15062026 CHAPEL GIOVANNI DE LILLO

Okay. Uh, one second. He will talk about probable autonomy, the trust layer for AI agents in regulated industries. Thank you.

Thank you to you. Welcome. been an event enables about AI. Unfortunately, I would have never believed because it was like almost impossible. But now I'm so happy to be here because someone gave me the opportunity to talk about AI and cutting edge technology heading the world.

Um imagine being a banker, a director of a bank, a director of an insurance uh company, being a director of a uh of an important company. So a regulated the industry or an enterprise or we can exaggerate a government would you implement an AI right now? I honestly don't think so. I had spoken with sign director and there is always the same problem that as everyone said yesterday and today we cannot trust the AIS. In 2022, OpenAI launched CHA GPT 3.

5 the state-of-the-art model it was great people started to believe that AI could enhance the world augment the productivity be better for the society. After that there were many companies who followed open AI such as Antropic, Google, Moonshot. But everyone has the same problem. They continue to develop a strong AI without having a secure one. If you are a bank, you cannot integrate these models into your software because if there is a problem, you will get in trouble and you have the responsibility for them.

they don't work for you. You are responsible for what they do. So every time you got to check them what they do, how they do, why do they do that? So it's like quite scary sometimes. Um the majority of director told me that they will be um very happy to integrate these models in the system, but now they just can't.

And I am happy to say that Hyperium Multid developed a Hyperium Hive mind which is a solution that is fixing this problem. I would like to go to the next slide because to explain what they made and how we are working on it. Thank you. Um then um the reason for why I have developed this product is because I had this problem too. I used to code with AI lots of times and sometimes AI uh starts to make decision by itself and this is not even possible.

How can a model take decision for me? Who are you? Why are you doing that? Um so I developed a system that will never do that again. Yeah, it will be great to go straight back.

Don't worry. Um after that we are seeing how the majority of companies are working with uh bank to integrate their models into the software but then they are not accepting it because they cannot audit the action that the agents do. So Hyperium developed I mind which is capable to audit every action that an agent does. Let's say for example you run a prompt that we call a mission. So we define again this word um telling to the II please make a uh deep research about my company how we work what we do and why we do that and tell me if my company is compliant to the AOI act and the I could tell you yeah you are compliant so you can go uh without problems but perhaps you could receive a bill of thousands of euro because you were not compliant and this is very dangerous and who is responsible within between them the AI or you unfortunately you will be responsible

yeah I prepared a slide absolutely

um and I'm I'm going to show you um how we made it first of all um we audit everything so if you ask to our model to our platform to write a code to uh be compliant to uh to be like to substitute someone they will not do that because we are uh under the AOI act article 9 40 50 and 30 and this is truly important for companies so they can finally integrated in their systems. Um for instance, you are in a courtroom and you want to show them uh that AI made a mistake. Now you can't you can go to jail because AI made a mistake for you and you have never asked to your AI to your large language model to do that and this is absolutely dangerous. So it had to be fixed. So we can say that hive mind is a layer a layer two between who develops this model and who use them.

Um then uh another important thing that is that we are compliant in the article 14 of the AOI act. So uh every like we classify the actions. So when you run a mission the AI will divide this mission in uh three parts. The first one is green, the first one is orange and the first one is is red. If it's green or orange the AI can go automatically without asking you what to do.

However, if they classifies it as red, they will never go without asking you for permission. So, you have to deny or approve the actions. And this is very important. Yeah, it is there. It's in Google Drive.

Yeah. Um, but no problem. So, I'm able to tell you what I made in two years continuously without showing you that. Um then um it's very important because we orchestrate AI too. So why should you like uh use entropic models in our platform?

Great, there are no differences. You can use your uh the entropic models in our platform. But we have a large language model gateway rooting. So you can basically choose between more models who can work together like an hive mind. That's the reason for why we have this name.

um which is great. Um a problem has been this problem has been finally fixed. If you are um the director of a hospital, you can finally work with the sensitive data of of your customers. Um the reason for why I'm telling that is because um I spoke with a lots of people and they asked me for that and we finally develop it which is great. I really want to show you the platform because it's great without just talking about my company.

Unfortunately, it's not possible. Um, but like uh imagine how awesome it will be to to have to finally have an AI which is auditable, which is traceable and that will be that will never be wrong with you because you will be responsible about AI. This is an important plot. It's an important plot because u um lots of speakers uh told you yesterday or today too that they um they talk about trust like Andrea did and how can you trust an AI? How how can you trust someone who is not human?

Someone who cannot take a decision because he's unable to do that. The AI has been trained for work under a human not substituting him. So with this new algorithm you will be able to um work with both of them. So it's like a new parigma which is great. Um can I please uh go to Google drive because I really want to show that is um important.

Um so I can finally uh show you what we made. Like someone deleted it. Great. Great is empty. No, never mind.

Um, could you raise your hand if you never trust the your AI? If you were working with that but you couldn't just trust it because you had no idea about was it was doing. Um, imagine like you are a product company was going to able to launch your product. Um, you could use AI for coding as as he said. Yeah, it's great.

But what happen if they if you ask to AI to like um develop your product with these ways and then just um and then like it it goes wrong it it starts to take some bad decision for your name for yourself. It is very dangerous and you cannot show it in a courtroom. Um furthermore um another function that we implemented in our platform that is great and could be used um is that um we have an algorithm called model agnostic. So we basically implemented personalized open source models for companies so that datas are now secure. Another question that the government uh are asking to themselves is okay great I can integrate I can finally integrate a large language model in my platform but how can I use them what happen if like open AI um steal my datas data of my population it is illegal it doesn't follow the the GDPR rules so if you open source a model an AI model you will be finally able to manage your paradas and nothing will go in training.

So very important I speak I have spoken with some politician yesterday and they completely agree with me because it is the main point of of the society right now. Um we are seeing that um the AI market is increasing a lot. We are talking about an impact of more than 10 trillion dollars. So uh enterprises will start to integrate it. Uh we believe that we are just starting with that and u um the the market will uh will increase a lot.

Uh yeah is not finished. This model will be are stupid if compared to the future models. That is great but um it's important to talk about safety. It's important to talk about um compliance and it is important to talk about how uh companies manage these models. For instance, we have seen that um the US government has just deleted uh the model of entropic fable five.

All right, they did not delete it but they uh the enable is not available right now. And why did they do that? They didn't do that because they cannot verify what the does. And this is not possible. Imagine how hard it will be to accept for for a government to use an AI that that is not like uh that has no safety for your population.

Um which is the main point. So if you like like AI if if you like this word perhaps you don't work on it but you are like enthusiastic about this word um take a look about what we are doing here in Europe um and not don't do not crit criticize sorry um the laws are uh coming like the AOA act because um it's for our safety and it's truly important and this is what Eperium is doing right now. Um if you are a lawyer, if you are a doctor, if if you are part of a government, um you can go from a company to a AI company to integrate your models, but they will steal your datas and that is not acceptable. Um furthermore, um this is not just an idea. I'm not pitching it.

I don't care if you will use my platform or not. You can just go to another company the same. But I just want to make you aware about how important is that. Um, nevertheless, uh, we are not alone. I'm not alone.

We are backed by the UK government. So, uh, the UK government gave us 20,000 GPU hours in the supercomputer in Cambridge. So, we developed a proprietary model and we um, this is an an executor model. So it's completely proprietary and we can use them to uh plan other agent. So it's a new paradigma as uh as they said before agents are now able to spawn other agents and imagine how dangerous it could be if not managed but if not managed well um like if an agent is able to spawn another agent and then this sub agent is able to spawn another agent.

We are talking about a new co-working system um without humans and this is not acceptable for safety of the society. Um, an interesting an interesting topic is that I've been into a hospital and they gave me the opportunity to see their software and it's funny that admitting that here in Italy uh they have like very old software old like uh I think that they develop it in 1990 so very old and they are not um able to integrate the AIS because um models are not prepared for doing that. So, um this is a usual overview, a deeper overview about uh uh how the world is is moving now. So, I'm I'm continuing to listen your speaker and they say, "Yeah, well, the AI is going to be uh more more intelligent than us." Great.

Absolutely. I completely agree with with that. But be safe with that. Um don't stop uh when AI is working see what they what it does because we had internal problem in cyber security working on a project and they start they are starting to um take decision alone um if you work with sensitive datas um you cannot use them for um these AIs for um for working along with them. So it is very important.

And then um a funny topic that I uh I have seen is that there are lots of companies who are starting to integrate AI in their system finally I would say but they use uh large language model with the um APIs. Yeah, we finally have it. Um but these models are not uh are not secure. So why do they do that? I have no idea.

Yeah. Um I go go down a bit so we can go faster. Um I I have written that a law um is not a is not mechanical but is a a requirement. I mean that is normal if a government implies something. It's normal if uh people uh if companies starts to be obligated to uh respect the new laws because it could be uh is is just to protect the population in general.

Um can you zoom a a bit because this is a screenshot of the platform. Um so yeah this is a hive mind we cannot see that properly but we can see that there are uh this is the auditing system. So here there are uh the all action that have been made by the agents and there is like a hash code with 64 numbers which is saying okay great this action has been successfully done with known with no errors because what what happens if you work in parallel with AI it happens that if an AI is wrong and in your end the eye is going well they will combine the other with uh the like the red action and they will destroy it again. So it cannot work right now and these models are very intelligent. So it's strange.

Um then um maybe we can see yeah um another important thing is the um is the uh article 14 of the AI act which implies that if an action is a pivotal is crucial for the development of the of the system. They cannot go without asking the human uh to go like automatically. They have to ask to the human always to approve or deny the action. And we develop a simple platform, a simple front end to uh to accept or not the action. And what happen if I do not give them an answer?

Great. They will just don't continue to do that. They will stop. And this is important because we can finally use these models orchestrating them as they are currently doing but in a uh in a way which has just more safety. Um after that um this is a strong platform.

We uh developed um lots of loc so lines of code uh to have a strong platform. Um furthermore, we also have some certification from the government. So it's very safe and they finally approved what they made and they are um like um confirming that what I say today, what I'm telling you today is real and they have this program right now and they will be glad to in integrate this kind of system in their platform. Um trust capability is the bottleneck. So like uh someone said before like Andrea said before we have to trust this model because this model are going to spur head in the society.

This model will be the state of art in everything. We are seeing AI a lot in coding. We are seeing AI a lot in program in I don't know engineering in healthcare but not in in everything. So in four, five or six year AI I I making you aware that AI will be integrated in everything and if you are going to integrate this kind of AI it will never work because it it's going to be very dangerous. Um then um I like to make aware that u I'm quite young um and I want to um tell you that I'm 17 years old right now.

and I have already developed a uh a strong product. We have done it. We are backed from the UK government. Uh we are backed from a venture capital. So um is is not is not not just an idea but it's something that is real.

So it's something that has been created properly and approved by two governments. the Italian one and in UK soon in United States of America too and this is something that I wanted to say not just to have a clubs I don't care about that it does not matter but to tell you that you don't have to uh give shakeless give limits to yourself if you have an idea if you want to do something and no one has done before if you want to pitch something if you think that it can work if you see the opportunity in your guys do that and do not stop because someone did not do that previously. Um at first it was very hard to develop this platform and sometimes is very hard to extract to to harvest this value because it's another platform and uh we are fixing we are resolving a problem that currently don't exist because companies are integrated it are using AIS but they have never had this problem. So they say like okay it's working we don't care but in but in like one or two years when this AI will will be more autonomous we will start to see models that will like deal datas will manage lots of sensitive datas and it's it's going to be catastrophic if someone do not integrate this kind of platform and we we're looking for like a new law to regulate it Um and it's great that we are also as I said before backed by venture capital. So there are people there are teams thousands of people who believe in what we are currently doing why we are doing that and it's great to know because this kind of endorsement allow us to work properly to continue to work also when someone tell us to that the product don't don't work but in reality it it will be the bottleneck of everything because yeah you can use AI you can use uh the best model the state of art model that cool uh Antropic could release uh Methos, OpenAI could release GPT 6.

5. It will be great. Yeah, absolutely. They will be just better, better and better. But the there are no limits for the better, but there are limits for the orchestration.

There are limits for the safety because um it's important that people start to think about that. Um furthermore I also went at University of Cambridge to um to tell them to show them what they what what we have to do what we have done and what what are we going to do because um someone before um was talking about the open sourcing model. What does open source mean? Open source means when a company release a model for free but not just model but the codes so people can just download it easily download it in their computer and then use them why does a company do that are they like philanthropic I don't know sometimes yeah they are sometimes not it just for their interest uh but it's great because we made it so it's available on on a game face our um proprietary model our AI I model um is is is tiny so you can download it in your computer. Uh it's not so uh yeah don't worry it's not so large that that you cannot it's not the next GPT6 but is a good model for planning other actions.

Um then um I want to uh thank you uh emerging ventures that was the venture capitalist that backed us. They are not here for sure. They are in the United States of America but we are recording. So um if I can thank you for backing us and giving us the opportunity to work with you and we will continue uh to uh we will continue your our our project. Sorry.

And then the bottleneck is not capabilities is wherever we decide to make autonomy pro approvable. So um if you want to um use AI in your ideas if you are finally uh happy and excited to uh use AI to start your uh own business it can just not be related to AI but AI can help you do that now. Don't worry about these problems. But when you become bigger, bigger and bigger and you have to manage sensitive datas, trust me that it will be very hard to manage them because you you could just get in trouble by giving AI too many permission as the majority of people do without just um looking at it. Um then uh if you like uh if you have this needs um I'm talking with many people who needed this that was so this event was a great opportunity for everyone to talk about that new topics enable is crazy to talk about the orchestration I cannot even imagine it uh here are here is my email my um private one and the business one if needed and I also open for some question without business stuff.

Yeah, sure. Uh there is the code of our website. There are lots of paper, free paper. So you can read them understanding our milestone, our our road map and how we are working that is uh very interesting because we started up with zero and now we have some resources. So we are going straight uh well uh my link.

So uh it's everything. Sorry if I did not show that u at first but we had not the possibility. I had to run a bit but yeah I think that everything is here. So if you want to use AI use it be happy with that but um keep attention of what they do and never give toi the permission to take decision alone. This is the main point.

Thank you.

Thank you. Thank you Javanni our younger speaker. Well done. Well done. Well done.

Okay. And