Transcript
Aidan McCullen: Today’s guest is a world-renowned AI researcher and generative AI pioneer. In her book, she explores the rise of creative AI from its earliest pioneers to the cutting-edge tools shaping music art.
And human imagination Today, she reveals the true capabilities and limitations of gen ai, drawing on psychology, philosophy, and her own cutting-edge research. She reveals how generative AI exposes both the brilliance and the blind spots of human society. It is a pleasure to welcome the author of Creative Machines, Maya Ackerman. Welcome to the show.
Maya Ackerman: Thank you so much for having me. It’s a pleasure to be here.
Aidan McCullen: It’s great to have you on the show, and it’s our first show on video Maya as well. So you’re our first guest that people, first time, they’ve seen me as well.
Maya Ackerman: Oh wow.
that, that’s extra special. I love it.
Aidan McCullen: great to have you with us and I thought we’d share four things today, and we’ll go much deeper than this, [00:01:00] but the unknown history of AI generated creativity, so we’re all using it and it’s almost like. It was sprung upon us and people started using it, but they don’t really know where it came from.
So that’s the first thing. The second thing is the capabilities and limitations of creative ai, your own framework for using AI as a tool to enhance our creativity as humans. And then. The really important aspect that’s often overlooked, which is the ethics of ai. So I thought we’d start with what you mean by creative machines, where they came from, going right back to Harold Cohen and the machine paint painter, Aaron.
Maya Ackerman: I love that we’re starting there. ’cause that’s where I started. It was 2015 and I was attending a little conference and I remember just feeling so disconnected and conflicted about my life choices to become a professor and study machine learning. And then I saw a little [00:02:00] a little segment in the conference called Computational Creativity. And I was like, oh, creativity, interesting. And so I went to the back of the room and then I see Harold Cohen starting to splash these pictures on the projector, these beautiful pieces of art. And suddenly he starts arguing with himself on stage about whether or not his machine painter is creative. And I was just, wow.
And in that moment, I’ve decided that I’m gonna study this and I’m gonna figure out what it’s all about. So that was really the beginning of my journey. 10 years before about 10 years before all this stuff meant went mainstream.
Aidan McCullen: It’s probably important for us to share the definition of creativity that most scientists rely on. Creativity is novelty plus value plus surprise. Because before we can talk about creative machines, let’s talk about what it is to be creative as a human.
Maya Ackerman: Yeah, so that’s one really funny thing about creativity, right? The moment somebody says something, you can just [00:03:00] change the lens. Are we talking about a process or product? Are we talking about the kind of creativity that changes the world or the kind of creativity that a 5-year-old should be engaging in? One definition that helps ground the conversation says that creativity is exactly like you said, novelty plus value. And then sometimes we add surprise as well. And here we’re referring to a creative object. So I can say that this painting is creative. If it’s novel enough, it’s not just a copy of another painting and it has value, at least aesthetic value in that context.
So value is domain specific. And when we have that opens the door to machine creativity because then we’re judging it by what it’s outputting. So it’s a good grounding.
Aidan McCullen: Maya, one of the things with this show is we look at the science and neuroscience of learning, and I thought we’d lean into that a little bit because of the creativity, and I’m jumping ahead a bit here to the risks I [00:04:00] suppose, of using machines to replace human creativity and not working together with them.
Because a lot of people are fast tracking the struggle of creation, whether that be art or music or writing, to use AI just to get there faster and in there we’re losing something important.
Maya Ackerman: Yeah, so it’s it’s a really interesting space because. In a way, the makers of the systems have certain expectations and different users have different needs and expectations. So as gen AI got pushed, it got pushed with the idea that we want to be rescued from the need to be creative. Here, I’ll do this instead of You Like a Don, you don’t need to worry about painting here. Here you go. Now look how creative you’re being. Right? Which maybe is okay if it’s really something you are just doing for work, something you’re doing for someone else that [00:05:00] you don’t care about, but if you are the one trying to be creative, creativity into a spectator sport. And nobody wants that. Nobody wants that. And slowly the industry is trying to move towards giving us more collaborative tools. And humans, human creatives are finding the most innovative ways to use tools, not designed to work with them.
Aidan McCullen: I thought we’d share beyond the book some of the things that you’re seeing, because you mentioned stuff like the AI painter you mentioned. AI music, creativity with EMI, for example, or Emmy, I dunno how you pronounce that, but let’s share a little bit about those and then how you see us working with them to create value as a collaborator or a cobot.
Maya Ackerman: So it’s interesting both in academia and in industry at first, people are excited about autonomous creativity or the agent, what can AI do by itself? For example, [00:06:00] Amy is always presented so that David Cope system experiments and musical intelligence back from the eighties presented as an autonomous system.
Whereas in reality, he did a lot of interactive work with it. Today we get really excited about agents that do stuff all by themselves. But really the beauty of AI when it comes to creative endeavors, where it’s really meaningful as part of culture is when it helps people to profoundly express themselves. Right? So there is this is literally what I do. So I have so many stories, but I’ll tell you just one here. We did a study. Quite a while ago, around 2020 with the University of Dhi on bereavement. And so idea that the idea was that people would write songs about loved ones that they have lost recently. And so the idea of having a machine writing instead writing it instead of you making it super easy doesn’t quite make sense [00:07:00] here because it’s about the process of grieving, it’s about that creative expression in order to what’s luck inside of you. And a lot of people said, I could never write lyrics.
I don’t know. I don’t know how to do this. This is too intimidating. And so then they used our system, Alicia, which is especially the part of it that has become Lyric Studio. And it would give them the confidence. It would write a little bit, it would give one line, give them some choices. They would pick a line or two, and suddenly they’re like, no, I know what to do.
I know what to do. And what would end up happening is that they would write almost. The entire RICS by themselves and the role of the machine was just to give them a little bit of confidence and give them a little bit of a push, this is where we need to go.
Aidan McCullen: There’s an artist called Olafur Arnalds. I dunno if you’re familiar with him. He is an Icelandic pianist. Have you ever heard of him?
Maya Ackerman: No, I’ve not.
Aidan McCullen: So the reason I share that, I know you were a pianist. I know you lost your piano when you had to move as [00:08:00] well. Sorry to hear that. Hear that. So this is when Maya was a child, but Olafur Arnalds has this beautiful thing he hacked.
There was this thing called the Moog Piano Bar, and it was created to turn your beautiful ground piano into a synthesizer. And nobody used them. They didn’t see the use for it, but he got a friend of his to create an algorithm, and what he did was he created two, what he called player pianos.
You know the way in a haunted movie where the piano starts playing by itself. So he used these Moog piano bars. And when he shared this idea it really sprung to me the real value of using AI with a human. So he said that he’s a trained pianist. So for example, he would play an A minor and then would say, naturally, when you play that chord, you’d go onto an F or something like that.
And he said What happened was that ai, [00:09:00] when he played. A note, the two pianos would play these sequences that no human would play because they’re trained in a certain way. And he said what the value was that it would lead him in a different direction next in a direction that he would not have chosen otherwise.
And for me, that’s the beautiful way that aI can bring you in a different direction, and I thought that’s the real sense I got from your work with Wave AI and indeed your creative passions.
Maya Ackerman: The way that you described it, honestly is perfect. It’s perfect. If you think about any creative space. A chapter, you could write a book, title, a song, the next line of lyrics. It’s this massive space of possibilities. Imagine it as like stars in the sky. It’s enormous. And most of us – even very successful creatives – tend to follow very specific paths within this sky of possibilities. And one of the most beautiful things that AI [00:10:00] can do is exactly what you said, help us turn a different direction, and then we start exploring there.
Aidan McCullen: Maybe we’ll share a little bit about The The background. So before we get into limitations a bit more, and I love how you bring in philosophy and psychology and bring in Jung and the shadow collective consciousness. We’ll get there in a moment, but you give us a bit of the history here of expert systems, really importantly Markov Chains
I think that, for me was a real aha moment when you wrote about that and then neural networks as well. I thought if we explain. The history, and you do this beautifully, you do this very simply to bring people on the journey, because then we can understand actually what’s going on behind the curtain.
Maya Ackerman: I hardly ever get to speak about that. Thank you. A lot of AI starts with imitation. So the goal of AI used to be to come up with an intelligence that’s on par with human intelligence. Then [00:11:00] we understood that we don’t have to imitate human intelligence in order to create something intelligent.
It can be differently intelligent from us, which is really important. But it’s the same thing about the process, and that’s been an interesting journey. At first, we started with introspection. So you know, you have a painter. Harold Cohen and he had his way of painting, and so he taught Aaron to paint using exactly his process. So it’s it requires introspection. Human introspection is limited in general. Harold did an amazing job with that. But if humans had perfect introspection, you could ask doctors, how do you diagnose people? How do you choose how to cure them? And they would tell us exactly what they do, and we would code it into a computer and we would’ve replaced doctors 50 years ago. But humans don’t have perfect introspection. That’s okay. That’s how brains work. They work, and we don’t know how they work. That’s reality for the most part. We don’t know how we do the things we do, not enough to code it into a machine, [00:12:00] and so gradually we gave our machines more and more freedom, went from being these hyper controlling parents, telling our AI exactly what to do into giving them gradually more and more freedom until now. We finally have the resources and the power, the real electrical power to be able to create actual machines’ brains that learn by themselves from data that we give them. And because finally the machine brain can be big enough, it can do a really good job learning from data, is why actually managing what the machine does. And guaranteeing its behavior is very difficult, but at the same time it can do these amazing emergent things that we don’t even plan for it to be able to do.
Aidan McCullen: Maybe we’ll explain the Markov Blankets or the Markov connections, because for me, that was a piece that when you described it simply or. You’re even [00:13:00] thinking about how they’re chains that just are the next link in the chain, literally like a chain, and that’s how AI actually builds. But then equally, it depends on what you’ve fed it as Chains.
Maya Ackerman: Yes. Yeah. Thank you for saying that. Thank you for yeah, redirecting me there. take like Emily Dickinson poetry or Dr. Seuss, or. The tweets of your favourite or not favourite politician, right? Any body of writing that has a distinct style. A really simple exercise that I always do with students.
Students of mine who study ai, is to show them how easy it is to do something that’s generative. Okay? It’s incredibly simple. You build this teeny, tiny little model, you literally look at. How often each word in your dataset is followed by another word. You look at all pairs that are consecutive and you just count.
How often is the word [00:14:00] creative followed by the word machines? If you’re looking at my body of work, for example, and you do it for every single pair of words and you essentially build this graph where each, where there’s like a circle for each word, and then there’s a line connecting each pair of words. With the probability of going from one word to another, and that probability is just directly copied from your data. It literally is the probability that these, that, that word follows the other one in your data set, so now we have these lines that connect these circles, and what we’re gonna do is we’re going to, for simplicity’s sake, randomly drop into one of these circles. any one of them, right? And now we’re gonna follow these probabilities.
So if one line has a probability of 0.5 then they’re 50% chance that we follow that link. And you just start following those links with a probability listed on the line and you generate. So you went from having a data set to be able to generate stuff [00:15:00] supposed to be in the style of this data set. It’s not gonna be perfect, but it’s stunning. How well something so simple works.
Aidan McCullen: And Maya, what changed in the last few years that we had the breakthrough with the ability to do this? Was it the data sets were big enough, or was it the power of computing was big enough? What? What was the breakthrough?
Maya Ackerman: So to some degree we were able to do this for a long time, but what is now enabling us to do it this well, right? Like really this link. Definitely there’s a big gap between what we could do now what we can do now and what we could do before. And a lot of the breakthroughs were gradual.
But a really big one was when OpenAI managed to convince Microsoft to make an unprecedented investment in doing the largest machine brain ever. We be talking about billions of dollars here. And they did [00:16:00] that. They actually wrote a paper that explains why bigger machine brains are gonna be smarter, which is obvious in retrospect. But that’s really what started moving things. GPT three showed incredible capability, and that’s before chat GPT, right? And already was starting to starting to show us how far machine intelligence can go.
Aidan McCullen: One of the things I mentioned earlier on was the Jungian elements. You’re obviously a Jung scholar. You’ve read about Jung and you talked about the collective and conscious and the shadow. And I often think about the show that’s on Netflix called Black Mirror, and the whole idea is the black mirror’s, like with the phone, it’s like a black mirror.
As a, also like a shadow of society. So it’s showing back to society how we behave, where we might not want to admit that. And sometimes we look in the mirror and go, oh, that’s ugly. And in, in the same way you say [00:17:00] that this is what ai, I love how you position this. This is what AI really is.
And it’s one of the problems with AI is that if we think about all those Markov chains, we’re the ones who have fed it, the chains. So if we’re getting back answers we don’t like, we’re the ones who’s fed it in the first place.
Maya Ackerman: We don’t like looking ourselves in the mirror. So at first, if you look at some of the earlier models. Even ChatGPT, but definitely the ones before that, you still have some of the less aligned models like Mid Journey. They’re just so honest about humanity. I did some experiments on how Jewish people are portrayed in AI systems and one of the most benign things was you ask for, which is a Hanukkah food, and you get bagels. Which is which is again, relatively benign. I don’t think that’s that’s a terrible form of antisemitism or anything, it shows how the West doesn’t really [00:18:00] understand Jewish culture. How they, they know bagels, but they don’t know the real Jewish foods that go for with the holidays.
And some of the stuff was much more serious. I just shared the light example here. When you give a picture of a woman and you add the word professor or CEO with Midjourney, you get facial hair on that woman, right? So it’s like saying that there must be something really masculine about her if she’s doing these things. And I think that there’s something really beautiful, painful, but beautiful about this honesty. Um, and the final thing I wanna add, which is not in the book, is that we have seen. This extreme alignment in LLMs today. Again, not all the models and the shadows still come through, but there is this really extreme alignment.
There is this effort that these companies are doing to eliminate this mirror experience, but what they’re doing instead is they’re just normalizing it [00:19:00] what’s appropriate in the west and actually in the process, in some cases, making things worse. So it’s very interesting how much we struggle with our shadow and how that’s reflected in this industry.
Aidan McCullen: You talked about hallucinations at one stage, and again you liken it to a human having maybe psychedelics and I was reading recently about psilocybin. So the active ingredient of magic mushrooms and what it does in the brain is actually it’s silences or it activates certain parts of the brain actually.
It turns off certain parts of brain. It turns on other parts of the brain and you start to make connections because the brain is almost got Markov blankets as well. It’s got these silos. And I was thinking then, ’cause you say it’s exactly what happens when the machine hallucinates because it starts to mix up its data.
In [00:20:00] there, you can have these massive breakthroughs, but you also can have what we call the hallucination that happens with ai. I I’d love if you share a little bit about that.
Maya Ackerman: That’s one of my favourite topics after the Industrial Revolution. We like to believe that the world is concrete and real and objective, that what we see is reality, but the truth is not the human experience. Is deeply steeped in imagination and neuroscience is showing that way that we engage with the world is through small scale prediction. You don’t really see the world or hear the world. You interpret it. sometimes the interpretation is so off that we notice, let’s say after a breakup, you might see your ex when they’re not there, right? You might mistake other people for being them from far enough away, which is something that happened to me. But it’s, even when it’s not that [00:21:00] extreme to us, are constantly projecting past experience onto the present. That’s almost what trauma is, but even when it’s not trauma, we’re doing it. And not. It is not good or bad. It’s the way it is. We are an imaginative being and the reason that our machines right now are so successful and so smart is because they finally learn how to imagine a way that’s quite similar to us. And just like we get mad at each other or at our kids when their imagination makes us uncomfortable, when it goes too far, whatever that means. It’s the same thing with our machines. No, you are wrong. Do what’s right. And so we’ve made these LLMs much less creative in an effort to make them behave. So there’s just so much more potential by letting our machines breathe a little bit.
Aidan McCullen: I was asking you about , what’s changed, and you mentioned open AI and the investment to [00:22:00] build a bigger brain, but also. You hear, and I don’t know how true this is, but the, we’re running outta data to feed these l and m’s because if you think about it there if you’re feeding it Markov chains, you have to have enough chains of data.
But I was thinking that how a lot of LLMs are actually using information on the internet to feed themselves. Yet a lot of the information on the internet is already created by ai. Ai, so it’s like self cannibalization. It’s eating itself. It’s like the, Ouroboros It’s feeding on itself and it means that the content becomes more and more.
Useless because it’s actually not coming from that beautiful, unique human place. Real research that you do, for example, as a researcher, that the origin isn’t there and there’s a real danger in that.
Maya Ackerman: Yeah, I. Really agree with you there. It’s almost [00:23:00] comical. Researchers have done really diverse research. We keep exploring new directions and it’s almost I keep thinking of California, how people came here to dig gold. Researchers went around, poked around and found these little pots of gold, and then OpenAI had a lot of money and so they drilled really deep and they found this one like enormous source of gold.
Let’s make chatbot. And then everybody else who has a lot of money to drill came and they’re drilling around like a square inch where open AI is drilling. The lack of imagination in industry is stunning. They have so much resources they can afford the best minds in the world who are willing to work for them. And, all they can do is make little iterations on each other, and it’s on purpose, right? They’re making a safe bet. This is not an accident. And so they realize more data, bigger brain makes it [00:24:00] smarter. So they’re like, oh my God, do we need to go to MARS to find more data? Like how do we create more data? But the reality is that if you constrain the machine less, if you make, if you let it imagine more, if you make it less of a copy machine and more of an imagination engine. There’s so much more than you can do with this technology.
Aidan McCullen: Just like in school, let the kids imagine, but bringing it back to education, I thought we’d use that collective consciousness. So you talk a lot about bias and ai, et cetera. We’ve spoken like about that before on this show, but I really wanted to share your framework for using AI as a tool to enhance human creativity and really importantly, your perspective on AI ethics.
Maya Ackerman: I appreciate it. So let’s imagine that you have a collaborator. You’re trying to learn how to paint, and there’s a person that’s assigned to partner with you to help you [00:25:00] out. Maybe they’re ahead in the art program or something. And they could say, oh, okay, so you’re struggling here and they take the brush from you and they turn around they fix the issue that you’re having.
Actually, they take a brand new canvas and they’re like, here, I fixed a few more things and here I did it. And they just keep doing that. You try to engage, you try to say, oh, actually I meant this. I was trying to make it come darker. And I was trying to experiment with that texture.
And each time they just throw the canvas away and take a new canvas and turn away from you and get it all done and behave like a text image model. And then you’re like, oh, what is wrong with me? Why am I unsatisfied? Why does this I don’t wanna deal with this anymore. And you’re just like, about it.
I don’t need this art class. Or there could be another person that you have the same challenge they’re like, oh, okay. [00:26:00] And they gently give a suggestion or maybe they give you an example on another canvas. But they prioritize you. when you don’t need them, Barely there when you don’t need anything from them, that’s like a good educator, right?
They want you to grow. The goal is your creativity. Not to give you a creative product quickly, but to help your creative process grow. And the goal is that after interacting with a person like that, will become permanently more creative. You’ll become a better painter even if that relationship concludes for whatever reason. that’s a kind of bar that we need to set for our technology. People have used who have used Lyric studio in my system that helps people write lyrics have told me that they’ve become permanently better lyricists. And I’m finally starting to think this idea start to take root a little bit in the industry at large.
Aidan McCullen: It’s beautiful. It’s also like being a good parent. [00:27:00] Nudge them along, but don’t do it for them. It’s very difficult not to when you’re under pressure, under time pressure as well. And finally, let’s share your perspective on ethics in ai, because this is something that we don’t hear enough of.
Maya Ackerman: So perhaps one of my main points is around AI and the all known Oracle. We have this desire to have life be more predictable. To have somebody know what the answers are, be it a religious leader, a God, like some kind of entity in this universe that knows what’s gonna happen, that knows the right way, right? It gives us comfort in a world that’s really incredibly uncertain. And whether it’s true or not, it’s a separate issue. But there is definitely this kind of pull. And so one of the things that I believe may touch this so successful is that it pretends to be that. It pretends to be this all knowing Oracle. And if you look at the marketing, if you look at how the industry speaks about LLM, [00:28:00] they’re feeding the science fiction narrative that this is our Wizard of Oz.
This is our all knowing Oracle, and we should be able to trust it. If it’s wrong, that’s a hallucination, that’s a mistake, that’s like an exception. And so we get super mad when these things are wrong, when they lead to terrible suicides and like work slop these things with different stakes. We get really mad, and perhaps we should, but an important piece of the puzzle is removing this idea that it should be an all-knowing Oracle. The idea that something knows the truths about everything is ridiculous. It’s absurd. It’s never gonna happen. We don’t agree with each other what the truth is.
There is no absolute truth. It’s not how reality works. And so we need to stop trying to create it, and we need to educate our children that this is just a machine that somebody builds. And you should take what it says with the same skepticism, with the same questioning as you would anything that a person has told you. [00:29:00] And that will remove a lot of problems. And if we can do this theory of mind shift.
Aidan McCullen: Let’s share where people can find you.
Maya Ackerman: our claim to fame is Lyric Studio. Yeah, lyric studio.net and lyric studio.com. And wave.net is the website more broadly, but if you wanna follow me specifically. I’m very easy to find on LinkedIn. That’s my main social, and today we spoke about the book Creative Machines, AI and us, which is available on Amazon and elsewhere.
Aidan McCullen: It’s been an absolute pleasure talking to your author of Creative Machines, ai, art, and us Maya Ackerman. Thank you for joining us.
Maya Ackerman: Thank you so much for having me
Please note that this transcript is auto-generated – please refer to the audio or video for exact wording.
Welcome to the debut video episode of Inside Learning – your go-to series exploring the intersection of learning, innovation, and technology.