Great video. I love these topics of conversation. I also love the saxophone playing in the background. I studied AI and neutral networks at university many years ago. Since then the possible size of neural nets has grown significantly. Huge data centres now exist to train networks that are very deep and have lots of layers of modelled neurons. Each modelled neuron has weights and bias values that are fed into an activation function to determine what its output will be from the set of inputs that are fed to it from the prior layer. The weights and biases are finally settled on after all the epochs of the training iterations have been completed. The training uses techniques like the backpropagation algorithm and the gradient descent optimization algorithm to do this. Other algorithms have been developed more recently to improve the speed of training. In any event the neuron's weights and bias dictate what it's output will be for a given set of inputs. You can think of the network as a very huge set of gates. Ultimately the outputs from the neurons in the final layer (the output layer) is what forms the final resulting output for a given set of inputs at the initial input layer. You have to do some preprocessing in order to convert the input and expected output data pairs (music, images, text or whatever) into a form that can be given to the network for training purposes. Also after training is complete, when using it for generation purposes, you need to preprocess the user's input to feed it into the network and then you have to do some post processing of the generated outputs to convert the output into the target medium. I think actually these conversion processes are the most challenging aspect of using neural networks and that is certainly where Suno and Udio's intellectual property primarily reside. You have to do a really good job of curating the training data so that it is broadly representative of what you are trying to train it on so that you get a a more generally trained network that can respond reasonably well to inputs that it did not receive in its training data set. It's possible to overtrain a network resulting in overfitting to the detriment of the outputs it gives after training. An overtrained network is more likely to be less general and have a tendency to produce outputs that were in its training data. While I agree with you that neural networks do not sample in the strict sense of the word, I do believe that it is possible for them to represent the samples as patterns in the weights and biases of millions of neurons that are then reproduced by the postprocessing operation on the neural nets outputs. It's not sampling in the strict sense of the word but they are still reproducing it nonetheless, and it's still problematic from a copyright point of view. Also I wonder how well curated the training data set is for the training of these tools. Could it be that the data set is poorly curated and they were overtrained and that they have a nasty habit of spitting out over prevalent copyrighted artefacts that were in the training data e.g. producer tags?

Thanks Paul. You’ve really inspired me to get into the both Suno and Udio and they’ve both been really fun and also eye opening. Your totally right though as you said in a previous video, It’s really a great idea tool more than anything, at least at this point. Thanks again.

Hi Paul, I'm a brazilian musician/producer. When testing Udio, I thought "let's see if this service has a pop/hip hop bias, or if it can swing to latin american rhythms". So I prompted for some sambas. What I received back where a few tracks, and among them there were the voices of brazilian singers João Gilberto and (contemporary singer) Céu. It weren't similar voices, it were THEIR VOICES, with their articulation and style of interpretation. A fellow musician asked for brazilian music and received back a track with Daniela Mercury's voice, a very popular singer from Bahia (northeast of Brazil). So, I guess we have a situation here...

The art that is from the AI obtained / contains none of the art on which it was trained.

That was a great video and thank you for clarifying that. The one thing, but at the end of the video you said it's up to the prompter to prompt AI to create an exact copy of a song. Can you do a video about how they would prompt the AI to create an exact copy of a song? I don't intend on recreating somebody else's song, but I'm just curious. I love the fact now that I can get all of the ideas out of my head with AI and as I left a comment in the previous video but it's due to my learning disabilities that I haven't learned how to play any instruments properly. But I would just like to know how a person would prompt an AI to create an exact duplicate of another song, I just mentioned I don't intend on recreating anybody else's music but I am just curious.

The biggest thing that gets overlooked or not mentioned with AI "using math" is "statistics." They're language models just like ChatGPT, but have music elements as the "language" they output. A random seed picks a starting point, then it looks at its training and "thinks", "based on the tags I was prompted with, what's the likely next thing to be played in a song?" Sometimes it'll select a less common element, a lot like a raffle where the common thing has 50 tickets and everything else has one or two. The video topic that was popular for a week where people were generating weird covers of existing Christmas songs is a great example. If "christmas music" is the prompt, neither the music model nor the text model used for the lyrics has a huge selection of Christmas music it was trained on. Sure, tens of thousands of songs, but that's hundreds or thousands of covers of the same song over and over, because there just isn't that much original Christmas music. It goes to Paul's point about users intentionally prompting it to generate infringing works, it doesn't do that if you don't ask it to.

People who say "Ai copies music" know that it isn't true.... but they are just trying to poison the well. They just want to reframe it as some sort of "immoral theft" so that they hope they can muddy everything and get it to stop or at least slow down. It's really sad how these people are fighting so hard against this amazing technology that offers the world SOOOO much in terms of art. All they care about is money and staying special and keeping the barrier to entry as high as possible.

So glad you clarified that AI and humans learn the same way! I remember learning music in school, and you’re right, it happened just like it does with deep learning now. I never once played a piano, or joined a choir, or improvised on my instrument, or spent hours mastering centuries-old harmony and counterpoint techniques - but boy, I sure did steal a lot of CDs from the library, label and define each entire track in bland, cookie-cutter terms such as genre and style, then run them all through a ton of linear algebra calculations. Backpropagation was the hardest part, but that’s just how we humans learn! And now, just like AI, I’m still unable to think, feel, have an experience, dance, play a physical instrument, respond or interact with a crowd, analyze a score, create original forms, or explain anything at all about what I’m doing, but of course it’s still music. In fact the humans are the real fools and thieves here: jealous, entitled, attached to their bodies, environments and communities, as if these things really exist or matter, and can’t be equated with data inputs to a computer model!

Well explained. Again ❤

An advantage AI has is that it cannot only copy a particular style of music it can copy how the music was recorded. To me it looks like the AI is taking the MP3 or wav form and manipulating it into something different. It doesn't play the instruments in the AI generations, it manipulates the frequencies in the original recording into a different song but with the same sonics as the original recording .

AI copies the sounds used to create music, but uses a mathematical formula to create music, because music is math.

No matter how much you like it or how you word it it is copyright infringement 101. But before keep arguing over wheter who's right or wrong I strongly suggest you find out the political agendas behind pushing Ai at all levels.

It sounds like copy alter paste multiple times from multiple sources much quicker than a human could do. Having listened to a Depeche Mode 'AI' track it sounded like a poor rip off with banal lyrics. I'll stick to the real thing thanks. Typing a few prompts can't compete with the joy of writing a great song yourself.