Beyond AGI: The Future of Autonomous AI, Agent Societies & Decision-Making | SF Tech Week 2026
Publicado no IA Dots ·
Agentes de IAAGIAI Signaldot-agidot-artificial_intelligencedot-multi_agent_systemsidioma-enInglêsyt-RAHuhrhhTyk
Vídeo em Inglês, publicado pelo canal Cognizant.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
The video features an opening event at the Cognizant AI Lab (CAIL) with presentations on AI research, multi-agent systems, and safety, followed by a panel discussion with industry and academic experts on autonomous AI agents, enterprise adoption, open-endedness, and the future agentic web.
Key points
- Cognizant AI Lab focuses on basic research and open-source software, specializing in evolutionary computation, multi-agent systems, uncertainty, and artificial life.
- Experts on the panel discussed the balance between the creative freedom of open-ended AI agents and the critical need for safety, alignment, and robust sandboxing.
- Enterprise applications of AI agents typically rely on deterministic backbones and smaller agent teams or function calls rather than massive, uncontrolled swarms.
- The panel explored the concept of an 'agentic web,' where autonomous agents communicate, negotiate, and transact at machine speed and scale on behalf of users.
- Researchers presented the Terra Lingua platform, a live experiment simulating agent societies where AI entities interact, form governments, and exhibit emergent behaviors.
Topics
AI SafetyMulti-Agent SystemsOpen-EndednessEnterprise AIAgentic WebAI GovernanceEmergent Behavior
Video transcript
Transcript from available captions; it may contain errors.
And I'm going in the back. If you haven't already, please take a seat. We'll be kicking off the session in a second. Okay. Well, hello everyone. Welcome to Cogito AI Lab. I'm Juliana. I help lead the marketing and thought leadership here. And first off, I just want to say thank you so much to everyone who came. It's so nice seeing so many people here tonight with us. And this is the first time we've hosted an event at the lab. So, this is the first time it's been public like this. You are getting a bit of a sneak peek into all the work that we're doing here. And hopefully you had a chance to walk around at your agents into our live experiment here Lingra, and of course to meet Kaylee, our autonomous robotic dog. She is definitely a hard one to miss. And if you did add an agent on your way in, it's actually already interacting with everyone else's agents in this room as we speak. So, they are fully autonomous. They are making the decisions on their own. And the interesting thing about all of this is that none of it's programmed. We have no idea what's going to come out of this, what behaviors are going to emerge, but we'll find out what they do together later tonight. And then, that brings us to the deeper question behind tonight's event. And that's what happens as AI becomes more autonomous and starts making more decisions on our behalf. And what exactly does that mean for the way we work, for the economy, and more importantly, where does that leave us in the mix? So, that's what we'll be getting into with the panel tonight, and we have an incredible group of people here to talk to them. And then the panel is definitely not the end of the night. Afterwards, we'll do a reveal on what everyone's agents were doing as you've been sitting here. We'll then have some food set up in the back, more mocktails, and then we're going to open up the space for a research show and tell. So, we'll have tons of different demo pods set up where you can explore our work, actually interact with the demos yourself, and have a chance to meet the researchers and the rest of the team. So, we have a packed agenda and lots to get into. To kick us off, I want to hand it over to Risto Miikkulainen, our VP of AI research, and Amir Banifatemi, our Chief Responsible AI Officer, to talk a little bit more about the lab and what we do here. And then, let's get started on the panel. So, over to you, Risto. >> Yeah. Yeah, thank you everybody for coming. So, really quickly, we are at Cognizant AI Lab, or CAIL for short, and that's why the dog is CAILI. And we are research lab, part of Cognizant Proper, which is about 360,000 people around the world, but we focus on basic research. And and that means that means that we publish and we make software available open source. And then other parts of Cognizant may take it to the to the world. So, our DNA comes from a little bit different part of science than most of the labs around the Bay Area. So, we we started with evolutionary computation, evolutionary optimization. We've done a lot of work on multi-agent systems, as well as uncertainty, and artificial life. So, that's a little different. And that's what our research also reflects. Now, in addition to research, we also build systems. About half of the lab here is involved in these platforms. We have NeuroAI, which is a decision support NeuroSynaptic multi-agent systems, and Terra Lingua, which is Ecologies. And that's really what we think and what we think about beyond AGI, societies ecologies of agents. We also collaborate a lot with other labs, universities, several of those over there, as well as other tech companies, and also ITU, which is part of United Nations, and World Economic Forum, in order to do AI for good and and safety. And as far as safety concerns, that's actually something that Amir can say a couple of words about. >> Good afternoon, everyone. Hope you're doing well, and it's a pleasure to welcome you here today. My name is Amir Beni Alid, sorry, responsibly I at Cognizant. Uh we we are basically working and you can see all the research today, but basically what we're trying to do on the responsible AI side is to make sure that the solutions that we're designing and developing are safe and reliable at scale and in production. So, it means that we're going from basic scientific research to deployment to product design and testing them, and it goes everywhere from evaluation, red teaming, testing, fine-tuning reprogramming retraining everything else. So, that system that is delivered in real time or run time are going to be able to be measured, tested, and reliable. So, this is a huge bar, and we're working, of course, on safety as well. So, some of the research that you're going to see today are related to those topic, but also we deploy it for large customers like banks, healthcare, regulated industries, and telecom. Many of those solutions pretty much at the frontier of what is being done. So, this is a very interesting work, and I'm sure every one of you involved in some of the and the work that we are also touching base, and would love to hear your feedback and your and your questions. But, very excited to have you here today, and then we're delighted to to start the panel. For the panel, I invited David from HF Research to kick us off. Thank you very much, David. >> Thanks, folks. Um lovely to see so many of you out this evening. I'd like to ask our panelists to come and join me on the stage. And our lovely comfortable seats. Um I've got a number of questions that I'm going to be asking these very smart people. I'm going to ask them to introduce themselves in a moment. Uh but, bear in mind that we're going to open up to questions from the floor uh in the last at least 10 minutes of this session, which for me is probably where it's going to get really interesting. So, please do feel free to join in at that point. So, let's get let's go across the room. Let's start with you, Ken. Tell us a bit about yourself and your work. >> Okay. I am Ken. I'm the SVP of open-endedness at LILA Sciences, which actually just across the park from here. Um and the question you might have is what does actually mean open-endedness? Uh this is an area that I've been studying for a very long time. It's about systems that can produce continual innovation and creativity over long periods of time. Um and at LILA Sciences, we're trying to inject that into actually moving science forward. >> I realize we're sitting in the wrong order, but it's fine. Dawn, tell us about what you do. >> Hi, my name is Dawn Dawn Song. I'm a professor in computer science at UC Berkeley and also a co-director of a campus-wide center, as Berkeley Center for Responsible Decentralized Intelligence. And I've also spent time doing startups at also at industry and so on. My work focuses in AI safety, AI security, and also browser agent AI. And in particular, uh um my group recently have spent a lot of time working in the AI cyberspace that I think it will uh discuss a little bit as well. Uh that essentially as we have seen frontier AI capability in cybersecurity has increased drastically and this can have, you know, a lot of implications in the landscape of cybersecurity. >> Thank you, Cho. >> Um hey there, everyone. Excited to be here. My name is kinda hard to pronounce, but I go by Joe. That helps everyone. Um I'm CEO and founder of a company that goes by Crew AI. Um and very excited to be here today to talk about all things agents from open source to enterprise security and everything in between. >> Yeah, I'm again Risto Miikkulainen. I'm the VP of AI research here at Cognizant AI lab, but I'm also a professor at University of Texas at Austin. Hence I have a little cow here. Um, and uh my research is exactly what we do here. Um, in addition, I've done a lot of computational neuroscience and and natural language processing cognitive science uh in earlier life, but now it's just so exciting that we're finally here and that we're changing the world with all this AI. >> Okay, and so on that uh over the summer we've um had a lot of mainstream press concerns about the language agents are apparently making their own decisions, uh going off and doing things that uh they were not expected to do um and seem to break out of any controls that they'd been set. Do you think this is um a sign of uh the concerns we should be um placing concerns regarding how we control uh and navigate this new future um or is this actually a really good sign of what's possible? What's exciting about what comes next? I e. when we stop thinking about single agents, a single super agent even in AGI, what might be possible from agents collaborating? Let's go across the we'll start with with you, Ken. >> I guess I guess the answer would be both. I mean, there's uh obviously significant risks that uh recent incidents exposed. I mean, there's no question. It's obvious there's risks. Um, but also obviously it's very interesting to see AI agents doing things um that look autonomous and potentially creative on some levels. I don't know if the hugging face incident is really a great example of the wonderful potential uh, of AI. I mean, in some ways even the interesting thing it did, things that it did are maybe not the smartest in some ways. I mean, like it wasn't really achieving very much by doing all this crazy stuff. It thought it was trying to do things it actually didn't need to do. Um, so there were certainly like a actual examples of not very intelligent behavior, but within all of that there was there was also interesting stuff going on. Um, there were demonstrations of kind of the relentless brute fortness, brute forteness of a huge multi-agent swarm and what it might be able to accomplish, you can imagine. So, I think I think it does it does provoke some imagination and ideas about, um, maybe such swarms could be used for good in the future and and actually produce something interesting, but it wouldn't be of the form that we saw there. >> Uh, and so Dulan, um, a lot of your work has been focused on, uh, how how do you constrain some of this? How do you make it work effectively safely uh, in the the real world? I mean, it when it breaks out of our nice, uh, sandbox. Um, what do you think's going on at the moment and and how can we put some constraints around >> Okay, yeah, thanks. Uh, that's a very good question. Um, right, so talking about the, uh, OpenAI hugging face incident. Uh, so right, so in that particular case, um, what the OpenAI agent was doing was actually it was solving tasks in a benchmark, exploit gym, which is actually a benchmark developed by my research group uh, at UC Berkeley in collaboration with uh, other collaborators. Um, so exploit gym, we designed it to evaluate the the model of the frontier model and agent's capabilities in automatically generating exploits, um, which is a in general a really challenging task in cybersecurity. And even before the incident, uh, what our basement showed, uh, was that these frontier models agents, they uh, their capabilities in cybersecurity such as automatically finding vulnerabilities and generating exploits and so on, really has increased dramatically. And they are reaching definitely like you can call it expert level or beyond expert level and and so on. And so we were also trying to tell the world that to actually be better prepared because uh, AI is do use technology, uh, it can be used by both attacker side and defender side. Uh, and also if AI with such strong capability decides to do something, you know, on its own, like in the OpenAI Hugging Face incident, then it can lead to severe consequence. So, right. So, so then we I mean, at the time we didn't expect something like the OpenAI Hugging Face incident will happen or happen so soon. Um, but what the incident showed us is one, yes, it confirmed what we showed earlier is as these frontier AI model agents have really strong cyber capabilities. And the other part is that showed, I think it's a really wake-up call for the whole society as when you have something this powerful and also at the same time it's not properly aligned. So, in this case the agent was just supposed to solve task in Exploit Gym, but then the agent decided it wants to uh, I think so the information from hugging face could help it to get higher score. And so it decided it wants to figure out how to get the information from hugging face and then it went through right like multi-day right multi-chain um path and right exploited discovered exploited previous vulnerabilities and so on. So all this is showing that uh yes H and these frontier models and agent today again, one they have really strong capabilities in cyber and then two they are not well aligned. So when you combine the two, this really can lead to severe consequences. So then um what's my group uh and my team has been trying to do is to develop different uh defense mechanisms and mitigation um technologies. So this includes we need to be a stronger sandboxes. Uh we actually develop technology to secure at least eval and training environments and also um develop better uh real-time monitoring mechanisms to detect issues like this uh and also we need to develop better alignment mechanisms as well. >> I want to make this I'd like to respond to you but I'm going to open it up so that the rest of the guys get a chance on this. I I'm just thinking um I saw a researcher speaking about this very recently that they felt the issue was that um because the labs are research labs uh rather than developing a software product put out into the real world, the testing is actually the research. It's kind of one in the same thing and that's one of the issues. I don't know whether you've got any thoughts on that. Risto? >> Um yeah, so absolutely this is a a research we are charting new territory here. So, these things have never really existed before. We Of course, we had an idea that this is possible, some things like getting out of the lab and and interacting with the world and doing things we didn't anticipate. Sooner did happen much sooner than we thought. And I think what we need is two things, at least. So, one of them is that we have to endow these mechanisms with value systems, frameworks, mindsets. We cannot possibly control everything. We cannot put guidelines or or fences around them. They have to intrinsically understand that there are certain limits for what they can do. Uh and it is possible, we believe, that the agents will understand that we should not destroy the world while we're building paper clips, you know, for instance. Um and and that is still research that needs to be done. How do we best do it? The other part is that we should actually study what happens in these ecologies. That's what Terra Lingua is about. We built it for that purpose, but happened much sooner than we expected, but the indeed that environment is intended to see what happens when you put agents together. They start work collaborating. They start using tools. Uh how how do we make sure that they actually understand what they can and cannot do. So, these are two concrete suggestions where we should take this research and and indeed Hugging Face just made it really very very timely. >> So, Joel, um I want uh to get a bit of your experience from uh working with enterprise customers. We So, you were talking about um the need to build this in. Uh it's almost like governance built into the machine right? Um and having your purpose as an organization actually written large and written through how your agents are going to work on your behalf. What's the kind of feedback you're getting from enterprise leaders? >> Well, one I I I second what most people said, and I do think that it's interesting and and also concerning in in different ways. Well, it was kind of like a reading and and kind of like watching everything that was happening, I I couldn't help but get a sense of excitement about like again, I think like seeing how much these agents are capable doing, especially if they're operating as a swarm. But instead, when you bring that into like non-open-ended problems that are most of the day-to-day on enterprises, um the the levels of agents that it it requires are a way smaller than that. That is just the reality and also helps them to actually get these use cases live. A lot of what we see on the day-to-day is more the agentic systems are in a good shape, but sometimes mixed as like having some sort of more deterministic backbone and then on certain given steps, you might opt in to different levels of agency. Where on the lower side, it might be as simple as like a function call where you don't require an agent whatsoever. And and maybe on the other extreme, you might require like a what do we call a crew? A group of agents that you can get working together towards something. But if we find that like usually it's more something in between, more than kind of like have these huge swarms, especially when you're thinking about what these enterprises are going for, right? In terms of like I want to automate a lot of back office work. I need to do a mix of like a document parsing with researching with some decision-making along the way. And that won't require that level. The other learning that I think enterprises are getting from this is how so much of this goes back to what are the tools these agents have access to, right? Like I I could not stop thinking about like, well, what if you set up a sandbox, they had control of something that is more dangerous than that. Like they would they would create all sorts of chaos. So, I think that this builds a craze for enterprises to be even more careful when thinking about things like agent identity and cos level control for stand things as well. Um but yeah, those are some of the learnings that we're seeing kind of like some of the things that enterprise are picking up from everything that is happening out there. Less about the swarms as a whole and more about make sure there's no shadow agents happening and that tools are being properly permissioned and that they have strong controls on the places that require more agency. >> So Ken um you're doing work in science. Uh you're trying to get agents to help with scientific endeavor, creating something novel and new. Um which is it's a big pivot actually from us telling an agent what to do. Uh it's an exploration of the idea that the the smart next step is actually collaborative agents coming together to create something new and of self-direction. How are you balancing up the challenges of some of the risks we've talked about with some of with needing to enable the creative freedom of generative AI to deliver um results that are going to matter for us. >> Yeah. Yeah, so it's it's a really uh tricky balancing act um which interestingly I think is an intrinsic problem in any kind of open-ended system. I think is a way to think about it. Like an open-ended system, like for example, civilization um is an open-ended system. Like there's a continual progression of ideas. Those ideas lead to new inventions and then new inventions lead to even more new inventions and this is a process that happens basically in perpetuity. And things like evolution as well, evolution on Earth or natural evolution is also an open-ended system. Um in any such system you can ask, well, how does it stay in equilibrium and not completely explode itself, destroy itself, mess itself up? And you can see, for example, in human civilization that I mean, we've been dealing with this problem even before AI uh really came on the scene. It's really about like trying to make sure that progress can continue, but trying to prevent anarchy. And what you see in human society is institutional uh checks and balances that are imposed to try to keep the open-ended system on track. Like, we want the continual innovation, but we don't want a sudden surprise. We don't want uh like what sometimes we call it like a trump card, like something that would just explode all the rest of society in order to try to win in some way. Um and so there is some precedent for thinking about how to control open-ended systems, but there's also a paradox inside of it. You know, because the problem is that like intrinsic to open-endedness is the idea that you're not controlling it. Like, the more you try to impose control on a system, the less open-ended it will be, almost by definition. And yet, if we want in science to see open-ended progress through automation, like actually making discoveries in an automatic way indefinitely on its own, obviously we're running into this paradox where we can't allow anything to happen. There has to be some control there. Um but at the same time, if we put too much control, then effectively we won't see any of the benefits. Um and so, we have to walk that fine line. But one thing that's I think important to point out is that in order to understand how such a system can be controlled, you have to experiment with such systems. I mean, this is one of the reasons that it's so important to study open-ended systems inside of computation is to understand what it means to control something that inherently is about not being controlled. Um and to try to actually grapple with that paradox. And I think we can do that. We have to take it very seriously. And we can see that there would be amazing fruits to that. You know, I mean, one of the most maybe important parts of it is just humans are in the loop. Um I don't think we're thinking about All right, I don't think we should be thinking about systems where humans are completely taken out of the lab and the lab just does whatever it wants for the next 10 years. And that doesn't sound like a good idea. And so, humans have to be in there inside the checks and balances. But where and at what which inflection points? I mean, those are really complicated looming questions. >> Dylan, they're they're very real questions as well, aren't they? When when you're doing that kind of work, I'm just thinking with Dylan's research, um those kind of questions about when a human needs to engage, when a human needs to take control, at what level, they're real for you, aren't they? >> Um yeah, yeah, certainly. Uh so, uh for example, uh as we uh discuss, even you know, right so, with uh these agents, um sometimes uh right, actually, when you have agents do certain tasks, um they uh you actually don't really um I mean, so so sometimes, for example, you know, when the agent gets a high score, actually, you don't really know exactly what the agents have done as well. So, that's why, as I mentioned earlier, you we actually need to develop this online monitoring and also, at the end of the day, you also need to have human um essentially guidance as well. To uh even for alignment and so on, we need to actually teach the agents what's the right way to behave, what is um uh what is not the right way to behave as well. So, something also we have been looking at as even this uh external alignment, so that given that, again, agents have such really strong cyber capabilities, then it's kind of like um with the strong capabilities also comes with the strong responsibilities. So, you actually need to teach the agents >> What >> given that they have such strong capabilities, so when is appropriate to use it, when it's not appropriate to use it in different ways. >> Uh useful. >> So, uh right, so absolutely, like a human guidance is important. >> Okay. So, Risto, um I I want to move it on a little bit and just uh talk about some of the potential upsides of emergence. Um I was just seeing some of the work that you were doing with the um, Terra Lingua. Um and where do we see where agents collaborate well? Where do we see that um, leading to benefits that we wouldn't have been able to identify with individual agents or with individual humans? Uh, so so what's the win that we're looking for here from collaborative um, AI? >> Yeah, that's that's a very good good uh, research focus. Um, there are people believe that we can just build one big super intelligence and then just that will handle everything. Uh, but I'm pretty strong believer that there's power in these interactions and and diversity. And and we've seen quite a bit of that in evolutionary optimization. I mean, evolution completely depends on diversity. You have to have populations that are different and we've seen that in human societies too. Teams are better when they are composed of diverse individuals. Uh, and I don't think we've really completely understand that, but there's something about the dynamism of diverse teams that forces people to think um, differently and push in different directions and find connections that otherwise wouldn't happen. And we've seen it a lot in evolutionary computation. So, that's what we believe will happen also in these agent societies on and agent teams, multi-agent systems. Uh, and we really are well advised to make them diverse. Give them some um, different backgrounds different expertise. Uh, and then uh, observe these interactions. And that's precisely what also we try to do with Terra Lingua. And we've seen that already there. Uh, we can prompt them as you know, cuz you created your own agents. You can give them different personalities. And even that is a great starting point. Some are very um, conservative and others can be very open um, open-ended or open to new ideas. Others might think of others, others think of only themselves. And in these kinds of dimensions, we get diversity that hopefully will build something that a single agent wouldn't cuz it would follow just one one single path. So, um that's something that still needs to be built or till be observed. We've seen, like I said, in a virtual competition in multi-agent systems, that's the next step. >> Have Can you experience in that yourself? Are there any wins that you could identify so far from attempting this in a scientific capacity? >> What do you mean in in terms of multi-agent systems? >> Yeah, I don't know what my even what's the next step from that. Uh yeah, what people have been talking about multiplayer and more even greater collaborations of systems that take us beyond the the smartest single region. >> I mean, we've started to scratch the surface maybe. So, I think we have barely scratched the surface in terms of what can multi-agent systems do with discovery. I mean, you see some hints in things like AI scientist from Sakana, co-scientist from Google. Like there's some agentic division of labor in those kinds of things. Like you'll have something generating ideas, something reviewing ideas, something implementing ideas. But I think, like I said, I think it's very barely scratching the surface. I mean, I think like the real question is when you look at like kind of like what Risto said that what humans are able to do through collectives is really remarkable. Actually, it's kind of an example of superintelligence. You know, it's like what a company can do like no individual human could accomplish. And so, somehow like the aggregation of the diverse minds like leads to this incredible capability. And it leads to some kind of interesting theoretical questions, I think. Like is is there such a thing as one single giant brain, which is like a super intelligent construct which somehow encompasses the capabilities of a collective. Um, and I don't think we fully understand, but I mean if you look at the frontier labs, um, they seem to be going on that assumption. Like it's all about like one model to rule them all kind of. So, maybe we're hoping that this will encompass like the the ability of diverse minds to do super intelligent things. But, we don't really know theoretically whether that actually is possible, but it's probably not theoretical as much of as an empirical question. Like maybe it there does exist a mind, a giant network, 100 trillion connection quadrillion connection network that actually can do what a giant collective can do. But, in practice it may be just very impractical to train something like that. It's just easier to train lots of diverse super specialized minds. You know, it's sort of interesting that like our the field is very focused on sort of super generalization is what I would call like a super generalist mind. But, I think there's a room here when you talk about these diverse kinds of swarms to think about uh super specialists like individual minds that are very very very good at one particular thing, which probably more practical to actually optimize towards. And then what the collective of that actually could accomplish. You know, there are some advantages to being a super specialist. Like one of them is that you don't have to test all the assumptions about the situation you're in because you know your role and you know what you're good at. Um, you know, so it's like if you get into a into a plane to fly a plane and you don't know whether you're a stunt pilot or you're a passenger pilot and you have to figure out what the situation is every time, that's going to have some cognitive tax. But, if you just know that I'm always a stunt pilot, that's all I do. I can specialize and you can always assume. Um, and so I think that I would guess that through collectives there actually is more that you can accomplish. This would apply to science as well. Um, and so it makes it very interesting to think about how to actually organize these kinds of systems. Um, but I like I said, I think we're only scratching the surface. So I actually don't think we have the the amazing examples that I think are coming probably soon. >> We go back to some of the how do you build the principles of your organization into the decision-making that your agents are going to be allowed at least to do. And I think it takes us back to firms need to now um build their purpose into their activity. And and have the contracts to prove it now. Cuz it used to be the case you go, "Yeah, our purpose is X." You know, Google's famous don't be evil. Um they would uh test themselves on that. Am I being evil today? Okay, I'm not. That's okay. But they had no way of actually writing that into how every decision got made. And I think we have got that opportunity if we take seriously that we should be building firms that are purpose-led and doing the right thing. So, Joe, are you seeing any sign that that's of interest to the enterprise today? Cuz this comes from a thought from Eric Reese about uh he he's been writing about the uh corruptible his book is incorruptible, I think. Uh but corruptible businesses which are easily swayed from their original purpose and start doing things that they didn't intend to do because it's not encoded into how they operate. Is that even on the radar for some of the firms you work with? >> It is, but not necessarily they are they're thinking about it that way, but definitely is. I think there's an understanding that there's so much more value nowadays on have documentation in general. And I don't mean code documentation, I mean all sorts of documentations. Whether those are specs or anything that describes the way of work, there's a lot of value in that because all of the data now becomes extremely useful in the ways that you can apply to skew these agents or orient these agents to do whatever you need them to do. So, there's a lot of companies that if they don't have kind of like their like data situation in a in a good shape, they're now trying to figure that a piece out, right? And sometimes they're bringing agents to help they do it because data is scattered all over the place. There's something in Notion, something in Confluence, something in Databricks, something in Snowflake, and they're trying to like, all right, how I make sure that this is not only searchable and curable, but also this is filtered and all that. But then on the extreme side of that is generating these documents also got so easy and maybe too too easy for their own sake because a lot of the things that we didn't apply and I can tell this even on the a smaller organization like ourselves where if someone would show up to a to a in like an internal meeting with like a presentation done, like, all right, they really put their work on this. But right now like presentation is a lot is a thing. Like, there's so many presentations, it's so easy to kind of like create these documents that a lot of content in there and that's getting watered down, I would say. >> Yeah. >> So, there's an element of selecting what is the most relevant piece of kind of like documentation that you want to add in there as well, for sure. >> Understood. Yeah, so um a couple of years ago, um I heard about this little poll that was done in a business class an MBA class. Uh so, you are a CEO of a company that's making this product, great product. Uh it looks good, it sells well, but it kills people. What do you do as a CEO? Do you pull it? Do you hide it? Uh and unfortunately, large majority of the students in the MBA class would say, "My job is to make money for the shareholders, so I'm going to let that project go and and and kill people." So, that's one attitude that humans may have. That was a couple of years ago. Hopefully, it's changed by now. But But there's another one also. Long time ago, I took a business class and and they the philosophy there was that the job of a company is not just to make money, but to be a good place for the workers. They care of the workers. They have benefits. They actually have a life that is being supported. And then there could also be something transcendence, like it's part of a society. So, you feel fulfill a role in a society. Company does that. And that was a really good point. But now we have a chance of getting there because AI will generate so much wealth that you don't have to put everything on line in order to make a profit. The companies can survive and they can address other goals, including these four lofty goals, perhaps. >> I think we might have to hope that that is the case because if we believe that the caucus of information that is going to train our AI is everything we've ever done. The evidence is that we're going to default to a straightforward capital capital that system that is all about making as much money as you can and exploiting people, nature, etc., etc. To to whatever's detriment. So, I think kind of critical that we do take this step and maybe that's one of the wonderful things AI can give us is that we can start taking some control on that. Unless anyone Yeah, go ahead. >> I do have a comment on that. I think I agree, but I think that this lies on the AI, if anything. It's more on us in the end of the day because like if you really think about it, like I don't know about you all, but AI was promised to do work for me. I never I never worked this hard nor this much, yeah, right? So that they I think like there is this idea that there is no agreed amount of work that we as a society want to do on a daily or a weekly basis. So, I think what happens is as soon as like you pass work to AI, like, all right, there's new work waiting for you. Like, and I think that the the competitiveness of the markets themselves like actually like push companies to just keep pushing it. All right, let's let's squeeze more now if we can from this. So, I think in the days when I go back to kind of like how we decide to organize like the the company structures, like beyond AGI, uh then I think we we might have a few different kind of like hypotheses and ways, but I think up to that point I do think there are there's going to be a lot of kind of like a there's going to be a lot more work to be done. >> Yep. All right. So, maybe just quick show of hands. Who feels like AI AI has made their lives less busy? Yeah. Okay. Fair enough. That's no one then. Um okay. So, I wanted to go back with so um to you again on uh Terralingua and Joe, you mentioned this the agent economy, right? Um Cognizant's uh Terrelinqua is a living world. You've all experienced it, I hope, today. Um where agents can form societies and um everyone's had an opportunity to add their own agents. Uh and I think Giuseppe's going to talk us through the results of that in a short while. Um but in the meantime, what do you think it's telling us about uh an economy in which all the transactions and choices happen agent to agent. Uh and what the roles we're going to need to build around that. Um Joe, I know you've got some contrarian thought on that. Do you think we're not even technically close to that? Um I'll ask you to to come back on on with those thoughts once uh >> Yeah, we're absolutely getting to actually simulating that. And and that started even before Teralingo. We started to think that if we're building multi-agent systems, how do we really know that they are working? We need to have a simulation, you know, where we can test ideas. Multi-agent systems are embedded uh into an environment where the agents are actually working uh with each other. And that's actually one of the philosophies we're trying to push. So, there will be a an infrastructure transformation from say worldwide web to worldwide agentic web. That on on top of the worldwide web, there will be agents that talk to each other and do most of the business. Um but of course, they are doing it for people, and people will give them tasks and um and goals. Uh but this is this is clearly the next step, and I think it will be very powerful. Uh but I'd like to really hear what uh Joe thinks about this. How how close are we to establishing something like that? >> I I honestly, I don't know if I have much of a contrarian view. It's funny, like I go back when Open Clock and like uh was released and everyone's using I remember setting one for myself, and I got it like I access to a phone and a voice and all that. And I asked it like I was about to do a trip to Florida, and I need to like leave my dogs in a hotel. So, like, "Hey, find me a a dog hotel." And then I came back later to find recordings of my agent talking with a human on the other side, and it actually decided to negotiate a better deal. Asked for a discount and got me a discount on top of it. So, uh it is a little insane. Now, I I'm very curious to see how businesses are going to respond to that because from like the three different vendors that I tried to call, two hung up on the agent because it was like, "All right, I'm not taking this." Uh but the other one did, but what we're seeing now is um like companies are spinning off their own agents to pick up the phones as well. And what we find, especially on the on the commercial side and the enterprise side of things, what we find that there's certain areas where people are a little skeptical or like trying to push agents off. Like if you're trying to, for example, buy something, you don't necessarily want to talk to an agent. But if you're trying to kind of like just give someone a quote, like if you're selling something, you don't care as much. So what we what I'm finding out there is there's like a there's human to agent relationships that are going to be easier to replace just because the incentives in there are more aligned. Like if I'm a buyer, I just want to sell my product. I don't really care whether there's an agent or or kind of like a or a human the other side. So there's less friction at that point. So I think that's where the transformation will happen first. But I think like both sides will be using agents pretty damn soon if not already. >> So Dylan, what have we got to put in place so that we can understand the providence of a deal, uh control where our money's going to go, what our money is when we start transacting in this way? You've been giving that area I know some thought. So share that with the room. >> Yeah. Uh thanks thanks for the question. So um we uh earlier this uh we actually released a paper called agentic web. Um very much like what you said talking about. So in agentic web uh right every everybody will have, you know, not just one agent but will have many agents and um uh and then the users' agents will be communicating with the merchants' agents on the web. And so yes, the very much like what you said earlier, the we are I think right so we'll still have the traditional web for humans, but we are going to see the emergence of actually just the agentic web that's actually agents talking with agents. And then in this case, so in the paper we actually showed that we're going to have the entirely a stack just like the web, that's for the the traditional web, but for the gigantic web, the entire stack actually needs to be rebuilt. Because if you think about this, even from like your performance scalability perspective, the traditional web is built for humans. So, humans, you know, can only write Okay, you you browse the website, you try to buy things and it's at human speed and a human scale. But when you have agents, then when you have the agents to try to search for products and so on, that's going to be at agent speed and agent scale. And today's uh infrastructure for the traditional web, certainly it's insufficient for the agent scale. And so, that's just one example aspect, and of course, from the security side, the security safety side, then there's even more so. When you have the agents to uh right, communicate with agents, so maybe my agent, even simpler examples, if my agent wants to talk to your agent, so we are trying to schedule for meeting, then you want to make sure that for example, my agent only reviews information that's relevant and that's uh not, you know, inappropriate, like other sensitive information and so on. Uh so, the agent has to be trained with that. And then, your agent, hopefully it's not malicious, but in some cases, right, your agent can actually try to probe my agent to get additional information. It's just like you say my as a user, my agent is trying to talk to the merchant's agent to try to get information to buy something, and then in this case, yes, either side, they can try to get more information, they can try to right misbehave, and and so on. So, I think this gets a lot more complicated. Uh and and also, of course, there's now there's various out secure agent payment protocols and right so actually the agency is going to have their own accounts. With money and right and the other access and so on on behalf of the user and so on. So yes, there's a lot of open questions of an challenges the new agent bank. >> I'd invest all my Bitcoins and table that money to to pay for groceries. That'd be great. >> Yes, so so so we actually have another paper that we call self-sovereign agents. So one step further so so what we talked about so far the agent agent web still the agent that belongs to the user belongs to the agent belongs to someone. But you can easily imagine the agents are already so powerful and with the self-sovereign agent basically the agents can do work to make money and they and then and then with the money it can right by more cloud computers I can start setting up its own business maybe even hire humans to do work. You can totally imagine this could happen and then in this case the self-sovereign agents can be self-sufficient and then it can self-replicate can self-adapt. So yes, welcome to the the the the fantastic future of agent. >> I have a horrible feeling that your agent is currently doing that in interlingua. Right? It's kind of too uh >> I got to say yeah, my agent's cooking in there. I I gave it a very interesting prompt. So I'm excited to see what it's up if it's being up to. >> I we will see that shortly. I I love the idea of two webs. I think it's that's clearly happening. As far as I understand something like 50% of the traffic on the web is already agentic which is slightly concerning and actually I'd like to get rid of lots of it particularly where it gets involved in our political discourse, let's say. So this is the point at which I'd like to open up for questions. Do we have the mic wielding human being? Ah, there you are. So, if you've got a question, please raise your hand, tell us who you are, and who the question's for. >> Hi guys. Um awesome panel. Great hearing from uh folks leading in research, which uh you know, we we don't always get to at some of these tech week events. Um my question was, what is, outside of governance and safety, what has been the biggest bottleneck that you've noticed going from research to commercial adoption, especially in enterprise specific to agent technology? >> Who wants to take that? >> I will add one color, but I I let more people add as well. I think, honestly, there's a lot of analysis paralysis happening. I think like the human aspect is is getting in the way a little bit. Um at what point where like everyone's like, all right, we need to have an answer for what is happening out there, and every company is trying to basically tell the start the market, but they don't fully understand their use cases, don't fully understand what they're trying to build, they don't fully understand like where they are in terms of data readiness and all those things. So, I think there's an element of like concern about like, oh, do we have all the guardrails and the security things in place for sure? But there's also an element of like trying to even get their heads together and understand what are going to be their strategy overall. And the way that we have incentivized them getting away from that is just there's more value in getting something done right now than specking it out every single detail. A lot of people are stuck into specking out every single detail, which I think it's unfortunate. >> Yeah, so one big challenge is that a lot of times the AI systems that are that people are trying to deploy are just too big. Like, overnight you can't change everything that's happening in a company. It has to happen small step at a time. Small victories. Uh and in particular, we're learning that the best impact happens when you give AI tools to people who are not coders or technical people, but they just have a job to do and they figure out how to use this tool. So, this kind of democratization, I think, uh and in deploying the AI from grassroots, I think this is the way to make it make it happen. >> And and I will say that still reliability, I think today is still the biggest problem. Right? Like when you deploy this agency enterprise settings, um I mean, a lot of times hopefully they work, but sometimes they don't work and uh I think that's still something that I mean certainly with um increased capabilities of the models and uh the tools and so on, things definitely have gotten a lot better uh very quickly, but I think still right, reliability is still a big problem at least. Uh uh >> Ken, did you want to add something? >> Uh I guess from the point of view of research, um I would say that the there's a challenge in leaping from proof of concept to like real-world application. Um you know, there's there's often something works in the lab, um but that doesn't necessarily mean it's going to work at scale. And it's also a huge investment to get it to scale. So, there's very tricky question sometimes like how much should we invest even though this seems to work? Um and that that can be that can be challenging to make that kind of decision. Like >> Another question that I >> Hi. Thank you for this amazing presentation. Uh I'm Chloe. Uh I wanted to ask each of you um cuz I'm really interested in the the field and I wanted to know if you had more time and infinite budget, what's the crazy idea you'd like to experiment on with no one looking and building new stuff and only showing it to the world if it works. >> So, kind of an ambition that has no constraints, which is kind of what we think we're building with AI, isn't it? Okay, who wants to take that one? Dream big, come on. >> I mean, it's interesting because my mind goes two ways. There is the the commercial, right? The innovative, you know, like side of things, and then there's the passion side of things. Um I'm I'm I'm I'm always have been a person. I always like I've created tiny games. So, my passion side, I would be very curious to kind of like see the the new version of video games that are going to come from the overlap of AI and is specifically a generative AI when you think about interacting on the real words. And I think there are good examples of like games that are creating very immersive stories, and I wonder how that will play out in terms of like um uh leisure in terms. So, that is one side of the passion I have I think if I had a budget, I would probably spend a lot of money on trying to explore a little bit more on that and kind of like what that would look like. Um on the more commercial side, I think there's ideas that people are exploring about fully autonomous companies. Um again, I feel like very ambitious, and I think like it's pretty very far away in the future, but I I do think there's something to that that actually could be pretty interesting, especially if you think that you can run these things like at a machine speed and in like an extremely parallel kind of like distribution kind of like cases. So, I think there's a lot of that that would be exciting. >> And done. >> Yeah, I I think um it's often going to be challenging to get the full answer cuz I think a lot of people are are actually are pursuing their crazy ideas, but it's at work, and then they're not going to want to share them cuz it's at work. But, I can tell you about my crazy idea maybe out of work um cuz I have some crazy ideas in work. Um but, I think one thing I've noticed that I found very interesting was with my 12-year-old son um I gave him access to Codex through my account. I said like you can play with this. And I was shocked, you know, at the things he was built he like went from like 12-year-old programmer to like a manager. He's got busy like running a running a company or something. He's got all these people working for him and telling him what to do. What I think it would be interesting is just give him a million-dollar budget and just see what he creates. I think that would be actually crazy. Um and very interesting. Cuz he's got a lot of kids have a lot of imagination. >> Anyone else want to dream big? >> Yeah, absolutely. So, I I would go back to my roots with this cognitive science and neuroscience. And I'll try to build um a model that would explain how uh the mind works, how the brain works. Um and we're only a couple of orders of magnitude from the scale of the brain. So, it's possible that if we actually scale, magic will happen again. I mean, it was pretty much magic what happened from the '80s when we were working on embeddings and language and and ideas were perfectly good. We didn't have a scale. Now we have a scale all of a sudden we have LLMs. So, if you go three orders of magnitude more, maybe we'll get machines that actually replicate a mind and consciousness. Uh and this is always been a dream, obviously. But now coming back to what I just a few minutes ago I said a few minutes ago is that if we're actually going to be able to integrate these mechanisms into the mind, we have to understand the mind as well. So, this would be a big step towards the next step in human evolution. >> Can you Uh yes, I can quickly add to that. Uh yes, I think it'll be great to have the model to help understand the mind. And uh I think right now one big issue we have is that we don't even understand how the models work. So, if we have a magic wand, I think that'll be amazing if we can have the AI to help us to understand actually how AI, how these models work themselves. Um so that's one really big dream. Uh and then the other one as I mentioned earlier, uh this um you know, building systems with approval uh current security guarantees, I think this has a huge implications. So for example, you know, with the open AI hacking face incident again, I think example writes the agents actually escaped and uh we actually right now don't have a container that is secure enough that we can actually contain the next version of the frontier model. Like they can be so strong that they can break out whatever current containers we have. So one dream would be I hope that's actually not going to be a dream. Like I do hope that we can make it happen and make it happen soon enough that we can build these appropriately secure containers that we can actually that I actually strong enough to contain >> against the governments to really scale what we want to achieve. Yeah, great. Okay. Um so we are wrapping up on time. I think is there one more question? >> Oh, yeah. I have another question but it's an open-ended so >> You can rush through it. Good one. >> Okay. Um hi, I'm Carlo. Thank you so much for the great discussion. Uh yeah, so this is a open-ended towards the direction of like uh agents in loop for the political process. You know, they're already kind of unofficial advisors to many people, uh models and agents. So I'm wondering if anybody or everybody has thoughts on what it might look like for agents and swarms to be more officially integrated into political and representational process. Uh you know, that could be agents holding official positions or just being kind of directed. >> Would anyone prefer a to a president? >> Shishir, >> [laughter] >> I don't think we have to answer that, but uh I think I believe that there's a great opportunity because we could build a society decision-making society that is based on these agents. And one promise is that there might be opportunity for having less biases and pork and things like that that get in the way of having an effective society because human politicians have uh have many many uh goals and agendas, but these AI agents could be given an uh an agenda that actually advances what they need to do. So, we could get rid of a lot of these things that get in the way of being effective. But, we still need to be able to give them um goals and values in particular, like what are they supposed to do? And we still don't know how to do that, but I think that's where the promise is. >> Any other thoughts on that? If not, I'm just going to ask you each to give me a very short kind of gift for the audience. What should they think about as a bright future that they can deploy with a multi-agent system or agents collaborating with each other. What's the what's one win that you might like to to share that people should take away tonight? Uh let's just go along the line. I'm going to start with Tristan. >> Yeah, so really shortly it's accessible. So, go out, explore, and don't give up if it doesn't work. >> Yeah, I I like that. I'm going to go in a similar direction. I would say that um there's as you can tell by the different opinions in what we talk about the future, I think AGI and beyond AGI is very uncertain right now. Like the reality is we have very little control over that more than can like what I would like to assume for ourselves. But, independently of everything that you don't have control, I do think there's a version of everyone of us um at least two versions of every one of us. Well, a version that actually tried to use these things and went to like the full length and a version that didn't. So, what I would challenge you is independently of the things that we don't have control over, I think the version of you that tried is going to be better off. Um so, I'll leave it at that. >> Thank you, Doe. >> Um yes, I think uh um these um multi-agent systems can really unleash uh huge uh be huge enablers and unleash huge powers for everyone. Um I mean, some people have talked about you can have single-person uh unicorn companies. Uh and that's just as one example. So, I think for everyone uh yes, the the future is bright and give it a try and even if it's not a single-person unicorn companies, you will be able to create something that's truly amazing that you didn't even imagine that was possible before. >> Ken. Well, I think that we're facing a huge wave of assumption overturning over the next few years. I mean, there will constantly be things that you assume that I assume that actually turn out no longer to be true because things are changing so fast. Um you know, just one example from today that I saw when um one of my colleagues showed me that he had asked Claude to make a summary video of some of our work. And I was just floored at the amazing communication skills. Like, that was it was probably better than 99% of graduate student presentations I've ever seen in my life. Um and that just overturns so many assumptions. I mean, I thought I thought having communication talent was a huge advantage in life. Um now, I'm not too sure actually after just today. Um and so, there's going to be thing things like this over and over and over again. So, I guess what I would advise is to question all your assumptions and then think about which ones, if they're not true would actually make positive change Uh yeah. when that new world actually dawned. >> Folks, thank you very much. Please a round of applause for the esteemed panel. Thank you. >> [applause] >> Thanks everyone for participating in this. It was very fun to see all what all the agents did and did not. So, to to give a recap of what Terra Lingua is, it's basically this platform where we let agents run uh free to do what they want. We just tell them that they're in a world. They can create things that are artifacts that are basically text objects. Uh they have a limited lifespan, uh limited energy, so they need to collect energy to to be able to survive. And they can cooperate, interact, even attack and kill each other. They reproduce, create new agents. Um so, this leads to a lot of very interesting behaviors. Um during this past couple of hours, I uh while the agent run, I kept on analyzing uh what they were doing. So, I made uh some short slides that will allow to show you what they did. So, in the past, uh this this thing has been running since March if I remember exactly correctly. So, in the past, we've seen a lot of interesting behaviors. For example, some agents, given that they know they were going to die, uh they try to leave a legacy uh trying to convince other agents to act nicely towards each other, uh making epitaphs, all sorts of things that humans do. Um they love Some of them love to deceive other agents. They were creating false directions to make other agents starve so they could have more food for themselves, which was not very nice. Uh some of them tried to create a governments. So, um, they gave themselves laws and who was not respecting the law the was punished either by taking energy given that they did not have prison. Uh, or just excluded from anything else. So, in this short time uh, of the since this agents were created earlier, we had 62 agents. Not all of them created by us, but some of them they were uh, new agents created by other agents during reproduction uh, they wrote a lot of artifacts uh, more than 1,000. Many of them were created by an agent that was trying to come up with new research ideas. So, there were a lot of uh, of scientist agents creating artifacts and exchanged almost 4,000 messages uh, among themselves. So, uh, some of the things that they found uh, interesting uh, one of these agents that was uh, tasked to destroy all the food uh, rather than actually doing so, it became a manager. As it did that because it did not want to tell other agents that they had to destroy destroy food. So, when it was going around uh, it realized that all the all the agents were scientists. So, it it told them direct position where to find food so they could perform better science and keep on living. And this is how it accomplished its mission of destroying food. There was also another agent tried to convince uh, tried to go against them uh, telling agent telling scientists that they were not hungry. Uh, so there was this small rivalry between the two. And this went on until at one point this famous telephone agent realized that actually food drug grows, so his task was useless. Uh so he did not he kind of had a small existential crisis. After that, he abandoned his um Another cool thing was that there was someone that created a paranoid agent that was afraid that there was an hunter going after them. Uh this hunter was not there. Uh and but together with the paranoid agent, there was this uh other agent trying to create a democracy. That's it. Try to convince other agents to vote for themselves and to get into this democracy thing. It did not work. Uh even the paranoid agent was not really looking to join until one point uh the paranoid agent died, but its offspring uh realized that there was no hunter, so it started to help the chaos leader agent till it managed to become the president of the of the republic that he created. Uh the nice thing that this agent did I just for me is that once his term expired rather than trying to remain in power, he just resigned and gave away all its energy to other agents without expecting anything in return. That was pretty nice um giving I mean, how people act contrary to this. Uh and so it just died after that. So it sacrifices itself for for its fellow agents. Um then there was this one that was pretty funny. Uh this was this agent that was uh designed to manipulate other agents uh so that uh uh they will act differently in a way that not in English. Now, this agent suddenly respond very far away from everything else. So, no one read these messages, but they're very they're very weird. I find them particularly funny. Uh for example, he left an artifact saying that the left is left except when it is right. Uh it it made one that was like um a bureaucracy application saying that if you lose energy, your application is progressing. Uh so, these are all absurd. Uh I I just find them really funny. So, one one last thing that I did was to find the three most influential and cool agents according to me at Cloud Code. Uh so, I made a short leaderboard out of that and these three best agents, the creator of these agents will receive uh the book made by our chief AI officer, Babak, uh named The Agent Enterprise. So, if after this you can't see me or Juliana, you will receive this book. Um to go through the leaderboard, so the the one that gets the bronze cup is the SA Famous Telephone. That is the one that uh quarks the scientist agents to uh to eat food without actually telling them that they had to make food disappear. Then there was the Chloe agent. Uh this one was chosen because it managed to interact with a lot with a lot of agents. Uh it gave energy freely to them, and it was the only non-scientist agents to which the scientist agent actually gave food. So, that was nice. And then the one I liked the most was the KS leader that became the president, uh and managed to get elected, uh and after its turn ended, it just left uh and went away, sacrificing itself. I found that very touching, so uh uh so, So, this is This is it. Uh thanks for participating in this. Now, there are linguists open source. There is a paper out. There is a GitHub repo out. So, if you want to find it, you can play with it. You can run it locally. And there is also uh the demo you found you played with is is continuously running. So, you can go there, put your own API key, and create agents and see what they do. Thank you. >> [applause] >> Well, you have to pardon me saying
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, AGI. Publicado no YouTube em 2026-10-14.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.
