 to compute because we can build relationships between everything, everywhere, all at once. And while computers can be trained on more data than we can read, it can't yet build links across disparate information the way we do. Music has long accompanied humanity, and yet after thousands of years of people making it, no one's quite sure how inspiration works. In fact, it seems impossible to scale, and we really need it to if we are to solve the types of problems that AI for good is aiming at. And when we try to interrogate these building blocks of creativity, we get surprising origins. For Gershwin, the sound of a clarinet warming up at an orchestra was the inspiration for that piece, Rhapsody in Blue, and I like to think of it as his prompt. It's the sound heard at the exact point in his life alongside his musical education, the socio-economic situation that he grew up in, and all that he was exposed to. The training data made Rhapsody in Blue. So, we all have our own training data. I've certainly had my fair share of input. I've been run over. I've fallen off a cliff. I even went through breast cancer during lockdown. I know, right? Quite a bit. But I've also filmed and performed on stage in over 50 countries. I've spent time with some of the world's most interesting people, astronauts, artists, hackers, founders, and you all here in the world of AI. It all adds up, but I just don't know how. So, this strange ability to connect the unexpected, does it give us humans the edge over AI when it comes to creativity? Well, it's complicated. Where machine learning currently tends to be, well, a bit recursive and finite, humans cannot help but seek out more. And though human programming is hard to understand, when it works, it's magic. Yet, whilst machines might not hold up against the best artists and inventors right now, they can get through breathtaking amounts of data with the right program. Generative AI now produces usable results at scale. Much more than one human could, even if they are a genius. And the tools are getting easier to use, meaning we don't need to be a genius for it to work. Now, I'm a musician, but this does not look like a musical instrument at all. About two to three years ago, this was how to make music with an AI. It's a coding environment that I, a non-everyday coder, needed to learn. It took multiple attempts to even work out how to use Google Colab running jukebox by open AI, and 10 hours overnight processing time, pressing a button every four seconds to create one minute of music. And even then, I found quirks in the prompting segments. For example, you needed to type lyrics in a certain way and leave spaces to help the output sing the rhythm that you wanted. Here's the result. I've got the Beatles to sing, call me maybe. That's made from the machine pulling together many micro Beatles units and then working out what could come next. Now, if we fast forward post pandemic, post AI being a household name, any of you can now compose music in much less time. No music theory skills, no instrument skills, and even no coding. You just need a sentence. The simple act of adding a natural language interface to generative models has changed the game, and it could signify a golden age of creativity now everyone can get involved. Here's some music I wrote with one or two sentences in Google's music LM. It took me less than two days to create over 50 pieces of music. Now, right now, the vast majority of publicly accessible AI tools use other people's models and training data. What's interesting, though, is power users are learning to manipulate these models by using special words in order to get special results, like cheat codes. Just like a few years ago where I changed the sentence format to make something more rhythmic, you can now do things like misspelling Mozart when you're trying to do a style transfer, or adding the phrase cannon 50 millimeter lens to your art generation, which automatically generates a better and more realistic response. I've decided that working out a prompt for these things is a bit like learning an instrument, but you're playing a scale almost. An AI co-creation, though, bypasses that physical and mental connection that we're making. So you don't need to learn an instrument in the same way. So though anyone can write a masterpiece, they still need to find the right prompt or incantation. And they've got to get lucky with that built in randomness. I don't need to say that the implications of a machine that can imagine can have this crazy response, synthesized symphonies, grand unified theories of everything. You get the idea. And so now it's time to play a couple of different tracks. This is stepping up from a single sentence AI creation with a couple of good examples of style transfer I made using Meta AI's music gen model. We're going to take a look at the first one. I'm playing an act music by Mozart. I love this track. Okay, so you know that one. What I did first was I made an mp3 isolating the melody. I started playing with the prompts. Would you like to hear a heavy metal Mozart? Yes, I thought you might. Okay. Now, bear in mind these are the best examples. There were quite a few that didn't work, but I do like the reggae one. Could you please play that please? It's not bad, is it? I wish I could take the credit. Okay, so this is where we talk about data diversity, because why do some styles work better than others? It's baked into the program as we all know. Music data skews towards western chromatic tuning. You know, major, minor. There's not that much outside of these major and minor keys. Secondly, AI-generated music is not challenging. True creativity needs more input. Like Einstein imagining cutting an elevator cable so that the elevator was in free fall that resulted in his relativity theory. If we included music outside the top 100 charts, it would make the model more inclusive, but it also makes the model more fun and more creative. So I wonder if machines just haven't been trained enough to be creative. It's like only learning four chords on a piano. You can play so many songs with this, so many songs, but you could not do... That's not going to come out of an AI, or even Beethoven's 9th. Okay, so what I decided to do is hopefully we're going to hear a few different versions of Beethoven's 9th Ode to Joy. See, many of these models are populated with data sets that users don't yet have access to. So let's go back and have a quick listen to heavy metal Beethoven, please. But it's fine. I guess also that Beethoven was kind of heavy metal at the time. So, okay. I wanted to include a failure as well. This is a dance remix. It doesn't work because I wasn't specific enough in my prompting. Let's have the dance, please. Okay. Everyone's having up of that. All right. The others sound okay, but to a musician, they are not musically accurate. Like the Punjabi dance track. Illustration isn't right. Right. You'll probably recognise that classical Indian drum sound. That's the tabla. But it's missing key components specific to modern Punjabi Bangra music, namely a particular kind of drum and a particular instrument that kind of does that. That's called a tumbi. So it's wrong. It's nice, but it's wrong. Gospel music as well. There's some piano, but again, the instruments aren't quite right. Gospel music has a swing to it. It's got a kind of specific feel to it. And if we can play it, I'd be really grateful, the gospel music one. The drum patterns are missing as well. This is, here we go. Okay. Thank you. Now, actually, the humans did a much better job of remixing Ode to Joy in a gospel style because it's actually the song Joyful Joyful. So, you know, humans are still around. Let's go to our next slide, please, because I put something in that was very familiar, very formulaic, and the machine does a brilliant job. I'll press that button and see what happens. It's going to do something strange, isn't it? I'm nearly there. I can actually play some old music while I do this. Okay, here we go. Let's play a romantic classical version of Despacito. Now, I did not feed this any chords, but it's got it right. What's going on? We're creatures of habit. We like to be predictable. The four chord structure that it's playing is so familiar, and it occurs many, many times in the dataset, unlike Punjabi Bangra music, right? Second, there's nothing challenging or surprising about this. Are we going to get a Beatles of Bob Marley or Billie Eilish? No. The majority of the training data is limited to a narrow suggestion, I guess, in the selection. So, can we train machines to know not just what new looks like, but what new and good looks like versus new and unusable? Ah, so, training a machine to recognize what's new and good becomes relevant outside of music. Some of the world's greatest challenges are hoping that AI in some fashion can help, from generating sustainable low cost energy, to increasing access to healthcare even, and equality and education for all. While these problems have legal, geopolitical, and, well, ideological issues to surmount, it's surprisingly similar to finding the right instruments or working out the prompt that leads to rhapsody in blue. And these issues we've spoken about over the last two days aren't solvable by companies or governments alone. So many people on this stage are wishing for more people to understand AI on a wider scale to see what happens next. And that's one of the reasons I'm collaborating to create this. This is going to be AI goodies. It's a user-friendly guide to take people beyond the fear of AI and start directing their thoughts to it intentionally, and we want everyone to be involved. We'll keep you apprised of the release date, and as we come to the end, we started yesterday with a powerful opening from ITU's Secretary-General, Doreen Bogdan Martin, a worthy quest to advance AI towards solving humanities problems while addressing the issues that plague it. And while we can't afford to treat generative AI the same as previous technological progress, we are not powerless. We have each other in this room, a network of more people beyond these walls and the opportunity to shape this next technological era together. And we also have the courage to educate ourselves. And most of all, we have the creativity inside all of us to imagine a future where we use AI for good. Thank you. Thank you.