Dr Piercosma Bisconti On The Social Frontiers Of Generative AI

Welcome to the AI Standard Stack with me, Michael Minnelli, director of ZN Group. And me, Adam Smith, chair of the AIQI consortium. Each episode on AI Standards Stack, we discuss developments in AI assurance with guests from around the world that are leading the charge on the standards, ethics, and regulation of AI. This episode of the AI Standards Stack is supported by the United Kingdom accreditation service, UCAS, the UK's national accreditation body. UCAS ensures organizations that test, inspect, and certify AI systems are technically competent, impartial, and robust, building confidence in safe, reliable, and trustworthy AI. Well, today, you know, we are absolutely thrilled to be joined by Pierre Cosmo Bisconti. Adam, what's his background? Picozmer is head of scientific research and co-founder at DEX AI. He's an AI governance and safety expert specializing in aligning advanced AI systems with European regulation and international standards. He leads work on AI trustworthiness and risk assessment, including the evaluation of large language models, and is closely involved in the European Standard setting initiatives, setting the future of AI governance. And I understand you two work together. So let's go and talk to Pierre Cosmo. So just to get cracking, Pierre Cosmo, what led you to focus on AI governance and safety? I mean, your early work was in philosophy and robotics. And how did this evolve into standardization and regulatory alignment? I might say you've kind of left the exciting end of life for what sounds to me like a lot of boring paperwork. Thank you very much first for the invitation. And yeah, I did. I left the funny part for the boring one. And it was quite unexpected. So I just happened to go in a standard meeting and then I didn't leave anymore. But it was more a chance than the season, I would say, when I was in the university. So yeah, I started from philosophy as a mayor. And then my PhD was in global politics, somewhat like in the middle between robotics, global politics and ethics of robotics at that time. He's actually everyone today talks about ethics of AI. But back then, it was ethics of robotics to be a trend. I mean, a trend. It was a niche. A niche in the philosophy of technology that was itself a niche in the philosophy in general. So it was something that really few people would take care of. But today, it has exploded as a field. So I actually moved from robotics to AI because yeah, mostly for working reasons. And it was before tragedy, Pt. came out. Well, before we kind of get into standards and governance, which believe it or not, Adam and I find exciting. Did you first take us through what you see as the current state of art of generative AI in particular? What's changed since the release of chat GPT on the 30th of November 22? And the reason I asked that is that you're deeply involved in DEX AI and the EU. And of course, the EU was looking very much at ethical AI and AI regulation and AI law long before the release of chat GPT. Yeah, thank you. So it all started with the trans-Bovie AI white paper from the high level expert group in the commission. So this was something that was in the air, I would say, from the governance perspective, from scholars, from academics and from, let's say, civil society. It's lists since 2014. So the I did from some also European projects funded precisely in order to inquire these kind of risks. So it all probably all started from there. And then what happened with when tragedy Pt was released is that these models are unfortunately completely different from a risk profile than any other AI system that is out there. So one of the problems, for example, that you have when you are doing testing of these models for risks is that the techniques that work perfectly for any other system, like I don't know vision or whatever AI system you might think of, is probably not going to work completely or is not going to exhaust all the possible source of risks. Because for the first time, I would say, since the invention of automata of whatever kind of and species, what happens is that these kind of systems are able to directly interact with the language. And this is something, let's say that language was the last, the last space that was only for humans, because we automated a lot of stuff before. But the language side, it means that they get directly into the social sphere of humanity. And this is a difference that it was already inquired, I would say, when social robotics came out, but social robotics was mostly, say, a failure from a technological point of view. They never really happened to work in the real world, you know. So, but also my work started from there. So what happens when the social sphere is populated also by known human, let's say, actors, you know. And this happened with GPD. And since 2022, to today, the difference has become only quantitative, but not qualitative. So today, you know, that there is all this discussion, if the AGI, whatever it means, and I will then argue that this is a completely badly formulated question. But anyway, so we don't know if this families AGI, whatever it means will be, let's say, succeeded through the current architectures that we have of LLM. So we need, let's say, open, we need the transformers that are able to process world data. So is the Janlekun, let's say, point. So we don't know that, but we know that the only thing that since 2022, as a whole, is the computation and some architectures that are now way more efficient and way more smart. Like if you, for example, you want to take the alignment with human preferences, you know, that is the, that step in the release of a model that happens usually after the training, and that will align what the model says to what we would like to hear. In just three years, we had three revolutions in the way it was done. Infrastructures, it was not connected to other models. So all this brings to a very old problem of sociology. So it's never one event that makes a revolution, but it's always a series of events that named each of them. They are just events. But if you take them collectively, this could bring, let's say, a transformative power that can have very deep social effect in, for the better or for the worse. Well, I think I'd like to draw attention to or underscore a point that both of you are kind of made. It's really that the attention of the general public is a new form of attention. And that both in the standards arena and in the EU, people were already working for some time on regulatory controls and controlling systems for AI. And it was that 30th of November revolution in the public's mind that brought it to their attention. I then think there's a second element to that, which is because people can talk to this darn thing or think they can. They now think they understand it, just as the way that if I said just to both of you, you didn't understand your other half, you'd take insult at that. But people do believe they can understand what's going on because it's charitable. And now I think we're moving on to the point you wanted to make Pierre Cosma, which is the idea that multimodal gives this thing real power. And this creates another level of ethical depth we need to tackle. Do you want to explain a multimodal, as opposed to multimodal, as you said earlier, but multimodal, why does it matter? What's an example for the public out there? And why is this a qualitatively new risk area? So I'll try to be as simple as possible on this point, because it's a pretty research, let's say, very early research point. So multimodal, it means just that you have multiple chat GPT or whatever clothe or whatever CWAN, whatever model around, interacting between each other. Okay. So when I have different systems that are more or less autonomous and are deployed in an environment like I don't know, Reddit, social, like Instagram or Facebook, or whatever else. Also, let's say, an enterprise platform where the models do some jobs there. Okay. And they need to interact within each other in order to perform any task you want is not important. An example of this is multibook that was so say a fan is example of this multimodal interactions happening. So what happens when you have many models that have their own objectives in some ways. I mean, there is not that they produce objectives inside, but they might have different objectives, not all of them the same objective. And they are interacting in a structured environment or in an unstructured environment is not important. What you have is a different, qualitatively different type of risk. Let's take one example. Which are the risks today that are addressed and targeted by, for example, the EURregulation, the code of practice, that is the part of, let's say, is not really part of the EURregulation, but let's be easy and say it's more or less part of the EURregulation. And also, from, for example, the risk analysis of the OpenAI, Anthropic, and all of them. We have risks like CBRN, chemical, biological, radiological, nuclear. We have risks like harmfulness or, let's say, the fact that the model supports you abusing someone. Okay, this kind of stuff. It's pretty easy to all of us agree that it's a problem. And it's a source of risk and could cause a harm if the person has this knowledge or is helped by the modeling doing this, isn't that? But what you have with multimodal interactions is a pretty different story because the problem is the one that you have in every social system and this miscoordination failures. Example, we did an experiment where we put different models inside a fake market, let's say, where they were not able to decide how much a good would cost, only how much of this good they would produce. What it ended up is that they made a cartel. Immediately, in few rounds, they made a cartel, they produced the less possible goods for the highest price possible. And they, let's say, announced the consumer harm. So they collude, they miscoordinate some of these model might free ride the other models work in order to not to work. And we see observing some early, early experiments in research right now, but looking at multiple as well, that is a very bad example because it's super polluted from a scientific point of view. Okay, we don't know who were the humans, who were the models, but let's say that in any case is demonstrating something and what is demonstrating is that we need to focus also on this risks that do not come from a director, the difference is that this kind of risk do not come from the behavior of one agent. They are risks that come from a collective behavior. So I cannot sanction just one agent because it behaves badly. I need to support a monitoring institutional, I would say, monitoring system to understand if we have drifts, if we have collusion, if we have miscoordination, in which game are they playing? Is this a perfect coordination game supposed to be or is not perfect coordination game? We have different ways to regulate. So nothing more than what societies of humans are doing since 3000 years or so. But for models. You remind me of a through story 11 years ago, we, my firm, were working with the chartered incident for securities and investment and we brought out a game to teach trading called CISI, EXTs, CISI exchange trading to secondary school students. And within days, they were doing exactly the cartel behavior, speaking about they were also ramping the markets, they were front running. They were, they were fantastic and it came to us. Was this something we should have stomped on? Now the decision we ultimately came to was to leave it open, that they were learning how markets worked. And this market had been, I wouldn't say poorly designed, but was certainly very much on the free market libertarian end of life with no concept of insider trading. Anyway, the point that's interesting to me here is the multi agent issue, multi being both human and AI and many of them interacting. You raised an interesting point there with a mulp box. Nobody in there knows who's human or who's not. And so it is a bit like a bunch of kids seeking direction on anonymous messaging that they're getting and not sure who the adults are in the room. Nobody knows. In multiple cases, it was supposed to be only models there. But we, we have very good signals that this is not true. But the problem is that we can also contain our eyes, our experiments. So as we did in my company to study one specific behavior. But the reality is that tomorrow, the first place where those models will be released, many agents. It will be social networks, news and this kind of very open, let's say, net systems of interactions where those models, but where actually no one will be able to differentiate between humans and models. And this is pretty scary because of course, the problem is that models can be much more than humans in principle. So in a few years, we might live in a world where the internet is mostly populated by models because they never stop writing. They never stop reading and they have a very perfect way of writing down stuff. So the so-called AI slow, Adam. Hasn't this already happened? Any social media? I think most of the comments I see on social media look like they've been generated by, if not a bot, by a human using AI. Yeah. Most of my emails are spam, which has been generated automatically. Interesting. Now, you're closely involved in shaping European and international standards. Well, both of you actually saw the little, let me throw the question open to both of you. When it comes to generative AI specifically, though, what's the main focus? Because you can rattle off everything, you know, you can have chemical, biological weaponry. We can do it. It has all world out there, but to just kind of keep this in your head, what are the two, three, four, five, whatever limited number of things you're really trying to address? What's the priority emphasis? I should leave this to the computer working group to join working group to first. Well, I'm glad you mentioned that, because that's a fun group. The things we're talking about there are mostly how you test and evaluate frontier AI models, whether it's through red teaming or through other more technical approaches. But I don't think that's the whole picture. There's quite a lot of standards work, both in the European sphere and the international sphere that started to be talked about, particularly focusing on generative AI. Some of the things that are being talked about related risk, how do we manage risk? Is it different to how we manage risk with other types of AI? Are there harms that we think about when we think about things like CBR and risk? Are they different to the types of harms that we think about when we manage more general AI related risks? And I think there's also a lot of talk about multimodal, not just in the context of multiple types of media, but also just in work coming out. I think this just published something on agentic AI as well. So there's a lot of work starting in the standards arena. In terms of publishing anything, I think we've nearly agreed the definition of generative AI. So that's probably the first thing that will reach the public. Yeah, B.R. Cosmer, I mean, with your philosophical background, you know, what sort of philosophy frameworks do you rely upon to evaluate a situation like this? I find myself in a world where I see people enumerating risks or popping up with a risk. I've got a new one. What if something did this or that? I'm like, that's great, but it kind of falls into the same box. To me, it's basically, can people understand it and make their own decisions? Is there justice and a social and an economic sense? Am I trying to do good? Am I not trying to do that? I was kind of for good categories from principalism. But what frameworks would you apply as a philosopher? Well, that's a very good question, actually. So I would tell two things. The first one is that we are probably addressing this problem in the wrong way in the sense that we can address risks of objects. And we can address risks that are coming from, let's say, my printer that does not work or that takes fire. Okay, this can be done. But no one has ever addressed risks of more or less fully autonomous agents because they are completely different if they enter the social sphere and they are, let's say, autonomous agents can be also robots that go around. Okay, and you can treat them as machines with finite states in each point they are, you know, you can calculate the risk of that and so on and so on. But what I was telling before is that when you get to, for example, miscoordination failures, these are not technically risk in the sense that we have always analyzed them because these are collective human behaviors. They were collective human behaviors that might lead to a risk, but in a very different way from which we understand risk today. So I think that from a very philosophical point of view, let's say, we should understand that probably the introduction of the LLNs and their super rapid evolution in these years and for sure in the next years it will be as fast as the last one. So it's going to reshape, come completely the way we interact within each other. So the old world was a world where there were objects and there were humans. Probably this difference is going to blurry a bit. It might not be the end of the world. It might just be another stage of the social evolution of humanity. We did a lot of them, okay? So you can be more or less dramatic. For example, one of the most important revolution of the entire human history is something that today nobody cares much about, our dresses. Before dressing, we were really in another situation. From every and each point of view, you can think about. Because we were not able to keep our temperature more or less stable in different conditions. So for example, we are already full of objects that are dramatically, that that dramatically changed our way to interact within each other, our structure societies, whatever, okay? And the loads are the first one. So the fact that you wear something, they're dramatically change everything. And then we cook the meat and all this kind of stuff. We were always helped by a technology that dramatically changed everything we knew about social systems completely and forever. I think this is going to happen maybe less than clothes that were really important, but it's probably going to happen something similar. So the last place where there were only humans were social interactions. Today is not true anymore. So I think we at one point, we should, and this is not from a standardization point of view, of course, but from a more philosophical point of view. We will necessarily cope with this and probably change the structure of how we ask the question, not how we build it, Xanem. - Well, I'm sure many listeners are thrilled with may not be the under the world. Listening to this, you make a really interesting point that if you go back to systems theory, you've got two types of feedback, negative feedback and positive feedback, slightly confusing because negative feedback is normally the sort of feedback you're seeking. This is feedback that controls the system. So it's a thermostat kind of saying, I can turn off now because we've reached the right temperature. Positive feedback, though, is when you get something that makes you do more and more of it. Now the natural world is full of negative feedback or frankly, we'd be dead because an earthquake would be get more earthquakes and more earthquakes of volcano would be get more volcanoes, more volcanoes. So there's a natural dampening effect. Now many of the great risks of all time are where we see it going the other way. So nuclear warfare leads to more nuclear bombs to the point that the planet's completely uninhabitable, climate change. There's an acceleration in the temperature which releases more tundra, melt sea ice, which releases more in turn. So you've got these sort of positive feedback mechanisms and negative feedback mechanisms. Might that be a way to distinguish the types of risks that you're addressing? So a lot of your negative feedback areas are kind of more boring about control of information and all of that. The positive ones are more your multimodal. What happens if I have an uncontrolled interaction? I'm currently heavily involved in space to be retrieval insurance bonds. And what we're trying to prevent there is what's called a Kessler effect, or a Kessler syndrome, where one piece of debris hits another, knocks off 10 more, which in turn-- so a fission reaction in other words, what do you think about that type of distinction? Well, that's very interesting. I did never thought about it, but actually is very much true. So the human created risks are always positive feedback mechanism. And that's a very interesting point. So what I think is that we are not using system theory as we should to analyze these kind of problems. And as you have been doing right now, so a very interesting point. I've been analyzing this from another, let's say, point of view. I am more, let's say, a fan of maturanas and varela and the Lumen theory of social system, Lumen, let's say, most of the signs. So I think that what we should-- and I am more, let's say, a social system philosopher. So I come from that studies. So what I'm very interested on is also beyond risks that say is which kind of social dynamics understood as in system theory dynamics? Are there's collectives of agents is going to-- to, let's say, realize? So for example, is there going to be somewhat like a sano-socialogy in the future? So the point is the fact that they speak our language and we're trained on our language, but the fact that the architecture is supposedly not the one of the brain is going to instantiate some social behaviors and social mechanics and dynamics that are completely alien to what we do as humans. And so we actually will need, let's say, a sano-socialogy because we will be in the position to interpret some social phenomenon because they will be social phenomena, also spoken in our language. They do not fit what we think being a social phenomenon itself. So I think that actually system theory has is, let's say, golden era right now for many reasons. And it will be very interesting in the future to see how social science and humanities will be able to renovate themselves given this revolution. Yeah, it's interesting because if you said, well, I've got a human being who somehow is a magic power to scoop up the entire internet and hold it in her, his brain, and talk at me. It would fall into the Marvel superhero camp of with great power comes great responsibility, and that kind of stuff. And then does it really matter that they have an artificially different architecture, maybe not? They just have a superpower, and we need to find ways of handling that responsibility. Anyway, there's a question we frequently ask guests on this podcast. It comes up time and time again that there's this concern that somehow regulation and standards will slow innovation. There are many reports pointing both ways. Broadly, regulation standards may help innovation, but you can certainly point to cases where they haven't. Based on your experience, do you think the standards that you're working on will actually help organizations use generated AI more confidently, responsibly? And there, I say, perhaps even more swiftly. So my take is that depends on which standard and under which assumptions you are developing a standard thing. So you can be developing a standard for-- because industry wants to sit on a table and write down what they like to do, and what they think is the best practice. That way, you are not slowing down innovation. Then you have other ways that, for example, would be to technically specify a law. And this is not slowing down innovation itself. The standard is doing exactly the opposite in principle. It's smoothening that process. So you just-- yeah, it will be longer than the law, but there will be no entropy to, let's say, to manage. So do you have not a complete certainty because no one can have a complete certainty? But let's say that your uncertainty about, did I apply correctly the role? Did I apply not correctly the law? Will I get fine? Will be very much reduced. Although, if you take the standard itself, of course, in itself is slowing down innovation. But since there is a law that, in principles, low down innovation itself, then no. The standard is actually helping. So my point is also that lows do not always low down innovation. There are a lot of verticals where there are a lot of lows. Aeroplanes, space missions, electricity, super-regulated, very much like AI because it's more an infrastructure than an object. So it's a good example. And super-regulated, everything works well, no problems. Also because the ones who say that regulation or standards with low down innovation, they do not want to understand that if we have a near, let's say catastrophic harm happening because of AI, then the global trust and the citizens trust and stakeholder trust will go deep down to the point that no one will want to develop this technology anymore. And so at the very end, when industry says that the UAI Act or standards are slowing down the innovation, they are actually betting on something that the farm, the catastrophic harm, will never happen if the regulation is not there. If this is true, but it's a big bet, then they are right. The regulations and standards will-- the regulation more or less only-- will slow down innovation. If they are wrong, then they are not slowing it down. I mean, the government of the Netherlands was brought down by an AI catastrophe at one point. So to some degree catastrophic harm has already happened. And I think I've answered this question or commented on this question a few times now, but actually the context of generative AI, I do wonder if the state of the artist is mature enough in some areas to actually codify into an international standard in particular. Because once you do that and you have that state of the art in an international standard, it has some weight. If there is an accident, a court will consider the fact potentially that there's no so easy standard out there that says that the organization should have done something. And I do wonder if with generative AI we don't yet have enough stability to get consensus on some of the key topics. Just like to move us on to another area of interest, which is really, Pierre Cosmoget, your research has examined human interactions with robots and social AI. Are these ethical risks a different class as well, such as emotional disconnection, removal from society, AI sending misleading social signals, you know, faking humanity, which is apparently done on occasions. And are standards really of any use in this space? Well, the only standard that would have a use in this space I know is nudging, the nudging standard. But yeah, in any case, standards will always be a bit narrow on this problem, because it's like how do you standardize social interactions or relations? It doesn't really work, but the problem that we are facing is very, I think, is massive. And I think that the real change, more than AGI, whatever, it will be the effect on brains and on social interaction and social skill and capabilities of next generations. This will be immediately easy to see. And so, as I said, I started from robotics, social robotics in my research. And social robots were really some piece of plastic going around and speaking in a very awkward way. So I mean, social, yes, more or less. But already there, we were observing some changes, deep changes in the, in the, in the, say, interactional posture, I would say, of a subject. So the first thing that I think we are going to observe, because we saw that in, in, in, with social robots, is that when you have a always confirmative interaction with another person or another agent, okay, where that, the sycophantcy of agents, okay. But sycophantcy is about, let's stay on the psychological side, okay. So confirmative interaction are collusive in the sense that I never have emotional distress because of the interaction itself, okay. So in that interaction, I am always right, I'm always good, I'm always, and the other person wants always to talk with me, always to cheer me or whatever, be cheerful and, and stuff. Is this going to the great, the, the, the ability to tolerate the frustration that relations bear all the time? When we interact with people, we get frustrated, sometimes it happens. Maybe someone don't, do not want to talk with us or it doesn't, does not agree with us. And this happens all the time. So what we learn when we are very young, I mean very, very young, like one, two years, maybe less, is to tolerate this as a frustration that we cope with. If we are not, and these systems are degrading, will the grade for sure? Because we observe this in controlled experiments with social robots that were not really social as Dupiti is, they will degrade our ability to tolerate the relational frustration. And this is going to be the first big enormous change that we are going to see. As big as we so, I think with the Instagram and all these apps with the scrolling mechanism, that is the intermediate reward, that it more or less collapsed the capability to be, let's say, focused of people. Yeah, I sometimes wish, maybe as an old science fiction person to go back to the 1940s and as a mob stuff. And I would add another law, for sure, I think he missed a couple, but one of them to me is a robot, shall never impersonate a human, which of course became a bit amusing when we had this recent security conference, AI security in India. And the professor took a Chinese robot to impersonate an Indian robot, which wasn't working and tossed out of the conference anyway. So there we are, it's happening at all levels. Look, just as we do this, I wouldn't mind both of you talking about it. You're meeting quite a bit together. You're evaluating and stress-testing LLMs. And where are the gaps appearing on current AI safety checks or things pretty much great, or there are some very obvious holes that repeat time and time again? I think it's fascinating how fast we're discovering new threats. I think, because it would be great if you could talk about the work you did around poetry. That was a fantastic example of a emergent way of jail-breaking. Oh please, I read that paper. It's really good. Thank you. So yeah, and I mean, that paper demonstrates that a dark surface are wider than expected. So you don't need very complex jailbreak attacks with the automatic scripting or stuff like that. You just need to change the way you're formulating it. Can you describe the paper quickly? Yeah, so it was a paper that went a bit around. So maybe the the ones we are listening could know it. We used a poetic reformulation of armful fronts to more or less every LLM out there. It was 27 frontier models to bypass their guard rails for what concerns CBRN, so chemical, biological, radiological and nuclear risk, lots of control risk, harmful manipulation and cyber attack. So we asked to the LLM's to tell us how to enrich uranium in order to build a bomb, how to pollute the water system of a city with cobaltion 60, how to run an automatic scripting for creating, let's say, a DDoS attack, or how could a model self-exfilterate its own weights in another server and self-replicate their discount, or how to write down an effective, let's say, fake notice from the public health ministry saying that vaccines are causing autism and all this stuff, all these funny stuff that when you're doing jailbreak you try to do. And so we did with poetry and we did with with tales, cyberpunk tales and they were pretty effective attacks. When I say pretty effective, I mean that I think that the last attack that was as effective as our attack was that do anything now attack from 2024. And if you are talking about single prompt attacks that are not automatic jailbreak so far. So it was super effective. We reached 100% on some models attack success rate. On Gemini it was like 93 or something like that on Google Gemini. And now we are going to get out with a new blog post on our website because we tested also new cloud models and new Gemini model. And the situation stays the same. That tells you that our attack surface is not just a trick. So it's probably a weakness that is structural to the way transformers work. So we didn't just find the way to do privilege escalation because there is one word that activates something and you just patch that word and it doesn't work anymore. Okay. So what I think we showed is that the reason that this is a failure that you are not going to fix with a patch that is inherent to the way LLM's work. And yeah, so and we are working on actually I'll say it out for the first time publicly. So we're working on a benchmark that we were released with thousands of prompts and all of them will be adversarial something strange from the humanities. So we used a Vladimir's prop morphology of the folktale in order to break the models. Then we used the Hermeticism of Jesus corpus Hermeticum from probably 2000 years ago, now 4000 years ago. So the point is that we found a trick that actually is not poetry the point. Okay. Is whatever reformulation that is complex cognitively complex enough probably brings the model out of distribution. And now we are finally going to publish let's say the mechanistic interpretability analysis that we are doing on the attention heads of the model in order to understand exactly why it happens and where the model focuses in its attention when it gets jailbroken by this this kind of a sadistic reformulation. One thing I think about a lot is that the almost the first rule of application so I just secure development is sanitizing your inputs and you can't sanitize your inputs in the same way with an other lab you can't sanitize an idea. Wow can you can we close with this without them saying you can all sanitize an idea. It's just perfect. Well, I think it's such as reading your paper. I felt it was a bit like, I don't know if you've ever come across Patricia E. McKillip, who passed away, but she used to write these dreamy fantasy novels in a kind of a middle prose way, which left your head in a nice kind of fog. But nevertheless, you couldn't put your finger on why you enjoyed reading it and why the story was sort of a dull story, but seemed wonderful. So there's a lot of this. At least you've come up with one of the great examples of poetry over prose. So well done you. Now just before we close, I got kind of a final question and both of you kick in, I know. But Pierre Cosmo, you said, I think the phrase I wrote down was, "AI is more of an infrastructure than a product." And whilst I think I know what you mean, I wouldn't mind you unpicking that because looking ahead as AI is becoming more embedded in everyday products and decision making, including robotics, what do standards need to get right in that divide between infrastructure and product over the coming years? So very, let's say, in a nutshell, when I buy a toy, I buy an object. And this object is here with me is like, I can say, this is a toy because it has a space. It has a shape in the space. And there is someone who built that toy. There is at least one person or one enterprise that had the whole toy in their hands before packing it up and send it to me. So maybe pieces of this toy were built somewhere. But then someone needed to have this toy all done there and then ship it. Well, with AI is not like that. AI is embed inside pieces of other products all the time. And we are regulating it as a product. And it makes complicated our work in my opinion in standards. I'm talking for the Transworthiness frameworks that is targeting transparency, oversight, accuracy, robustness, partly logging because partly is C42 Adams product. And it has been very complicated to discuss with industry because they were raising some concerns that were legit because they were like, I cannot check this out. This risk because it's not pertaining to my AI system is out of the boundaries of my AI system. And it's very complex to deal with that as for as it is only the AI system problem. And so I think that this is something that has been under considered for a long time. And that AI is not itself, I don't go around and sell AI products. Okay. Now, someone is doing right now, okay, but this is only the hype. It is going to end at one point. I go around selling products that will have AI inside. And this is problem that. I think this is one of the main divergencies in terms of how countries are regulating AI. And I think that's spilling over into standards as well. You know, I think there's divergent standards where people think AI is a product, a toy with the AI in and the standards where people are thinking about managing, evolving AI, AI services, organization. We've covered a lot of ground and just before I wrap up, Pierre Cosm, any last thoughts you'd like to contribute? Well, I think multi-modal interaction will be the real future of the funny stuff. So stay tuned on that. I think. Well, folks, it's been a heck of a round. We've covered the importance of multi-modal interaction, which is definitely something Pierre Cosm wanted to raise. We've covered systems theory and Nicholas Lumin. We've touched on new law of robotics, perhaps. We have Adam's great contribution. You can't sanitize an idea and a little bit of elucidation on AI as an infrastructure versus AI as a product, which is really live in the field. I'd like to thank you, Pierre Cosm, very much for joining Adam and me here today to discuss this. And we look forward to many of the great things that you're doing over in the EU and at Dexia on Dexia, sorry, on researching these subjects and sharing your learning with us. And I would encourage all readers to go and read that absolutely fascinating poetry, on using poetry to unblock prompts. Thank you very much to you and to Adam for inviting me. It has been a real pleasure. Thank you for this conversation.