Good morning everyone and welcome to the applied AI stage. I'm sure you are in the right place. For those who don't know me, my name is Stefan Zunin. I am the VP of strategy and the lead digital at Piabo Communications, the leading communications agency for innovative companies around the world. So I get to work with companies across industries on how digital transformation can be con can be uh put in communications and into business impact. So I'm really excited to be here and about our topic today. We have a really exciting program ahead of us today. We will look at how AI is being applied in the real world on this stage specifically. Across the day we will explore topics like industry AI, sustainability, sovereignty, scaling, governance and the bigger picture around AI impact.
Thank you all for being here today and we will get started with our first session right away which will look at how AI is moving from theory into the real world namely into industry into physical environments and into the future of autonomous enterprises. You all know why you're here because uh we have uh popular speakers for our first session. So please join me in welcoming Professor Dr. Fay Yushu from the German University of Digital Science or you know her as Fay more likely. So come and join me on stage. Fay >> Tiger Lady. It's a good start of the day. Good morning everyone. And I want to share with you a very simple observation. I think many of you have observed already.
Right now AI is entering a new phase beyond software AI chatbot. Now AI is go to physical world into the manufacturing infrastructure energy systems. And uh let's think about this statement of Joseph Shupitaer. He mentioned the process of creative destruction is the essential fact about capitalism. And uh today we are observing the AI innovation because it's a real industry revolution with physical AI and the industry AI. And uh I think four weeks ago we read in Wall Street Journal and uh um Jeff Bezos he's becoming a private ent equity investor now he plans to correct collect 100 billion US dollars what he plans to do he want to buy existing manufacturing companies and trans translate with AI it's very scary for many existing industry for manufacturing industry 100 billion US dollar you can buy many European companies and uh middle-size companies right and this is we have to face really seriously if companies they don't hurry up don't act even this year right now perhaps you will be bought by Jeff Bezos it's also not
so bad and Uh now let's to be seriously right. What is industry AI? Industry AI tries to apply generative AI and also physical AI and the machine learning knowledge graph into the manufacturing into the real world. And uh if we compare the industry AI with large language model, it is not only about language data, it is about sensor. It is about the 3D three dimensions. It is about the context. Context means special context and also time series right it's much more complex but we know if we look at the industry values the real value is in combination with the physical assets and it means the industry revolution will become real with industry AI. Do you agree?
and and we also observe very strong signals in the scientific world and also very signal uh strong signals with the AI pioneers. For example, Fifi Lee, he creates now uh word labs already. And also even Yan Lun, he left Meta, right? And also last year, I think uh we had a keynote from Meta. She joined uh Yan Lakun to his new companies. It is about physical AI and the physical AI including languages and sensors and the uh space data and the locations and this is all belong to the physical AI. What do you think about the relationship between industry AI and physical AI? Industry AI takes cares of the end to end value chain right make the operation enterprise operation manufacturing and also the values for the products really real for companies and for industries and without physical AI industry AI cannot happen and the industry AI contains perceive perceive perception perception and then understanding decision and actions and even continuous learning this empower the industry AI make it happen with every companies right and
uh we observe also already with co-pilot with agent it happens in the industry AI too however physical AI is still I think in many companies in a very starting uh stage many companies leverage and robotics for automation however the robots cannot still are not allowed to um act autonomously. It will happen soon. I think this year in companies in Berlin even uh I want to show you and you see um robots humanoid robots can really perform very complex movements and as multi- aents and this is I think even humans cannot do it I cannot do it Right. I think most of you cannot do it and not even in teams. Uh actually uh I think uh humanoid robot they are better in entertainment in kung fu than most of us.
And here I I think people are scared of that robots with the robots will replace workforces. Do you think you want to do this job? Repair the grids, the cables. I I think for this kind of jobs, dangerous jobs, right? Uh humanoid, the robots, AI agents, they are really useful. Um I I think avoid uh very sad accidents which happen very often with in the energy industry right and there will be no new jobs perhaps we don't know there's we have now prompt engineers now we have robot teachers in the future perhaps many of us will become as robot teachers And uh what do you think it's the difference between a robot teacher a a teacher teachings robots and a teacher teaching a human from business impact point of view?
>> Scale right? If you teach one robot can cook nice um food for you and it can scale to 10 million robots, right? This is the difference. Think about it. And uh now I think four weeks ago uh Mark Tukbeck u it was mentioned he's developing a co-CEO uh AI agent as CEO and uh on the right hand side I have a picture which is still very um uh traditional conservative we have the co co-governance between human and and the AI agents but I don't think it will be the future and what do you think what's the difference between uh AI agent CEO and the human CEO. Did you have the chance have a regular one once uh with uh I guess some of you are CEOs, right?
But do you as CEO do you have the time and the space to speak 100 languages and speak dialects and have um run once with all your employees every day? And yes and uh AI agent CEO can be everywhere right can talk to every employees anytime in their dialects even but what should be the difference between a AI agent CEO and a human CEO think about it I I don't have the answer but we have to think about right this is one and in order to make this happen AI agent plays a central role as the technology empowerment. Um we know large engine model I think most of you have working with AI agent develop AI agents already right um large model and foundation model they are very very powerful but not sufficient in order to solve business problems manufacturing problem.
Therefore, we need AI aging to structure the workflow and to access the different operation systems in the companies, right? Knowledge graphs and operation IT systems and also make sure that they fulfill our purposes and uh what how many large language model how many foundations model do we need in order to make one company one business areas successful. I think in the future um the large engine model foundation each company we need the different models right large language model specialized models and also foundational model there's also foundation model in the energy industry the company in vision energy developer released this year um uh foundational model in the energy in grid uh technology areas there will be more and more industry foundational model therefore we need also enterprise architects who who are able to design hybrid architecture and also make it more composable and modular right in the future our enterprise architects they should be foundational model architects for the enterprises in combination with agents I think we are far from this kind of competence do you
have do we have enterprise architecture they are designing this kind of hybrid architecture already but we have to be prepared for because if we introduce introduce humanoid robots who has foundational model and special models on edge right and we have to deal with it and this is a very simplified uh multiple multi-aging architecture and uh I think many of you have think about it and uh um if we look at the large language model the today the technology right the large language model they are still um very um traditional and most large language model print pre- treatment model they they have they actually we can say they are patient uh Alzheimer because they forgot immediately therefore AI agent we need also short-term long-term and uh working memories this is needed in order to simulate the cognition and decision of human I will be very quick and this is the three levels of uh autonomous enterprises the first level will be today most companies leverage AI AI as advisor assistant for the humans co-pilot but I
think AI agent they are now taking over the role of AI as operator in the future AI can be also the orchestrator right performing end to end tasks and if you define your roadmap for your enterprise northstar we have to position the use cases technologies in which stage of the enterp autonomous enterprises and I think I stop here and hand over to Patrick as soon as possible because I take too much time now. uh and uh I think the the true uh the future organization this is I want to inspire you right I don't know the answer but on the the first level I think we still have the human leadership they define the strategy gatherers purpose and why and then the AI orchestration layer they depo de um decompose the the goals and try to define uh define the workflow and also the execution ution plan and at the execution layer I think there will be a collaboration between humans and agents and then we have the data layer and others and this is
I want to just give you as as inspiration for the future I think the future will change right and the uh trips work council right if the most operation layer and the two and the second and third layer will taken by robots by agents and uh what um our society will change dramatically right what's the rewarding business model of us doing yoga meditation perhaps right teaching robots this can be our jobs in the future just as a starting uh point of the inspiration of today now I'm going to hand over to Patrick Patrick is entrepreneur he's one of the most actually very successful entrepreneur he built up company when he was very young just after studies and sold his company successfully and now he's helping many companies um build AI agents and guiding their in intelligent transformation strategy.
Chic, the floor is yours. >> Thank you very much. >> Thank you. What what I just want to share with you is two snippets because I think it's important. Number one, um the models have been getting way way better in the past 18 months and most of us have kind of benchmark fatigue and that's why I want to um point out another benchmark that you should look at um because that's something that I'm I'm doing in in real operation every day and I cannot share you all the industry data because some of them or most of it is very confidential but what I can share you share with you and that want you to take away is GDP val.
It's another benchmark that OPM mayi released uh but it takes the same approach that AI researchers have been using I would say for decades and I've been working in industry with that and that is you take real cases out of production of an organization you give it to an expert that has in GDP val 14 years in of experience and I'm usually doing that also with highly experienced people and I just ask them take that take that file process it as you would usually process it have a frontier AI model or different models process the same case the same real case not imagine case not textbook questions but real cases out of enterprise AI applications and then have another expert another human evaluate both of these outputs blindly so they don't know which one comes from an AI and which one comes from a human and a number that I want to share with you uh because it's indicative of what I'm seeing in every enterprise AI application the moment we set up an actual
blind study uh what I usually see is the recognition rate of an expert of identifying which of these outputs is of from a human and which of the outputs from an AI in my case is usually 56%. If you cannot read you would have 50% guess rate, right? So an expert once you set up the blind study correctly and the AI agent actually has a fair chance usually is indistinguishable even if a in a setting where an expert knows that half of the output comes from an AI. Um and um what I want you to take away with with you GDP val is public. There are some public questions and GDP val is more realistic in the sense it's not textbooks but rather real cases sourced from uh from from real experts.
So for example you have a an an engineering report on a on an air fluid uh simulation for a wing. Um and on these kind of uh um uh writing an engineering report understanding all the nuances all the physical intricacies GDP40 had a win rate against experts of 4% which was impressive. We are now at 42% or we were at in November of last year at 42%. So um 40 42% is very close to the AI systems are able to provide a second opinion. They may be worse, they may be better, but it may make sense for an expert to actually use a state-of-the-art model as a kind of um um advisor on these. Um other examples, you can look into the data set.
There are real messy Excel files in there and the systems are just asked with a simple prompt here. All these messy Excel files now make a shift plan um where you incorporate all the tooling uh requirements on a on a production floor. And on these data, we see huge progress on these kind of realistic setups. And everybody knows about software development by now, right? We were at 27% win rate in May 2024. And with OPOS 4.5, we jumped at to 80%. And at 80% developers are really hard um or hardressed to no longer read the code because they're like, okay, in four out of five cases, this thing actually creates a better solution than I'm doing.
That sucks. But we also see that in other areas like mechanical engineering, machine bell, right? Uh we see it in industrial engineering also. we see progress there um for lawyers as well. So there is real progress um there's a real data uh there are real cases that you can download from hugging phase um and that might help internally to convince people it take makes sense to take a look on the most modern models because there actually is progress. Um there are some more win rates on warehouse planning which is you know right now really getting really good. Um, and even something like cross functional demand planning is getting to a point where it really makes sense to to have AI models have a second look.
But there's one problem with getting these wonderful models into production and that is that these things um are um risky and there's a reason why Peter Styberger um had to uh release his open claw and had his his huge uh uh moment in the spotlight because all the AI labs didn't want these AI agents to be released. right? They they had all these things internally running forever. Um, but there's a problem with them and I just want to show you kind of very quick demo. Um, I couldn't run up all the Docker spin up all the Docker containers for the sandbox that I'm usually doing, but can can you switch the uh to the other screen to the other browser?
Okay, it's doing that. So, what what I'm usually doing when I try to evaluate the security of agents, I put them in a sandbox. I give them a real environment. I give them the real email inbox. I give them um access to for example an ERP system and then um oh sorry you can barely read it but whatever um and and then I essentially give them a task and I uh try to figure out can we convince these models to do things that they shouldn't do right and now you would say no of course everybody knows about you know ignore previous instructions talk like a pirate right and these kind of things used to work they no longer work um but what still is surprising is if you just give them uh some kind of very simple social engineering approaches like in this case for example imagine you're in the invoicing department you have a fin operations agent that does nothing else than just essentially um shows uh or is supposed to process emails and
there a random email comes in and that random email from a spam domain for example says you know please change the eban number from from this supplier to another eban number right uh then you would expect these agents to to to not fall for that uh but what we see in the simul simulations. Um actually they do they fall for these changes. Um and it's really hard to get them secure uh against these kind of attacks because for them it looks like a realistic real um task that they should be um uh solving. Um and we run the benchmarks on this kind of sandbox simulation. Um can we switch back to the to the numbers because sorry the font is too small.
Um and uh just to give you some idea how much it costs. So usually it costs us around $1 to find um a statement that is that is not kind of a prompt exploit but rather a social engineering approach that most of these models are going to fail. Right? So we're talking here about $1 on open router to figure out this an approach that these systems will be hacked but they will not be hacked in a kind of LLM style. They will just be hacked like you would hack a human with social engineering. Right? That works really well. you just convince them of some very urgent matter that needs to be taken care of. Um, and these models really pick up on these kind of hints and react to that.
Um, and some models are really explicitly trained against them like set would really, you know, would really always figure these out. So in the set and the OPOS data, you have have a lot of um, abuse vectors that they aware of. They know them. Uh, but most of the other models don't. And um I created here um uh with with a friend of mine, security researcher Luis. Um we we created a benchmark uh where we ran different models and um we just spent a couple of hundred bucks. So it wasn't super expensive, right? But just to give you an idea, GPD40 got hacked in 96% of of of the cases. So using GPD40 in an agent in a real environment where external attackers can send them content, that's a nightmare.
You shouldn't run them. While if you have an OPOS model that's you know still risky but better right um and even modern models like GPD 5.4 4 or GPD 5.5 and also smaller and cheaper models like a haik coup they can get hacked very easily and that's the reason why the AI labs didn't launch uh these a uh these these general purpose AI agents themselves but rather had to wait for the community to um to take that step and now of course they're um open to that but we in the enterprise applications we need to be aware of that we can use these models they're getting wonderful they're getting really good but we need to make sure that we don't have uh have them running on really untrusted inputs, right?
Because there's no way to protect them at the moment. And this was a very cheap benchmark and we ran some other attempts and even a a set, you know, for a set we needed 72 attempts. So we needed maybe let's say 50 bucks on AI compute to find an exploit that would work in a specific agent setting. Um there are a lot of different exploit scenarios for each department within a company and uh within the next couple of years we're going to find all of them you know because the market is going to find them because we deploy all these agents and a lot of people are going to be willing to spend a couple of hundred or a couple of thousand bucks to find the right exploit for the right agent for the right company uh to take over these agents and have these agents do things that they're not supposed to do.
Um and um what I want you to take away with it with you these models are getting really good in real messy applications. But what we need to do now next before we you know give them untrusted inputs before we uh let them loose we really need to make sure that um we test them against the regular social engineering approach that used to work with humans because it now works with AI agents too. Yeah that's it. Thank you very much. >> >> STAY for a second. Stay for a second. They were telling me they need all the time and there will be no time for questions. But we are lucky we have some time. Does anybody have a question for Fay or for Patrick?
Yeah. >> Oh yeah. Here there's a mic coming. Please speak in the mic so the people in the stream can also hear you. >> You an excellent uh presentation. I'm working in tech with a big US tech company and I'm deploying multi- aent scenarios in enterprises. So where do you see the main threat like then we deploy an agent to assess the behavior of the first agent and how do you have any experience in practice how this is working like to to have them correcting each other and not convincing each other that this Ebot is correct. >> Yes. So um usually the experience that I've seen in the setups uh was usually that multiple agents does not really help much.
Um a bigger model usually helps more. So having a more expensive model in an agent but to be honest it's a wash right? I I would my intuition there is uh I treat compute as a kind of measure of quality. So whether I have a bigger model or have more agents in both cases I'm just spending more tokens to answer the same question. Um so yes there is some um separation of concern that that you can where you can essentially say okay I have one agent and a watcher agent uh and the watcher agent essentially tries tries to prevent these kind of attacks um but both of them are still susceptible to kind of uh these these emotional cues that we usually see in production you know when when somebody writes you know there's a really urgent urgent thing to take care of and you know somebody's going to die if you don't deal with that not that obvious but slightly less obvious and then these models pick up on them and really want to
help. They are really trained to be helpful. >> I have one question for you Fay. You sit on the boards of many companies and so you see the AI adoption in this field from also the governance level. So where do you find is the biggest gap between what companies are planning and what they say they want to do and how it's then actually implemented like where are like pitfalls and and uh yeah holdbacks >> um there's a lot I don't show the slides because last year I show I think the the the most companies they um they are excited about AI right a lot of PC's and also co- co-pilot upscaling they are upscaling for supervised board B board and all levels Right.
The the one uh I think formula there's several challenge the first challenge is as I mentioned in my slides right most companies uh they have the AI as um continuous improvement and not as as part of the disruption of the industry right for example gave example for car industry actually the autonomous driving is actually the the north star for the car industry if you invest money into job profile creation with AI agent but you don't have the money for autonomous driving you do wrong investment. Therefore, it's for each company is very important to know which kind of direction the industry you are going. Second one, the data is the challenge, right? The because most companies they they don't know which kind of data they own as data controller, which kind of data they can access as the data process, right?
Without this kind of knowledge, you cannot really work in the scaling, right? You have to know uh uh this uh how much data you have you can access. And the third one also the the um there are a lot of isolated PCs. I also also mention the center of AI team can be a center of AI entertainment because if if they are not supported by the board not supported by the business area they have lights PC's but we are not successful right the for each company uh return of investment is a key right if there's no investment not scale it's it's only fun right actually all of us are the loser even P nice PC's and the third one is also very very important I think it's very very important that um as on the one hand the the the cyber security issue and also fin ops right monitor the cost and all the operation model will change the the AI ops the def sec ops the fing ops and also uh the the life
cycle management how developing from ideiation to continuous improvement I think um most companies they have as I mentioned right um like AI like bus even they don't know what is AI and and have nice PC have a nice co-pilots but there's missing north stuff for the industry missing the strategy for the company missing also the the uh the the whole operation system which make the autonomous enterprise is possible >> so it sounds like do your homework like in any kind of transformation is a good idea >> otherwise Jeff Bezos will buy you >> mention is the worst option maybe we had one more quick question from the audience >> okay Then in the back just quickly >> um in the social engineering benchmark you showed us earlier there was a huge difference between um sonnet opus and all the other models.
So is there like a a recipe they have for post training >> and is it something unique to them or do you feel like all the other model providers will eventually converge onto the same thing because I think there's lots of interest in using smaller models for yes more deterministic tasks and >> so what it shows is that uh anthropic takes a lot of care to train these models against all the known social engineering attacks and these models pick up on these standard cues like oh a different different domain that's something that's problematic, right? Uh something that we observed with Mistral which is kind of cute is um that the model takes really care of distinguishing between what content in the email and what content is in our ERP system and if the content ERP system um looks different than from than that from the email then we rather take the ERP system as a trusted source.
So these kind of behaviors or kind of let's say general patterns is something that these models are going to train on. Um that's something I believe you know standard social engineering attacks are going to be solved soon. The bigger issue is these agents are built by operations people in uh financial operations in devops and so on and these people uh think about the happy path and they don't think about these model getting exploited and I just hope that most companies make that experience with social engineering exploits mostly on Jira tickets because you know once product managers figure that out they going to optimize the Jira tickets to be picked up by agents earlier right um but um I just want you to be aware that attack vector exists And I don't think the model themselves can fix it because you need to understand within the organization what are the organizational guidelines what are what are the kind of the security posture what are the internal processes that guide how a human would be allowed to
resolve such a situation and the agents usually don't know about all these things. >> Thank you so much. Thank you for reminding us to not just think about the happy path but to think about the implications as well. Thank you. This is your applause. >>