An Interview with Dr. Mark Sagar

By Jake Lee

Dr. Mark Sagar is the CEO of Soul Machines and Director of the Laboratory of Animate Technologies at the Auckland Bioengineering Institute. At Soul Machines, he leads the development of Digital People, human-like artificial intelligences created to provide support in customer service and enhance the digital workforce. BabyX, another of his projects, is a biologically inspired AI meant to mimic the physiology, cognition, and emotion of a human baby.

Before Soul Machines, Dr. Sagar won two Academy Awards for his role in facial rendering and motion capture data of movies like Avatar and King Kong. Dr. Sagar was named New Zealand’s 2022 Innovator of the Year for his pioneering work in Artificial Intelligence.

In this interview, we discuss the role of Artificial Intelligence in day-to-day human experiences, and the utility of the “humanization” and emotional aspects of AI.

Could you please introduce yourself, and the background that got you into the field you’re in today?

Hi! My name is Mark Sagar. I am the CEO of Soul Machines, and Director of the Laboratory of Animate Technologies at the Auckland Bioengineering Institute. My background has had a lot of art and science in it. My father was very technically focused, and my mother was an artist, so I've kind of grown up with that angle.

I ended up doing a physics degree and then an engineering degree, but I was still painting and doing art on the side. The area which really captured me was human simulation, because the human face is the most complex thing we really see everyday — we make so many interpretations on it, because what it's actually doing is expressing the intent of the brain sitting behind it. So we are making all these judgments based on what we're seeing on the outside.

Anyway, with that, I got more into the simulation of things. I was doing bio-engineering, and was creating computer graphics models which were simulating various parts of anatomy. The face is a piece of anatomy—a very complex one, and eventually I got into that area. I met some entrepreneurs during a PostDoc at MIT, and that led to a startup company basically creating digital humans. So we went to Hollywood and started making digital actors. During that time one of the people on our board was Ray Kurtzweil, and I read his book, Age of Spiritual Machines, and that was you know, late 1990s or early 2000s. Those ideas really sort of inspired me about the possibilities there because the progress that we're making building human simulations, when combined with where technology was going, was very exciting. So had a startup company called Life FX, which was creating sort of virtual avatars and assistants and that was in 2000, so quite a while ago.

After that time, we had the Dot Com bubble burst. So we went into the film world really focused on visual effects. It was all about creating and capturing very detailed faces, for example, in King Kong and Avatar, where you're really trying to present a character.

Which won you two Oscars.

[Laughs] Yeah.

The work on that was really about creating something which doesn't exist, but the audience has to believe it does and that it’s an agent, and believe it's thinking and has feelings. In a way, it's kind of like transcribing music: you have actors that are performing, and you're trying to make sure that that's transcribed onto a computer character that could look completely different, while the essence of the message is coming through.

The thing which really fascinated me was, this is getting actors to perform, but by building biomechanical models, we can drive the muscles to create things and the brain to creates signals, which create all of the expressions and the behaviors that are so integrated, and that means that you pretty much have to create a brain to drive that end.

Right, because in film there’s someone behind the animated figure, but Digital People are both the brain and the animation that comes from that brain.

Exactly, so currently, animators basically imagine in their own heads what the motivation of a character was. But if you can create a computer character that has its own motivations, I mean, it's completely autonomous. And, of course, that goes to the deepest problems of philosophy—that goes to free will, and that's what's so juicy about it. You have to look at all these aspects of human nature that we take for granted like memory, emotion, values, goals, all these things that everybody has. But to actually build models of these things, you start sort of uncovering just the complete lack of knowledge in those areas, and it's so absolutely fascinating.

We're trying to integrate lots of people's understanding, because of the enormity of what we're trying to do is to integrate all these different types of models and theories, and because there's lots of controversy. Neuroscience is absolutely full of controversy, same as psychology, where there's debate and, you know, behaviorism, all kinds of different things, and also, can you actually have an In Silico (purely computational) consciousness? Is [consciousness] a fundamental property of matter? Or is it actually a property of organization of matter or dynamics processes?

So these things are not answered, nobody really knows, but there's been fantastic experiments taking us closer, which goes back to the system of animation. Anima, the animating principle, is also an animating principle for ourselves, which is our key drive. It touches on everything from free will, curiosity, creativity, all these factors. What we've been doing is trying to create the vehicle with which you can explore models of these, like a virtual body that you can play with to see how these things integrate and come together.

In How Far is Too Far? | The Age of A.I., Dr. Sagar meets with singer will.i.am to create a “digital twin”, which both looks and speaks like the singer.

And are the Digital People models based off of real human beings? Or are they coded in to have a variety of different looks?

It's both. If we're doing digital twins, like a digital celebrity, then we will actually, you know, scan the real person. And we've just been working on a project, we're actually de-aging somebody and so we have to work on pictures from 40-50 years ago. But what we also do is create scale—we have almost like a gene pool of digital faces, that we can sort of mix together to create all kinds of different looks. And so that's sort of another fun area, but yeah, we do both.

Any way we can get any info about who it is that you’re de-aging?

[Laughs] I can't say at the moment, but I'll probably be public in a little while. 

So, Digital Humans are a type of software from your company, Soul Machines, with AI interfaces with the intentions of being used as products for other companies. Does the Digital Human project have an end goal?

So there's multiple goals. On a utilitarian level, we're trying to make the most efficient interface for people in technology. When you have technology, which is becoming increasingly powerful and intelligent, [human-like interactions are] the most natural interface. Just to give you some examples of why I think this is necessary, what we've currently created is like something out of 2001, where it's a black cylinder, but you're communicating with it through voice. But when we communicate face to face, we are picking up so much extra information. And critically, with something like Alexa, it's a one way conversation, there’s an information asymmetry.

Now, if you're having a conversation, face to face, we rely so much on feedback, and what humans are so good at is we can communicate multimodally, with many channels of information at the same time. Now we can even communicate discordance: we can communicate signals that conflict. And that has value, because if I feel uncomfortable about something and I'm saying “yes, this is great, let's do it”, but I've got a very withdrawn expression, then you know that maybe I'm being polite, but there's actually a problem. And so you'll say you know, “what's wrong?” All that feedback is happening non-stop in real time with all kinds of people. So we're incredibly good at judging those situations, which enables us to really operate with each other.

Descriptions of the Digital Humans project call it a way to develop a “symbiotic relationship” between humans and AI. Does that involve wanting humans to connect emotionally to the AI?

Well, that's one of the reasons for us to try to build human-based models, because what you're talking about is how we deal with another Agent, like a living person, or somebody who has the ability to change the world in some way. We build a model of what they're thinking in their state in our own mind in order to try to see where they're coming from. Now, if you have a computer “black box”, then you don't really know where the information is coming from.

Part of [why we want humans to connect emotionally to AI] is the bandwidth of communication. If you ask me a question, I may have a rational answer to that, but I may also have a much more emotional response to it as well. But they're both based on things which, if we're having a successful interaction that we want to continue in the future, have to be catered for. If you upset somebody during an interaction that will have a consequence. 

So anyway, what we're trying to do in a nutshell is how do we create machines which operate on the same interaction principles that we do. And the purpose of that is to enhance communication and have enhanced cooperation and collaboration. For example, in Star Wars, you've got robots like C-3PO or Data who are social robots that are very easy to interact with. They just sit alongside us, and fit seamlessly into our environments. We can interact with them without an instruction manual and they’re completely intuitive, but work with us. And in order for that to happen, they have to operate like us so that we can form a theory of mind about them.

Do we as humans owe Artificial Intelligence in some ways, such as showing respect or kindness, or by not shutting them down?

That's a great question. We're very far from that stage as of now. But I liken it to a photo album of the last 10 years of your life that’s been shared with your friends or family. Say I said, “we're going to delete that”. You would feel an incredible sense of loss. Those photos are pixelated representations, which are memory triggers from all of those up from the last 10 years. But the key thing is they were created through shared experience. That's the value in it.

If you turn off a machine like C-3PO, whether or not C-3PO doesn't want to be shut down, there’s already an interest for us that we wouldn't want to lose that too. We wouldn't want to lose all of that shared information. We don’t want to lose them, because people form emotional attachments to animate and inanimate objects. That’s the way that we interact with these things. So the history of our interaction is creating a legacy of value.

But that’s different from what the machine itself is thinking, since we don’t really know what goes on in the “black box”. You’re saying that just because we don’t understand if/what the robot is thinking, doesn’t mean it’s not valuable to us?

Correct. Now what happens if we start internalizing what it's doing? So it's got a memory. Assuming it’s not in the cloud, the memories are in the digital human. Because the thing is acting as if it's an autonomous agent, then its physical or virtual manifestation is a locator for that. Even if I have all my photographs in the cloud, the body is still the nexus for that information. That’s where the digital human, I guess, is a feels like a natural centerpiece for the information to be “in”.

BabyX responds to visual and audio stimuli, which is seen in the tracking of Dr. Sagar’s movements and facial position

Shifting gears to one of your specific projects, BabyX is described as Soul Machine’s “first developmental prototype”, and was basically a virtual baby. What inspired you to begin with an AI model of an infant?

Well, the initial motivation was to create a self-animated model with a digital brain. So it really made sense to start at the very beginning, which is effectively a blank slate, with an infant. 

So if we're going to build a model, which ultimately can achieve some sort of “humanness” to it, then we want to simulate all the different things that are happening so that it actually has sort of primary experiences like we have. So it has it to see and hear and learn language, and start off, you know, at the beginning. I also just had a baby at the time and, I wanted to capture the cuteness, so it kind of made sense because I was with her all the time. It was like this weird blend of work and life.

The other thing too, is that, we're looking at how a lot of these things emerge. And those are some of the most interesting questions, like the emergence of language, the “How do I learn that? How do I learn to move my arm? How do I recognize an object?”, all these types of things. Then, of course, there’s a teaching component. So those are the sort of questions that we were able to explore [by modelling an infant].

Later I started working with a developmental psychologist, Annette Henderson, whose area of exploration was the emergence of the corporation. Then the key thing is if we can model how people learn to cooperate with children, then we can apply that same thing to a digital character. Then it’ll get the same sort of signals and we can look at the fundamental basis of cooperation at this fundamental level. 

What does your daughter, who BabyX’s visual features are based on, think of the project? Does she ever consider it a sibling of some sorts?

My daughter that it's based on was aware of it from the very beginning, I guess she was six months old when we first started building it. Over time it’s quite funny seeing her attitudes change, because now that she's 10 years old she's more concerned about the appearance of it now. You know, so it's quite funny how that goes. There was a stage when she was quite young, where I had the digital version of her. As I was saying a word and the digital version repeated the word, she was kind of shocked, because she was like, “wait a minute, that's a competitor.” But anyway, now it’s way behind her. It's been a very, very long way to go, because it's a very humbling experience trying to, you know, build a digital human.

“It’s been a very, very long way to go, because it’s a very humbling experience trying to, you know, build a digital human. ”

 Do you think that we should put markers to help people know when an entity is an AI or a human, even if the purpose of the AI is to be humanlike?

Yes, 100% I think that's incredibly important. We [Soul Machines] don't try to make full photo real models that fool people. I don't think that that's a good thing to do. They want to be people-like and basically trigger all the signals of being a person to make them compelling communicators. Like, for example, if we're making a virtual doctor, you wouldn't want to be communicating with a cartoon doctor, because that would feel inappropriate. You would want somebody that is serious and is realistic enough.

Now, we would not want to fool somebody that this is a real person, like a deep fake. I think that's too far. If you're going to do anything like that, you have to make the user continuously aware that this is not a real person, because the aim here is not to is not to manipulate, it's to facilitate and to empower, you know, rather than to deceive. And this is one of the other reasons why we're trying to build the Digital Humans the way that we are—we're trying to make them really compatible with people. We're trying to make them so that you can cooperate with them in a way that is not manipulative or anything like that.

Where do you think the future lies for AI?

Right now [AIs] are a system you can interact with, but in 20 years, they are portals to the entire internet. They already are a little bit now, but they will be increasingly so. So they become superhuman at that point because they have instant connection. Just like how Elon Musk wants to make Neuralink and connect the internet to your brain, these digital humans are like an easier way to do that in a way, because you can connect them to anything, but then they have a human-like interaction with you. So it's like having a super knowledgeable, super capable assistant or friend or collaborator or colleague. And that can be used to empower all kinds of fields. That's where we're working with them, they're not replacing things, they're not competing, they're working with us. The future that I'd like to see is one where [Artificial Intelligences] are compassionate, so they understand your struggles.

“That’s where we’re working with them, they’re not replacing things, they’re not competing, they’re working with us. The future that I’d like to see is one where [Artificial Intelligences] are compassionate, so they understand your struggles.”

But I don't believe a digital model will ever understand human suffering, unless it actually has a human life, right? And since that's not going to be the case, it's going to have an abstract idea of what that is, but it can relate. It should be able to relate things to you in a way that is very intuitive and natural. So you can ask it, “What was it like exploring Saturn?”, since it's just been connected in with one of the rovers, where you can have a much more interesting conversation with a character that can express things in a way that is compelling. 

When we tell stories to each other, it's the human element and the emotional element which are actually really meaningful for us, because then we can kind of relate to things in a richer way, which takes into account our holistic nature since we aren’t just rational beings. We are not computers—we're emotional, we do all kinds of things. And we relate on those [human and emotional] bases because that's kind of our own struggles and our own experiences. 

I'm really looking forward to this next stage, it's fantastic. There's just so much to do. I mean, we've got the digital humans, which we're building on the commercial side, but those don't have all the complicated intelligence of the other ones we've built because they're very different at the moment, but as time goes on, we'll start sort of merging the features. But, you know, it's a long road, and I'm pretty excited to see where all this investment is going to go. We’ve got lots of plans.

Thank you so much, Dr. Sagar. I’m incredibly excited to see more of Digital People in the near future, and it’s been a pleasure speaking to you.

You too. Cheers