The DALL·E 2 of MUSIC?
Вставка
- Опубліковано 3 чер 2024
- Over the past few months there's been an explosion in AI text to image engines. Could the same happen for music and if so, what does it mean?
00:00 Twoset violin review Huawei's AI smartphone music
02:00 Nirvana song AI
02:40 Text to image AI
04:34 The internet and past 'future technologies'
05:53 How text-to-image AI works
06:51 The Foundation Model
07:50 The diffusion Model
08:34 The current stage of music AI
08:50 MIDI-based models (Musenet, Magenta)
10:50 Audio-based models (Open Jukebox)
12:00 The next steps in future Music AI
13:44 Next steps in visual world (video)
14:30 Implications for future of Music and Composing
16:10 Creative solutions (Reeps One)
SUPPORT THE CHANNEL ON PATREON
/ davidbruce
MERCH:
The Lick Midi Sequence T-shirt I'm wearing:
david-bruce-composer.creator-...
other merch:
teespring.com/en-GB/stores/da...
Follow me on Twitter:
/ davidbruce
Follow me on Instagram:
/ davidbrucecomposer
David Bruce Composer Spotify Playlist:
tinyurl.com/y798swcy
My 2nd UA-cam Channel:
/ @dbc2
Reeps 1 documentary
• We Speak Music | Epi...
David Bowie on Newsnight
• David Bowie speaks to ...
Twoset - Artificial Intelligence Composed this symphony?
• Artificial Intelligenc...
Dalle 2
openai.com/dall-e-2/
Musenet
openai.com/blog/musenet/
Open Jukebox
openai.com/blog/jukebox/
Dreamstudio
beta.dreamstudio.ai/
Midjourney
www.midjourney.com/home/
Imagen
imagen.research.google/
Research:
www.economist.com/interactive...
ryanschultz.com/2022/08/11/co...
Huawei
consumer.huawei.com/au/campai...
Reeps One
• Reeps One ft. A.I. 'Se...
Huwawi
consumer.huawei.com/au/campai...
Nirvana
www.rollingstone.com/music/mu...
Animated instagram
/ ch7av2mjwod
/ make-music-with-artifi...
Google/toronto uni images
www.cs.toronto.edu/~emansim/ca...]
when people think AI art will replace "real" art, they're confusing the practice of art with artistic commodities. sure, maybe AI can one day make art that rivals what we make in quality, and maybe people will consume that art. but people won't ever stop making art, because what's more enjoyable than experiencing art is the practice of making it. people will always want to play music, paint, etc. it's part of our nature as human beings. the point of art isn't just the end product, it's the experience of making something.
So true. An ai can replicate kid's drawings all it wants, but a child will still make doodles if given some paper and a marker.
I think most concern is professional art - we want to make art/music and always will, but the struggle is how to make a living - I tend to follow David Bruce's perspective that professional artists will explore how to integrate ai into their work.
Spot on I think - what's more, I think all the best art/music, is what is created by artists who make what they themselves want to hear/see, rather than catering to an audience. Maybe it will eventually happen, but for the moment at least, I have a very hard time seeing that happen with AI. If I am right, the Max Martins might well be replaced, but the Frank Zappas will remain competitive. Which, IMO, would be a win..
yeah - creators have nothing to worry about. if anything, more people will have access to artistic practice and there will be more variety of aesthetic
people also wont stop consuming human art either. Sometimes what you want out of art is a piece of the artist, not a bunch of pleasant sounds/shapes. I would say that is the majority of the fine art market.
A big part of why we enjoy consuming art is because it gives us insight into someone else’s perspective. Also people will still want to see live music, performance, theatre etc
When radio came along, the music industry panicked, saying "that's the end of anyone buying music!". Soon after came the most prosperous decades for recorded music. I wonder what parallels, conclusions, and opportunities might be drawn from our current historical inflection point. As a full-time screen composer, this is definitely top of mind for me.
Brilliant essay, well done. Put everything into clear perspective and context.
Making inductions based on completely unrelated past situations and circumstances is not really logical.
When radio came along the music industry had a party, a new media to promote music was out.
These things are unpredictable. Not just the music industry either; no one can predict the future, because it is inherently unpredictable. One can envision a whole range of possible scenarios, one of which is that human composers (or at least ones that manage to make a living from it) disappear. But I wouldn't bet on it.
I find the general public is quite blown away by Dall-E and Midjourney. Professional artists, not so much. Similarly, professional composers are far less impressed with the current state of AI music than most other people. Of course, this says nothing about the market - maybe the public will be happy with AI art and music, and since these will be free or virtually free, it may be that no one will be willing to pay for human-composed music anymore. We don't know.
As artist, my worst future dystopian scenario is actually not one in which I no longer sell any art, but one in which I am forced to retrain as prompt writer, and then spend hours sifting through auto-produced imagery to choose the best. What an utterly mindless, boring way to create art, and I suspect the same goes for music. Frankly, I'll go live under a bridge and eat out of dumpsters before I go that way.
But it remains to be seen what will happen. Your example of the radio is not the only such example by a long shot. David Bruce mentioned portrait photography. Yes, it led to portraits being painted in new styles, but to this day there is actually still a lively market for traditional, realist, handpainted portraits. Some of these are done from reference photos!
When digital photography became available, everyone predicted the end of the professional photographer. Why pay someone if you can just take a million shots and then choose the best one and delete the rest? That one shot will be as good or almost as good as the professionally done one! And indeed, to some extent that was true, but instead of professional photography disappearing, its standard absolutely vastly improved, to the point where no inexperienced amateur with even a good camera can hope to compete. Even now that cameras contain a great deal of AI.
Here's an example that knocks my feet from under me: when CDs became available, everyone predicted the end of vinyl records. Now CDs are gone, and vinyl has made a big comeback! Who could have predicted that 30 years ago? But sometimes, a somewhat awkward, gritty imperfection is precisely what people want, rather than the slickly professional. That is how Van Gogh became famous.
There is quite a bit of both visual art and music that is, I think, way out of the league of AI, and to the extent that AI can do it, it remains to be seen whether people will want it, or whether it will really push the human product completely out of the market. No doubt some low-end composing and illustration jobs will disappear. But sophisticated sonata form? Maybe not so easily.
Or page through one of Hergé's later Tintin books: sophisticated and often strikingly original stories, expressive art that matches the narrative, every page ending on a minor cliffhanger, subtle, layered humor and satire that can be enjoyed by young and old alike because it often functions simultaneously on two levels... the kind of AI that could ever produce that is an AI that will not just put artists out of jobs, but also computer nerds.
Then streaming came along and everyone really did stop buying music.
However, when streaming came along, the music industry did panic and say thats the end of people buying music, and they were right.
AI will consume all sorts of artists jobs. The top end will remain, but all the run of the mill jobs producing relatively low value stuff for the bulk of projects outside of Hollywood movies and chart topping artists, will be farmed out to AI, in just the way many will farm out album covers, as he suggests, to visual AIs. And ai will slowly creep into the workflow of even the best artists, as they use it for brainstorming, polishing and exploration.
I think the only upside of infinitely capable music generation AI would be to reframe music as a participatory art form rather than a consumptive one. The implications for professional composers would of course be unfathomably devastating though when industry higher-ups realise they don't need actual humans involved
My take is that what AI (and technology at large) does to art is liberate tem from being product oriented activities to process oriented activities. It makes us 'seek' the human in the art.
@@someguy5261 Yeah exactly. I think that's the threshold at which we should define something as art or not art- there should be a sentient consciousness at both ends. It's a horrifying, alienating and worryingly popular view that art needs to be nothing but a sensory spectacle. I think art should only be considered meaningful as art if there is any reason at all for us to speculate about the thoughts and intentions of its creator, that's something we should always value. Of course we cannot really define what art is in a 'true' sense, but we can choose a definition that makes human culture far richer and more interesting if we adhere to it.
I agree for the most part. I’m a programmer and producer and for the last two year I’ve been working on a project called Neptunely. My primary goal is to bring the joy of music creation to more people. A secondary but not unimportant goal is to empower artists to shape what these new generative ai tools look like, to build and use and distribute their own models. A unique bot for each artist with attribution rather than the murky space we are entering with digital art currently. I think for better or worse, this is the direction we are going so I’m hoping that we can shape it in a way that is more human and kind, and less exploitative.
Industry high ups being the key factor - technocrats get bigger houses - arguably the main driving factor for ‘industry’ through history. Already much music sounds like Sunlight soap suds - little innovation in pop - heavily computerised and ersatz like fake cream - ha ha to me at least. Certainly humans will continue to create new and interesting, emotionally charged music Anyway. However a nice informative video Bruce!
Nah. Lowest common denominator. Music has already been cheapened to a thousand bucks for a million plays. Tech "innovators" are usually men who neither understand nor like people very much and regard humans as inferior computers. Their aim is to replace all human endeavor with computers and convince us all to feel the wonder and awe of our elimination. Hail the misanthrope messiahs.
Wow. I’m a software engineer. When I clicked on this video, I wasn’t expecting a music composer would take me to school about machine learning.
Impressive explanation and analysis.
Very interesting video.
I am very active in deep learning (AI) and I must say that your explanation about how these systems work and their possible implications is excellent and to the point. Better than most I have seen from people who are supposed to know.
One more evidence of the quality and depth of this channel.
Thank you for your superb quality and inspiring content.
Exactly. I’m a software engineer by trade, and I didn’t expect to watch one of the best explanations and analysis in this technical area from a music composer.
Someone posted a share pic on twitter the other day: "For Dall:E to give you the exact image you desire, you have to give it a precise description of what you want. We artists are safe."
I first saw that joke being framed as a programming joke, when worries about AI taking programming jobs was high. The worry still exists, tbf
ai + artistic person will always be greater than ai alone. The latest shiny tool will always be the solution for everything and later easy to get bored of.
Assuming the AI gets even a bit better at what it does, that statement applies just as equally to someone commissioning an artist, or an artist setting out to produce something they want to do themself. As AI gets better it only ever gets closer to (and way beyond) the input-output structure of a real artist.
@bigfishh Yes, we are already several generations past DALL-E 2 just in the last few months
@bigfishh yeah.. humanity is just destroying itself . Because art was just a way to express our feelings and share our perspective of life.
It wasn't necessary to replace that.
Is funny how stupid can be humanity and it takes just one idiot to press the atomic button.
It would be fun if in the future humanity destroys thereselves because they create an AI to win the war...
And everything takes out of their hands of course lol
I think we've already hit the point long ago where musicians can think of recorded audio as raw material, and AI generated audio could be extremely useful to that effect. I could see myself going to an AI to generate a handful of generic drum breaks, then dropping those into a sampler and using them in a composition, rather than just letting the AI handle the drumbeat entirely on its own.
Also, to be perfectly honest, I much prefer it when the AI goes off the rails, because at least it's producing something I haven't heard before. It'll be a damn shame when they smooth over that stuff.
Open AI's Jukebox is great at going off the rails, the problem is it doesn't return to the original verse or chorus. It's A A B C D E F G... song form, much like GPT-3 continuing a story that just goes off on endless tangents.
I like off the rails too. Hopefully we just get sliders to play with, where we can decide how smooth or off the rails the audio generation will be. I’m picturing a workflow similar to Stable Diffusion with many advanced settings and options.
Midi CAN capture pitch bend, articulations of various kinds such as aftertouch. As well as controller settings played in realtime. Not sure if the public domain midi files routinely contain that.
so what? using the actual sound file is still far superior
I don't hate twoset, but I've been waiting so long for someone to subtlety roast them about literally never doing proper research
You explained text to image better than 99% of people, and just to use for comparison's sake with audio generation, very well done. Everything you said in this video is pretty spot on. I think your optimistic view in regards to creative AI will be proven true in the coming decades. This kind of tech is just a tool that democratizes creative expression after all, the backlash is almost entirely based in fear and a sort of sunk cost fallacy from established artists. If anything they should learning how to use this technology when it is still new, they have a massive advantage by being able to use traditional methods, having a better trained eye for art, etc. Smart artists already are, and it will be the same for musicians. Fantastic video!
"Democratizes creative expression"? What is that suppose to mean? Since when was art some sort of an elite thing? Art isn't inaccessable, ANYONE can do it with enough time and practice. This A.I. doesn't even improve the creative process, it basically does the whole work for you. It's not JUST a tool, its an automated art machine. How is that a good thing? Just to get results faster? Just so you don't have to put the effort of actually creating something?
Yeah, people already ARE using this technology in art. In regards to paiting, some people have already won art competitions using A.I. Do you seriously don't see a problem in that? Are results and the raw product the only thing that matters? What about the meaningful process of creating something new? You know, the thing that actually makes art fun? That doesn't matter either?
You are not "smart" for using A.I. You are desperate. For as long as we see art as nothing but a product to consume, companies will require artwork to be better and faster. And automating the whole process is the only way to do it. Ignoring the fact that this takes away all the joy out of art, which is the reason artists do art in the first place, it will end up making all the works derivative.
The fear is very much sensible. A.I. in general has been stealing more jobs than it has been creating for a very long time now. Kurzgesagt already made an in depth video about it and people STILL think everything is fine:
ua-cam.com/video/WSKi8HfcxEk/v-deo.html&ab_channel=Kurzgesagt%E2%80%93InaNutshell
I would rather listen to the 1% who explain it better than him tbf.
I agree completely. I’m a programmer and producer and for the last two year I’ve been working on a project called Neptunely. My primary goal is to bring the joy of music creation to more people. A secondary but not unimportant goal is to empower artists to shape what these new generative ai tools look like, to build and use and distribute their own models. A unique bot for each artist with attribution rather than the murky space we are entering with digital art currently. I think for better or worse, this is the direction we are going so I’m hoping that we can shape it in a way that is more human and kind, and less exploitative.
@@TachyBunker he _IS_ the 1% - that's what Vala was saying..
@@SineEyed woops then i prolly meant the 0.1%. He has put through a good perspective but it's all surface level things, there is so much to the 'iceberg' (if this metaphor isn't too old lol).
One pianist which name I don't know once said something like:
I don't listen to people playing, to hear what "I want" to hear. It'll be what I'm already imagining, I would just hear to what's already in my head, that would be boring
What I realy want to hear is what others think and interpret, so I can grow and learn from it. To see something new no matter if I agree, it's not about it being "right", just being able to see somones visions is beautifull in it of its self
Maybe something like that will be exactly why human composers will always be wanted, but who knows
It feels to me like this sentence only works if you're already a somewhat talented musician.
Most people don't have the luxury to hear the "right" thing for them in their mind, and will enjoy having it shoved in their ears
With AI-generated music the expressive and communicative aspect of music is lost. For uses where the music is not the primary content, such as most movies and games, AI-generated material will probably achieve significant success, but I believe that for people that view music as both an art and an activity, it can never replace our own creations.
But as I see it there is no real contradiction. The pianist wanted to hear a different interpretation and AI would be able to give that. The AI perspective could be very different and more interesting than what a human would be able to produce. But I also must point out that the pianist has the view of a musician. Most people don't hear music as musicians do.
I reckon it's a year at most until musicians find themselves in the same existential crisis that visual artists are experiencing now around AI, this is coming from someone who fits into both catagories. Exponential growth creeps up on you.
I agree, apparently the team behind Stable Diffusion is already working on a music AI that will come out soon. One of their engineers said that realistic images will be "solved" in a month or two, and to me music seems like the next natural choice.
Without any question. I'm on a team working on exactly what he wonders if anyone is working on. And we're making crazy progress. This is just as tractable a problem as image synthesis, if not more so. I think we're only months away from brilliant ai generated music, if google and openAI are working on it.
@@theantiantichrist Well there's the answer then. I know machine learning, but I didn't have any specific knowledge here. All I could see was that music appeared to have far lower dimensionality and would require much less training data. I'd actually be a bit surprised if the final results don't jump clear to the super-human level like they did with games. Though the fitness function must be the hardest part... I've been wondering this for a while now: Is it possible to train a "quality" discriminator that separates the large quantities of songs that didn't become international successes from the ones that did? What about how long something stays popular (possibly hundreds of years)?
Dude, thanks for this comment. I love art and music and I make both and I’m also exhilarated with this new technology and have been having an absolute BLAST with it!
These things are tools most powerful in the hands of folks who know how to use them anyway. I feel like it’ll only bolster each. It’s weird to me how anyone keeps insisting they can never make “real” art or “real” music that humans can’t connect with. Already I’ve seen a million fantastic intricate art pieces done purely with AI by people dictating what they want to it. I’m sure music will be no different.
100% agree
I wish I could give this 50,000 likes. It's an outstandingly clear description of how the relevant AI technology works, and a very thoughtful discussion of how it could be applied to music. Thank you so much for posting this.
[Will bots ruin AI Music too?] Anyway, I second Bruce's comment. I've had a career that's been half AI and the rest regular software engineering, and I've found most descriptions of this new technology rather unsatisfying. By the way, this sort of thing (in both art and music) is something that I kinda-sorta predicted 45 years ago, but it's been a longer time coming than I might have guessed back then. I suggest that there's still a key component missing for the AI of Music - a workable and as-complete-as-possibe Ontology of Music. I suggest that one or more schools like Berklee start a joint major in Music Ontology and the Artificial Intelligence of Music. Maybe should be endowed by Spotify (to which I'm listening at the moment, and which does a pretty good job of categorizing music genres).
I have the same optimism that you have. I really don’t see people switching to just listening to AI music with AI art on their walls while they read their AI book. AI won’t replace any of these fields of art, artists of all types will still receive commissions
But what if AI art becomes 2x, 10x, 100x more enjoyable, interesting, moving than what any human could produce?
Jazz and Classical also "still receive commisions" but they're nearly irrelevant in the general culture. The same could happen to human generated pop music. It's not like the music that's dominating the streaming platforms (Spotify top 100 etc) isn't already so formulaic and basic that could not just be replaced with AI and nobody will notice...
That's because you won't notice it. I mean, the fact aside that any intuition about "my field isn't going to be outdone" is almost guaranteed to be bad and subverted in a heartbeat, it's a tool others use to make compositions in the first place. So we end up with folks generating riffs and hooks or simply lyrics to a degree where curation is everything. And we actually have been at this point for a long time now, all that's going to change is its feasibility - which is skyrocketing dramatically, even as we speak.
People will listen to whatever they like, regardless of who created it. Most of us barely even know or care about the specific artist, all that counts is what it is and what recommendations based on it we can accumulate. And yes, AI is definitely replacing "these fields." Just check out RTX remix presented by nvidia and how it potentially completely sidesteps the arduous, manual modding processes we've gotten used to. Stable Diff and friends being used as full-blown toolsets practically overnight because they so severely reduce even the most accomplished digital artist's turnaround time.
It's not even bad. What people like Adam Neely and Jazza appreciate, which I like a lot about them, is that this isn't freaking competing with human creativity. It's supplemental, and what really counts at the end of the day is the product we curate. You still get to do music as a composer, much like recorded music spanning more than a century didn't, in fact, end up displacing live music and bands, quite on the contrary. Nobody requires you to hang up your hat, it's just that the point at which we can end-to-end produce novel compositions with all the bells and whistles is rapidly approaching us, effectively displacing cheap labor composers (which clearly aren't even a thing, asking for this kind of work and "cheap" is mutually exclusive unless you're really lucky and know some bored professional).
You should still call it optimism, but you're 100 % wrong about our art not being saturated with at least AI-powered works. It's not like shoddy pulp writers aren't already saturating our markets anyway, same for direct-to-DVD nonsense and such. All it's going to do is democratize creative processes that are unfathomably difficult to pick up, and it will have massively exciting and far-reaching outcomes for all types of education, creativity, and our general sense of what constitutes culture. Embrace AI - everyone else will.
@@foljs5858 Right! „Fahrstuhlmusik“ or the typical background music in supermarkets will be the first. Supermarket music will at first be tunable by the programmer‘s company and later by the renting supermarket, depending on the place it will be placed to influence us buyers and so on.
I have AI art on my wall already
That was fascinating, thanks. One of my favorite software toys in the '90s was Fractal Music, which used fractal mathematics to generate complete musical pieces. You could tweak innumerable parameters and it occasionally produced surprisingly good results. Most of the time, though, it generated little more than a collection of random pitches. But I wouldn't be surprised if within a generation, we'll wonder how we lived without computer-generated music.
Imagine feeding an AI with UA-cam videos from music educators and never seeing again David, Neely or 8bit... can't imagine anything sadder
Definitely. Some people seem to equate change as a good thing, but they're separate. change =/= progress.
It'll only be a matter of time before they automate prompts (midjourney probably ain't storing them for shits and giggles.) Completely taking people out of the process. Perhaps just becoming like targeted ads in the future. It'll only become a hobby in a few years to a decade unless you're some sort of online personality. No doing it as a purely skilled craftsman which is the saddest f**king thing ever, venues are dying and that's sad. This will make also the tools to do it insanely expensive due to the lower quantities being produced, think paints, a guitar or a new synth or something is expensive now? Oof. You could argue live stuff would still be good, but how could you trust anyone actually wrote their own music in the not-so-distant future?
I guess there will always be a market for hand crafted things, but the way it is now is dead forever - at least very soon it would be. Much how people buy Ikea over a hand-crafted wooden table that'll last them their whole life. Only very few will do it full time, and the others that do probably just make one piece a week on the side of the job they now, probably won't ever be able to escape. Well done futurists, you've killed what was great about humanity and culture.
@@stinkypete9070 You’re welcome. 😁
Adam Neely recently did a Q+A video that touched on AI music, and my first thought here is the same as there: sooner or later, I think that AI-generated art (both visual and aural) is gonna get hit with some copyright laws, similar to what happened with sampling. Training neural networks' datasets on copyrighted works with the intention to imitate them raises meaningful questions with "intellectual property" law, and I could easily see that requiring a license in the future.
But I disagree with the following: "and eventually we could all end up with our own weird little playlists of AI-generated pieces, all catering to whatever quirky niche we fancy: you start to see a picture of the world where most of the culture you consume is planned out for you, specifically tailored to your needs and desires, and all generated by AI" - I don't think individuals want to be in their own isolated bubbles of content specifically tailored to themselves: I think a lot of us turn to art for very social reasons, including both a desire to connect with other people who understand the same references, and parasocial or vicarious connections to performers and creators who might be seemingly living out others' dreams. While it's very easy to dismiss "popularity" as a measurement of quality, I do think a certain amount of shared experience is essential to any Art holding any level of Cultural or even subcultural relevance.
@dfasht "Every algorithm on every platform proves otherwise." - does it, though?
1. While big platforms like talking about feeds being tailored to individuals' preferences, I think it's worth asking how individualized those feeds really are. (Jacob Geller's recent video about "individualized" therapy apps touches on this a little bit, though it isn't its main point)
2. They don't magically read your mind to determine your preferences, they're just making suggestions based on seeing correlations between liking content X and liking content Y.
3. Because of that, algorithmic aggregation favors pushing things that are already popular, creating some of the same "shared experiences" that I was talking about. Most people who like David Bruce are gonna get recommendations for the same set of popular music youtubers, not necessarily individualized recommendations for what they might like best.
4. While algorithmically driving content aggregation to maximize engagement is undeniably profitable for the platforms, that doesn't mean it's what the users actually WANT. Most folks I know actively resent platforms trying to push algorithmic suggestions over the things they actually want to follow.
@dfasht I agree that people enjoy being in bubbles, my point is that they don't usually want uniquely individualized bubbles; they tend to prefer larger bubbles of shared cultural experiences.
One of the main things that has kept full generated text out of ai's reach (and I don't know if it will really ever get there) is the understanding of context that is crucial to language and communication and to convey actual meaning. That is, as Tom Scott pointed out some time ago, computers currently can't consistently parse out the sentence: "the trophy can't fit into the suitcase because it's too big." for a myriad of reasons, including not knowing the how "big" applies in this situation, what "it" refers to (could be the suitcase or the trophy), or what sensible reasons there are for the situation, and so on. There's a LOT going on in language to create consistent and meaningful communication.
I think the same is also true of music, but perhaps less concrete - meaning in music is much more subjective than in writing/spoken language. All the same, the contexts we use without thinking about them will only be interpreted as patterns and the recycled music given back will likely feel either absurd, hackneyed, or just inhuman.
I think ai works with visual media so we'll because context is largely displayed in the same medium as the object itself. We know elephants are in the field because it is shown in the field - there's less abstract context needed to reproduce an image on that advanced pattern. But I am certain you will get more absurd and disconnected images if, for example, the prompt is an entire story, not just an object (I've tried in limited ways and unless I'm very simple, it just ends up amalgamating it all rather than parsing out a discreet scene - which leads to some fascinating images, but certainly not what would happen if you asked a human to illustrate the story, for example).
In all, I think ai will become a crucial tool for people to use in their own creativity. And just like how the proliferation of photography led to more people taking photos but there still being a demand (perhaps an even higher demand because people got used to photos but also understand a good photo better) for professional artists who are skilled in it, I think it may be likely ai will open the doors to creating music to many but not at a professional/artistic level.
I thought it was going to be about ignorantly dismissing the power of AI. But it's neither ignorant nor dismissive. Good job!
The Guardian printed an op-ed "written" by GPT-3, and then mentioned in a little section in italics at the end that the editors had actually patched together a coherent op-ed from a large number of attempts by the AI
High praise! Most every article is written by someone and chopped up by an editor.
@@timchapel77 most opinion pieces are not written dozens of times and then chopped into pieces to be assembled entirely by editors. You could sooner claim that trees build houses
@@mywifebeatheroin your example is that of an ape learning to use a new tool. Not understanding the tool is on the ape, not the tool.
@@mywifebeatheroin Trees build all types of things, including homes.
@@timchapel77 construction worker? I prefer Wood Editor
TwoSet and David Bruce would be a crossover episode I'd pay extra for
Yes! Our man DB should include them as the artists for the next 5 Composers vid. And then TwoSet can play them during the encores of their next tour?
Yes, that'd be an epic cross-over
Actually not. They behave kind of idiotic if you ignore all the Classical music part.
Please don't. They are so tiresome, they might as well be AI-generated.
@@hugobouma
Yeah. David Bruce makes it pretty clear that you can disenchant that Huawei crap with TwoSecondsOfThought...
I’m an amateur painter and a professional programmer and I’ve been using stable diffusion and a couple of other models in concert (rather than dall-e2 which has gone the wrong way IMO in its restrictions and charging model, which discourages the interactive conversation), and I’m finding it a deeply creative and inspiring experience. I also write fiction and being able to quickly visualise and illustrate scenes is bringing a new perspective on my writing too. Like the beatboxer, I find it feels like having a conversation with the AI, collaborating with it to spark ideas. I’ve used the same analogy about photography and music recording when talking to friends about it - I don’t think this will destroy human art but it will change it, and I’m fascinated to see how. You should probably sell your shares in stock image companies, and I wonder about the effect it’ll have on the value of reprints of non-famous artists’ work, but I don’t think artists who paint for the love of it need to worry, until the AI picks up a brush at least! Even then, both in music and art, what we love about art is the way a person can express their humanity in a way that reflects our own - the fact that a person put their own sweat and talent into it is part of why I value it, otherwise why would I prefer live music to a recording, or a painting to a photograph.
Glad to hear there are others using SD in this way. I started with trying to get some sci-fi vehicle concept art for a game, but was not expecting the feedback loop of ideas and the journey into my own imagination it caused. I'm not a fiction writer by any means either, but giving SD guidance and themes rather than strict controls, and allowing it free reign to draw scenes in those themes, it's just... indescribably amazing. Here's a little trick... give it a "stylistic" prompt you are going for, but flip it negative, see what comes out. Then, describe any common elements from the results as best you can as a new prompt, and flip that negative, so instead of naming a particular artist, you're asking for "the opposite of the opposite of Dali" or whatever.
I'm not so optimistic. It's only been a few months and people are already losing work in the visual field.
I think the collaboration phase will only be temporary; it'll only be a matter of time before they automate prompts (midjourney probably ain't storing them for shits and giggles.) Perhaps it becomes targeted content like ads perhaps, or something. No living off it, unless you have the capital to advertise. It'll only become a hobby in a few years to a decade unless you're some sort of online personality, no doing it as a purely skilled craftsman, and that's sad. This will make also the tools to do it insanely expensive due to the lower quantities being produced, think paints or a new synth or something is expensive now? better hold onto those and take good care of them if you have them, might be worth a bit one day.
There will always be a market for hand crafted things, but the way it is now is dead forever, at least very soon it would be. Much how people buy Ikea over a hand-crafted wooden table that'll last them their whole life. Only very few will do it full time, and the others that do probably just make one piece a week on the side of the job they now, probably won't ever be able to escape. Well done futurists.
David, thanks so much for putting the time, creative energy and serious thought into creating this video. Absolutely first class.
Great video. I feel like the explanation of all of this was so good in this. It is a very complicated topic, and this is honestly one of the best and easily understandable I've heard.
I think that art has two parts; there's craftsmanship, and there's the message. If you're an artist, you've got something you're burning to show or tell the world. But there are many who have something to say, so we use how much effort someone put in (craftsmanship) as a sort of proxy, to figure out if someone's message is worth listening to. Surely it's not worth spending time on someone who isn't very good at conveying what they're trying to say (I realize that this may hit me right now...)
Craftsmanship is going to get democratized. Everyone will be able to say what they want to say *very* effectively. But still, some people will have something more interesting to say than others.
Exactly, people seething over AI art is hilarious since it's essentially a future abolishment of AI scarcity which will make art more accessible and less elitist.
@@danielsurvivor1372 Elitist?
You think artists were somehow just born knowing all the art fundamentals?
It took years to master their skills.
Art is just like any other skill. You don't think programmes were just born knowing how to code right?
Honestly anyone can learn how to draw but unfortunately we live in a world where people don't want to work hard for something.
They just want instant gratification.
@@danielsurvivor1372 just go watch steven zapata's new video about it if you have time, he is a bit smug but his points are still at least worth thinking about.
@Lord Vader yes this seems like another production of utopian anti-capitalists and anarchists to "democratise" something that will just make it even harder on the people they claim to help, because it's not democratic for a reason. People used to think that phreaking and crypto anarchism would "give the economy back to the people" and destroy the power of the financial elite, it just ended up as a great way for people to scam and defraud people. Before that it was the anti-psychiatry movement being turned into Scientology and "Care in the Community".
I totally disagree, I believe the dichotomy you've drawn between "craftsmanship" and "the message" is confused. I don't think you can separate the two. Music is a great example for this. The "message" of (instrumental) music is the music itself, all of the decisions and aesthetic considerations that went into this note being this one here, this harmony being this way here, rewriting the chord structure here etc, all of these acts are not simply someone exercising their "craftsmanship", but are the crucial creative and emotional decisions that makeup the "message". AI would have large swaths of these decisions be eliminated, in other words, large parts of the real "message" coming from an algorithm, not a mind.
Really interesting topic, and a great video on it. Thanks for taking the time to make it!
I just stumbled upon your channel and am thrilled. I am in music therapy school after graduating from recording arts school and am back in classical music after 30 years (my school doesn't have jazz piano teachers), so analysis of all I'm doing - blending old 18th century theory with new school is more than fascinating. You are the first classical composer that I've seen breaking stuff down in this way. Thanks!!!
I think the only reason why we don't have "DALL-E 2 for music" is that it's harder to commercialize.
Jukebox was 2 years ago, and it seems to be a viable approach. There was a tremendous progress in image and text generation in the last 2 years, so it seems like a scaled-up Jukebox might just do it.
But training large models is expensive (on scale of millions of dollars) and the music industry is notoriously litigious. So a production-quality music generator would likely end up shot down by lawsuits if it's trained on copyrighted music.
So OpenAI probably doesn't do this for strategic reasons.
Agreed, it's a definite consideration. The tech to create music on the level of the current text to image generators is already here, the constraints are not really technological, and they may well be simply legal. The one point in the video I disagreed a bit with concerned the complexity of Wav files, to a massive neural net computer there would be no problem teasing out the details of melody/harmony/timbre/etc, such things would probably just emerge organically, as long as the dataset and training time were large enough. So I believe this could easily be done already, and the only reason it hasn't could well be our mess of a copyright system. BTW, look for a slew of lawsuits against those image generators any day now, it will happen for sure. If the end result is a win for the AIs, perhaps only then will we see the true AI music generators built.
no, the reason there's no dall-e 2 for music is the music industry will actually sue if you violate copyright. Visual artists and photographers won't. There you go, easier to pillage data from the visual artists than the musicians
there will have AI who will know " what other AI Copy " if u use sample , u will have to pay. in a song like Sisqo - Thong Song , he singed " vida la loca ' he had to pay Ricky martin for this sentense.
Hey, I wonder if you've heard about DDSP, part of Magenta? It's not about composition but it's definitely AI-based and music related. The idea is to do tone transfer, either in trained ranges (ie map a piano to a guitar), or _outside_ of them (kinda the equivalent of outpainting in AI image generation). And you can play the models through MIDI. It is very, very interesting, and already available!
I think it's definitely here that AI shines in the musical world, for now. Can't wait for more advanced AI-assisted sound synthesis.
Think something like a VST with just a prompt box and a few sliders (fidelity/creativity and so on), you type in for example "a warm pad, light distortion, heavy chorus, long lingering delay, ethereal" and then you just play.
oooh, wow now THAT would be a cool soft-synth! The prompt is the preset.
Thanks David for featuring our work w/ Reeps One :D
Several years ago I watched a video (which I have since forgotten the name of) that talked about how a group of vehicle engineers were deploying A.I. design aids to help think of more aero-dynamic designs, but the A.I. struggled with aesthetics and human functionality (it would sacrifice seats or leg room for aero-dynamics). Basically the engineers were playing a game of intellectual soccer, kicking the idea back and forth with this A.I. software until eventually they found the Goldie Locks zone of functionality and human use. Since then, this is how I've viewed A.I. I'm sure painters thought it cheapened artistic integrity when manufactured paint hit the scene and younger artists no longer mixed their own paint anymore. Like all things, there will be benefits and there will be draw backs but it's coming either way so best to brace for it and try to find oppertunity.
In spring of 2017, I was up at night watching a computer beat the best human Go player for the first time ever. Now just 5 years later that same technology is already doing science fiction stuff.
This. This is why I'm subscribed. I have no idea what the IV chord is, but this blend of music, art, tech and philosophy is both fascinating and enlightening
in music, a scale is essentially a palette of (usually 7) notes that work well together, and a chord (which is 3 or more notes stacked on top of each other) can be built off of any chord. The most common type of chord is a triad, consisting of a root note (which can be any note in the scale), the third (the note two notes above the root) and the fifth (two notes above the third.) A IV chord is simply the triad built off of the fourth note in a scale.
You now know what a IV chord is and your rhetorical device has been ruined
destroyed 😎
Excellent perspective David, I think you're spot on. I subscribed to the Two Minute Papers YT channel a year or so ago and it regularly scares the hell out of me!!
You need to hold those papers tighter !
This was an excellent video in so many ways. An effective breakdown of the technologies at play, some interesting history, compelling case studies. I was already curious and excited about the future of AI in music before, but I'm even more so now wanting to see where we push ourselves to next.
Thanks again for all your hard work David!
Great video, great sumarization of what a AI model is. I'm a developer, and the same thing starts to emerge with AI for software development and it is in a much more advanced stage than it is for AI assisted music composition in my opinion, pretty mind blowing i'd said. But the questions that it raises are very similar.
As someone who's worked in machine learning, I'm impressed by the technical detail and accuracy. Well done!
I think the comparison with photography is apt. Art is at it's best when it challenges our assumptions, and music generated by existing pieces only stays relevant when it constitutes a minority of the contemporary corpus
Start switching those titles and thumbnails David!! This video is bloody incredible and more people NEED to see it!! Inspirational, scary, informative, entertaining all at the same time, you are still undoubtedly the greatest music educator on UA-cam and the work and research that went into this video is just unbelievable! Another masterpiece 👏👏
Damn, the editing has gone next level since the videos you out out years ago. Hats off. Looks amazing.
I don't think this is particularly new, but something we've been struggling with for some time. Is art merely a commodity or something transcendent? At some point most of us long for the connection to others that art provides. Our lives are expanded when we find the commonalities we share with others and experience the strangeness of their understanding of the world.
Well, thanks to these "futurists" it'll only ever a be a product.
It’s not just composer and artists. I keep watching “Humans Need Not Apply” and it becomes just that everyday.
Thank you so much for the optimistic note you ended on. I didn't realize this fully.
Just exceptional! When art and technology come together, magic happens. And it certainly did in your narrative! Thank you so much!
Can't wait to ditch composition to become a "prompt engineer"! What a prestigious, respectable art field.
And cheap pay too, yay
exactly
AI is frequently overstated. Computer-aided, yes, but not AI. Even at the bleeding edge of AI, "stochastic parrots" is as good as it gets right now.
AI doesn't actually exist. No Algorithm is intelligent. They're just algorithms, made by humans, and are only able to work because of human creative output they are trained on. Wish people would drop the AI fantasy and call it what it is, generative art, or algorithmically generated art.
Absolutely an amazing video! very well pu ttogether, and you put a lot of interesting perspectives in here, i was drawn in throughout the entire video!
I really appreciate the “Implication for future of Music and Composing”. Thank you!
The coolest AI music stuff I've heard so far is PROTO by Holly Herndon, where she brought in a bunch of choral singers to train an AI and then used the uncanny results to compose jarring and fresh otherworldly music
I've been using Open Jukebox in some of my own productions lately, but in a way that I haven't seen others doing yet. I take my own music that I produce, use a piece of it as a primer, and generate a new section of the song. For example, my next track that's coming out is a remix of Blow! by sineila. I took a clip from the end of a chorus and fed it into Open Jukebox, telling it it's Steely Dan. I then took the clip it generated, and edited it back into the song at the point I took the initial clip from. The end result is the song kinda devolves into insanity for a bit (especially since the rest of the song sounds nothing like Steely Dan) before coming back around into the song proper again for the ending. I've found the most success using artists that have complex and diverse music (such as Steely Dan and Meat Loaf), gives more variety to the generated clips. I'm working on a track right now that's going to have 4-5 different AI generated sections built off of other parts, all of which will be based on different artists. It helps that my music is largely a fusion of prog and hyperpop, both of which make heavy use of non-traditional song structures and musical adventurousness.
I think of the rise of AI music as akin to the rise of synthesizers. People initially were worried that synthesizers were going to replace actual musicians (which did happen to some extent, though not nearly to the level the biggest naysayers feared), but they wound up having their own musical niche. The exciting part about synthesizers is not trying to replicate the sound of a trumpet, it's to create entirely new sounds that traditional instruments can't (and though of course there are synth trumpets, even those have their own unique sound that's desirable for uses that real trumpets might not be). The future of AI music is not to replace human musicians, it's another tool for musicians to open up a whole new world of sonic possibilities.
This sounds like what Adam Neely said about AI music.
@@tweer64 A very level-headed approach - which I honestly am seeing from most open-minded artists, safe for maybe the occasional stubborn musician (Rick Beato's hot takes come to mind) throwing a fit about the good old times. Jazza did have the same intuition about it, saying it is bound to supplement our way of working and we will be way better off. Which is a no-brainer for anyone doing VFX work and arduously rotoscoping boring footage all day long.
And I guess much like fretting over synths back then was already silly on account of the arguments being made, the same can be said about any ML-based technique: it is not even ambiguous. We already know how much good all this does in the creative department alone, there is a clear value proposition almost everyone will embrace day one, like they did with Stable Diffusion and the rest. It's democratized education and abstracted artistic ability. Anyone painting a bleak future is missing how empowering this is going to be for literal billions of people, many of whom might have severe cognitive or physical impairments - or even might lack all the tools except for the phone (which has gotten quite ubiquitous).
AI music is going to be fantastic. What machine learning did for mixing mastering alone is insane, shoutouts to Ozone9 and all the cool tools out there.
Maybe in 30 or so years, the sound of clumsily-generated AI music will be retro and people will be nostalgic for it.
@@goodlookingcorpse Watching Vinesauce Vinny quite a bit, I guarantee you plenty of us already are nostalgic for clumsily-generated AI tunes
Time compression hits pretty hard, no need to wait nearly as long lol
That may have been true in the 80s/90s but synth/sampled instruments are used very often today to sound like the "real thing." Many movie soundtracks, game music, advertising music, etc. is made entirely on a computer but sounds like an orchestra or band.
David. Thanks for your great analysis. It helps me think and learn theory. And it flows easy.
Now the request: make a video with commentary on Polythia please
Your explanation of the diffusion model is the best I've seen so far for people who don't understand how this technology works. Great video!
Big tech hell bent on putting every single human being out of work, assembly lines, A.I. data analysis, self-driving trucks, concept artists, music makers...
I would be interested to see a collaboration between you and 3 Blue 1 Brown about this, particularly around Fourier Transforms, which I think are pretty cool.
Where do you see a relevant experiment to this that either people can contribute to? 3B1B did a bunch of FT already, I don't think there is a lot to be gained from it (and 3B1B is kind of not the huge collaborative type to begin with). Not that there won't be a compatible subject matter, but FTs are kind of extremely basic in that scheme, what 3B1B would be asking for is a novel question about it being asked.
@@minhuang8848 I think each would likely have knowledge of their discipline that would highlight things that the other might not know to highlight. Obviously a lot of that includes ideas I myself am not sure I would know about in advance of them making a video.
Super interesting, great video and I love the loop daddy t-shirt.
Speaking of Schubert's unfinished symphony, Luciano Berio's "Rendering" is a magnificent work based on it. It's truly a masterpiece. One of my favorites ever. ❤️❤️
Kudos from a Russian composer in Istanbul. :)
As a software developer, I must thank you for showing me exactly what I want to work with in the next decades!
These AIs so far just take huge amounts of data fed into them and mash them up together, then form something out from that for you on request. But they can't really leave the space that's been fed into them, similar to how Google search will only return you something that's on the internet, it won't make a new scientific discovery for you regardless how inventive you're with your prompt creation.
Will AI eventually get over that? Probably? But I don't think we're anywhere close to that happening.
This is just terrific. I work with AI and with Music .. albeit separately... and this summary of the Generational field as it stands a, at this exciting point of commoditization, is as good as any I could imagine. Considering both music and art together has led to a more insightful views of the implications than I've seen from companies focussed on just the visual. Props for David Bowie's quotes too. This all relies upon the volumes of cultural data which we have only because of the Internet after all. THANK YOU, Sir !
thank you, a very well made video about a topic that I was wondering about for a while
It's strange to me that music is apparently so difficult for AI compared to visual art, when a piece of visual art requires so much more information (at least if you're talking about scores rather than sound files). From what I've seen, any few bars of a piece of AI-composed music sound right, but the larger-scale structures aren't there. It just sort of wanders around instead of developing a theme or doing what human-composed music would do.
I think the main hurdle is the lack of a massive way to produce annotated datasets.
For Dall-E and similar algorithm, they used datasets composed of images and captions produced by humans, collected all over the internet. Billions and billions of images. This is were the algorithm derives it's ability to mimic similar features of images produced by humans so easily.
We don't have natural ways of massively tagging music with descriptions in the same way. How would we produce annotated datasets for music on that scale of billions of works?
Notice that the images weren't captioned by specialists or by a particular process aimed simply at producing the dataset. The idea is that our natural usage of images online produces those captions. Frequently we publish a image on the web accompanied by a descriptive caption.
We don't do that with music. We don't habe any natural way of producing a dataset of billions of pieces of music annotated with descriptions or labels that would allow us to train an algorithm on that scale.
It might just be due to market forces. The amount of money you can make from fairly bad and incrementally improving art (which is what you'll get starting work in that direction) is much more in images than music. Perhaps for the same reason, image manipulation software was much more advanced than sound manipulation software even before AI.
@@RafaelSCalsaverini is quite right. Another thing to take into account is the complexity of time. AI image generation is generating a static result.
Because music encompasses that one thing that AI cannot intuitively fathom. It's to create symbolic rules and deliberately operate on repeatable motifs without human intervention
You'll see the same thing happen when AI is used to end to end create a coherent movie of a long enough length
I can see one dystopian possibility is the danger that everything becomes a derivative of a derivative. Worse yet, copyright becomes extended to “sounds like” Hendrix- and giant streaming services will charge for instances of AI generated music “in the style of” and these tools will be behind a premium paywall exclusive to one or the other rights holder…
I don’t think we will ever loose out on the possibility that people will compose new music- but I can imagine a world where those skills can’t be marketed to compete with machines for mass consumption.
Everything is already a derivative of a derivative, there's never something that is truly new. Even in very weird outsider art cases like The Shaggs (who are usually called "something new", for better or worse), what they do is mostly derivative: they are playing songs, using regular instruments, the members are arranged in the shape of a band, and so on
What if the AI scans the reaction to its art and optimises for surprise, perceived novelty?
I mean... here's the thing: we already live in this world. People are derivative, so, so derivative. It's great because that's where inspiration stems from, but we already are dealing with cashgrabs, asset flips, direct-to-home-media movies, covers upon covers as well as badly produced low-tier music - if anything, that sounds utopian on account of how much better even the worst compositions might end up sounding; amateurs might not know how to compose, but they sure can curate music and determine what we should listen to.
Maybe it'll be annoying in places, but we already are dealing with atrocious tiktok-songs skewed into the depths of hell, awful uplifting trailer stock music, generic soundtrack beats made under time pressure. Even this case I just don't see happening, personally, and if it does, it'll establish itself as a particular genre niche I might be interested in.
Great review, very layman-friendly across fields. Keep up good work
That part of Beethoven's fifth makes a great chord progression from the notes. See "no new tale to tell" by love and rockets
I think it will speed up and improve the composers work if they embrace it and not see it as a enemy. Like think if you have an AI you can try ideas on and really fast get a response. Maybe ask for advice and improvements.
Fantastic essay. I do think more AI generated music is bound to happen fairly soon. Especially things like what the google doodle tried to achieve: Generate a piece of a certain form in the style of this composer. I even think they’ll be capable of “genius” whatever that is, but that is such a difficult thing to pin down, that they aren’t likely to know when they’ve achieved it.
Great video David. The content on your channel is very high quality. Subscribed :)
Thank you David!
I think spectrum analysis is the perfect way to teach a computer about timbre, rhythm and style.
There are many research to isolate an instrument from audio file using FFT, of course the ammount of data the computer has to store is much bigger in comparison to the image.
Thanks again for your videos, I am a big fan of yours!
When I heard the AI might take many painting artists' work thing, the first thing that came to my mind is how people feared that invention of robot will take over workers' job. I think the same apply to AI and music makers: it will more likely provide tools(as it is already doing right now) for artists, and they will come up with new ways of making music. Art is not like games, e.g. chess and such, where there is winning and losing, in which AI will just dominate. Art is made by and for people. And even in chess, AI didn't eliminate the game at all!
Exactly, it's the standard textbook luddite argument, pops up any time new tech threatens to disrupt any establishment, regardless of what area of human culture it happens in. Wherever an established standard exists, there will be people coming out of the woodwork to insist that it's the final evolution of humanity's treatment of a topic and that nothing new will ever be worth giving serious consideration to. (And they turn out to always be wrong)
Basically your from the previous generation. Unable to accept the incoming future. Also prejudiced and in denial.
AI is getting better and it's a matter of time, say 10 to 15 years, computers will be making full fledged music. It does not matter if you like it or not.
Actually mainstream musicians like yourself should learn from Emily by David Cope (A program that wrote Bach type melodies out of machine learning)
Now can you write Bach type melodies? No. End of Story.
@@seedmole to be fair to the actual Luddites, they weren’t just irrational people who had irrational fears, their perspective was actually rooted in a real and serious impact to their lives, and they had every reason to take many of the stances that they did. I realize that what it is “to be a Luddite” in popular culture, or modern vernacular, is more akin to someone who just has an irrational fear of technology, or is afraid of certain progress, and such along those lines, but I do think when one studies the Luddite movement in the context of the socioeconomic/sociopolitical times in which they lived, and in the context of the changes they (and most of Western Europe) were experiencing, that they make a lot more sense. I also think this study reveals that (hashtag) not all such concerns are irrational, or merely based in foolish prejudices, and ignorance, and that in fact sometimes it is very rational, perhaps even a bit necessary, depending on the context (particularly the political economic context).
To be clear, I don’t mean to contradict anything you are saying. In fact, I largely agree with your points, and the OP. I’m simply pointing out that the OG Luddites weren’t totally ludicrous, and that current and/or future persons with concerns over the effect of technological advancements might also not be, that it really depends on the details of those situations.
Ultimately, new technology needn’t *ever* be inherently detrimental to the lives of anyone, but unfortunately it’s historically been the case that in certain instances it’s utterly devastating for some persons. More than anything, it’s ultimately politics that determines if/when it will be great for someone, or not. Here’s to hoping that social and technological power will be wielded in the best ways possible, bc I’d love for humanity to see a kind of Star Trek level of tech (and society) in the future!
@@nikolademitri731 Hope doesn't get us anywhere in that respect. The ballot box does. We would all do well to remember that.
But robots HAVE been takign jobs away. In fact, they have been taking more jobs than they have been creating. Kurzgesagt made a video about it:
ua-cam.com/video/WSKi8HfcxEk/v-deo.html&ab_channel=Kurzgesagt%E2%80%93InaNutshell
A.i. is not JUST a tool. A tool is supposed to make the creative process easier. A.i. bypasses all of that and does the whole thing for you. It can actually have it's own ideas. All you need is a prompt. This has never happened before in the history of mankind.
And I think ONLY the performative aspect of art will survive, but the creation aspect will be harmed.
The thing is... A.I. is "growing" *geometrically,* it has advanced more in the last 4 years than it did in the 10 years prior. I don't think we're really that far from AI being able to compose truly enjoyable pieces.
And it's progressed more in the last 6 months than it has in those 4 years..
Really glad to hear an informed, nuanced opinion on this matter.
Great video, thanks for not pretending to know the future and all the answers.
My ringtone is by OpenAI's MuseNet, "Beethoven’s Für Elise in the style of Disney." The piano plays the theme, then a meth-addled drummer comes in and goes nuts and the rest of the jazz combo tries to keep up. It's a wild 47 second ride.
You point out music files are larger than image files, but there is much long-term regularity in music. That's the biggest problem with OpenAI's Jukebox: it doesn't restate earlier motifs and themes, it doesn't even return to the verse after a chorus. That MIDI drummer is now a coked-up producer demanding "Give me something new!" every 25 seconds to a crack team of session musicians, and they dutifully comply. It's entertaining (go listen to all the ideas in "Never Gonna Give You Up, but an AI creates more of the song" including killer lyrics “Kiss the boat Denny I’m Satan’s pirate arrr... You know the rules and so you have to die”), but it's not satisfying unless you want endless instrumental noodling for your yoga meditation instead of structured songs.
The mindblowing thing about AI image generation is that the AI training phase can compress every visual thing that can be named, every medium of visual art (from "ukiyo-e print on heavily textured paper" to "taken with a disposable camera"), all art tropes (rain is sad, clifftops are sublime), and the styles of tens of thousands of artists, into a mere 4 billion bytes of weights and biases for its artificial neuron network; just one DVD of data. I suspect the latent space of all of music will likewise be small, but it has turned out to be harder to get the AIs to compress music and pay attention to regularity. OpenAI's Jukebox was announced April 30, 2020. *What* has OpenAI been doing since then?
Reeps One sums up the problem for artists. 16:45 "He found it inspiring and used it to take his art in new directions. 'I am having a conversation with this technology which is something I think everyone is going to be forced to do.'" More like anyone who's motivated can have this conversation with technology that results in pleasing creative works, *no musician required*. And those who aren't motivated will have their musical and visual preferences satisfied by AI.
17:47 "If anything the experience of hearing human performance is likely to grow ever more precious." Optimists make a similar argument for visual artists: the AIs only generate a digital image, which is a pale reproduction of actual shaped 3-D blobs of material on canvas, let alone sculpture. Human music-making and mark-making will continue, but both will be further devalued by the creative output of machines that have absorbed more existing art than any living artist can possibly experience.
I'm not sure it's only formulaic music (like call waiting tunes and elevator musak) that's at danger. Avant guard, ambient atmospherics, and improvised music can be emulated much more easily... it's not like it has a big Mozart-like coherence to begin with, so hearing an AI version of some meandering jam and a real "avant guard" musicians jam, wont be that easy to discern even with basic AI
I really am not sure whether we're heading into a retro dystopia scenario or an era of fortune for the art.
You just took the words out or my mouth. I still haven't watched the video but in the future I expect pop music and more simple forms of music to be produced by AI since they are so formulaic in the first place.
@@placeholdier Yeah I mean we can go further. I would bet anything that some big corp is looking at AI development and trying to work out an environment for a customer where everything is AI generated, the music/images/videos/voices... to put them in a bubble. This is really one of those sketchy and distant-sounding dystopia scenarios but erm I can't help but think, if this is possible, why wouldn't it be?
@@TachyBunker perhaps this is what they will do with all our data? Use it to create a constant stream of highly personalised visual and audio media, designed to give each customer pleasure all day long. A creative echo chamber of their mind.
@@62serpens Oh yeah sounds logical.
Your video editing is fun and conveys your message so well. Thanx :)
OF course AI is the purest BS. Then again, organizing words, dictionaries, subject domains (lingo for specialized tasks) and then information about words into huge databases can solve some problems as Google has definitively demonstrated - but Google of course has no value without humans. Consider people who advance AI as the smith who detected that a mill-run hammer could do more and better to create sheet-steel (etc.etc.etc) AI as a hammer? Yes but one it takes a lot of skills to master -- and it is not worth much without humans.
But when you say "composers like me can feel reasonably safe" from being outcompetitioned by AI then I laughed a bit, because humans are creating meaning and also creating AI - I sense a linguistic problem. When we want to expose criminals (AI as criminal? haha, but it is a parable) or reveal dishonest politicians, then we must make sure that we do not give them more publicity so that the "weaklings" believe that after all it is like the politician said, ... we must believe our politician more than this new stuff ... or something. Do you follow my gist? Sorry for this tooo long comment.
You mention at 4:05 that these generated images are all copyright-free and free to use, but that may not be the case. Stock image provider Getty recently banned all AI-generated art from its platform, citing the fact that the copyright status of these images is _unclear_ and they don't want to take the risk. "Copyright laws are currently unsettled with regard to imagery created by those tools".
How would you detect them?
IMO the AI creator could have a clam on all art created by the AI
@@tweer64 I'm not an expert in copyright law, but I think there are two main problems.
The first has to do with the fact that they're trained on copyrighted material, and that anything they produce is only a remix of that. Unlike a human taking inspiration from multiple sources, an AI model isn't using its own creativity on top of its inspiration, but merely remixing the various things it has consumed. And depending on what it's trained on, there could be questions of how this impacts the copyright ownership.
The second is similar to the issue of copyright issue with art created by animals (mainly photos, but other forms of art can apply). Does the animal itself have copyright? The person who gave the animal a camera? It's not clear how the rights would apply in this case.
@@JimCullen I meant, "How do you know which images were made by AI, and how do you know which AI made them?"
@@chuckcrunch1 Not really, as these AI algorithms in most cases were trained on copyrighted material, without artists consent, or even knowledge about the fact that their work is being used. The text-to-image tools to me are a borderline scam as in some of the images generated you can still see parts of or even full watermarks. I imagine with musical AIs it might differ a bit but still the question remains on how those algorithms were thought to do what they do. There is also an argument that artists already are feeding of each others work, and while that's certainly true, there is always a human adding something unique to their work in the end, their perspective, feelings and experiences. In the case of these AIs it's simply mimicry and it cannot be anything else. Those algorithms are just mixing the preexisting elements supplied to them, nothing more and since those parts are copyrighted I think anything that comes out of them is copyright infringement.
If using AI to compose I would advise to use only piano midi files
If you give a whole piece or song like that to Musenet by OpenAI, It'll have an easier job and spit out realy great results
I'd also advise to tell the AI via advanved settings that it should do so in the style of Chopin 1, Beethoven 2 or rachmaninoff 3
While the AI is supposed to just continue the piece it won't always do that, it will often just use the piece given as a ruff gideline with elements off the piece thrown in
It will take a lot of time but the results are likely to be pretty awesome, not yet human level though
I think it could probably easily pass off as atonal music - even atonal "composed" by a human lol..
Hi David, The company behind SD have a diffusion system under the banner of harmonai, and they are working towards some of this. good video. Enjoyed. Balanced, thoughtful.
An amazing video. Interesting and well made. Really puts things in perspective.
As an artist, having an ai that can infinitely generate copyright-free composition, texture, and pallette inspiration for a prompt WOULD be a big time saver
You're not doing it yet? Why?..
As a non-artist, having AIs that can generate gorgeous creative images and music from a prompt WOULD be transformative, but WILL decimate the employment of artists.
It sure will be a big time saver when you’re replaced by ai
@@lp712 how do you know i haven't been already
@@nutwit1630 😔
I sort of suspect, though I could be completely wrong, that "artists" of the future will spend most of their time coming up with interesting prompts for whatever AI they're working with, and combining the best images/music pieces/etc. into stuff that is so good and innovative we can't even imagine it.
Alternatively, we'll live in a future where artists have become completely obsolete, which would be interesting I guess, and kind of depressing.
Personally I think that by the time artists are obsolete, most jobs are. Which could either be a somewhat utopian society where everyone does what they want and work on things AI cant do, namely argue over things that cant have meaning and doubt the objective results they’re given, or it’ll be a cyberpunk dystopia. Or something in between. By the time we get there we’ll already know what to do or we’ll already have accepted our fate.
Just sounds awful tbh. I hope an asteriod hits earth or something and the people left have to start from scratch.
@@princepark1013 would value still be a useful concept if no one worked?
Awesome analysis David! Also love that you're repping Marc Rebillet!!
David, excellent video. Thanks so much!
I'd be careful to describe the work of text to image AI as "totally unique" -- in a lot of AI ethicists warn of "copyright laundering", where image models can just regurgitate source images almost identically and all of a sudden you apparently don't need permission from the original artist anymore
The maodels don't regurgitate anything. That's not how they work..
But they don't "regurgitate source images". You just can't do that with just a prompt. It's like asking a painter to paint the Mona Lisa from memory. It will have some resemblance of Mona Lisa, but it will not be the same.
@@itsbazyli we are already seeing visual AIs produce images with remnants of actual artist signatures, indicating that it was trained on their artwork, very likely without permission.
If this space continues, it will only get worse for artists vs art thieves.
@@doublinx2 no permission is needed to use images to train ai models. The images used are all displayed publicly. Do art students need permission to do artist studies? Should we not allow students who are aspiring artists to critically examine the works of others to gain insight into the craft? That would be silly and counter-productive, I think. Well that's not very different to what's going on when they train these ai. The ai engage in a comprehensive and technical artist study. I don't see much of a difference between a human student of art and an ai student. But feel free to make that case if you'd like.
Also, these ai's are not reproducing artists' signatures when they generate images - remnants or otherwise. That's just not they work. No part of any image used in training is reproduced or reconstructed as part of any image generated by an ai. So any worry you might have which stems from thinking they do, is completely unnecessary..
@@itsbazyli that's still the same problem though, you can get an existing piece with slight 1% variation. under normal circumstances that would be a clear copyright violation, but with AI people can pretend it's not and rip off artists without their permission
Honestly, people say "AI isn‘t quite there yet" but I think it‘s exactly where it should be - right in the part of the uncanny valley that‘s extremely funny. My best friend and I spend ages generating ridiculous stuff w/ Dall-E and sending it to each other (Hint: Give it incredibly abstract prompts, something absurd or single word / gibberish prompts, tje results are great), and let‘s not forget when Open AI Jukebox tried to recreate country roads but fucked up the lyrics to the point that "Life is old there, older than the trees" became "Life is over, for the damned bees". This is way better than AI doing things perfectly!
One time my best friend prompted midjourney with "Hilarious meme" and it created a poorly rendered hand pointing at a near photorealistic chicken
@@SubsonicNoise if you like the weird and abstract stuff you should be using Disco Diffusion. I'll drop a link to a colab notebook if you like..
Your work is absolutely necessary 💕
love this.. great commentary and forecast
I would actually appreciate an AI being able to finish my orchestral sketch.
More like the ai makes the sketch and you finish it?
@@nekokna Yeah, you could probably do that as well
Honestly, i hate this. Why does A.I art have to become a thing?
It takes so much time learning and creating art, but now you just have to click a button to be an artist? Who will care about all the abstract artists in the future?
It's not like they are overpaid and need to be replaced.. it's excactly the opposite. I hate that this is happening.
its just a another new tool for us humans to use to make art. dont worry about it.
It's literally how they make money. That's like looking at an app that can literally do 90% of your job and saying *"don't worry about it."*
Just yesterday I was live at a talk between composer Claus-Steffen Mahnkopf and author Hanno Rauterberg about AI in the arts, and today you upload a video on just that topic, wonderful coincidence :D
AI-Human teams, or collaboration with AI, is going to integrate into nearly every step of the production process of digital art. Rather than thinking of AI as replacing your own work, it can let a creative swap between rough drafts of major elements, experiment in further directions, and selectively alter any detail of a work until it feels intensely personal. Prompt engineering is (mostly) a stand-in for plugging in many methods of AI fine-tuning into familiar kinds of creative software. It's exactly accurate to think of future art as a conversation with the AI, a conversation that can push a creator to improve their skills and contemplate their tastes. Artificial neural networks have also made it inevitable that we'll have programs that give personally tailored coaching to vocalists and instrumentalists, and coach painters on how to apply different techniques, with or without the collaboration of a human teacher.
AI assistance will be trained for most aspects of life and shake up culture and the service sector at least as much as the internet already has. It's impossible to know how culture is going to adapt to this, but I think that art from small teams and individuals will become more intricately crafted.
That first example... I mean, I'll have to look into it, but even with how limited it sounds like that AI was... I just simply don't believe that it runs on a smartphone. Maybe you can access it from your smartphone (like with programs like DALL-E, that run on an external server), but that's not the same thing. lol
Yeah, that was a pretty sensationalistic claim. It's like using your phone to facetime a doctor and saying _"my phone does medical appointments!"_
One of the most brilliant and in-depth video on AI that goes broad and deep
Bravo! Thank you 🙏
I’ve been thinking about this a lot lately. Right now, I fail to see a future where a “music producer” in its current sense, doesn’t become an occupation of the past. Why learn how to make a beat on a DAW when an AI can generate 10 of them faster than you can click in a simple MIDI chord progression? All fully mixed, mastered and arranged? There are already sites that can do this. People say that these tools will “enhance human creativity”, more like replace.
I thought for a moment that you had a collab with Twoset! Still loved the video :)). Very good analysis!
This is amazing quality content non-ai generated that will be decades away of producing in any automated manner. Congratulations for the channel that I will be following from now on!
I liked the video very much, David, thank you for that, and have a wonderful day :)
Fully 20 years ago I helped on a Media Lab project with computer-generated composition tools for children to give input into graphically …I had the task of translating the kids’ assisted “creations” into legible scores for orchestral musicians to perform from. David, this js a very lucid description, but I hope it broadens beyond prompts of text to image or text to music.
great video, top arguments! thanks, Bruce.
David, enjoyed the video. You might be interested in the jazz pianist/programmer Dan Tepfer's experiments with computer programs that can respond to his improvisations.