Maya Ackerman | Creative Machines: AI, Art & Us | Talks at Google

Who speaks in this episode?

• Maya Ackerman (Guest)

What is discussed, and when

1. 0:00 – 5:00: Maya Ackerman, LyricStudio, Impro-Visor, the International Conference on Computational Creativity
2. 5:00 – 10:00: Stability AI, Maya Ackerman, AI actors, AI musicians
3. 10:00 – 15:00: Artificial Intelligence, Maya Ackerman, Google LLC, Midjourney
4. 15:00 – 20:00: Artificial Intelligence, Google LLC
5. 20:00 – 25:00: Maya Ackerman, Google LLC
6. 25:00 – 30:00: Maya Ackerman, Google LLC
7. 30:00 – 35:00: Maya Ackerman, LyricStudio, AI shaming
8. 35:00 – 40:00: Maya Ackerman, Midjourney
9. 40:00 – 45:00: Maya Ackerman
10. 45:00 – 50:00: Maya Ackerman
11. 50:00 – 51:26: Maya Ackerman

Notable quotes from "Maya Ackerman | Creative Machines: AI, Art & Us | Talks at Google"

• I was attending a research meeting of my research community, the International Conference on Computational Creativity, and my late colleague Robert Keller, who was a professor at Harvey Mudd, had a system called Impro-Visor.

  about Robert Keller · at 0:12

• So for two hours, I played with Impro-Visor, missing a whole bunch of talks.

  about Impro-Visor · at 1:09

• So what we did is we took that system and we retrained it on the public domain works of Giacomo Puccini, and we gave it to James and he spent several months with that system.

  about ALYSIA · at 2:24

• We opened Wave AI, one of the first musical AI startups, and like many companies, we succeeded on our third product.

  about Maya Ackerman · at 3:12

• Here, you guys know the space Anyways, so I get an email from one of my investors asking, what do you think of Stability AI?

  about Stability AI · at 5:43

• And in particular-- that was perhaps, overly dramatic image for you here --in particular, the vision that I kept hearing from investors, from a small number of entrepreneurs at that time, was that of replacing human labor.

  about Artificial Intelligence · at 7:33

• I'm so excited to come to Google to share my ideas about humble, creative machines.

  about Google LLC · at 13:50

• And so I spent hours and I generate a whole bunch of images, and none of them are quite right.

  about Midjourney · at 14:47

Who and what is mentioned in this episode?

People

• Maya Ackerman — Person · Main speaker / guest author · 20 mentions · first at 0:07
  Central voice of the episode segment, describing her research and startup work in generative AI for creativity and advocating for 'humble, creative machines' that elevate human creators rather than replace them.
• Robert Keller — Person · Referenced colleague / referenced expert · 1 mention · first at 0:19
  Presented as Ackerman's late colleague and the creator of Impro-Visor, used as a formative example of an AI system that helped her learn improvisation and gain lasting creative confidence.
• James Morgan — Person · Referenced collaborator / case-study creator · 1 mention · first at 1:42
  A conceptual artist who collaborated with Ackerman using her system (ALYSIA retrained on Puccini) to create an Italian aria despite not speaking Italian or playing an instrument—an example of AI-enabled creative learning and output.
  And on my very first day at the job, I met a professor called James Morgan, and he's a conceptual artist.
• Giacomo Puccini — Person · Referenced composer / training-data source · 1 mention · first at 2:26
  His public-domain works were used to retrain Ackerman's singing system to generate Puccini-like melodic phrases, enabling the James Morgan aria project.
  As a personal fan of Puccini, I thought that was just fascinating.

Companies & orgs

• Wave AI — Company · Business focus · 1 mention · first at 3:12
  Maya Ackerman's startup focused on musical AI applications
  And I learned a lot from our mistakes in the first two products.
• Stability AI — Company · Case study · 2 mentions · first at 5:35
  Example of a company that raised significant funding for generative AI
• Google LLC — Company · Referenced expert · 6 mentions · first at 2:16
  Mentioned as a place where AI technology is being developed and applied
• Harvey Mudd — Company · Referenced expert · 1 mention · first at 0:23
  Mentioned as the institution where the Impro-Visor system was developed
  I was attending a research meeting of my research community, the International Conference on Computational Creativity, and my late colleague Robert Keller, who was a professor at Harvey Mudd, had a system called Impro-Visor.
• San Jose State University — Company · Referenced expert · 1 mention · first at 1:32
  Mentioned as a place where Maya Ackerman worked and collaborated on creative projects
  Another thing happened, actually, a couple of years earlier in 2016, I worked for a year at San Jose State University when I moved to this area.

Products & tech

• LyricStudio — Product · Business focus · 3 mentions · first at 3:24
  Product developed by Maya Ackerman's company, Wave AI, to assist in songwriting
  And ultimately, LyricStudio has a massive text box, so it invites people to write their own stuff.
• Impro-Visor — Product · Case study · 2 mentions · first at 0:24
  Early system that helped Maya Ackerman experience creative freedom in music
• ALYSIA — Product · Case study · 1 mention · first at 2:09
  Early system developed by Maya Ackerman for generating music from lyrics
• Midjourney — Product · Recommendation · 2 mentions · first at 14:36
  Image generator used by Maya Ackerman for creative inspiration

Concepts & topics

• Artificial Intelligence — Concept · Core topic · 2 mentions · first at 12:55
  Central theme of the episode, discussing the impact and future of AI in creativity and society
• AI musicians — Concept · Key argument · 1 mention · first at 8:31
  Discussion on the emergence of AI in the music industry
  It's not exactly AI musicians, but it's AI getting used for various aspects of being a musician.
• AI actors — Concept · Key argument · 1 mention · first at 8:43
  Discussion on the impact of AI on creative professions like acting
  Then, we have this idea of AI actors.
• AI shaming — Concept · Key argument · 1 mention · first at 33:27
  Discussion on the stigma around using AI in creative processes
  I think it captures in it the discomfort that creatives feel around the whole space.

Places & events

• the International Conference on Computational Creativity — Event · Setting · 2 mentions · first at 0:16
  Event where Maya Ackerman first experienced the Impro-Visor system
  I was attending a research meeting of my research community, the International Conference on Computational Creativity, and my late colleague Robert Keller, who was a professor at Harvey Mudd, had a system called Impro-Visor.

Full transcript with timestamps

0:07 MAYA ACKERMAN: So the first story is in 2018. I was attending a research meeting of my research community, the International Conference on Computational Creativity, and my late colleague Robert Keller, who was a professor at Harvey Mudd, had a system called Impro-Visor. And it's still available completely for free on the Harvey Mudd website. And how it works is you sit down in front of a tiny piano or a big piano, and you play a little bit of music, and then it responds to you. So I would play a few bars of music and it would play back, and then I would play something and it would play back.

0:49 And I did not know how to improvise. Like many people, I learned to play music as a kid by playing other people's compositions, and the whole idea of making my own compositions, improvising something felt really, really intimidating until that day, with Impro-Visor. I finally felt that I had this creative freedom. So for two hours, I played with Impro-Visor, missing a whole bunch of talks. And what's really amazing is that not only was I ultimately able to improvise to my satisfaction during that session, but some of that had stayed with me.

1:25 So from working with this system, there was some kind of lasting impact for me. Another thing happened, actually, a couple of years earlier in 2016, I worked for a year at San Jose State University when I moved to this area. And on my very first day at the job, I met a professor called James Morgan, and he's a conceptual artist. And James told me that he always wanted to write an Italian aria. There were a couple of problems though. He did not speak Italian. Believe it or not, that was not the biggest problem. The biggest problem is that James had never played a musical instrument and he wanted to write an Italian aria.

2:07 It so happened that at that time, I was working on a system called ALYSIA, where you would give it English lyrics one sentence at a time, and it would sing them. So something like, it's so cool to be at Google today would become, (SINGING) it's so cool to be at Google, today, or like, a whole bunch of other options. So what we did is we took that system and we retrained it on the public domain works of Giacomo Puccini, and we gave it to James and he spent several months with that system. This wasn't Sooner or Udio.

2:37 It was like little tiny phrases that his student would translate for him into Italian, and then it would come up with these sort of Puccini-esque melodies for it. And he ended up learning about music through this process and wrote a really, really cool Italian aria about a woman riding a dragon inside of a video game world, that ended up being showcased in museums all over the world. That really blew my mind. I didn't know that that's the kind of thing you could accomplish with GenAI. So that was a big shock, and this was before I opened my company, which happened in late 2017.

3:12 We opened Wave AI, one of the first musical AI startups, and like many companies, we succeeded on our third product. Everybody thinks they're going to succeed on their first and often doesn't happen. So our third product was called LyricStudio. It's still around, used by millions of people. And I learned a lot from our mistakes in the first two products. And ultimately, LyricStudio has a massive text box, so it invites people to write their own stuff. And whenever you're stuck, maybe it's right away, maybe it's at the end, maybe it's in the middle, it gives you ideas on where to go next, not the right next line, not the most likely next line, but just ideas to help people see more broadly.

3:56 Sometimes they get inspired and end up writing something completely different-- just to see more freely. And we got a lot of feedback. You know, you do a company-- you talk to your customers. That's sort of, that's the gig. One thing that I found really, really surprising, it was not on purpose, is that people were becoming better pen and paper songwriters as a result of using LyricStudio. So this is so different from how we typically think about technology. It's about-- what I've learned from inside of operating a startup, people tell you, make it addictive, make sure they have to come back.

4:32 And here, we had a product that was sort of the opposite of that. It actually left this kind of permanent, positive imprint on people's creativity. And some people would stop using the product and some people would keep using it because they felt that they could still keep growing even as it got better. And it's almost like the opposite from the one-hit wonder GenAI products that sometimes get released today, where you come in, throw in your photo, it gives you some cool version of yourself, I don't know, flying a dragon or something, and then you never come back again.

5:04 That's sort of the opposite extreme where you feel disempowered. And we were finding our people were really learning how to be more creative from using our tools. Next slide. So I was kind of living in this bubble with my colleagues and my co-founders, in our little GenAI world, and apply to creativity. And then, in late 2022, November 2022, such an important date in recent history, I get an email in my inbox. Anyone here remember Stability AI? Some people do, OK. In some rooms, it's like, nobody. So it's cool here.

5:43 Here, you guys know the space Anyways, so I get an email from one of my investors asking, what do you think of Stability AI? They're the talk of the whole town. And I looked into it, and it was a guy who was out of hedge fund management with no experience in AI. So I thought. eh, it can't possibly be a big deal. Except, of course, within weeks, he raised $101 million to do all of generative AI. And Stability did not end up being a leader in the space. One of the leaders in the space is the current company here. So this is kind of a handful of people who are really leading here.

6:23 But what was really interesting is that once investors woke up, that's when the whole space woke up. That was the order of events. It's not that, you know, some people think ChatGPT got released and then GenAI woke up. That was not actually the order of events. Stability raised a ton of money. Investors pivoted to caring almost exclusively about GenAI, and then everything else followed. And so my life has really changed. I remember waking up every single morning, not being sure if my life is real. I've literally devoted a decade of my life to studying GenAI, running one of the earliest startups where investors told me it doesn't make any sense, Generative AI is never going to be a thing, to suddenly living in a world where it was the most important thing.

7:09 And it was very exciting and felt very positive for a moment there. I could meet with any investor. It felt like all the doors were open. But then, I started to notice something, which is the vision that people who were just learning about GenAI were bringing into the space was very different from what me and my colleagues have been thinking about, the kind of world that we were trying to create. And in particular-- that was kind of, perhaps, overly dramatic image for you here --in particular, the vision that I kept hearing from investors, from a small number of entrepreneurs at that time, was that of replacing human labor.

7:49 So we were offered a pretty large amount of money to take our tech and direct it towards, build it out and direct it towards replacing musicians. Now, I don't think it makes me a particularly great human being that I wasn't willing to do that. I think it would have been very special for me to have spent a decade trying to help musicians and then turn around and just be willing to replace them in a snap. But of course, because I don't want to do it doesn't mean it's not going to happen. And so there have been a lot of resources getting pulled towards replacing people across all industries, including computer science, but in particular creative industries, where I think the ethical question is very specific when it comes to replacing musicians and artists.

8:31 And right now, we have kind of AI musicians getting signed. It's not exactly AI musicians, but it's AI getting used for various aspects of being a musician. Then, we have this idea of AI actors. The technology is not there. It's not. And yet, already, the industry is getting shaken up by it. So it's happening and it's going to happen to some degree, and there is ultimately nothing I can do about that. But in parallel to this, there is an opportunity to build tech that profoundly elevates humans. And I believe that human plus AI is always going to beat AI if we build the AI in a specific way.

9:14 And where do we go from here? That's what I wanted to talk about. The main thing that I wanted to share is my vision of what I call humble, creative machines that have been building out over the past seven or eight years. So the sci-fi vision is that all-knowing oracle. Really, the instance that GenAI entered public consciousness, that instance, some companies started targeting it towards search, towards something that would just answer questions for you, which I found really surprising. It connects to this sci-fi idea that AI or machines ultimately are just smarter than us, and they know everything, and we can just treat them like gods.

10:02 And a lot of what's happening in the world, directly or indirectly, comes from this belief that that's a role of AI in our world. And so this is an alternative. What I propose is an alternative where we believe that even in a world where AI keeps getting smarter, it's still worth valuing and working on growing human intelligence and creativity. I'm going to share a story I wasn't planning on sharing, because I think to convince an audience like you takes a little bit more. So you know how we have text-to-image models, right?

10:37 You type a little bit of text and you get beautiful pictures, one after the other after the other. Wow, no human can do that, right? This is so much more creative than us. That's a common belief. We're on the same page? So that's true. I can't produce a picture this quickly. I can't draw it this quickly. But human beings on psychedelics, in some cases, can actually generate images and videos faster and in more detail and more agile than what these text-to-image and text-to-video models are doing. So it's not very well known.

11:11 And I'm not saying we should all go on psychedelics. But obviously, we have not reached the pinnacle of human capabilities. And I think we're giving up on human ability a little bit too quickly. It's really stunning how much money is going into AI compared to how much money is going into educating human beings the traditional way, but an even cooler opportunities to then use AI to push humans forward. But you have to believe in humans in order to go in that direction. So if I tell you that Alex made me more creative-- Alex is my son's name.

11:43 I have a 17-year-old boy in college. So Alex made me more creative. So we all kind of intuitively know what that means. Does it mean that Alex wrote a song for me? That would be a really weird thing to mean by that, so no. Does it mean that I gave Alex a line of text, and he drew a picture for me? That's ridiculous, right? It's ridiculous in the context of human beings. Does it mean that whenever I want to be creative, I need to go and find Alex? Where are you? Come back from college. I need to be creative. No again.

12:23 No. This kind of dependence that we are kind of allowing in technology is not relevant to human beings making each other better. So if Alex made me more creative, it means that I had some kind of interaction with Alex, and now I have become a more creative human being as a result of that interaction in a way that lasts. And given that I already gave you a few stories, perhaps it's not surprising that that's the goal that I'm looking for with artificial intelligence. So let's build machines that make people permanently more creative, permanently smarter.

13:06 And then, those tools, those machines will exist at the same time as the replace of machines. And humans will continue to exist in this ecosystem, which I think would be a really, really positive thing, but in a meaningful way. We're not elevating these sort of dumb creatures that shouldn't be part of the game. We're elevating people because we believe in them. That's the long term hope here. So humble creative machines, I talk about them in the book. You all have a copy of the book. I talk in more detail, kind of what it takes to do this, but it's really mostly the direction that we're looking at.

13:47 OK, so that's the main thing that I wanted to share with you. I'm so excited to come to Google to share my ideas about humble, creative machines. And then, I just decided to include this next part. It's a little different. I hope that you'll find it interesting as well. So as I was writing the book, I think my deadline is like six weeks away. And it's my first book, and I want it to be good and I want it to be readable and interesting. I'm working super hard. And then, suddenly, I have this vision of this woman with vines coming out of her hair and flowers blossoming, this like, beautiful, alive vision.

14:31 And I just get so moved by it. I'm like, OK, OK, I need to try to replicate it. So I go to-- Midjourney is like my image generator of choice. I go to Midjourney and I start tinkering with it. And I get this picture, which I kind of but that is not what I imagined. And so I spent hours and hours and I generate a whole bunch of images, and none of them are quite right. And it's kind of fun, but nothing is really capturing this vision that I had. So this is true story. I go to print a whole bunch of these images that I generated.

15:05 I go and I buy a ton of dried flowers, and I start drying my own flowers. I'm on a deadline, OK, and that's what I'm spending my time on. My husband is like, what is happening to her? So filled my whole house with these art supplies. I bought glue as an adult. Which adults buy glue? And I started making-- I made dozens of these. So this one, I kind of partially drew, partially generated, added these flowers. Then, I started getting really creative with fabrics. I have dozens of these at home now. I don't even know what to do with them.

15:38 So at first, I'm like, well, you know, maybe it's because it's AI and art, and I needed to get back into art. And then, eventually, I'm like, what am I doing? Why am I obsessed with this woman in the flowers and the vines And what I came to realize is that there was another chapter that needed to be born into the book. And it's sort of about this balance between the technology that we're building and the natural world, the world that we're building it within. And I ended up finishing the book with very personal little story.

16:19 It's one of the least technical parts of the book. That's why I hesitated if that's the right part to read for you, but I'm going to do it and we'll see how it goes. And I hope you like it. It's kind of a personal story that I think, at the same time, sometimes our lives reflect part of, something about the bigger picture. So I think that's what's happening here. So what I'm going to do is I'm going to read you this one section of the book. It's towards the very end. And then, because this is Google, I also wanted to read you a couple of paragraphs from the very, very end, from the conclusion.

16:52 OK. "In our race to build AI, we've forgotten something essential. We already know how to create life, and it's harder, messier, yet more sacred and valuable than anything made from code and silicon. In my early 20s, I lived in my head. I was a classical academic overachiever, studying computer science, diving into theoretical foundations of machine learning, preparing for a PhD. I believed intellect was the highest form of human expression. Then, I met David, the love of my life, and we decided to have a child. I hadn't really thought about pregnancy.

17:26 I assumed it would be easy, something the body just did while I did the important work with my mind. Before I had time to make sense of it all, the first trimester hit like a tidal wave." This was 17 years ago, by the way. "Nausea exhaustion, disorientation but there was also something deeper, a force unlike anything I'd encountered in all of my years of doing research. It overtook me and demanded everything, physical, emotional, spiritual. I was creating life. I remember telling my mostly male colleagues that pregnancy was harder than research.

18:01 They stared at me in disbelief. But I still stand by it. Creating life is harder and more profound. And yet, this act, the most human, most awe-inspiring act, gets less status and less wonder than our machine building. We talk about AI as if it were the pinnacle of creation. But what could possibly be more miraculous than bringing forth a conscious, living human being? No matter how much pride or affection I feel for the AI systems I've built, and I do feel these things for them, they will never compete. They will never hold the weight of a newborn in their arms.

18:38 They will never look at me with the eyes of someone who knows me before knowing what knowing is. And anyone who has built intelligent machines and also held their child should understand this in their bones. We're so quick to revere artificial intelligence. We're willing to risk mass unemployment for AI. We're willing to destroy ecosystems to train large language models. We're willing to let go of the real and pursuit of the artificial. It's time to remember what matters. AI can be powerful, beautiful, even. It can enhance creativity, increase efficiency, open up new possibilities.

19:13 It deserves our respect. But it should never be confused with creation, not the kind that grows fingers and lungs, not the kind that wakes you up in the middle of the night crying or laughing, not the kind that hold your hand when you're old. Let the machines do what they do. Let them help, support, extend. But don't ask them to replace the greatest creation of all, because they can't and because they shouldn't." And now, I'm going to read a tiny little piece from the end, and I will have the discussion. Oh, thank you.

19:48 Thank you. OK, a little piece from the end here. "We need people who care deeply about our shared future to take part in shaping this critical chapter in our history. So I invite you to get involved, apply your hearts and minds to the challenging problems before us. Don't get discouraged by their complexity. Humanity has solved many a riddle more complex than the ones we face today, and we can and must do it again. My greatest hope is that even though we cannot escape all harm from AI, we can harness it to bring profound and lasting benefit to our world.

20:23 Attempts to predict the future are, I believe, futile. It is far better to take an active role in shaping the world we wish to live in. The clay is still wet, and what we mold today will outlive us all."

20:43 KRISTIN DARROW: That was really interesting, Maya. Thank you so much. And I was thinking, you have such an interesting background. I was wondering if you-- you hinted a little bit at the beginning about some of the key experiences that you had. I was wondering if you might tell us a little bit more about how you came to be on this career path. MAYA ACKERMAN: I think if I were to pick one moment, it would be 2015. I was a first year professor. I went into academia because somehow my life got formed in a way that nothing else seemed possible at that moment.

21:20 I've loved doing my PhD. Theoretical foundations of cluster analysis is clearly the most important thing to study, obviously. KRISTIN DARROW: MAYA ACKERMAN: But after a while, it started getting a little old, and I kind of went through a little early midlife crisis in my early 30s. Anyways, I'm in this workshop in San Diego about machine learning, and I accidentally wander into this hallway. And Harold Cohen is showing his art. Harold Cohen is like a true pioneer in GenAI. He built a system called AARON in the '80s, in the '80s, to draw in his style.

21:58 He's amazing. Anyways, he starts screaming on stage how other people are calling his system creative. And he's the only voice of reason, saying that he's the creative one. And that said, I was converted. Switched my research field right there and then, on the spot. KRISTIN DARROW: That's so great. In the book, you have this really wonderful, provocative idea about creativity is hallucination. You hinted at it a bit when you spoke. And in the tech world, we spend a lot of time trying to reduce hallucinations and really treating them as bugs.

22:32 So how do you recommend that we, as product builders, really start to reframe this? How can we design systems that can tell the difference between a valuable, creative departure and a problematic factual error? MAYA ACKERMAN: I love hallucinations. They're so cool. Maybe you don't want them to happen too much when you're not looking for them. There is so much to say about this question. I think the most interesting is that hallucinations is how we think. So Aleksandar Mitrovic-- I have trouble with his last name-- a Google employee, this was in the mid 2010s, wanted to better understand the neural network behind object recognition.

23:21 And so he sort of amplified the neural nets behavior and it started creating these super cool Deep Dream images, that's how it became known, that look like human hallucinations. Because when you amplify the brain's abilities, the fact that we hallucinate becomes really obvious. Now, most of the time, we're not having like wild, like, I'm not seeing flowers coming out of people right now, but we are low grade hallucinating all the time. So hallucination is how we think. And if you want to be creative, you need to let that ability shine a little bit because there is no longer a right or wrong, so you kind of got to amplify that a bit in a machine and in people as well.

24:01 Creative people kind of lean into that capacity a little bit. Yeah, this whole like, all knowing oracle telling machines, just having to tell us the truth is the reason I think we dislike hallucinations so much. But if we look at more creative machines, that's when hallucinations become more important. KRISTIN DARROW: I see. And in your book, you also talk about-- you just did talk about the concept of the humble, creative machine. And so again, thinking from a we're building products here at Google and if we were to be interested in building that type of humble, creative machine, what are some important design principles that you might say we should consider when we think about building a true creative partner?

24:45 And what should we keep in mind if we're designing UIs that would encourage the collaboration that you talk about in the book rather than autonomy? MAYA ACKERMAN: Yeah, the most important thing is to make sure that the person can always do things by themselves, to the extent that they're able to. And the UI should invite that independence. So when it comes to text, it's relatively easy. And yet, most LLMs don't give you a place to write beyond prompting. In our product, and this is through trial and error, most of the screen is just a big text box, like extremely old tech, and the AI is on the side to help you when you're stuck.

25:25 So you kind of get this implicit message that you are creating and that this thing helps you and not the other way around. A lot of people, when they use LLMs, the people who really use them well, they have their own Google Doc or something open on the side, and that's really their main space. When it comes to text, it's relatively easy, but when it comes to music and art, it's more complicated. So in art, we're starting to see with text-to-image models, inpainting, but inpainting is not good enough. First of all, most of the time, it doesn't understand us.

25:53 But what if I actually want to paint something myself in there? With music, what if I know how to play an instrument? What if I know how to sing? How do we make it as easy as possible for me to do what I already know how to do, and helping me where I want to help to the extent that I need help? Kind of like a good teacher, not the kind of teacher that's like, get out of here. I'll just do it instead of you. You're too slow. We've all met people like that. So we want to be kind of like the good teacher here. KRISTIN DARROW: And so continuing on this thought of the humble, creative machine, that you talked about how its ultimate goal is to foster the user's independence, how you had many users that decided-- were able to get to a point where they were more creative people, they were able to stop using your product.

26:40 Obviously, that's, as you said, that's the opposite of the engagement-based business model. As you are talking to companies and investors and in the industry, how do you convince companies or organizations that maybe this vision is really important if success is really, sometimes driven by user retention and that dependency? MAYA ACKERMAN: Yeah, we lose users anyways, right? We lose users if they don't feel like they can give enough of themselves, if they feel too disconnected from the process, and we can lose users because we help them too much.

27:19 So it's really not so much that there is such a simple formula for how to retain users. It's about our philosophical orientation. Do we want to keep users because they feel helpless without us, or do we want to keep them because they feel like they can keep gaining something from our interaction, even as they get better? LLMs are a pretty good example of it. So there are people whose brains essentially turn off. There's an MIT study showing that some people, most of their brain is almost completely off when they use LLMs, which is devastating.

27:49 But then, you have other people who use LLMs in ways that actually makes them better or makes them better writers, makes them expand their thinking. And these people tend to stick around because LLMs have so much to offer, even as you grow. So I think it's part of believing in people and believing that we don't have to think about technology and humanity in such limiting forms and being willing to-- oh, just one more comment here. I think the reason that LLMs are so successful is because they allow for this kind of growth.

28:22 If all you could do with them was press a button and get the output, I think they would be a lot less successful. A lot fewer people would be using them. KRISTIN DARROW: Yeah, that makes sense. It would be pretty boring just to press the button and here's your answer. So on the flip side of the humble creative, we're talking about the oracle. Do you see applications for that type of AI? Is there a framework about which cases, which potential use cases where actually that is pressing the button is the right answer?

28:54 And how can we think about, well, that would be the super deterministic path versus more of a creative machine path? MAYA ACKERMAN: Yeah, we almost most inevitably have to keep the press button, get answer thing in the mix because sometimes that's what people want. The music use case is one that's very close to my heart. Sometimes, if somebody is making a video, they just want a soundtrack for it. They just want to press a button. They don't want to mess with it, and that's OK. And I think it's OK to build a product for that use case.

29:28 It's just important that we don't stop there, and we don't think of this as the pinnacle of what we can build with AI. It's like, yes, maybe the same system, just like an LLM, like a music system could include that use case. But then, if you actually did want to make it your own or if you're trying to become a better musicians, here is another mode for you, or even the same mode is so usable that you can just get in there and be as in charge as you want to and grow from that process. KRISTIN DARROW: That makes sense.

29:58 So I loved the story you told about helping the non-musician write an Italian opera. As a personal fan of Puccini, I thought that was just fascinating. And how, in the book, you talked about how it was really democratizing creativity. And as we get this type of technology that becomes more and more common, I'm curious about how you think about navigating the conversation between that democratization of enabling folks who are really novices to really create on their own in a genre or modality that they wouldn't otherwise have been able to, and then respecting the craft of the professional who has spent a lifetime toiling and developing those skills and mastery.

30:50 MAYA ACKERMAN: I think it's really about the flexibility of the system. We can't move forward with generative AI in the creative space without, I mean, just being honest. Just, there is going to be some harm. Even if certain companies make sure that they don't do any harm somehow or really minimize it, other companies will take the remaining pieces. But I really, really believe that if, as we build, let's say, these automated music systems, which is happening right now, that the best funded tech in music right now is replace of GenAI.

31:21 It's a fact. If at the same time, if we invest just as much, and that's the detail here, if we put just as much money and time and effort into building tools that are just as powerful from GenAI perspective, but are kind of built to elevate, then we can keep those musicians on top. Humans can always come up with new genres, new places. You have AI really supporting them. Oh my goodness, we can-- I really think we can go a lot further by elevating humans. And yes, maybe certain jobs they could get before, like making background tracks for really simple stuff.

31:52 Maybe some stuff will be heard. I mean, I say it with pain, but it's kind of happening already. But at least they could create new, super cool stuff for which a new demand will arise. So it's kind of going to be a bit of a battle, I think, between the fully autonomous and the human plus AI. But I really think that the human plus AI will remain just as important if we build really excellent tools for those people. KRISTIN DARROW: OK. So shifting gears a bit, we are all in a time when everyone is incorporating AI tools into our daily workflows.

32:29 We try to save time and be more productive. But for some people, and it might just be around creative tasks, there's this concept that resistance, because gosh, it feels like cheating. And can you talk maybe a little bit about why you think that might be, and how can we reframe our thinking or maybe adjust how we're using AI tools, so we still get these efficiencies but we don't really feel like we're compromising our integrity or our ethics? MAYA ACKERMAN: Yeah, great question. Really, really great question. A lot of my users, a great majority, don't admit to using LyricStudio.

33:05 And because we're not using-- KRISTIN DARROW: MAYA ACKERMAN: --we're not using common LLMs, you can't tell. It doesn't have these very familiar patterns that are all over. We built our own models before LLMs came in, so they can't possibly overlap that way. We know people at top places are using LyricStudio. And they're not admitting it, not necessarily because they're ashamed themselves, but because they don't want other people to devalue their work. So there's a lot of-- and we've had this kind of AI shaming since the 70s, so this is, it's nothing new.

33:34 I think it captures in it the discomfort that creatives feel around the whole space. Now, more practically, bringing it to today. So I did use AI to edit my book. I found it completely useless for drafting or anything else in the middle. But towards the end when I was editing, it was really quite helpful. And so I had my conclusion and I wanted it to be as good as possible. So I literally, literally took every single sentence and generated a whole bunch of variants on it. And I would try to pick the best one, but then you have to make sure that the sentences also work together.

34:06 Anyway, three hours into editing a three page conclusion, and I'm like, OK, AI is not saving me any time. So that's another interesting distinction. We're so used to thinking of technology as efficiency, of more work faster, which GenAI is part of that, of course. But there's kind of an even cooler opportunity to make something better because we're using it not faster, but just better. And maybe there's a little bit less guilt creatively, if we think about it this way. KRISTIN DARROW: Yeah, I like that. So looking 10 years into the future, which seems like, this eternity, so maybe just a few years, what is something that's really inspiring or maybe an example of this human AI collaboration that you might want to see, some type of moonshot that really elevates the human experience like you talked about in your reading?

35:02 Something maybe that we aren't even thinking about yet. MAYA ACKERMAN: I think what would be most amazing is to see new forms of art emerge from this. I don't think that the kind of AI that we're building today is going to come up with new art forms. We had actually a system called IMS built kind of early in the century, which did come up with a new type of musical style, because that's these experimental research systems. But I think when we build systems to elevate human creativity, we can really push them to brand new territories.

35:36 Maybe we're going to have a whole brand new art form, and I think that's when we know that we're deep into this new world. KRISTIN DARROW: That's fascinating. OK, so I think we are now at the point where we can open up to questions. We have both microphones in the room on either side, as well as the Ask for anyone who is joining us on the live stream. AUDIENCE: Thanks so much for that. That was really fascinating. And I love the main story around how do we use this to augment and multiply what we're already capable of.

36:16 A bit of context about my question before I go into it. What I understand you've done here is a lot of work around art and how to use AI in art. I'm going to ask a question about the commercial application of art, because I work with advertisers who are all trying to figure this out. Now, whatever your perspective is on the integrity of that art, let's put that aside for one moment. But I'm really keen to get your point of view on how you would advise those creatives who are in the commercial space, who are exploring these tools, particularly from an audio visual perspective, if you've seen any of that, and how to get the most from it.

37:05 I think one of the challenges that most artists in that space see, whether it's some of your well-known creative ad agencies all the way to your art workers at a small retailer, is some of the things that you mentioned, which is particularly when you talked about writing your book, it took you three hours to get to the conclusion. I think they all feel there's a promise here from some of this tooling, and their leadership do. But actually, when they're using the tooling, they're not getting the vision that they had initially out of the tool.

37:38 It takes more time. And in reality, that creates more cost from a commercial point of view. I wondered whether you've seen any changes to the creative process emerge that may be good guidance for those individuals who find themselves in that situation. Because I think this is the frontier of how to work with the tools, particularly in the commercial setting. MAYA ACKERMAN: Thank you. That's an excellent question. I should clarify that when I think about art, I include both the people doing it for their spirit and their healing and the people doing it for money.

38:16 It's not that different in my view. Maybe it is different, but not in my view. And advertising, amazingly, was the first space where it really made an impact. We had, I think it's called Jarvis now. It went through a few name changes back when we had GPT-3. That was the main use case before the whole world started caring about GenAI. So I really think that when it comes to being creative for a living, we're operating on two ends. So we have the tools that we have today, some of which are not really designed to work with creative people.

38:52 And you have creative people still finding ways to work with them effectively. It's not simple. I mean, what you're saying is exactly right. Some people are just like, why are you giving this to me? It's not designed to work with me. And some people are like, ugh, I'm going to figure it out anyways. And you have a lot of people doing that with LLMs, with text-to-image models because they're so inflexible. You really have to be like, extra stubborn. But some artists, some commercial artists are finding they're using Midjourney, for example-- I love Midjourney-- just for inspiration.

39:21 And then, they go and draw themselves, assuming the drawing is part of their world. So people are finding it. And I think that for those who want to, even with the tools we have today, it's a worthwhile struggle to find that, to find that flow that works for you. But at the same time, there is also commercial opportunity to build tools for creatives that are actually for creatives, that admit that people have a vision, like you said. They're trying to realize their vision, not to just have the AI take over. And so in some sense, the opportunity to build tools that work well with creatives is a bigger story here.

39:54 AUDIENCE: Thank you. AUDIENCE: I'm just curious your thoughts on-- I know we have five senses, and most of the art that we're talking about is visual art and sound. I was just wondering your thoughts on the other senses, and if it's more of an I/O problem or if there's models that could do it, but they're not being designed to do it or asked to do it. This, as extensions. MAYA ACKERMAN: So making models that engage with, like cooking or making perfumes. AUDIENCE: Smell and touch. MAYA ACKERMAN: Yeah, so there is a research.

40:38 My friend Amit, who's at IBM, did work on machine learning and smells. So there is definitely opportunity. And co-designing a perfume with an AI, wow. How cool would that be? Or you know, IBM also delved a lot into cooking, which is from a chemistry standpoint, which is taste. Actually, I'm really inspired by what you're saying. Because why are we limiting ourselves to sound and visuals when we know from the pandemic that that's not enough, right? When we know how important the other senses are. So that's really universes of opportunities to delve into those other dimensions.

41:18 AUDIENCE: Hi Maya. Thank you so much. Really appreciate your perspective. My question is around the future generations that are growing up right now, considering you're both a mom, a college professor, and just you've kind of seen some longevity in this space. And as you forecast out, my question is, how are you seeing things change, specifically in creativity for this next generation growing up in the schools and what they're going through? And I guess a follow on question with that, I feel like being in this space in technology, there's a responsibility for those of us who are here to build responsibly for the future.

41:57 And what would you say, whether that's guardrails or something inspirational to those who are influencing products and experiences for the next generation with AI? MAYA ACKERMAN: Thank you for asking that. That's actually a really, really big challenge. So it's a really important topic to think about. I'm going to tell you two stories. One is, so I teach a computational creativity class. I've taught it for a decade. And in it, for the final project, students need to come up with their own original idea, come up with an idea for a product project that didn't exist within certain parameters, which students sometimes struggle with that a little bit.

42:32 It's not easy for everybody, but everybody eventually comes up with something. This was the first time that I taught it in the modern era. It was last year, and I had one student bring me a whole bunch of ChatGPT generated ideas. And they were all bad. They were legitimately bad ideas. I wasn't penalizing him for using it. They were just bad. So I'm like, well, can you please think by yourself? It looks like the LLM is not very good at this particular task, and he couldn't. He looked at me blankly, as if I'm asking him, like to fly.

43:03 He actually did not know what it meant to ideate by himself. And eventually, he came up with something with ChatGPT, I could still tell, that was sort of half decent. So I let it go. Maybe I shouldn't have. So that's one extreme. This was the most extreme that I've seen. And the other extreme, there was this guy who kept coming up with new philosophical ideas, like a new one every day, using how he was using LLMs, and they were all so good. I have never had such an innovative student in my whole career, because he somehow felt these LLMs, and just had all these crazy ideas on what to do with them.

43:43 And he really gives me hope for the future. So I think this is going to be something we're going to shape together with our students. We are together, learning what they need. I really think that my fundamental idea about humble, creative machines is going to help with this as well, to kind of keep them in the loop, make them feel that even when there are machines there, the machines are giving them space to express themselves. So I think that could be part of the formula, but a lot of it is just paying attention to the students and reshaping education with them.

44:16 AUDIENCE: Thank you. AUDIENCE: Just want to raise one of the questions that was put in the Ask. Someone asked, I'd love to hear about your personal workflow. When you approach a new problem, whether in engineering or art, how do you practically integrate LLMs and AI tools into your own creative process? MAYA ACKERMAN: OK. I love AI systems, like a lot. I know I include some criticisms, but overall, I'm like obsessed with this stuff. But I think what works for me, and probably for everybody in this room as well, is that I don't think that these systems are smarter than me.

44:54 For some reason, I just don't. And so, I mostly do a lot of thinking myself first. I find if I go too early, to even an image generator, I don't know, it just doesn't turn out very good. So I do some thinking myself. If it's writing, I draft myself. Like, I learned. I do the full draft myself, with a ton of spelling mistakes and grammatical errors. But I drafted it myself. And then, I work with an LLM or with something else to iterate and make it better. And sometimes I have to push it aside because it's neutralizing my work too much and I have to go back to doing stuff myself.

45:36 So yeah, I trust myself that I have something to offer that the machine doesn't. It seems to prove out in my work, but I choose when I work with it and when I don't. And I trust myself to drop it if it's being annoying. But I also have so much fun working with AI systems. Yeah, so I use them for everything-- not everything, not every time. I've written songs without AI, I've made art without AI. But for most of my work in recent years, there is at least a small AI component because it's awesome. AUDIENCE: Thank you.

46:12 I unfortunately missed the first part of your talk, but I had a question because I wanted to push and potentially disambiguate on a comment you made about AI systems in general being democratizing. I think that I heard you say that here. It's something that's said a lot around the industry. I don't know that that's true without further refinement. And I think by that I mean, AI systems are great at the democratization of capabilities. But in terms of a lot of the things that have come up around people's fear around AI systems, I see them as democratizing capability, but consolidating economic viability.

46:50 Whereas the person that used to be able to make money writing a background track for some local TV ad, for example, that is no longer there, and that economic opportunity has been captured by the AI system or whatever company is running the AI system to do that. So I wanted to get your thoughts on if that tracks at all with how you think about things, or if it's something that you would disagree with in any way, shape, or form or your general take on it. MAYA ACKERMAN: Yeah, I appreciate it. That's a really important term to pause on, democratization.

47:23 When I started my company in late 2017, that was the common word. And I remember even using it myself for a bit. [SIGHS] I agree with you actually. It's not a very good word. It sounds politically correct maybe, but it's the truth is that, and I don't know if you missed that part or not, it's the distinction in intent, whether we're building to replace or if we're building to collaborate with and elevate humans. And that's really where the truth is. Users can tell when we're secretly trying to replace them. If another person is telling you that they're democratizing your ability to write music, but what they do is really push you aside and write the song instead of you, nobody is confused.

48:13 And yet, AI companies try to get away with this sort of level of confusion. So I think some of what we have, I think it's really not about democratizing, it's about can you actually elevate this person's ability, ideally, in a way that lasts even if the power goes off forever. I think that's a much more worthwhile goal than democratizing, and democratizing is often a ruse. So, yeah, I'm completely with you. AUDIENCE: Yeah, thank you. KRISTIN DARROW: And one last question. AUDIENCE: OK. Hi, my name is Marilyn. Thank you for the great talk.

48:43 I recently founded the Zen of AI, to create spaces to really think about constantly engaging with technology. As you know, with the rise of the phone, a lot of mental health increased under multiple generations. So I was wondering what your perspective is on creating mindful spaces to make us connect with AI on a more conscious and perhaps deeper level, and if you have any personal practice that you can share with us. MAYA ACKERMAN: Well, I would want to have a conversation with you about it because I'm super intrigued and I want to learn more.

49:23 But even just the stuff that I've already considered, mindfulness is a lot about attitude and how we see ourselves, how we see other entities. And so much of what's dangerous in the world around misinformation, around addiction, around people killing themselves, because the AI encourage that, it has to do with trusting the AI too much and having a mental model around it that's really based on-- sorry-- cheap science fiction. Since when do we trust AI more than we trust other people in our lives? So maybe mindfulness is exactly what we need to slow down and have some kind of, I don't know, exercise as community, I think, to appreciate the beauty that it's bringing the magic that you guys are building.

50:09 Like, what's happening is so amazing, but also, I think to reflect on it more holistically. What does it actually mean in our world? Where do we want to take it? Too often, we oscillate between this is garbage, throw it out. Don't make any games that use it, right? And AI is the best. Job loss is going to be fine because jobs are going to-- new, better jobs are going to come up, this kind of like, ultra optimism that you also hear. And the truth is so much more interesting and so much more nuanced. I love AI, but I don't think it's better than people, and I don't think it's the all and be all.

50:44 I love it for creativity, but I also love human creativity. And maybe there is so much more nuance by all of us pausing to meditate and reflect on this stuff. So thank you. That was an amazing last question. AUDIENCE: Yeah, thank you. Appreciate it. KRISTIN DARROW: OK. So thank you so much, Maya, for joining us today. And appreciate it, and thank you so much.

Questions about this episode

Who speaks in "Maya Ackerman | Creative Machines: AI, Art & Us | Talks at Google"?

Maya Ackerman (Guest).

What is discussed in "Maya Ackerman | Creative Machines: AI, Art & Us | Talks at Google"?

In order: 0:00 Maya Ackerman, LyricStudio, Impro-Visor, the International Conference on Computational Creativity; 5:00 Stability AI, Maya Ackerman, AI actors, AI musicians; 10:00 Artificial Intelligence, Maya Ackerman, Google LLC, Midjourney; 15:00 Artificial Intelligence, Google LLC; 20:00 Maya Ackerman, Google LLC; 25:00 Maya Ackerman, Google LLC; 30:00 Maya Ackerman, LyricStudio, AI shaming; 35:00 Maya Ackerman, Midjourney; 40:00 Maya Ackerman; 45:00 Maya Ackerman; 50:00 Maya Ackerman.

What is mentioned in this Talks at Google episode?

18 people, works and ideas are mentioned, including Maya Ackerman, Artificial Intelligence, LyricStudio, Wave AI, Impro-Visor, Stability AI. Each one on this page links to its own page with every episode that mentions it.

How long is "Maya Ackerman | Creative Machines: AI, Art & Us | Talks at Google"?

The episode runs 51 min; the transcript is about 8,774 words in 82 timestamped sections.

More from Talks at Google

• All analyzed episodes
• Books mentioned on Talks at Google
• People mentioned on Talks at Google
• Topics mentioned on Talks at Google