Introduction

So, yeah, the introduction was great, and I realized that some of these presentations might be more technical, but I think two of them will be slightly more technical.

So I like it that I go up front, so I set up a bit some theoretical part of it so that the next speakers can actually work on it.

Speaker background and motivation

So, introducing myself, I'm Luis Pirit Santo. My background is on IT engineering at Tecnico in Lisbon. I also composed for and rehearsed several music groups, and that's what brought me to the world of automatic music composition for my master thesis at Tecnico as well in 2019, when no one was actually so into music and AI, it was back then.

And then I went to CERN. I worked a bit with data engineering. But then I realized that I really wanted to do research. So I came back to Portugal.

I worked at Casa da Moeda as a deep learning researcher for biometrics. And then I really realized that I really wanted to study a specific topic, which is the foundations. So, theoretical foundations of creativity and learning in AI and in general, to be honest. And that's why I started my PhD in Coimbra and Brussels.

More recently, I've been in Tokyo for eight months where I worked a bit on large language models and one specific technique for accelerating large language models called speculative decoding. So if you want also to touch on that later, we can talk about it. it. But yeah.

So another thing is that recently, I became part of the steering committee of the Association for Computational Creativity, which is an international association studying computational creativity, which means creativity in computers. And this is before creative AI.

Defining computational creativity

So yeah, and that's where I want to start my presentation, basically, because I offer for you the definition of what is computational creativity according to the Association of Computational Creativity. And they say, it is to model, simulate, or replicate creativity using a computer to achieve one of several ends.

The field’s three goals

So there's three ends. One, construct a program or computer capable of human -level creativity, so actually having computers being creative,

creative, to better understand human creativity, or to actually have programs that can enhance human creativity, okay, without necessarily being creative by themselves. So all of these are studied inside of this area.

Focusing on human–AI co-creativity

I'm going to focus on the last one. one.

So the last one is the setting where we have one human, one computer, and we want the setting of them together actually being able to produce more than they would be able to produce alone.

So the case, there are several questions which are mainly what does that mean to produce more to be more creative in this case and so that that is hard what does it mean to enhance creativity or what creativity means it's hard to answer that it's just because it's hard to answer doesn't mean we cannot study

the same thing as intelligence right we cannot really define intelligence but everyone here is trying to understand artificial intelligence so several ways ways to do it is just to define it or actually try to understand what creativity means.

Why “creativity” is hard to pin down

And that's what Ana Jordanos and Bill Keller did in 2016. They went through a corpus of descriptions of creativity and just tried to collect what those people referred to as creativity.

And they collected a bunch of things. So it could mean a lot of things.

It can mean active involvement, development, dealing with uncertainty, general intellect, independence, originality, development. So you can see it can mean almost anything, to be honest.

And there's several ways computers can actually enhance this in all these areas, right? Right? But we also know that they can actually make humans worse in all these areas as well. And that's what is hard to balance. And there's techniques to do that, to balance that.

Designing tools that enhance creativity

Human-centered vs. user-centered design

One of the most current proposals or the most accepted one is human -centered design. design. In human -centered design, it's not user -centered design, and there's a difference. Let me just cover a bit on the differences between these two different approaches.

On human -centered design, we identify the right problem to solve before trying to solve anything, while in the user -centered design, we basically have already something and we're trying to to identify what is the best feature to add to our proposal already that there's, or we assume that there's something and we're trying to find the features for that.

Also, it is centered on all people that might be impacted by the software, while this is centered on the user, while in several software doesn't only impact the user,

right, it can impact a lot of people that actually are not using the software, right? Right? So human -centered design includes all of that.

Everything can be a system for them, for someone that actually, it doesn't need to be a digital system. It doesn't need to be a software. It can be anything that solves the problem, while in user -centered design, usually we

go more to mechanical or automatic or digital solutions, and usually human -centered design design tries to not provide a lot of, to disturb a lot of the processes. So small, effective

solutions while the user -centred design usually includes, even though they're trying to make it easy for the user, it can entail some onboarding and some explanation of the user interface. face.

Artists and creative AI: tensions and open questions

So regarding, I'm talking about creativity, and one of the many areas are the arts, right? So, and that's one of the things I explore the most.

So we already know a bit what the artists don't want AI to do. We've seen probably many people complaining about AI just writing, writing, for example, or producing images, or saying something can, can we get some AI to pick plastic out of the ocean? Or do all the robots need to be screenwriters? It's so they don't want AI in their field.

But my question is, how much of this is a reaction to actually the way AI developed and was was put in place. How much did artists want from software before this happened? Can we actually know that now? Can we collect data for that?

What artists wanted from an “almighty AI” (pre-ChatGPT study)

Well, we cannot collect data for that, but we have some data from that, which is a paper that my colleagues and I published published with data just before ChatGPT was known. And it happened, we were like seeing these tools, no one really knew about them, and we were like, okay, what do artists actually want from these tools, from software, actually?

And basically what we did was just we asked several artists in different domains from design, cooking, writers, all kinds of artists and we asked, suppose you have an almighty AI system, what would you ask it to do? What would you like it to do for you?

Basically we didn't implement anything, we didn't want to bottleneck anything, we didn't want to provide a prototype. No, we just provided a WhatsApp group and said, okay, now suppose that this can do whatever you feel like. Talk to it as if it's doing whatever you would like.

And they basically started talking to it and trying to ask things, and then we collected all of that and made some categories of things that artists want or wanted. Maybe some of of them would not appear right now if we asked them.

A taxonomy of artist needs

So back then, we found all of these recorder, generators, completers, instigators, gatherers, variators, analyzers, organizers, operators, mappers, critics, and enablers.

You can read more on this on our paper or ask me.

So all of these needs were things that they wanted. But not all of them appeared in all contexts as well. so some of these needs appeared in other contexts but the the artifact developments so sometimes they wanted these needs in productivity for example

not actually working with their art but other things or social interaction or teaching there's a lot of artists there are also teachers so there's a lot of different contexts of use where we can find all these needs for artists so this

This was my first year of PhD. It was very interesting for me to have a good landscape of what artists wanted.

Toward a mathematical formalization of creativity

But what I really wanted was to actually work on formalization, so mathematical formalization of creativity.

Because I thought, if we can do it for machine learning, and there's several mathematical formulations of machine learning, why wouldn't we be able to do it for creativity?

There are some approaches in computational creativity. By the way, if you like computational creativity, the conference for computational creativity this year is in Coimbra, so you might want to pass by.

And I started working on it, and I'm just going to go through it a bit to see if you like it or not.

I'm not going to get into too much theory, don't worry, but you cannot see properly. that's interesting to know.

Formalizing agents, experience, and models

Basically what I define is mathematically what means to be an agent, what does it mean to actually perceive data, and what does it mean to have an hypothesis or a model of the world based on data, basically. And And I formalize these.

I say there's different agents that take the same experience, what I call an experience, which is just a set of data, a sequence of data, and can output different generators or different hypotheses that can generate actually different data, right? So basically, this means that on the same data, data, agents can actually take different conclusions from the same data.

What teams can learn—and why definitions matter

My theory is based on a theory from the 1990s, and back then what they were able to prove mathematically, which was very interesting, is that there are mathematical sets that can be learned by a team of agents that cannot be learned by a single agent, which is a very interesting theorem. It feels like there's something to a team that makes it stronger.

Recently, this theorem has been disproved, or at least it was not really... It's not that the theorem is wrong, it's the definitions that led to this conclusion, namely, what the team of agents means in this context was not really the best idea of team A of agents. agents.

Creative interference and augmenting potential

So one alternative thing that I want to prove in my PhD thesis, one thing I'm focused is what I call creative interference, which is related to co -creativity. And what I want

to prove is that if an agent has not reached its full potential yet, then there is another other agent that can take what agent A produces and can provide and can generate based on that new things that will make A better or can produce more.

So basically, B can interfere on what A can create. So this is the kind of theorems I'm trying to prove based on my formalization.

basically I'm trying to prove that it's possible mathematically to have augmenting potential of agents. So this means that co -creativity might be possible but there's also negative impacts as we already talked and one

When AI undermines collaboration

thing that concerns me is for example this one this is about collaboration how how computers and automation can actually make us not do something.

Automation complacency: evidence from a braking study

So in this case, the Irokazu Shirado's paper was about two people driving against each other and they would just try to avoid collision, of course.

But when introduced an automatic brake system, they would just trust on the automatic brake system and not do anything. think.

And once they removed the AI, the automatic system, they were not communicating anymore. So they would collide much more often than before, actually. So this is very interesting.

There's several kinds of studies like this where when you introduce AI, suddenly people start to stop collaborating, even when removed the tools.

Creativity, autonomy, and human rights

And this brings me to the idea that, OK, I I don't want AI to take over creativity because I think everyone likes to be creative.

A rights-based perspective on creative work

And this is a paper of my colleague, Alight, where she looks into the Declaration of Human Rights and she tries to understand if there's actually something that we can say in the Declaration of Human Rights

if there's any kind of idea that we can say that creativity should be a human right. So there's several of the points that actually touch on creativity.

activity, so actually automating it and removing it from people could be considered an attack on the human rights, which is a very interesting point of view.

Conclusion

And basically, I want to leave you on this note. Thank you for listening, and if you have any questions, and I will just have five minutes to answer questions, so if you have any other questions that I cannot address, contact me on LinkedIn.

Yeah. Thank you so much. Thank you very much.