how people react to the environment,
objects, and each other. It can effect their motives and even their
skills temporarily. For example, you might be able to clean the
house better and faster if you put on some up tempo bouncy music.
The player should be able to assume the role of disc jockey on the
radio, and play from another larger library of music and
commercials, that effect the peoples moods and buying habits. The
TV of course is another source of mood altering temporal media,
with commercials and shows that should effect different people
differently. But the most important part of this idea is instead of
the game effecting the music that’s played, the music effects how
the game plays! The ultimate way for the user to effect the game
via music, is to insert one of their own CD’s into their real
computer’s CDROM drive, and the game would recognize it, and
start playing it (maybe with a simple cd player interface to select
the song). There could be a database associating the unique ID
number of the CD with a table of contents and “moody” tracks that
tell how the song effects the peoples emotions over time, with
"percussion" events at dramatic moments of the music that can
trigger arbitrary events in the game (like provoking a fight that was
brewing, or triggering an orgasm at just the right place in the
song). We hire monkeys to listen to well known CD’s, and enter
time synchronized tracks with semantic meanings in Max (like
note tracks, and user defined numeric tracks) or some other
timeline editing tool). Put the database up on the web for instant
retrieval, so when somebody sticks in a new CD, it downloads our
“moody” tracks that go with it, and it starts playing and effecting
their game! Streaming emotions over the net! Eventually there
should be an end-user tool so people can record their own
responses to music as moody tracks they can use in our games.
This mechanism could be used in all kinds of games, to varying
degrees of effect. I’m not saying that music should be the only way
to control the game – it’s more like a subtle background effect, but
there certainly could be a scenario where you try to accomplish
some task (like taming a wild beast) by using only your musical
taste and timing. The real bottom line benefit is that you get to
listen to your OWN cd collection of music you want to hear,
instead of being driven crazy by the repetitive music bundled with
the game.
In hindsight it was quite adjacent to MIR, affective computing, adaptive soundtrack systems, and some of the ambitions that Echo Nest represented. That's why I was so excited about The Echo Nest in 2010 when I was working with Will Wright at the Stupid Fun Club on a music spatial organization and navigation system called MediaGraph.
MediaGraph Music Navigation with Pie Menus Prototype developed for Will Wright's Stupid Fun Club
>This is a demo of a user interface research prototype that I developed for Will Wright at the Stupid Fun Club. It includes pie menus, an editable map of music interconnected with roads, and cellular automata.
>It uses one kind of nested hierarchical pie menu to build and edit another kind of geographic networked pie menu.
> [list of 20 classical artists] I’m aware that many people are unfamiliar with this musical tradition, but it forms one of the sturdiest pillars of what we casually refer to as “western civilization.”
> Unfortunately, this tradition is not much respected
> The use of the word “song” for instrumental music — that is, music that is not sung and hence is not a song — is borderline illiterate.
This writeup is insufferably pretentious. It almost reads like a caricature of someone that listens to classical.
Prompted playlists is a beta feature designed to cater to most users. They are likely using a heavily quantized model, fine tuned on common use cases.
Is it really surprising that it doesn't cater to the fringes of Spotify's user base from the get-go?
Clearly the author believes that their taste in music is the superior one, and so Spotify not designing their product experience around their tastes is "appalling."
Then you get absurd rants like this:
> I’ve heard people claim that an AI can compose music. But how can that be when it can’t even grasp basic concepts in music?
Almost like these are two completely separate models, in two completely separate products.
"I should mention that my perspective might be a little different from most people’s because I don’t listen to pop songs. I prefer music of the 500-year tradition that encompasses (in roughly chronological order) composers such as Tallis, Byrd, Dowland, Gesualdo, Monteverdi, Lully, Blow, Corelli, Purcell, Vivaldi, Rameau, Handel, Bach, Scarlatti, Haydn, Mozart, Beethoven, Rossini, Schubert, Berlioz, Mendelssohn (Fanny and Felix), Schumann (Robert and Clara), Chopin, Liszt, Wagner, Verdi, Brahms, Puccini, Mahler, Debussy, Strauss, Beach, Schoenberg, Ives, Ravel, Stravinsky, Berg, Price, Copland, Shostakovich, Carter, Boulez, Gubaidulina, Pärt, Reich, Glass, Eastman, León, Adams, Saariaho, J. L. Adams, Wolfe, Higdon, Adès, Thorvaldsdottir, Mazzoli, Shaw, Fisher, and many others."
This is not about AI, the author is mostly just pointing out that Spotify was not designed for classical music.
This is a product issue. Spotify DJ is essentially “shuffle with some voice interludes”. There’s probably some non-AI code in there to explicitly prevent it from playing an album end to end.
Besides, AI is not one thing. It’s weird to generalise “This beta spotify feature doesn’t serve me, hence AI is useless”. For example, when the author says “if it can’t do this, how could it compose music?”, that’s a category error.
Honestly the whole post and tone are just baffling. It’s mixing up all sorts of opinions and trying to put them under one umbrella, and about 50% of the text is just name dropping specific classical pieces.
I happen to agree that the Spotify DJ feature is terrible, but I think this is a very ineffective way of presenting the argument.
reply