The Rise of AI Agents: Do They Really Work? | EBC11

So, um, you hear it in my voice. I've been networking hard for the last two days. So, um, I welcome everyone and give you my opinion. You've made it to the last panel of the entire conference. So, you've been through days of intense discussions.

I was giving keynote speeches and I must have had too much coffee. So give yourself a pat on the back for that, okay? You can do better, I know, but we'll figure it out. So, um, our topic today may seem like one of the most intellectually challenging of all. So, the rise of artificial intelligence agents, do they really work?

But don't worry, we, um, with our brilliant minds on stage, we'll make it insightful, maybe a little provocative, and most importantly, human and interesting. So imagine it more like a conversation at a bar and then at a final exam. So before we start introducing ourselves, I'd like to start with a quick, um, warm-up. I'll ask each of you to answer, perhaps in one word: yes, no, or perhaps, are AI agents already working today? At first you might be like, yes, yes, yes, depending on whether I can say it.

Okay, thank you. So, um, great, that already gives us an idea of where you all come from. So before we get into what's behind these answers, let's quickly get to know who's here. So please keep it short and natural. Maybe just a few words about who you are and what you're working on in the context of AI agents.

Maybe you first. Good. Hello everyone. I am Asim. I am the co-founder and CEO of Kernel Labs.

We built a product called Openbox AI that focuses on solving the black box problem of AI and opening it up to make it more transparent, manageable, and transparent, manageable, and , ahem, reliable. So, we are working in the space of trustworthy AI. Hello everyone. My name is Max Lee. I am the founder and CEO of ORT.

Er, ORT is trying to create the world's largest AI data collection platform. It's actually a startup that came out of my lab at Columbia University. I'm a professor at Columbia University, you know, in the department of electrical engineering. My research interests are artificial intelligence and blockchain. Hello everyone and sorry for my fake photo.

I thought it was fashionable. I didn't know I was the only one, but I'm a real person, not an AI agent. I'm David, CEO and co-founder of a project called AMO. Simply put, we are bringing AI agents into group chats. So not as a manager, but as a teammate on Telegram, Slack, and Discord.

And before that, I was an engineering manager at Google for about ten years. So, as one of the first participants in the Google program, I like your image and your agent. In my case, I have been an entrepreneur in the cryptosphere for many years, and for the last three years we have focused on education. We are a leading technology education platform focused on Spanish, and I am Naruta H., the CTO for Giga at UNICEF.

Many of you may be surprised that UNICEF will be present at the blockchain conference and on the AI ​​panel. Um, Giga is a partnership between UNICEF and ITU to connect schools to the Internet around the world. Um, we want to connect every school by 2030. And for that, um, one thing we're doing is really using open data and source technology to speed up the process, make the whole process more efficient and transparent. Um, and these technologies include blockchain as well as artificial intelligence.

Thank you very much. And for context, my name is Danila Bubak, I represent the Bundis Fabant blockchain, also known as Bundesblock, the German blockchain association. So the voice of the German, ahem, blockchain community. So now that we know who's here, let's get down to business. AI agents.

I would perhaps give a definition of what an AI agent actually is so that everyone is on the same level. And maybe you can correct me if I say something wrong. So, an AI agent is like a smart little assistant that can think, learn, and perform tasks for you on its own. Something like a digital assistant that understands things without always having clear instructions on what to do. Agree.

Perfectly. Maybe one thing I would say is that it also depends on the definition, right? For example, AI agents appeared already in the 70s and 80s, and at that time they were AI tools for performing specific or solving very specific tasks, and I think that now when we talk about AI agents, we mean AI agents based on LLM. So I think it's worth noting this difference as well. Thank you.

So my first question would be: What specific task have you seen an agent do that really impressed you, and what is one task that agents are absolutely failing at right now? So who wants to start? I can start. So I'm going to try to focus on the web trio, because when you talk about AI agents, you can apply them everywhere. And sometimes they work, depending on what you want to do; sometimes not.

And one thing I find amazing is that recently, like, they recently, like, they , um, were published in Anthropic; For example, there was an agent that was able to program for 30 hours, so that was really amazing–it could keep working, keep working, and keep improving, using the same thing it was working on, but focusing on Web 3. Recently, there has been a wave of new protocols that use AI agents to help users invest in DeFi. And there are some that are able to consistently offer much higher APY ( much higher APY ( Approximate Annualized Yield) on USDC and USDT than Yield) on USDC and USDT than perhaps very well-known protocols like Year Finance and all the optimizers we had in the last cycle had in the last cycle . And now we have the same , but higher APY. This is consistent and fully automatic with the agent.

Of course, the agent is a kind of LLM, but also machine learning, because it is based on information and data, but at the same time I think it is very good. Um, and I think that will develop in this part. And what worked didn't work. Maybe things are different now. Actually, we had, uh Actually, we had, uh , dinner two days ago, and someone told me that it works much better now.

But, at least in my experience, a few months ago, a kind of artificial intelligence interface was also launched to make it easier to navigate DeFi. So you could ask the agent, " ask the agent, " Hey, exchange these tokens for other tokens and make this bridge and do this again." And it seems to be working now, or at least someone here told me so, but in my experience it was a terrible experience. So the UX UI improvement I expected from AI. Cryptocurrency is very non-deterministic.

Therefore, you need to learn a lot to use it. AI is like talking to a person. That's why open AI has already been implemented by 800 million users, and we're still aiming to reach 500 million in cryptocurrency. So I was hoping to see this easy-to- use AI interface in DeFi and cryptocurrency, but it has n't worked yet. I guess I'll understand.

So, um, I think the moment I saw the title of this panel really made me think that people are n't really enjoying themselves. Um, but on the other hand, speaking of, um, our adaptability. Remember when GBT 3.5 just came out, there were so many "wow" moments, and I think the AI ​​agent worked on a lot of different areas, um, but there's still the question of whether it works or not, and I think that's really interesting. So, in my opinion, there are so many areas where an AI agent, or an AI, or a big programming language is already modeling itself.

For example, like the ability to synthesize content and organize it, summarize content and also generate it, and I think that's what works, and there are many, many other different areas; but, and I think for those areas think for those areas where AI agents don't do very well, maybe one example is the ability to pick up on emotional nuances, the ability to read a room and be more fundamental or real, um, common sense reasoning . So these things are more related to emotional intelligence. Yes. So I think that's my main takeaway. Yes.

Can I have a little? So, um, I want to share two examples. Uh, one really good example that struck me about AI agents is that the AI ​​agents were, um, you know, uh, people who had already used the services of a law firm, a patent law firm. So patent law firm. So , you know, it used to be, for a patent lawyer, it was a very, uh, difficult task to check if your innovation or a certain new technology was already covered by some patent before, right?

So, the patent is huge. And now we're just using artificial intelligence to, artificial intelligence to, you know, through a document in AI, say, document in AI, say, "Hey, this is my innovation, uh, groundbreaking thing, I think, uh, is it covered by any patents yet?" You know, uh patents yet?" You know, uh , you know, an AI agent can give you an answer in a minute or two . And you can give a hint by saying, “If my innovation is already covered by certain templates, just give me details about those templates, links, and everything else links, and everything else .

” And the AI ​​agent will enter everything into the document and give you a report in PDF format, something like that. So, you know, before, it could be a job, uh could be a job, uh , weeks from a law firm, and you had to pay money, right? But now it's just one or two minutes. You don't need a law firm to do this; you can do it yourself, very easy. And, uh, that's how an agent really helps a lot, uh, in terms of, for example terms of, for example , patents.

And one agent, uh, example, another example that I want to share that an agent doesn't work, is, I would say, it all depends on the numbers, it all depends on the numbers. Because at Columbia, which I have, in New which I have, in New York, basically, you know, I have a research lab; We 're working on building an AI agent with a focus on finance, and finance is all about numbers, right? So we're trying to use all the big models today, like gro, like chat GPT, like, um, to do an analysis of the SEC reporting from Nvidia, from Tesla, from these companies. They file quarterly SEC reports with lots and lots of numbers, characters, lots of tables, and you see that the results of all these AI agents are, I would say, probably about 90% say, probably about 90% wrong. You can't trust them at all.

You need to go back to the file and check all the numbers there and see how to, uh, you know, do your own checking. So, uh, when you're dealing with numbers with your current AI agent , it's probably not going to work. So, we spend a lot of time figuring out how to deal with it. But in general, it's really difficult because, you know, in science, in academia, we always say that when AI works, we don't know why it works; why, when it does n't work, we have no idea why it doesn't work. That's the problem.

Okay, maybe I'd ask the audience to raise their hands if you've ever used an AI agent. Okay, not bad. Does everyone use it in everyday life? Perfectly. Okay, great.

You just mentioned the word trust, right? So when we talk about AI agents, trust is one of the big words that always comes up. But what does trust really mean when the agent making the decision, the decision, is not human, let's look at how governance governance , transparency, and decentralization come into play, and who is responsible when something goes wrong. Right? So my question is , should agents be decentralized and interoperable, or do users just want them to work seamlessly?

So how can you actually trust them and who is in charge when something goes wrong? So maybe I'll start with that? I'll just move on from the last question to this one. So, last week I was talking to a very senior executive who reported to the CEO of an organization with 150,000 employees. They laid off 18% of their employees and asked everyone in the organization to start using agents.

This is all research work, and this is a big technology company. They say, company. They say, "Okay, start using AI agents to make us internally efficient, but whatever works, let's take it to market and start selling it as a consultant to other people." So, it's important to note that everyone knows where this is going, but we're still in the discovery phase. Even the biggest organizations throw spaghetti at the wall and wonder what will stick, right?

But when it works, you need to make sure it's reliable. You need to make sure it works for big financial use cases, healthcare, defense, and everything else, not just for us who do things in GPT chat. And then if the data is wrong or not, you know, sometimes, you know, it's our fault if it's, uh, if we relied on it. But the organization is not going to blame AI because we are not at the stage yet where organizations are using it in a way that they are telling everyone that you should use it in your organization. So one of the problems that we have in AI right now is that we're building isolated systems, like we built Web 3.

Um, you know, you have agent-to-agent protocol, you have agent network protocol, and a lot of these systems are built in isolation. So as a result you get trust gaps. So we need to make every, uh, individual system accountable. You need protective barriers, and you need to make them effective, but then you need to ensure that system-to- system interaction, agent-to-LLM interaction, all of these things, all of the trust gaps, are filled, because having (and for me being) decentralization, trust itself is an issue that needs to be addressed for organizations. Decentralization, in my opinion, is not the most important thing, and this is coming from a venture capitalist venture capitalist who has been investing in blockchain for seven years.

Good. Thank you. Maybe someone wants to add, of course, maybe add to this from other industries as well. You mentioned web-three, but I think it's something similar that applies to other industries, the telecommunications industry in my case. I think the most important thing is transparency, like really promoting open source agents, uh, making sure we have open protocols so they can communicate and be interoperable, uh.

And again , in the telecommunications industry, usually, uh, operators, providers, they work in isolation, uh, and they can't really, uh, communicate with each other. One of the things we're doing on our side is we're launching an initiative called, uh, "Open Agents" for, uh, telecommunications companies, and the idea is to , uh , uh , make sure that you're building a foundation for open AI agents. We have a benchmarking platform so you can also compare, uh, I really think that these, uh , open benchmarks and platforms can also increase trust in these systems because you really know, uh, what's going on, and also going on, and also , again, promote interoperability between different systems. Uh, I would say yes, the same goes for different sectors. So I would add that there was, uh, an article about how a lot of companies that tried to implement AI and AI agents ultimately didn't work, and I think one of the reasons is that reasons is that you can't predict exactly how it's going to work, how it's going to react.

So it usually works very well within your company, where there is verification. So you're working with AI, but ultimately you're doing the validation, but when you're handing over all the responsibility to the AI ​​agent, I think it's hard to trust it 100% right now, at least when you look at DeFi in the web 3 space. So I think a good way to set the limits is to move it to the smart smart contract level. So the number of bad decisions he can make is limited by the number of actions he can perform on the smart contract. So I think this is a good way.

As for the question, it looked like this : do we want the agent to work, or do we want it to be decentralized? And I think that maybe if it's decentralized, it will work better. And I look at the Pit tensor ecosystem and all of its umbrellas, like the subnets that it has, and there's a subnet called Ridges, which are kind of AI agents for coding. And in some benchmarks the improvement was even higher than that of Claw and some agents from Code; and that was because the incentives were very aligned. And so they were able to create an agent that was even better, not at the level of the model, but at the level of the tools they were using.

So I think maybe the best way to make agents useful in DeFi is to make them decentralized and, um, kind of go against incentives, instead of just trusting a model, like an open AI from the cloud or something. So I would advise using smart contracts to limit actions and maybe reduce the tension a bit, and a decentralized way to make the AI ​​eventually work correctly. So, um, I guess one more point to add to that : um, I guess I don't have an answer for you on whether I would prefer centralization or decentralization. But there are two considerations that come to mind. The first of them is mind.

The first of them is , um, controllability. So, when you give autonomy to AI agents and work on behalf of people or/and some organizations, and you don't want to be in a situation where there are some bad consequences because of some actions of the AI ​​agents , and you, um, it's an AI agent, it's not me, and there's no responsibility for that. So you need to be able to track and have a record, and be able to get information about what happened between the time AI agents interacted with humans, and ideally it should be trusted and immutable, and that record should be auditable. So, that's the first consideration. The second consideration is that, um, it's about ownership.

So I think an AI agent is by its nature something that runs on different platforms, and so I think sooner or later we'll start to see more and more AI agents running embedded in different systems. And what's working behind this is the memory and the context associated with this AI agent, and it's a treasure trove for users, and that's probably the most important thing. So how do we make sure that the right people have access to that memory and context, and control permission to that context, and have the ability to monetize or have ownership to claim the revenue from that context or memory? And I think those are two things that immediately come to mind: and whether centralization or decentralization is a better solution, to think about those two dimensions. Hmm, I guess those are pros and cons.

So in fact, trust has already become a bottleneck for the mass adoption of artificial intelligence agents. That 's why I visit Silicon Valley many times a year. You know, I've been interviewing a lot of places in Silicon Valley, which is in the United States and is already a hub for artificial intelligence agents. So when they sell their AI agents to customers, the main obstacle is trust. For example, let me ask you this question because Chat GPT is very powerful right now .

If you allow Chat GPT to have full access to your phone and your laptop, CHP can become your very powerful personal assistant. But would you like to give Charge full access? Who will say "yes"? No, of course not. Right.

I just use Charge for editing or writing something like that. But I would never give my phone access to Charge. Why? It's about trust, because I don't trust...I have no idea what to do with the information from my phone, because tragedy for us tragedy for us is a black box.

We have no idea. Uh, we don't know anything about that , do we? Um, so trust is a very important topic, and there are several ways to address the trust issue if we want to mass- promote AI agents. So I would say that decentralization is one way, one approach that uses blockchain technology, blah blah blah. But this is a long-term effort; it's impossible to do in a year or two, I would say, because blockchain technology itself is not mature at this point.

Since I have been teaching blockchain technology and artificial intelligence technology at Columbia University for almost 10 years, I know that from a technical perspective, I can say that it is still very far from the maturity of blockchain technology itself. And I think right now the only way we can do that is to still build trust in centralized organizations like Microsoft, OpenAI, Google. We have to trust them because we have no other options, right? So we can rely on the government to make these universities implement more regulations. Let them be a little more transparent and tell us as users some information about AI and how things work behind the scenes, and then we might have some trust.

And later, maybe next year, you'll see some AI agent that you want to use in your databases, and the argument is like, " argument is like, " Okay, we're built on the Microsoft Azura cloud, which you can trust, or we're built on the Google cloud, which you can trust," so that's where the trust comes from for now. But in the future, I would say, in 5, 6, 8, 10 years, maybe the decentralization technology will become mature, and then people will start creating AI. We always say that an AI agent should be built by the community built by the community , used by the community, but that's a very nice vision that can't be realized in one or two years. Uh, for now, I would say we still need to trust big companies. Thank you, I'm so glad you're on my panel because you build bridges to the always- following questions.

So, we touched on trust and governance, but trust alone does not ensure the proper functioning of the system. Therefore, agents also depend on access to data infrastructure and connectivity. So let's talk about what it really takes to make these systems reliable and globally usable. Because not everyone has the same resources or digital foundation. So what should come first?

Is it connectivity, management or trust? So what do you think? Shall I start? So I guess I'll keep it pretty simple. Without a connection, you can't do anything.

What is a dataless agent? What is an agent without computation? So, connectivity without connectivity–you don't have artificial intelligence. This is the first. With this, it's like walking without legs without a connection.

Then comes management. Management is, in fact, one of the levels of security. You have a connection. You have some checks and balances. But still, it's like walking on your knees.

Trust is, in essence, what you strive for. This is something we should all strive for in the long run, because that's when you start working. That's when you start getting examples of AI being used in enterprises. That's why startups in Silicon Valley can't force big companies to implement them. And you know, we all believe, and the market valuations demonstrate the expected, you could say, uh, disruption of AI.

But we're still early in this curve where, uh, the trust element needs to be addressed before, you know, large organizations that would be classified as having over $500 million in revenue per year would implement this type of technology and take on this kind of risk that could have some negative consequences if it's not done right. I just wanted to quickly ask you, I agree to some extent. However, I believe you need to have the right governance structures in place from the start. Otherwise, the damage or loss could be enormous by the time you get to it. Er get to it.

Er , maybe it's also worth talking from the perspective of children and vulnerable children. Hmm, for the types of populations for the types of populations we work with, there's usually not even any data about them. Hmm, either there is no connection or data on these vulnerable children we are trying to serve. Um, he's gone. And, as you said, if there is no connection, it is simply impossible to provide them with digital services, including artificial intelligence.

And we certainly cannot use the potential and take advantage of artificial intelligence to provide or ensure even better services for these children. However, if we don't have proper protection, laws, privacy, governance, once the connection reaches, exploitation can occur and you can do a lot of damage. I would say you need to make sure you are laying the right foundation from the start. So what do you think is most needed at the moment? What should be taken care of first among these things?

About all this. I think it's a chicken and egg question chicken and egg question , right? Like you need, uh, proper management proper management , but you also need communication. You have control, but you don't have it, if you don't have communication, then you have nothing, and vice versa. So I think it's a and vice versa.

So I think it's a chicken and egg dilemma. Does anyone want to add anything? Yes, it is possible. I think it's pretty complete, but the connection is critical, because if it's not there, you have a brain that can't do anything. So it must have the ability to interact with other programs and other tools.

But I would add that perhaps this is what is necessary, and I have felt this in the cryptosphere. It's more entrepreneurs, more people trying something new. I think agents are capable of so many things that it depends on our creativity, and there has been a kind of boom in artificial intelligence agents in cryptocurrency. And I like this kind of growth. I thought it was amazing; I thought it was really cool thought it was really cool .

There were, like, a lot of agents, and then the number disappeared to zero, but even completely autonomous, they were able to create their own IP address and collaborate with other agents. And there was an agent who paid $1 to another agent because he needed a favor, and it was very similar to the original stuff, and it's okay that it didn't work out. But I would say that the more people get involved, the more people create these connections, the more people create these MCPs and connectors between applications. Yes, more entrepreneurs are creating technology are creating technology . Thank you, so I think we're at a crucial time now, so when are we actually going to...

B. So, we've looked at what's working today, what's not working, and what needs to be created, right? So let's look ahead, say, 3 years. Will AI agents work quietly in the background of our daily lives, like invisible assistants, or will we still be trying to get them out of beta get them out of beta mode? What do you think?

And I can probably take a step. I actually don't feel like it's still in beta mode. I think in beta mode. I think it's already changing many industries and already generating a lot of revenue. So, in 3 years, I don't know, I think there will be different formats, but I think a lot of platforms could change.

So the way I think about it, we think about current platforms, they're like flat services. So, this is a flat interface. For example, when you work with Instagram, you are part of it, and when you interact with it, you are part of it, but when you live, it has nothing to do with you , but what AI agents really allow us to do is that they really enrich the user interface. It's kind of like how technology is embracing different corners. So in your conversations, in your different kinds of integrations of other platforms...

So, I think that's going to change a lot of things and change where intent is captured and served today. So, even now we see many companies and many startups. They are already making a lot of money automating blog posts, automating cold contact emails. And this is, to some extent, I think, the initialization of the existing Internet. And how long will this last?

But people are still looking for human values. So where will these human values exist in the future? So, in my opinion, in three years we'll see the next wave of giant companies giant companies , like, uh, existing Tik Toks, like YouTube. They are appearing YouTube. They are appearing , but they will no longer have the same format as they do now.

Another thing that's really exciting about large language models is how much information is collected in such a small model that can be distributed in a fairly decentralized way, and that will allow for things like content creation, targeting individuals, and recommendations. You don't need a big data cluster to make recommendations make recommendations , and this can actually happen for you, for your groups of people, and is going to change how we interact with the internet in the future, and I think that's something I'm looking forward to in the next three years. Uh, Max, maybe one question for you one question for you . What backbone infrastructure is still missing before agents become mainstream? What do you think you mean by backbone infrastructure in terms of centralized or decentralized, which would you choose?

Okay, I think I want to talk about decentralized infrastructure because as you know, this is a normal access stage, let's talk about some things related to blockchain. We're already seeing a lot of decentralized, we call it deep contact in general, right? Decentralized physical infrastructure... So infrastructure... So , as I said, from an infrastructure technology perspective, this is far from a mature infrastructure.

When we talk about the performance of the backbone of this infrastructure, we're talking about SLA, if you understand that's a performance metric for cloud data centers, right? Latency, availability, and reliability are three key performance indicators for data centers. But if you use the cloud in a decentralized way, latency is a big problem, right? Accessibility is normal. But reliability is another issue.

So how can we handle reliability and latency comparable to AWS or Google Cloud? From a technical point of view, this is very difficult. For example, you know, in terms of latency, AWS offers you a latency of 100 milliseconds. But in a decentralized way, if you use a decentralized GPU for logical inference, for example, for the logical inference of your artificial intelligence agent, will the delay cause the application to stop working? Then you will have to deploy the application again to AWS or Google Cloud.

That's the thing. But, you know, for many years now, there have been many research groups in academia that have been working on this and training an AI agent in a decentralized way to reduce the cost. And, uh, for learning, it's fine because you don't have to worry about lag while you're learning. But for logical inference, for logical inference, uh, delay is key, right? And reliability is also key.

So decentralization is quite a challenge, and how we can build a decentralized infrastructure foundation to meet that availability, reliability, and latency metric comparable to AWS or Google Cloud is still, uh, you know, I would say, an uphill battle. Yes, I wanted to add that it's also about the power and energy required to power all the GPUs GPUs , the data centers that you just talked about. Um, we all know that I think energy demand is going to grow exponentially. It's already growing, and I think it's a already growing, and I think it's a similar debate with web3 blockchain. Um, now there are those who say that data center companies will have to invest or, um, build the same infrastructure as they have over the last 20 years, over the next few years.

Um, some of these companies are building their own solar power plants, their own energy power plants, and so on. Even nuclear energy is coming back into the conversation and discussion because I think the energy issue is still an unresolved, um, issue that people are trying to explore because we know for sure that it's going to grow exponentially and that we need to find a solution for it. Yes. This will happen over the next three years, which I think is very exciting. So I'm absolutely certain that AI will be everywhere.

I don't know at what level and at what speed, but it will be everywhere, and it will have artificial intelligence, and it will work on all devices. But if we're talking about three years, I think we'll see decentralized versions of AI competing with centralized ones, and I hope that's the case, and I really, really like the bit- sensor approach. I know there are a lot of projects that are trying to do the same thing, but I think the way they've created incentives and the way they're addressing them, like logical inference or even decentralized learning, and it works; so I think it's important that this is, um, an extremely complex project to understand. But, for example, the energy problem energy problem , like the largest supercomputer is Bitcoin, and they built it using only incentives, so I think incentives are key, and if you match them well with, um, the exact problems, I think you can create amazing things. So I hope that in 3 years we will have decentralized AI products that are more competitive competitive , better in benchmarks, and cheaper than centralized options.

Can I just answer the same thing? So, I think it's really scary and at the same time incredibly exciting how AI is going to be everywhere. And, you know, we also talk about also talk about it being in the background, but I think with the development of chips, models, and also energy, maybe we're talking about a little bit more time about a little bit more time than just 3 years. And also, if you think about this new topic around robotics, AI is going to be physical, right in front of our faces, right? So when it comes to the physical or digital side, trust has to be woven into the very fabric of AI, right?

We already have something like dark factories in China, where robots work 24/7 in the dark. So, you know in the dark. So, you know , this is a topic that only moves in one direction. But we've also seen scenarios where agents don't recognize who their master is, and as we move toward intelligent AI, you and I, we're entering a world where trust will be fundamental to this technology. Right now, innovation is ahead, regulation is ahead of trust and governance, but ultimately, as we move towards mass adoption, I know there is a decent revenue for startups, but I would still call it decent.

It's not, um, it doesn't matter on a global level right now. Aside from, you know, things like GPT in chat, we don't really see it in everything we do , but that's the direction we're moving in, so it's exciting, but it's scary at the same time, it's a fine line between, you know, innovation and regulation, which totally reminds me of that movie "E-root, me of that movie "E-root, " oh my God, okay, my voice is almost gone, I'm so sorry that time is flying, and I want to end on a note that will really give everyone a conclusion. So I ask each of you to complete this sentence with one line. AI agents will only succeed if you go first, without any pressure. I think for me it's almost like a sentence structured like, "Uh structured like, "Uh -uh," is wrong.

I think AI will succeed. The point is made after the AI ​​agents succeed, and it's just a case of if not if... great. Good. What about you, Max?

Uh, if we fully understand how AI works. Oh, my God. Same answer. I was just going to say something about explainability. So, given the opportunity to explain, I would answer the same way, that he is already succeeding.

So just keep doing the same thing . Good. So, to change things up a bit and get back to my topic again, if people are put at the center and are the ones who will try this. Good. So let's see Good.

So let's see what our collective mind says. Good. Uh, great. I think this is the perfect note to end the conference on. So, uh, maybe I have one last question.

What is the funniest or strangest use case for an AI agent you have seen so far? The funniest. The funniest or the weirdest, or both ? The strangest drawings of fish. I saw a website that deals exclusively with this.

Anyone else? Nothing funny in your daily life? I'm thinking, um, an AI agent who became an artist. He self- published his own albums on Spotify, and they were pretty good. They were not bad.

That was fun. Interesting. OK. So thank you to everyone who stayed with us until the very end. I thank our amazing panelists for their insights and all of you for being a part of this journey.

Enjoy what's left of the day left of the day , and let your next AI agent actually do what it's supposed to do. So thank you all, and let's support our human agents here on stage, please. Thank you. Thank you.