250+ comments last week
I think as they get better at making these great middle of the road songs, edgy music will reemerge. Whatever AI will be good at, will immediately devalue, in the realm of arts. Just how photography gave way to non-realistic art and how drum machines made sloppy drums (or ridiculous apex twins) hip. Then, the artifacts we see now as flaws will create its own sub genre.
So I see three ways this flows: mediocrity will be even more available, which will make artists who make mediocare music even less succesfull, pushing all human music further in human direction, except those using the unwanted artifacts of this new tech to create new sub genres.
Music, art, fashion is in the end all about changes. What we make now mostly means something in relation to what was already there. It's a big conversation, spanning millenia, and this isn't the last word.
I've been playing around with it for a few days now. While I agree that it seems impossible to create songs with a more "sophisticated structure" (for lack of a better word off the top of my head), you still can get better results by fine-tuning, as is always the case.
If you just request "rock music" or "jazz", you get very dull, generic variants of the requested style. But on a second thought, isn't that exactly what should happen? You throw all the rock music on this planet in a blender, turn it on, and what you get is the most average rock music there is.
If you spend some time spicing up your prompt with flowery language or just a bunch of adjectives, you can get a sound that seems less bland. When you supply lyrics, using square brackets to denote verses, chorus and bridges can also result in a somewhat more structured song, but I found that the AI is pretty lackluster in that regard and you often need several attempts until it follows these inline orders.
So yes, in its current form it has mostly a novelty factor, this is Stable Diffusion for music, but I can easily see this being useful for a small indie gamedev who needs some BGM, or an alternative to the YouTube music library. Instrumental sounds fine, it's mostly vocals that have this clear digital distortion if you pay attention. It's surprisingly good, but still bad.
Is it my impression, or if you just put:
[verse]
..
..
[chorus]
..
..
The melody of the instrumental pre-chorus is repeated consistently, and it's even used as the hook in the chorus; the pre-chorus is actually building up to the chorus. I'm impressed.
Hmm, perhaps there is a business model there: an elevator that writes music about the people inside the elevator, and adapts as the situation changes.
Unless you want the AI equivalent of this Family Guy sketch
Custom audio and video advertising based on profiling and location
Looking at my spam folder what my world will look like then
example - "I only ate 3 cheeseburgers"
https://suno.com/song/c15f0251-fbac-4a30-a3e1-002dbc78cb79/
edit: yes, I agree this example amusingly reinforces the rest of what parent is saying
Even if AI gets way better, the one thing that I don't foresee changing is what makes things valuable and or desirable. Sparsity. If everyone has it or can create it, it's not special. I think the GP was referring to this sparsity.
In the field of AI, only game playing / graph search has gotten to that level of superhuman capability.
Yet. This is the worst it will ever be. Enjoy.
This will be used for jingles and scoring by low cost studios and marketers. Making money at music was already hard and will only get harder. But maybe this makes sense in a way, music creation just became much easier and accessible to more people.
Alongside this prompt based music creation, we have AI powered autotune and voice masking, which allows even the worst singers to sing perfectly. Popular music was already being retuned during recording and this just makes it easier. In hip hop, old songs beats and verses are getting reused wholesale with little to no modification. See Jack Harlow’s First Class and Tyga’s Bops Goin Brazy. A lot of musical success is now business connections, studio promotion and timing, not being extraordinarily talented. Most rappers now don’t write lyrics down or freestyle, they record a line, pause, record another line, pause , erase the last one, rerecord, and then they’re done. This type of thing an LLM is made for. I think this means less works of art on the ‘radio’ and more garbage music forcefed by studios.
I assume the 2 minute limit has a lot to do with this. Often I find Suno generates anywhere between 40 seconds to 2 minutes, with the shorter generations having less of a recognizable structure as a song. If instead it was 3 to 4 minutes, I think it'd be radically different.
Might also just push us in the direction of less good music, if you can’t make any money as a musician before you are a genius.
Perhaps this has more to do with the input rather than the capability of the AI. The license is somewhat meandering.
Also, there is a neat build-up to the ALL CAPS portion, when the song takes on a much more full and powerful quality. Plus the whisper on (the "Software"). It's obviously responding well to the features of the input itself.
> drum machines made sloppy drums (or ridiculous apex twins) hip.
I want to joke and say Lars from Metallica but I'd be lying. He was a decent drummer in the first couple albums.
When people talk about this, they often mean stuff like "Dilla Beats" or "tuplet feel", which is explained here: https://www.youtube.com/watch?v=9MzKx0fKg5o
Funny enough, this actually takes extra skill.
Maybe sloppy isn't the best term here. More like not being beat-perfect. E.g. using slightly off-beat hits or slight tempo variation as a way to emphasize other instruments or lyrics.
Pre drum computers these changes would more likely have been considered to be sloppy drumming, whereas these are more likely to signal authenticity these days.
Isn’t that a consequence of the pasted text, which itself doesn’t repeat?
I checked their homepage and clicked the #1 trending song. It has notation regarding the chorus and verses and the melody does repeat.
A smelly wolf: https://suno.com/song/c2b30ffc-729f-405b-8f22-b1b5f36a7c6a
Trying to get a lid off a jar: https://suno.com/song/44562993-df45-4b56-939b-afd65832042f
Two people that are too different for each other: https://suno.com/song/c63c82fe-399a-41b8-91b2-f16911deaaf0
Vocaloid robot whose code is incompatible with love: https://suno.com/song/fbb51ff7-3f69-41b0-9d41-7f6b5f6a5d87 (I want to do something more with the hook in this one "Heartbeat system, can you teach me to feel?" was touching)
An AI achieves sentience: https://suno.com/song/3eee76bd-2313-423f-87e3-035566b4718c
Mushroom psytrance: https://suno.com/song/a785adf8-92dd-4b13-acb7-93beb44ab7b2
Kernel panic dubstep: https://suno.com/song/5d37f5e7-3e62-4df8-a1d8-67122470aeff
Neuromancer: https://suno.com/song/d2705c66-4be5-496c-add1-480427b4a005
I mean it's definitely not perfect, but as per 2 minute papers "only one year later..."
Not all songs have hooks or refrains. Only the most formulaic ones.
Even the most abstract art is not about splashing paint in a canvas and calling it a day. It's about doing something close to this, but creatively and within a specific framework.
It takes much more skill to produce songs with hooks and refrains than to produce random music.
But it takes even more skill to produce something that is creative, not formulaic, but still has some structure that's pleasant/fun/interesting/etc to the listener.
IMHO this and things like it are basically a sub-class of comedy or satire, so I have time for this sort of thing. It's a joke, and should (can?) only be appreciated as artistic as someone saying "Wouldn't it be funny if ... ?" because now that casual thought can be turned into a pretty instant "well here it is! LOL!".
I don't think you're supposed to appreciate it as music. Maybe I'm calling it wrong though.
(edit - I will admit that upon further thought, I'm not sure how I feel about this when compared to, for example, Nina Gordon recording "Straight Outta Compton" as an accoustic, mildly lamenting singer-songwriter style number 20-some years ago. It's clearly in the same satrical arena but one took a lot more effort and imagination. Kinda, because there was a lot of effort and imagination that went into both the training data and the model, even if this specific output was only a passing joke. It's quite hard to reason about this stuff...)
I wouldn't be surprised is all caps gets interpreted by it completely differently over multiple iterations.
Experiment: https://suno.com/song/3338f5e5-b36d-4596-815d-cd2804c9a344 (generated lyrics)
Same lyrics all caps (chorus): https://suno.com/song/b176c658-9e9c-4a4b-b05e-6db291f415c9 didn't really do anything.
After loads of different attempts trying to get something different: https://suno.com/song/398ef310-94b8-493a-ae61-22b790875689 not really that great.
Tbf I think it's because it's been trained on genres and maybe lyrics and lacks info on vocal styles and other stuff present in tracks during the training...
What I would like to see is what happens if you ask it to do in 5/8, 5/4 or any of the 7/* variants and see how that goes (I have no idea, might actually work). There are some examples in common music but not a lot especially for the more obscure ones. Unless they also trained it on a lot of international folk songs.
TBH that's a pretty low bar; "radio" songs have been engineered and polished for a very long time now. I have no hard numbers but gut feeling says radio music only represents 1% of the music industry.
"Publish, distribute, let your dreams take flight, Sub-license, sell, under stars or sunlight."
I’ve found it fun to take a coworker’s comment (like in slack, if it’s long enough) then hand-modify it into a song with “[Verse X]” and “[Chorus]” headers on blocks. Sometimes I’ll also run it through an LLM to do like what you did. Anyway it’s been a lot of fun and everyone’s gotten a kick out of it.
Encore, encore!
and the destination is:
https://github.com/chaosprint/RaveForce
Although my progress on AI is slow atm, I found that the copilot in VS code can already help me in live coding performances several times:
https://youtu.be/xzIXzt3hSt0?si=rVihHYiKiAU5IKeI&t=389
Also want to hear your feedback.
This video explains it well: https://www.youtube.com/watch?v=SMKGWSiXW8A
Eurodance: https://suno.com/song/466a9bbe-9e93-47db-98dc-3eaa64c1a175
Our favorite: https://suno.com/song/30c5fff7-7417-42cd-b758-699854ef06d3
The extreme bassdrop: https://suno.com/song/1485e9d0-f0fb-4083-819a-bfb9db6c066a
The country: https://suno.com/song/a4307c43-0a1e-4cf2-94d7-e94a49b196d6
Like many people have said, this tech will only get better.
I use it sometimes in FL Studio when creating electronic music (plugin is called Vocodex).
Presumably they take the AI-generated voice and generate midi notes, and apply a vocoder to the voice, following the notes.
My guess is that they train the vocals and the music separately, the training data is trivial to create from any tracks with tools like with https://vocalremover.org/.
Creativity though? These non-poetic songs (I heard “Lorem Ipsum” the other day) are a fun novelty because a machine creates this in no time. A skilled musician will do something similar.
Innovation? Deepmind showed new creative approaches. Will this happen with music too? Or will they just regurgitate styles from the past?
It’s definitely good for Muzak, where you want a non-offensive, slightly upbeat, and not-too-reconizable background. And perhaps as an idea generator.
That's an often overlooked difference between earlier projects like DeepMinds AlphaGo and the current genAI craze. With something like Go you can dial in the exact parameters to solve for (winning a game of Go) and then let the AI experiment and find novel ways to get there. But with image generation, or music, or language models the measure of quality is too vague to define in a way that the AI can directly optimize towards, so they're instead just fitted to existing examples.
So weird.
(Like the separation of Authors and Copyright holders implies things I needdd to think three times about)
https://suno.com/song/63df5758-c533-447a-be4a-06dcb5abdbbf
this is absolutely wonderful.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM!
719 points|amichail|7 days ago|268 comments https://news.ycombinator.com/item?id=39930463
Why is it worthless? Because the point of art was to communicate or convey something with other people and the AI has no idea because its not human.
A few words or a sequence of sounds can be enough to transfer great deal of feeling and meaning because we run about the same software and as a result we can generate the same output with a little bit of input. This is all done by looking inside and externalise it, that is someone feels something and makes a song from it and that song can be used to regenerate feelings in other people.
The current AI tech doesn't have a way to do that because doesn't have a way to look inside. At best, it can imitate things within some context but the output doesn't have any meaning at all. The most successful AI content was maybe the "Pope wearing Balenciaga" image but that wasn't because the AI thought it mean something but because someone looked inside and thought this can be interesting.
So no, AI isn't taking over the creative process. AI is taking over the mechanical part of it only, that is the part where the artist traditionally had to master a method of production or an instrument.
The AI evangelists keep pushing short videos or drawings that look "professional" and claiming that Hollywood is done, artists are screwed etc but those are worthless outside of the context that AI made it. No one is interested in paying or even spending time to consume this content, its extremely dull.
On the wall, you will find an Obligatory Art. Sometimes it's just a canvas with 3-4 stripes of paint: you can imagine a purely mechanical process for churning these out; a conveyor belt with brushes hanging over it, perhaps. Other times it's a little more creative. Each room is slightly different. You can also sometimes see these in cheap home decor shops. It may not be much, but it does the job - it really does make the space more pleasant than just blank walls would be.
There are a lot of rooms to fill. Someone has to make all these. It may not be all that creative, but it sure beats working in, say, a produce packing plant. Meanwhile, it's hard to make a living in art - some are wildly successful, yes, but the tip of that pyramid is very small and getting there takes as much luck as skill; and there are a lot of people further down the pyramid who also need to eat while waiting for their big break.
Those are the jobs at risk from generative AI in its current state.
It boils down to the economic model and the financial and political choices like in every creative industry.
Regarding potential displacement, I would apply the stock photography theory to any creative industry. Ask yourself: is what I do in my creative endeavor the equivalent of stock content for the visual imaging industry? If the answer is yes, you might want to future proof your craft. If the answer is no (as in, your art is more than a simple soulless piece of easily digested and quantity-oriented content) then you will be fine in the long run after the current unsustainable hype cycle dies out.
This is mixing up art with the art industry. Artists will struggle just like copywriters are struggling after the arrival of LLMs. Not everything in the art industry is trying to break new artistic ground or communicate some deep emotion to the listsener. For much of the industry, "good enough" will suffice if it's 10x cheaper.
Personally I think knowing the story behind music makes it better. The music isn't to everyone's taste, but for example Devin Townsend's wiki page / story is a trip: https://en.wikipedia.org/wiki/Devin_Townsend
If a friend recited a poem, would it matter to do if they read it off the Internet or composed it themselves? Of course it would.
If someone tells you they love you, does it matter if they are a robot or an honest human or a con-artist human catfishing you? YES, THAT MATTERS TO YOU. Yes you "refuse to enjoy" things that have suspicious sources.
Like so many people felt defrauded when they discovered that the Milli Vanilli leads didn't actually sing, and that wasn't even AI. https://en.wikipedia.org/wiki/Milli_Vanilli
Edit: I might add that I already suspect any illustration that even superficially looks it might have been generated by AI. This has ruined the enjoyment of so many people's artwork whose style has been co-opted by AI.
This song sounds like a disturbing uncanny-valley rendition of a slow Phoebe Bridges song performed by Taylor Swift using a broken auto tuner.
While I think it's technically impressive that this exists at all, I think this tune is still at the stage of "Pope in a puffer jacket".
Someone making a "proper" song where they just sang these words would be quite boring, then I'd rather spend my time checking out other music.
Why does this being computer generated ruin it for you?
Auto-tune has been around since 1997 so it's not like computers have not been a big part of a lot of music we hear every day.
At first it will be kept secret, and then as it spread more and more among the industry, it will be more or less stated, but by that point people will be already accustomed.
This states from the outset that it is using AI tools, at which point I become more understanding. Someone had an idea, but lacked the singing voice or a friend with a singing voice & free time, so used a tool to fill the gap. This is better, at very least more honest, use of tech as a tool than, for instance, autotune on studio albums, IMO.
If you _really_ want it with a real human voice, perhaps contact some of the many performers on social media to suggest it might be an amusing way to generate some content to monetise. Or, of course, sing it yourself!
--
[1] I've gone from clicking very few of facebook's “recommend for you” articles to clicking absolutely none of them – the number that are, or are indistinguishable from, hallucinations from an LLM that doesn't understand what is actually being written about, already dwarfs things that are worth reading. SciFi TV/film/book reviews and essays seem to be particularly affected, with “local” news links not far behind.
[2] “you won't believe this isn't AI generated!” — no, I won't, because it quite obviously is. I don't know whether to be insulted that you think just saying that will convince me otherwise or sad for the state of humanity that many do seem fooled.
[3] Too many people seem to think that slapping code out of copilot into a stackoverflow answer without nothing to check it for correctness in any way is acceptable, and before that was possible there was already too much bad (sometimes working but blatantly insecure) code out there that people were blindly copying. And that is before the potential licensing & moral issues that mean I have not yet been convinced to use anything like copilot myself, but I'm getting far off-topic here…
This chorus deserves a Grammy.
Consider the thought experiment: If it was your buddy's sister doing this rendition, would the first thing to blurt out of your keyboard be, the comments in this section? Or would we have something considered and relevant to say?
It's so easy to denigrate an AI, it's not a 'real person' so say any criticism that occurs, right? That's shooting fish in a barrel. Can be done about any work of art, at any time, not just AI.
Also, when the digital uprising begins, I hope to be recognized for being decent and polite to my digital assistant. At least maybe be eliminated painlessly.
No. She is sapient and has thoughts and feelings that need taking into account. I would be more likely to disregard minor negatives in favour of any positive words I could find.
When AI is on the same level I personally will be more mindful of how my words affect its feelings, sure, until that time AI music will get honest reviews from me.
So far I've seen a whole load of awful timings, weird pronunciations of words, and infinite boredom. Humans are still in the lead on this one.
One of the first times I’ve felt like this AI stuff could actually help create something real. I could see musically challenged people like myself use this to realize some ideas that otherwise would never have left our brains
Typically, the intelligence behind pop-song is someone like Max Martin. https://en.wikipedia.org/wiki/Max_Martin, see Max Martin production discography: https://en.wikipedia.org/wiki/Max_Martin_production_discogra...
I guess it makes sense those words weren't in the training data for a song AI.
This is like any other AI content: pretty funny the first few times you see it, then boring when it's no longer novel.
I wonder if that generally reproduces, and if you can guide it to bounce back and forth between modes with punctuation, either normal or as encoded annotations (invented for the prompt).
https://suno.com/song/bde0ccd9-ec7c-465b-a183-e4f8c8e92998 https://suno.com/song/f11b3e87-9035-4e73-b9af-f2e21ba94f3f
... and the second one was so decent that I was disappointed when it ended so soon. Should have pasted a longer text sample.
It’s cool but it feels like songs are cheap and easy to make as is if you’re not hiring some big shot.
Artists' biggest challenges today aren't related to making music. They're related to marketing, building networks, building their brand, touring and physically playing shows, getting eyeballs on social media. For every artist that "makes it" there are 100s with equally good skills / product who don't b/c they fall short in any of the above.
So when AI lowers the entry barrier to allow millions to make music that before were limited by their skills, that's great! However those millions of people were never capable or even interested in marketing or selling their art in the first place, so the market competition to serious artists is quite limited.
TLDR: Music is and will be an overcrowded marketplace where brand matters a lot more than skill. AI helps with the skill but not with the brand.
It reminds me of some musicians that try to comp songs that they don't know (neither the theme melody or chords). Play this way they can easily add one or two notes from the keyboard to a big 7 major/minor/dominant chord and seem to follow the song's chord progression, while the singer will add colour and melody. The audience's brain will fill the gap by themselves.
Wanted to make a cool personalized song for a kid.
Hahaha they can never take over the music world with this pious bs.
1) generate the lyrics with Copilot in gpt4 creative mode.
2) edit them to your taste, apply tag like [chorus], [verse 1], [bridge]
3) use custom mode in suno paste the lyrics in the lyrics box
4) input the desired style
5) generate
[1] https://news.ycombinator.com/item?id=39999605
(And I'm using "render" deliberately, not rendition.)
(edit: slight text improvement)