Why the next generation of AI must be human-first

Okay, so now that people are coming in from the waiting room, thank you so much for joining us for this virtual event that we're hosting on designing human-centered AI. As always, we have our amazing moderator, Mark Esposito, our chief economist, also Harvard University and Northeastern professor. We're also very excited to hear from Aurelie Jean, chief AI officer at Infra, and Jamie Pike, director of the Reconfigurable Robotics Lab at EPFL. And today we're going to explore how algorithmic transparency and soft robotics can shape AI that's adaptive, accountable, and built for people. So we're very excited to have you all. Please feel free to use the Q&A chatbot on the bottom if you have any questions. Mark will be answering questions throughout the seminar. Similarly, if you want to enable captions for this event, you can do so in the show captions icon on the bottom as well. So again, thank you all so much for coming. MicroOne is super excited to be hosting this AI for Good virtual series, and I'll pass it off to Mark to then continue the session.

Great. Thank you so much, Sydney, and welcome back to our series. And today, particularly happy to have Aurélie and Jaime on the show. Sometimes we do invite people that we know of, but we don't know in person. In my case, I'm pretty lucky that I know both Jaime and Aurélie in person, and It's just an absolute pleasure that we can organize this. You and I really have been doing several conversations when you come to Harvard to do a guest lecture. And, Jaime, we met in an event in Dubai, and then I came and see your lab in Switzerland. So I have been able to see the level of brilliance that is really making kind of your job so unique in the world. So I like, first of all, as these are like lots of people are watching our series also afterwards. If you could like in a nutshell share a little bit of what keeps you busy in your respective fronts, what would you say? And that's our way to introduce your work. Aurelie, I'm gonna start from you since you're a computational scientist. And so therefore you kind of within the radius of what MicroOne has been doing in terms of advancing algorithmic capacity on AI and recruitment. And you've been working with AI from even before AI was becoming so popular, right? So share with us, right, what really is the nature of your work.

Sure.

Thank you, Mark. And hello, everybody. So it's— yeah. So what I'm doing on a daily basis is really to— I'm building algorithmic models, AI models to simulate the phenomenon from the reality in order to make a prediction, to detect weak signals, to make strong analysis, or to generate content, for instance, in any any type of broad range of areas. And so today I'm working, so I'm president and chief AI of InPhra. It's a startup, deep tech, med tech startup that we created a few years ago that started from an independent research project. And we actually tackle the holistic approach to the women health, women preventive health approach by using AI. And especially we target, we tackle two main challenges in women's health. Which is the early detection of breast cancer, as well as the risk evaluation of cardiovascular disease based on the analysis of the same mammogram using machine learning AI models based actually on the paradigm of LLM.

Great. Thanks so much for the early— we're going to go deeper into the conversation about algorithm and in a particular area where you've been writing extensively on fairness, bias is something that I'd love to have your thoughts on. Jamie, look, when I met you in Dubai, we started to talk a lot about the work you're doing at EPFL. For people who are outside of the scope of the science university, EPFL is one of the most advanced science universities in the world. It's one of the top in Europe. It's definitely one of the few in Switzerland that are renowned worldwide since it's not too many universities in the country. So you, you run beside you being a professor, but you run the Center for Reconfigurable Robotics, or the lab, right? Same question I asked you earlier, but in the context of what really keeps you busy, what's your take on that?

So I've had, well, I found my lab 15 years ago when I first joined EPFL. And it was a difficult challenge, not just to figure out what I'm going to work on, you know, to build my lab, but how to call it. And after a long period of time, because they gave me gave me about a week to name my lab. I call it Reconfigurable Robotics Lab. So it's been about 15 years I've been trying to define what does it mean by reconfigurability in robotic hardware and how is it gonna impact our future lives with technology all around us. And in the beginning, it was a huge challenge because robots were meant to do one thing and one thing really well. And they're measured by their efficacy, return of investment, speed, torque, and economically valid. And the concept of reconfigurable robot is a little bit different from there. How do you make robot a single platform, little more adaptable, safer, and resilient and versatile for multiple tasks that we don't know exactly what we're gonna give it immediately or in near-term future? So our lab really looks at how to design and program robots so that they are more adaptable to known and unknown tasks in known and unknown environments. With known and unknown risks. So they usually do not like, do not look like immediately like humanoids or robots that we can think of, like WALL-E, but they take a very different form in the form of wearables, in a form of touchable, squishy, soft balls. But they definitely can be controlled, programmed, and actually help with your daily lives. Wow.

You know, one thing I learned from you, Jemmy, and I'm going to go afterwards, right, a bit deeper into that is the concept soft robotics. I didn't even know that you can call robots soft. So I'm going to engage you with that conversation a bit. And, you know, before I continue with you two ladies, I invite any one of you that have questions to put them in the Q&A part. I will look at that towards the end of our conversation so that we can integrate some of your thinking into our flow. So early look, I have I appreciate this so much, the work you've done on trying to create, I think, a more human look at artificial intelligence from a perspective of what people sometimes like to think. Like, there is this urge and sometimes almost like exasperation of thinking about superintelligence and a technology that will take over. And you have always been very, very good at making sure that people see that we have hyperbolically looking at this, not necessarily considering really the real power of the technology, the way that technology works. So in terms of technology, as internal AI and biases, this is a big conversation. Tell me what is, in your opinion, the major limits that we have in considering that it's not that today the latest models of AI are less biased, right? We might have changed the language. We might talk about hallucination and fabricated data. But this is a problem that we don't seem to be able to eradicate. And it will be great for us to understand What is, in your view, this biases in the sense like introducing people to your way of thinking about this? And where do you see this eventually going moving forward as AI is becoming much more diffuse and accessible?

Yeah, that's a very good point, Mark. And also, we've been also working together for so many years on those topics. You know, it's interesting because I'm very conscious on the power of the words we're using, the element of languages we're using when talking about AI, even the word hallucinations, you know, leads to a certain type of anthropomorphization of the technology, which makes people think that there is a consciousness, that there is a soul behind those machines, which is absolutely not the case. Obviously, it's only a statistical model. And I say only, but which is obviously already difficult to build. And biases are in different areas. They are in how we use technologies, how we talk about technologies, and how we debate about those technologies, which is very important. Important. Then it's about how we are building those technologies, which is literally what I've been working originally on, which is we have to make sure when we build an AI model that we don't introduce any biases from our own cognitive biases in the choices we make when building this algorithm. For instance, the choice of data, the choice of assumptions that we make, or either the choice of scenarios on which we're going to apply this algorithm, etc., etc. Then there is like biases within the dataset that are not directly related to the choice we make, but that are related to how the data already exists around us that is biased because obviously our society is biased as we know. And most of the time we build our AI models on past data, sometimes synthetic data, but still it's always an extrapolation of the past data that we had And obviously there are so many consequences of those AI models that can contain algorithmic biases and that can lead to technology discrimination, which to me is one of the main threats of AI today. I'm a strong believer of the benefits of AI, as you know, I'm working in that area. And I'm fighting against those few people who are always talking about the long-termist apocalyptic vision of AI, which is not going to happen, like as you say, or AI will destroy humanity, AI will blah, blah, blah. The reality is that there are 3 main challenges: technology discrimination, environmental impact, and social impact on the democracy, on the workplace, etc., etc. And those challenges need to be articulated in our conversations in the right way. And by the way, there are already answers and solutions for some of those questions that we should use, some others that we should explore, and some new to build. And so it's very important to understand that. So I'm a strong believer of what we like to call the AI governance that introduce the players in that field to the best practices, the rules of conduct, including all the tests we have to run on those datasets, on those AI models, the best practices in terms of programming, in terms of deploying code to production, or even in terms of which type of information I provide the end user with. To make sure that he and she understand the technology that they're using. And all those things together is important. And we have to have this AI governance vision to make sure that we take advantage of all the benefits of AI while avoiding the threats.

Wow. Aurélie, I always find it fascinating. I'm always thinking eventually when people listen to you, they'll be able to contextualize it for what it is rather than kind of buying into a mystique of what it is. Right. Which is continuing to happen. Interesting on that. And I'll go to Jaime now. Like, as much as we know, we talk about a mystique about AI, robotics has not been immune from the same. If we think about AI and superintelligence, we do think about robotics and kind of the exterminators. We think about— well, I think Boston Dynamics has been able to do a lot of really visual commercialization that looks like— so, you know, I really would say anthropomorphic. I would say the same. They really look like humans. And so we see this conversation going in the direction of the superpowers of robotics, right? In the same way as kind of Horley helped us to see the limitation from a perspective that we are transferring some of these biases in the cognitive processes or in the data. And of course, it keeps on building on the stratification the same. What are some of the limitations in robotics in terms of where is the reality of robotics considering you've been working on this for so long in comparison to kind of the fancy part of it or the fantasy or the a fictitious one?

I can't be, I guess I am biased because my doctoral degree was an anthropomorphic robotic arm and hand. And my question, or my research question was, why can I build a Terminator? So this is completely away from what people want to say, you know, Jamie, you're building dystopia. And then my question is, why can't I? It was not because I can't build dystopia, but I just wanted to see if we can really approach and physically recreate what Terminator could do on a movie. And it was just not physically possible. It may look it, but it won't be able to deliver everything. So there is a reason when you see a movie screen, since you're on the movie metaphor, so if you see a movie screen, the very good-looking actor or actress never really executes all their stunt. If you— unless you're Tom Cruise, I guess. The reason is when the hardware is doing one thing, that hardware can do one thing real well, either acting or looking good, but executing those extremely difficult stunts is not compatible with that same person. And bring that analogy to the robots. What we see from Boston Dynamics is amazing, and many other robots that we see on the video clips are absolutely amazing. But it's because they are very human-like, either they get you transported to this movie scenes of dystopia, well, they're They're gonna, if they can do that, they can definitely come kill us all. But the thing is that often these movies are the best scenario cases where everything is perfectly staged with a very good lighting, with a very good environment where there's no other Bluetooth or Wi-Fi confusions or anything. Obviously technology evolving so fast that we are seeing the humanoids being considered as manual workers, replacing a lot of the workers in the automobile industry or in every other, you know, manual industry. But we are not quite there yet. First of all, one of the— simply put, the batteries are not there yet. The motors are not there. The understanding or understanding how humans work very smartly around the environments is also a study we have to go further or dig deeper with this new advancements in AI technology. What we know from AI or from learn from AI is based on the existing data. And then we can project this pattern, how robots can work in a certain environment with and without risks. But these are the physical environment. We already know how much work that needs going with working with images and languages. And that's in, I don't know how we call it, 2 dimensions, well, it's on the screen, so let's call it 2 dimensions. Imagine you need to get those words or images or sounds to be transferred to an action command. Now you're looking at 4 dimensions, adding time and actual XYZ dimensions. So this additional dimensionality, it's not just proportional to the complexity, this just exponentially brings us up. So even my lab, I can definitely say all the video footage that go out is the few successful cases that shows a proof of concept of ideas, but it's definitely not the representative of what is possible now. But am I saying it's not gonna be possible in the future? Definitely not. But it's not the level that we expect them to do. And I think we are somewhat more, so we may be a bit more gullible and a little more naive because they look like humans. And we've seen a bad scenario in Black Mirror and Terminator and all these dystopian movies. But obviously we need to be aware of what technology can bring in the future.

You know, Gemmy, similar to when I talked to Orly, right? Every time that you talk, and this happened when I met you, when I saw the lab, when we had a conversation, when we want to write together, I learned so much. First, from the way you phrase it in a way that I can actually really relate to that. Which brings me to a kind of commonality between the two conversations is that we tend to have a fascination for what we would like it to be. It's almost feel like divine, but we are measured by the limitation. And I think if I understand this correctly, in the concept of AI is really our cognitive limitation that are largely impacted by the biases that are refining our reality. It's just that when you digitize that, that becomes problematic. And in the robotics part is the fact that we don't consider that the act of a human body dexterity, it has a degree of calculation, right, that is so complex to even think about replicating at the level we would like to see taking over. And so this is something that you both really made it very clear in their respective field. And I really appreciate, Jeremy, early on this. So look, I'll keep on drilling some of this conversation because I think our major problem will be just not enough time to really get as much as we possibly can. So I'll try to go to the next one. So, Orlé, look, one of the things that I guess, regardless of the fact that we feel the limitation of AI, I guess the narrative about AI for good has finally become something you and I can talk about, right? When we designed this series for Micro One, we were calling it AI for Goods. When I started to see the work you were doing, even before the infra was publicly, right? And you and I were talking, I could see that there was an orientation about AI serving important problems. And you mentioned it before, you know, we have to consider that AI is not a neutral technology. It does come with repercussions, but equally AI can be a catalyst for positive transformation. So if you can share your thinking, I always think you, I always felt you were a techno optimist. Although surrounded by, I think, a lot of negative vibes sometimes with people that really want you to see what doesn't exist. How much progress have we made in really taking AI for good to the next level? I'm really intrigued because—

Yeah.

Infr is a good example, but you're exposed to more, right, by the nature of the work you do.

Absolutely. So first of all, as every single optimistic, I consider myself as a realistic person. But, you know, but people think I'm optimistic. But you're right, Mark. And in fact, I mean, You know, it's very challenging today, I have to say. And I said that because I'm currently doing my second fundraising. So I see the challenges of the deformed, approximated vision of AI that people have today. And in fact, you know, I would say that the way people see AI, the way people talk about AI can be a challenge when you really want to use AI for good, when you want to tackle and solve large-scale impactful problem. And I would give you some examples is that today, literally, the world is divided into two groups of people. If I had to do like a model like that, there are, I would say, the vast majority of people who really want to tackle large-scale impactful problems for good, you know, literally the vast majority of people. There are a few people who usually are the strongest, what they have the strongest voice in the public debate, who might not share this vision entirely, who don't think entirely that way, and who might usually propagate some myths about AI, and who actually damage the other group of people who is, again, the vast majority of scientists, engineers who want to do things for good. So first of all, then there is a matter of perspective. I'm, um, I live in the US. I'm originally from France. Two different continents, two different cultures, obviously. And there are other cultures in the world. And I can see that depending on your own perspective, your culture, where you live, where you work, etc., you might not see the good as the same thing. You might not see a company who is not from your country developing and let's say like a generative AI based on their own values, culture, can actually differ from your culture, and then you don't see that as good as it could be. So I would say that for the first time we have a global innovation that literally interferes with our cultural differences, which is the good news because I always say anything that reveals our differences is the way to be more tolerant and to make an effort to understand each other. So I see, I saw in the chat that someone was talking and asking about the race in AI, global race. And in fact, it's true, you know, there is a race. And but we should not see that race as something to, um, that would push us not to invest into the AI for good. But instead invest into the AI that is more profitable in short-term period. So it's a challenge, but I believe, and especially in medicine, but obviously in any other field, including finance, education, transportation, energy, etc., etc., AI can have huge benefits on those fields by transforming them into something more precise, more predictive, more personalized. And this will literally change the paradigm of many, many domains, fields, you know, of expertise. But again, we should be very careful because there are a few voices, usually strong voices, who again, you know, do not align entirely on, you know, with the vision of AI for good, and, and who tend to orient the public debate onto the wrong topics. But this is my personal view.

Yeah, you know, it makes me think that that in many fields, the majority of scientists, they tend to have usually a much more realistic view because that's the nature of what science does.

Yes.

There's only a very small fraction that tends to look at the more sensational side because, of course, it's more monetizable, right? When you're basically making sensation out of that. That's fascinating. So, Gemma, one question that I'd like to explore with you in the robotics field. Look, AI for goods is a conversation we can have. Robotics for good, we don't hear it as often, right? But somehow the work of a reconfigurable robot, especially in the context, for example, of artificial limbs or giving autonomy to people that maybe lost some autonomy, right? When I visited your lab, I saw basically the vest that has an extendable arm. And I would imagine, right, really the idea that technology in general can be a form of elevation. Robotics is a form of the physical limitation we might have, right? So where is this conversation about robotics for good purposes, right, going? And I connected to what I said before. I think the idea of soft robotics was really the idea that, you know, robotics shouldn't just be a rigid structure, right? But more adaptable. Where do you see this eventually going in terms of really seeing an impact of robotics in improving the state of the world? Across the domains that you study.

Right. I think the reason that we don't hear so much about robots for good is because lack of marketing and the fact that it's been very, very slow and comparable to how AI has been advancing. Even the first version of machine learning, if you would call it AI back then, it was from 1960s. But we did not see that large leap of advancement because they did not have as much of data that was cheap, and we were not as connected where we were sharing everything. So it was not advancing as much. Robots existed then, but we didn't call it robots per se. It was called automation. Why did you do automation? Because it was less dangerous and it put people in less of a danger in assembling things, and it gave advance to the industry. So definitely it was good, and that's why everyone invested in it and wanted more things to be automated, but we never really thought it was being good. It was natural advancement of technology. So what does soft robotics have to do with this? This is a new paradigm of understanding robots. Can we bring these automation or technology closer to our daily lives? Because you definitely do not want your Honda Accord manufacturing robotic arm to come into your home. First of all, it won't fit. And second of all, it'll be a lot of threat to you. You won't es, you don't want to be having around you. So the natural progression of the idea from a lot of roboticists was how do you render robots quote unquote softer? So immediate question was maybe the material itself can be softer, maybe we can control it to move softer. And third is any material as hard as diamond can be soft if it's designed thin enough. So how do you make a design conducive to softer movement? So we, the one in the soft robotics community, and it's been very strong following since we had our first soft robotics conference back in 2011. And it's been grown to be one of the biggest community within robotics community. And this is what we look at. How do you make robots more compliant, adaptable, so that it can work better with humans and within our lives in contact or within, without, with a very close proximity?

Well, it's fascinating. I think the premise, is the same, right? If you think about how artificial intelligence was a way of recognizing technology that is statistical and computational, that is going to help us with cognitive processes. What you're mentioning now is how technologies like robots are designed to support us with more physical processes, right? But I think the premise starts from a very good foundation of using the tools around us, right, to support humans' life. And this is where I think we have established this across the conversation that we had. If that's okay with you, generally, I'm going to do one more question that is more specific on me learning from you, and then we'll see how the question in the Q&A are populating. But I, I look forward also to that side of the story early. My kind of next question is related to a conversation that we had about data and kind of— and Jamie equally talked about data before, I guess, data in the concept of integrating more AI and robotics together. And one of the challenges I think we see is that the data, when it's historical, might be very stable, where we're dealing with events that tend to be predictable. And likely there are some of those. But if, for example, you're looking in the medical application, a doctor is trying to eventually diagnose a cancer before it is visible on a mammography. Probably cannot depend entirely on historical data, right? So how do you integrate the fact that we have data which is solid and stable but limited in the application of future fitting, and the fact that today we have more needs and more questions that require a different typology of data? If you can help us understand what you always do in such a great way when I ask you this kind of question in the classes at Harvard.

So it's a very interesting question. So first of all, in medical application, the answer is clinical validation, because the regulation enforces us to do strong clinical validation to demonstrate that the dataset on which we build our model and on which we validate our model is actually a dataset that is relevant, efficient, and accurate. Now, that being said, you're right, there are two things. Because we have to So we still train and calibrate and validate and test our model on historical data, but not only. We have to provide historical data, but are different from the ones that have been used in the past AI models using medical applications. So we actually use different metadata, we use different influential parameters in our model. And so the real difference is which type not on raw data, but structured data or metadata on which we're going to train our model. And actually, I realized that the challenge for us at least was, and you mentioned it at the beginning, how we're going to train the doctors not only to use the AI, because they can use AI, you know, they are pretty smart and it's not, you know, they can adopt the solution easily. The real challenge is How do you train them to learn how to make a decision from new parameters that they've never seen before? So the real challenge is that we are capable to extract from the raw data, historical data that is stable, as you say, some new type of parameters that might not be as stable because we have not as much experience on manipulating this type of metadata. But the biggest challenge, it's to train doctors to learn how to make decisions from those new parameters. And I realized from such a long time, I thought, oh, I need to train doctors to understand AI. Wrong answer. I need to train them to actually understand the new parameters on which they're going to do their diagnostics, pronostics, et cetera. And in fact, I think it's fascinating because it forces you to rethink about the training in grad school, in medical school, you know what I mean? And I always say something that we taught, we doctors have been learning how to treat a disease. Now with AI, we have to train them how to actually maintain health. Which is very different. And this is the preventive medicine. And this is a concept that was initially developed by Dr. Fabrice Brunet, who is the— who was the former director of the Hospital of Montreal. And I love this concept. We should go from the disease to the health. Yeah. Using preventive medicine.

While you were talking, Jeremy wanted to ask you a question.

Oh, sorry. Go ahead, Jeremy.

No, no, Jeremy, please.

I really like how Aurelie went to like what doctors need. I'm currently suffering from shoulder and neck pain a lot, so I've been seeing multiple doctors and we understand orthopedic surgeons have their view on how the bones should be aligned. Neurologists have ideas on how the nerves should be aligned. But I'm a single body and the symptoms are I need painkillers and they're not coming together. And what they need to understand is what Aurelie said. They need to know how to maintain health. And then to do that, it's not database. The database currently just says, oh, this person is in pain. Let me see how I can fix this part and that part. The pain comes from— it doesn't matter where it's painful. It just comes from— it doesn't matter where it's coming from. If you're in pain, you're in pain. So where am I getting with this? If we were to have new raw data that's not based on how they're collected, this will give a better understanding of the position or condition of a person. Currently, these datasets are what? What's your scale of pain from 0 to 10? Look at the frowny face. Is that frowny face number 10 with that red, with like all the yelling out voice, to frowny face is one with a little discomfort. Currently, we cannot understand humans' understanding, or we understand the mechanism of how we understand pain or interactions, but currently it's impossible to communicate because they pretty much assume you're a crybaby or you have a somatic problem to understand. Musculo-neural connection is dismissed, which means you just don't feel pain. So doctors, I don't want to generalize, but that's the sort of bias they come up with. So what do we need? I believe we need a way to communicate sensations digitally that's not based on surveys. And that's one of the things we can achieve with soft robotics. We can understand what is your level of sensitivity by changing the resolution of actions. And we can quantify this. And these are the datasets we never measured before. Somehow we measure this when the doctors are doing robotic surgery. Because doctors, they are basically operating on a patient remotely. Okay, it's a room away. And we do have technology about doing surgeries miles away with robotic technology development with master-slave control. But that's another, besides the point. From that, we understand how we perceive our actions on a patient is completely different when you're doing just open surgery, because how they touch that tools are completely different. So you have to make an abstraction of a doctor's feel in a remote patient using robots. We can make this much more intuitive and tactile. And that's where the soft robots come in. And that's one of the research topics I feel really passionate about. Wow.

You know, and I would say, Max, sorry, I interrupt you just to jump over what Jamie said and to answer deeper into your question, Mark. I believe that robotics are our best ally in this paradigm because with robotics in a general matter, which would be sensors, very sophisticated, advanced IoT, etc., etc., we are going to collect new data that we've never collected before that will allow us to build novel AI models to make the medicine, the health more predictive, personalized, and precise. And this will definitely change the paradigm of medicine. And it's— there is a big challenge, really. And it's fascinating.

I think it is. I think I have to say, you two, Aurélie and Jérémy, you simply get it, right? And you get it from a point of not complacency.

We get the problems, Mark. I'm not so sure I have the solution.

Get the problem. I have to say, the people who understand the problem are able to structure it differently and very likely change the outcome that comes out of that. So I think it's great that you get the problem, Jamie, because I find that a lot of time we get to have that kind of pseudoscience on data that seems to be clean and neat. But as you rightly say, they've been trained with an assumption of what we should be measuring rather than thinking that reality might be necessarily different. And I think in robotics, you are exposed to this combinatorial power of the fact that many things has to occur. While I think in data science, sometimes we go by vertical, right? Because that's how the data looks like. And so all I'm thinking is, is how we are really in front of us starting a significant way of re-challenging the assumption of what we consider to be legacy systems. And I think technology is definitely helping to accelerate the conversation, either because we see the limitation of data when you're trying to apply this in a field like medicine or because you see the limitation of data when you're trying to integrate this into a robotic infrastructure that now needs to be trained differently. And somehow I love where this is going. I hope we will have an opportunity to keep on building on this, like I would say, symphony of the minds, considering that what we share beside everything else is our friendship. So this is great. If that's okay with you, ladies, I see there are at least some questions that are coming. The first one was from, for you, Jaime, from Antonio, which is basically something you kind of answered when we were talking about the limitation. But in terms of domestic humanoids, right, you went to an event in Japan. We all know from the conversation on media that Japan has been advancing robotic nurses and robotic assistants. So we had a feeling that that robots could enter into the homes, right? What's your sense of that? And thank you, Tony, for the question.

They definitely can. It's a matter of what your expectations are. Do you want this domestic robot to do everything or just clean dishes? Or can it do the mowing the lawn? Or do you want it to do babysitting and take your dog out? It depends.

Yes.

So again, robots are really, it's easier and we can see the predictable future of how long it will take to perfect that task if we know a very specific task. The difficulty comes when you want to have a single humanoid because it, and have it do a lot more things. Because it looks like human does not necessarily mean that it can do as many things as human. So if you lower the expectation of this humanoid robot just to do the, clean the dishes, it may come much faster than if you wanted to clean the dishes and play with your, or changing diaper of your child, which probably you don't want to risk in the near-term future.

Wow. Fascinating, right? So it's mainly the nature of the task that we're designing that would define the likelihood for us to use this robot. So of course, Having a humanoid robot that is calling for support if you're feeling unwell because you change your posture, that's one story. But thinking about an all-rounder, Jaime, from what I understand, it is an entirely different set of tasks, right? Again, it puts a lot of emphasis on our thinking rather than thinking that the answer is in the robots, which I find it to be so educational just to hear that. So in this kind of back and forth we have, Nikhil Aurélie is asking you a question. It's actually a fascinating one. So you are somehow collecting data in the medical field. You definitely understand the fact that you want the doctor to start getting trained on the way to react to parameters that never happened before. But it's also true that to some degree there is stability on data, on the biology. There is a lot of data that is still very good that we can use for modeling. Where is this dataset shifting? To what degree do you find or let me phrase it differently. How would you know what we can consider as stable and something that we can rely on versus what we need to integrate in terms of, of a different way of thinking of the data? Is there a balance? How do you define it in the work you're doing with infra? So how do you deal with dataset shifts?

And thank you. Thank you, Nikhil. It's a very interesting question. And I would say, obviously, be or more stable, because if not, we wouldn't be able to deploy it. But the big difference between us in the medical field and, for instance, someone who builds like a recommendation algorithm on social media is that from the regulation, we have to do what we call a post-market follow-up, which means that we need to follow for many years, decades, how the solution is reacting, which enforces us to actually tackle this shift over time. The shift is based on, as Nikhil mentioned, the type of mammograms, also the type of imaging centers where the technology is used, as well as the prevalence. And the thing that is interesting is that we, by developing and deploying AI solutions, we actually influence the prevalence. So we're gonna have to to tune our model over time based on the increase of resolution of the data on the images, the raw images, by actually introducing new type of data as well. We didn't talk about that, but we want to introduce new, more data because we'll be able, thanks to robotics and other technology, we'll be able to get new type of data. And also based on new type of mammograms, based on new types of doctors as well. You know, so there is a shift. There will always be a shift. The big difference between healthcare, medicine, medical application, and other types of applications is that from the regulation, we have to tackle that shift. And I tell you that because, for instance, if you look at recommendation algorithms on social media, there is an amazing— I mean, amazing— there is a massive shift on the dataset. They don't really tackle it all the time, to be honest with you, and they just assume the model to be stable even though society is changing, even though our behaviors are changing. And again, because the algorithm is retrained over time on the set extracted from the use of the solution, in the case of social media, you actually increase the influence of those biases and errors and changes in behaviors. And so the good news is that in my case, because we're in medical application, Thanks to the regulation, we have to tackle that shift.

Really, what I love about your question, first of all, is that not only you address it, there was another question from someone that— this is anonymous here— that was talking about really where do we move from here, right? And I think that your question equally covers that. So I really appreciate that. There are two questions for you, but there's one from Paula that I think is an area we haven't discussed yet. And Paula, thank you for asking this question, and I appreciate your engagement. So, you know, we talk about governance of AI. We're trying to make sure that we are protecting ourselves against a very aggressive or to some degree predatory use of AI that might generate even more challenge in terms of social media, content information. But in terms of adoption of robotics, right, it's actually a question that I will equally ask you, right? What is the framework of adoption? To what degree do we decide what is permissible in the adoption and what is actually still just part of the research lab?

It's definitely done. I don't know if you remember, right before the ethics of AI and the AI for Good committee was formed in the UN in Geneva, like a year before, 6 months before, what they decided was no drone space. They created how the drone should be controlled. So definitely, this is a key, one of the aspects they talk about. But what they understand now is, okay, let's talk about body a bit later because they already opened up the question. AI is moving much faster in advance of robots. Obviously, we cannot really separate body from the brain. And currently AI, as we see how it's propagating in terms of their efficacy and their utilization, goes much, much faster than how the robotic body has been developed. But definitely robots cannot move unless the brain is there. Currently, we show how the machines are used, but at the end, you're gonna tell it to make decisions. We do it already in a way that if you're gonna go to find the best trajectory code from A to B. So these are the level of AI we're using, but it can get a lot more complex. Pick an object and put it in a box, and we don't tell where the box is. And you will start looking for it. And these are the research-level things that we were already doing. But the question of how the robot should be controlled, it's never going to be just about the body, nor just about the brain, which is AI, but it will be a combination of all. So it will be a lot more complex questions to answer. So I think it goes in parallel with how the ethics and then the regulatory measures are taken toward AI.

You know, you make me think about how the promise and kind of the, the idea of a driverless car was flagging around for a long period of time.

Since the '80s.

Since the '80s, right? And we still are. I still drive the car myself and I still find that most of the time that I am driven by, by either a taxi or Uber or whatever, it's still fundamentally a person. And I think it's really where you're mentioning The complexity of defining the accountability of something that is so difficult to regulate from our traditional system of regulation, I think, is just going to be even more difficult when you're adding robotics that has a physical impact with the AI, which is really one of the questions that has emerged from this. Aurelia and Jimmy, you know, we are not far from finishing this. So there's a question. So I have to have this hard job of trying to select the next question, which probably will be the last one. The question from Rahim that I think works for both of you. And so AI in pharmaceutical, AI in, in rewarding pharmaceutical. Let me give you a premise of my understanding. I would imagine early in the context of medicine, there is proximity with the kind of work you see probably in pharma. Love to see what you notice considering you're working with doctors. And in the question with you, Jami, All I can think of is robotics in the context of maybe the production of pharma. But I can also imagine this distributed into pharmacy where rather than having a pharmacy, there could be some form of automated process that is robotic in the sense it's mechanical, right? In this sense. But I'd love to have your sense on this. I think it's a topic where I wouldn't have an answer other than asking to experts like yourself. So Aurélie.

Yeah, sure. So I'm not— I've never worked in pharmaceutical chemistry. But I know a bit about it. And what I can say is that we've been using AI a lot to actually develop new medication, new molecules. And in fact, people don't know about that. But we were talking about AI for good. The COVID vaccine that companies developed and universities developed in collaboration would not, could not, they were, we were able to build those vaccine in such a short period of time because we use actually AI to pretest some molecules and proteins in silico using computer simulations, AI simulations, before testing a smaller set of molecules in a lab. And I tell you that because literally people, you might not remember, but it took only a few months to develop a COVID vaccine. Obviously, it was FDA approved a long time after, but we had a vaccine ready in a few months. In reality, like 10 years ago, we would have needed maybe at least 2 years. And I tell you that because the real, to me, it's the real benefit of AI in pharmaceutical chemistry, pharmaceutical, you know, medication, bill, etc., is really to device new set of molecules, vaccine, medication in a shorter period of time, more efficient, and in a more inclusive way, with a cost-risk-benefit that are optimized.

Great. Fascinating. I didn't know about this part about the molecules. And thanks for—

And we are confirming these formulas that suggested by AI tools. But you cannot have enough manpower to do this all at the same time to test out all these. So whatever that AI shows are most promising molecules, they can be mixed with a robotic arm and mixers and the chemist, because this is a very tedious process of mixing these. So, or mixing and combining those. So this is actually being ongoing research, not from my own lab, but in other labs.

Labs.

So this is actually a very pertinent process in advancing pharmaceutical industry. And this is quite exciting for not only for their ROI, but also people who are suffering from very rare diseases. Normally, these processes take so much time that if it's a disease that's only suffered by very marginal demographic, even though there's technology there, they're never going to be developed. But if the cost of this development of drugs go down because of the AI and the advancement in robotics, then we can also bring this treatment to marginalized populations.

Yeah, ladies, I should be a good referee and make sure that this is coming to an end. It's just that I can't help it but ask— use the other two questions to ask one question combined. And I promise it's the last question. And for anyone listening, guys, I won't be able to take more questions. I want to be respectful of Aurélie and Jemmy's time. So kind of both answers if you can in a short one. First is on your sense of confidence that we're going to get to the level of artificial general intelligence. So this is like super intelligent that we see distributed all over. And I think not from a perspective of what we see in movies, but from a perspective of the development of technology where it's becoming so like a general purpose technology, I guess, if I understand your question correctly, Antonio. And then on the other one, right? So Aurélie, you know, you are familiar with the European system, originally from France, you go back to Paris several times. So you know how the Europeans have a lot of kind of ring-fencing regulation to protect. Jaime, you are in Switzerland, so you are kind of exposed to that too. What's your sense about this regulation? Are they helping you as scientists to move forward? Are they holding you back? So if you can give me a reaction on both.

Okay, we would need like a few hours, but let's do it in 1 minute and I will start if you don't mind, Jami. So general AI, it depends on what you mean by general and the matter of words, very important. If you go back to the meaning of general intelligence, general human intelligence is actually an intelligence that masters all types of human intelligence, including analytical, logical, creative, emotional, and practical. This is all the theory developed in the '80s in the US and Europe on what we call multiple intelligence theory. When we talk about general AI, we don't really refer to the general intelligence that we talk when we talk about cognitive science and neuroscience. We actually talk about the fact that we are going to build an AI model that is general enough, meaning that we're going to train and calibrate this model on a set of data. But this model will be able to solve a problem on which it was not initially trained, calibrated, and validated. So that's why when you talk about general AI, I would say we have to be very clear that if we talk about the fact that we're going to build models that are capable of solving problems on which they were not specifically devised for, this will happen, obviously, and we already started to do so. It's a matter of correlation, statistical correlation, comparative analysis, etc. Now, if you talk about general intelligence, meaning, oh, an AI that's going to master all types of human intelligence, including creativity, emotional, practical, it's not going to happen, which is good. So I'm fighting for that, as you know. I completely agree. Yeah. And then for the regulation, very quickly, We have a few questions here. Do we need a regulation on AI? Yes, we do. Then is— this is my view— then it's how do we have to build this regulation? Well, we have to build this regulation to make sure that we protect fundamental rights and we encourage innovation. The GDPR that was on personal data in Europe that inspired, by the way, the American text, the CCPA, California Consumer Privacy Act, is actually much more liberal in a way than the AI Act. The AI Act is a bit more conservative, I would I would say. But the good thing about it is that it's a framework, and then this regulation will have an impact when cases raise and when we'll apply the law and we build some cases on which to refer to when we're going to build our, our system in the next years. And the good thing about this text is that it's based on the risk evaluation of the given AI. Which depends on the type of AI and the type of application. And there are other considerations. The text is not perfect, but I'm a strong supporter of the AI Act. Now it's a matter of like, are we going to inspire other countries in the world as we did with the GDPR to build our own regulation? So despite the executive orders developed by President Biden in the past, nothing happened. So I hope we'll have soon in the US a draft of what could be the equivalent of the CCPA for AI.

Great. Jemimé, on your side, on any level you want to—

Right. So in terms of regulations, I completely agree with Aurélie, even though I'm not completely involved as my role as a director of a lab that's part of a federal school. But what I also understand is because technology moves so fast, obviously it's relative where they To me, it's not moving fast enough, but for others, it may seem going much faster. But the regulations should be built not based on fear and precaution because they get to see and punish those who's pushing the envelope. It should be based on predicting how this can have— it's going to be an enabling technology that can be used in both ways. Portuguese. So we are humans by nature, creative beings. We cannot prevent bad by doing nothing because it will happen. So to understand how to make things better and preventatively, I think that's where the regulation should come in, not to block technology or block the advancement, but to how to use it for better. So that's how I think we can prevail using this framework work that would help more people. And then the first, the other question in the beginning was, uh, yeah, artificial— yeah, no, I completely agree with Worli. Uh, as I hope it will not happen 10 years, much later. Why? Well, think about this. You have the smartest kid in the school who aces every exam, and he will or she will be the best student ever, valedictorian. Guess what? Is that person going to be able to swim because that person learned how to swim by reading? Would it know how to survive once it dipped into the water? It can take it as a figurative metaphor or literally. Because I think the question of AGI was because you want to see when the robots or any other technical beings can actually do things what humans can do. Obviously, I'm never going to be able to memorize everything in encyclopedia or anything under Google, but I can swim because I have a general intelligence. But are we going to be able to expect this from a robot? Probably not. And would your AI know that? Probably not, because it never swam in life. So you won't know what to do once it gets into water.

Well, you know, I have to say, I'm bringing over to a closer now, Aurélie and Jemimah. First of all, my humbling kind of recognition of how lucky I am that I have you as friends in my life. I think this is beyond the just conversation is how rich you are able to make a conversation that is becoming more and more critical, but also by providing us with real sense of direction, which I think is, is an act of rare leadership that sometimes The scientific community is too absorbed by its own sense of progress and time that they don't hear enough. And I think you becoming kind of, you know, voices, right, of that community in your respective field, that's a treat. So thank you for this. For whoever has been able to watch this, and there were many all connected throughout the session and then more will see this later on. This is really where we want to engage the best minds moving forward. It's not in limiting or in just assimilating technology to what technology is not, but defining a specific place for technology to play a role without having this ambiguity that we want always to kind of generate in conflated model, right, between humans and machine in terms of thinking, humans and machine in terms of robotics or mechanical. So this was honestly, thank you so much. I have loved—

Thank you.

More than what I was anticipating. I was ready to start it from a pretty good high bar. It's just the beginning of our ability to continue to talk and do things together. That's what our responsibility is, that we have to continue to disseminate knowledge in the way that is needed out there. I know that, Jamie, for you it's 3 o'clock in Asia, and I appreciate the effort to be be part of this. Thanks so much for doing this. Aurélie, it was a pleasure to hear you.

Thank you.

I am proud of the work you've been doing since I met you. It's just beyond expectation on how much you've been able to really change for the better, right? The narrative on AI. So my dear friends, thank you so much for being with me. I appreciate. To anybody else connecting, this is the end of our AI for Good virtual series that MicroOne has been curating for a while now, and I've been lucky ask their chief economist to be directly exposed. At the end of the day, I'm the one who's absorbing most of the learning from all this conversation. So thanks again. All the best. Enjoy the rest. Well, good night, Jeremy.

Good night.

Los Angeles early. And to anybody else, thank you so much, guys.

Bye-bye.

Thank you. Bye-bye.

Bye.