Good morning. Oh, you can you can also I have Yeah, I have a warm uper with me. Yan. That's his real name. Thank you so much. We already um acquainted each other at breakfast. Yes, it's his real name. Hi, I'm Kathleen or Katen. Um I'm your MC. Yesterday someone told me they don't know what an MC is. What is an MC? ceremony. Thank you so much, master of ceremony, but I'm also the CTO of this room today, the chief time officer. So, I am also sometimes then the party pooper, which means when the speaker is too late and they're running out of time and I don't want to say we're running out of time, then I come on stage and try to be gentle and help them to finish their sentence.
Although they are all really accomplished speakers that's why they are here. So we want to hear what they want to help us to understand but unfortunately we have a lot of speakers today. So we have to be a bit um they told me I have to be the timekeeper here. Um my background is in science communication. I worked a long time in the Fraunhofer Association. So um I know a bit of AI um and virtual reality and augmented reality and all these things. So um I can also ask a couple of questions but only if our speakers stick to the time and we have uh Faben here in our room who is helping also in the room to uh get your microphone when you have a question at the end.
Um do you need anything else? You have seen the program you know where the restrooms are. Anything that is catering please um ask Faben or someone in the room. And without any further ado I don't want to take too much time from our first speaker. So let me introduce her for those of you who don't know her. Um you must be new in AI. Please uh join me in welcoming to the stage and thus opening our stage professor Dr. Fuku. She is a distinguished leader in AI research and industry uh with a wealth of experience spanning from academia to the boardrooms of major technology companies. Well, some of this is for instance, she was the senior vice president of S SAPSE where she defined and orchestrated the company's global AI strategy and she was a vice president and head of the AI lab at Lenovo in Beijing.
um in this session. She told me this morning she will also be accompanied by a long-term collaborator Patrick Monk um who is working with her since uh over a decade and his passion is and well the seeing technology possibilities and transferring that into real world applications for enterprises but also for the common people. So we see in the session um how they explore how ecosystem collaborates, strategy, governance and real world applications are shaping the future of the industry of industrial AI and prepare for actionable insights from one of Europe's foremost AI expert. The stage is yours, Fuku. Good morning. Thank you Katherine. We really have a lot of content for this uh presentation. I really want to have this content as a working session with you together.
At first, I want to share with you the most recent uh releases and development in generative AI. For example, I think you know uh Llama 4. I think Llama 4 uh going is going to the right direction for the industry adaptation and applications. Why? For example, the llama scoot. The llama scoot can run on one H00 GPU even on devices and with the large content almost to 10 million tokens. It means because I uh work with Mendy companies together, we always worry how can we do long-term and short-term demand forecast predictions with u open source open weight and also uh a big context 10 million you can do combine the long-term forecast with short-term forecast with real world um datas.
this is really the right direction and the second and also for for industries the uh generative AI in combination with physical world with robotics this is really important I'm really looking forward the physical world with generative AI in the near future and also of course who cannot talk about the generative AI and Patrick will show you very cool applications demos and uh Now let's start to work together. I know in the past two years many company start to define AI use cases and it's some company have even more than 800 use cases. Some company has ah I have 200 use cases but the company and the board is not uh ready to invest 100 million and uh therefore how can we encourage the bottom up uh innovation and the use cases.
on the other hand also know this all contribute to the right direction of the company. Therefore I think it's right time now to dis uh define the AI northstar for every company. Do you have a AI nostal for your company? Let's verify later. And uh then without digitalization, without the data strategy and AI is only a good will, good intention, right? Therefore, I it's very important to find the answers together. Do we have a comprehensive um catalog or inventory of all the data we own, we produce and uh we can access and we allow to access we need to access right these kind of u questions are urgently need to be answered otherwise we cannot be successful and what is the current state of the data uh ecosystem in our industry can we work together right if you talk about uh supply chain demand forecast without ecosystem you cannot do it and the AI stack the IT uh uh uh the IT team does does not really know how many stacks we need and also
the industry standard in the business unit uh how many AS stacks do you think you need in your company for a company like Boeing how many stacks do you think you need two or three at least uh I I think at least three one I I will continue later and uh also um since uh the AI uh experts are limited the number of AI expert limit they are expensive right and the use case so many use cases you can implement in your companies how can you organize the teams how can you upskill your teams and how can you attract the talents right um this therefore the how you organize uh your central AI hub and in combination with the business unit and I want to also get your answers your suggestions.
Furthermore, collaboration with regulators and we always complain about the EU AI Act and also the German government. However, I think as industry, as companies and we have we should proact proactively work with the regulator because many of them don't know how to do the modify their regulations for industry standard. For example, if you want to define how you produce a airplane, the in particular also airplanes which are mission critical for helicopter and the manufacturing process have to be uh regulated. If we um integrate um visual inspection into the regulation sometimes the regulator they are scared say oh if we integrate visual inspection into the manufacturing process the humans they will be lazy perhaps the quality will reduce right but uh they don't know how to do the regulation therefore we as AI experts uh if we work with companies or in the companies we have to proactively work with the regulator in order to make the innovation happen and also thought leadership.
Of course, we cannot be thought leader in the in every uh components in the ecosystem, right? But we should think which is uh how can we be thought leader in our industry with AI and I hope I can find answers with you together perhaps we not today but I I think these questions are very relevant in order to make our companies successful and uh let's start with a Nostar what do you think about a car company I will not uh uh let's go back perhaps what can be the Northstar for a car company like a BMW, uh, Daimler, Neo. What are the major components of a, uh, car company? You, you know, you you mentioned you have Nostar already.
Perhaps you can think about for BMW. Right answer. I want to be more precise for the northstar of car industry. I think at least two components. One is autonomous driving. AI is the key technology to autonomous driving. But autonomous driving is not only technology, not computer vision, not radar and leader uh technology. This is also about ecosystem. Do we have the uh right infrastructure? Do we have the right legal framework? Do we how can we share the data among the um ecosystem? Right? Therefore, autonomous driving is not only limited to technologies, right? And second one as you mentioned, right? Uh embed AI into the manufacturing processes and embedded AI in the business processes. And with not this northstar you can then make decisions of your road maps, milestones and also investment.
And among these 800 use cases perhaps you can choose the top 10 and also monitor the progress. And what can be the northstar for a company like Boeing similar right? Yeah, like uh AI pilot, AI co-pilot, right? There's I think a lot of automation intelligence in the in the uh cockpit in airplane already, but safe landing is very important and the sustainable uh uh airplane is very important. AI can really contribute a lot, right? Because the the human pilot, they cannot forecast if there's a bad weather ahead, right? We know a lot of accident and in particular with helicops, right? uh in the news this year and also uh for um manufacturing a airplane you have a like um Airbus A 3320 and a 350 there are so many components so many supplies how can make it uh um successful.
Do you think what the joint technology needed for the manufacturing company in order to make the manufacturing successful digital twin and the AI together right? Yeah. No, last week I met uh two week uh um airport the uh B and what can be the north star for airport like Berina airport US passenger right the passenger experience I think often in particular in Berlin people were not happy when they uh try to land or when they want to leave the airport and they will try find the use case right passenger experience how can make AI contribute to the and also uh airport is so complex right the the safety security customer services all these kind of things therefore also a northstar for uh for uh airport uh is needed with AI and now let's talk about data I have three questions do we have a comprehensive catalog or inventory of the data we own produce can access and need to access.
Um it took me uh I have my dear SAP colleagues friends here. It took me um many months in order to define the uh legal framework for the customer data access. It was a big effort but it's worth it. You can be successful. Don't use GDPR as your excuse. And of course u with generative AI you can improve your u data generation and data quality and even generate synthetic data right and uh also as mentioned how can we um uh industries companies work together for our ecosystem sharing the data and take advantages of the sharing and if we look at industry and business data it's very complex it's not simple like only text and images right perception data You have a uh tree data structure for the business processes, production processes.
You have a lot of transaction data. Patrick will mention how difficult to deal with transaction data. You have billions and um data. How can you do it? Put into all the large language model. How to can you do it in an intelligent way? And often the transaction data does not have enough semantics in it. Right? And also if you are in the chemical industry, you have chemical formulations. If you are in the manufacturing company, you have the geometric data, temp physical data. How can you do it in integrated this kind kind of knowledge data into the generative AI context make successful? Now I want to share with you I'm very proud of uh the company tricker.
Um Philip is here in the audience and uh I'm the I am in the board of tricker. Tricker is a innovation company. um help the companies in the beginning not in in the beginning of their intelligent transformation journey define the AI strategy and also implement uh use cases for example there's one very cool use cases uh since uh I was global head of AI at SAP when I give talk people always say how can we deal with the master data and um ta has a solution how can use um generative AI in order to repair the data uh check the consistence and also enrich the data and make the data searchable. Therefore you see it's so complex AI make company successful right and also I talk with a small real estate company I check how challenge the data right in order to sell make your business successful for a small real estate company if you want to be intelligent you need data about the market about properties about the rental you need to know the
geographical uh um uh development social develop in your city and your country and also regulations right? Even a small company therefore data is so important and uh this is um I have met so many companies in the last one year I have learned so much right therefore I also want to learn from you this is um I discussed with Patrick right that uh the the the construction uh companies they don't have enough data for the do uh uh uh u right uh documentation of the their real world construction production right but with drone with AI it's also with sensor with camera it'll be very easy how much time do we have left one minute okay uh 10 minutes oh I can work with you even longer I have still five minutes right okay good just let me tell me to stop uh now let's talk about asc from five u uh F5 the uh this is standard Ast every company if you want to build AI applications because we are on the end of the
value chain this ASC is needed right starting uh from the infrastructure I also put companies uh which support this um ASC um I think the hyperscaler control the tech of course if um of course if you have a Microsoft Azure you can cover really infrastructure uh data processing also AI foundation model, AI model like framework many of them ptor and they are also open source right however each company you have to control the application services either you work with the third parties make the your deployment and delivery and the continuous improvement in a safe and secure way this part you cannot hand over to other people don't be too lazy and I I think the hyperscaler cannot take over this part and the applications you have to do yourself or the ask companies like tricker in order to support you and in Europe uh we have very limited infrastructure provider companies like GCO provide infrastructure and also the lower stacks therefore I want to also introduce u the I don't know whether you know the
company GC Yeah, GCO provide because for many companies as I mentioned right you need perhaps ESC for cloud applications for on premise and also for on on your machine like if you have a guest to be have windmill right I'm in the board of Zimmen's energy big machine I'm so excited about it however if a big machine in the in the ocean offshore it is often is disconnected from the cloud also for the security you need to embed them um edge solution. Therefore um if um Elena from the GC is also in at rise of AI perhaps you can exchange with her and also one company no land AI they try to embed u um uh domain knowledge and enterprise knowledge in combination of the traditional technology with the new technology like also I think u uh uh good is here you also uh your team is also as team is also working on this part right yeah and And uh the industry knowledge why ontology and corporate knowledge uh graph and with rack and
uh prompt engineering with context in context learning they are smartly develop this kind of aging architecture in order to make the application um uh with uh compliant and of support industry applications. Um now uh very quick um because many people ask ask me uh Dr. Do we need a chief AI officer? Do we need a a central AI team? They are useless. Uh and uh but I I told them if you say they're useless because you don't know what is needed for a central AI team. A central AI team had at least three important tasks and this is needed for the company. One is the strategy part, right? And also investment decisions and the partnership make buy decisions.
The second one very important build up the ASC build up the development life cycle essential in order to make sure the application development will be successful and the third one most people don't know that the third one is important make sure that the application will be successful scalable you need a small product team in work with the business unit in order to make and with these three pillars I think the chance to be successful is much bigger than before and uh I know many of My friends they are in the board supervised board want to be board and as a board member you have to do risk manager for yourself for the company for the board member.
These are the six components which you need if if you become a board member or if your board will really uh have the AI governance if you want to you build AI applications you need AI governance in the board and supervised board. make a picture. I was and in the future we talk about AI software, AI hardware, but in the future AI will be part of our body because in the future AI will become the sensor in our bones if we have artificial knees, artificial hip and artificial liver and in the future I think in the near future AI will be part of them. This means AI will emerging and the syn uh unify with our human body.
I stop here hand over to the exciting part part of Patrick J. That's kind of uh a great starting point um because agents when you talk about agents um we had them in sci-fi in the 80s. So when you think of neurommancer which talked about what happens when a machine and a human intersect right in that book we already had the idea of agents encapsulated and everybody was thinking about it in the cyberpunk era era of the 80s and 90s. Um so why is everybody talking about agents this year? Um because in the past we had chatbots and chatbots are really kind of interns that are waiting for you to give them instructions and agents give us the option to actually get work done which is kind of valuable in the enterprise.
Um why does it work now? Because I mean the baseline is we now have systems in AI that have common sense. That's kind of the big unlock. We have instruction following that works really well. And one of the biggest unlocks for this year is actually was was initially R1 uh by Deepseek because it was so incredibly cheap and so incredibly good. I'm going to show you some numbers on that. Um and that forced the hand of OpenAI to release 03 Mini which uh was quite competitive with them and now O4 Mini. Um and I think we are starting to see a changing perspective on trust. Uh let me explain what I mean by that in a second.
So agents as they exist today or as you can build them today mostly work for swivel chair applications where you have to handle multiple user interfaces in the enterprise where there are no APIs or the APIs are not well you know not well documented um and all the robotic process automation graveyards are also kind of a good uh um a good pick for things so things that you couldn't automate before you may be able to automate them now um and the outsourcing playbook is actually applicable for agents so if you don't know anything about AI I just think about your business process. Think about it how would you outsource it to a business uh process outsourcer write an SAP standard operating procedure.
So that's the process you want to do. Um define the barriers uh define the escalation path define the service level agreements and expect 8020. That's a great paper from SAP researchers that essentially showed the current models even in the simplest question of SQL um so um text neuro text to uh SQL they have a really high precision. So we're talking about 99 to 100% precision on being able to correctly predict given an SQL structure given your human text natural language what you what kind of SQL code you want to write but only when you look on the recall only on around 80% on the tasks. So you need to be able to identify tasks that are business processes where you say okay I'm fine if I can only outsource 80% of these tasks the 80 easy percent of course right um and I want to give you a demo how that looks so what you see here it's Gmail because it's easier to integrate than Outlook most of the time if I'm sending an email
uh to to one of my agents that I've built where I say okay please order an iPad mini or in this case order two iPad minis in space gray with 256 gigabytes and they just write it as free text and I can write as much crap as I want into that email then essentially You can build an agent that receives these emails that checks for these emails continuously. Whenever they find a new email, they try to understand what's going on there. They're trying to encap to extract from that email in a structured form. What is the thing that the colleague is actually ordering? Checking in an internal database. Does this person have the budget? Does they have the authority to process the order?
Um check what kind of task do they have available? And then when they see okay, I understand the task. I've seen the task in the past. I'm allowed to do that. then they can log into a service now portal that's kind of an enterprise application and um reason about how to get this job done. So what you see here is not nothing that I'm doing and this is also not a lot of AI building work. All I'm doing is I'm giving the DOM tree, so the HTML content of the website of that enterprise application. I'm handing it over to a large language model and asking where should I click and they can do that out of the box.
And even the problem is when they can't scroll down, they have to scroll down to actually find the the right checkbox. They're going to find it. Now they have to figure out. So they space gray is already selected 256 GB. Now they have to adjust the quantity. Um and they managed to do that. And now it has to click order now. And after I clicked order, now it has to understand it's actually done with the job, right? Um let's see. It's in real time. That's slow. Um but I I I um sped up the loading time of service now. And uh then of course it sends uh an email telling you, okay, I I process that.
And of course, you can create a lot of more um stuff about it. But the question is, you know, we all know LLM, hello, they do stuff. Do they actually work? Right? And we had the first agent ideas since April uh 20 uh 23. Um but the answer is uh no they didn't work. So uh human 99% quality let's say GPD40 in summer of last year 67% cloud sonnet 75% you cannot you cannot roll it out that's not doable right? Deepseek R1 87% and 10 times cheaper which is interesting right for mini 98%. and let's get even better. Um, but before thinking about getting better, what you have to think about, companies know how to manage risk.
They know how to deal with manufacturing defects in the manufacturing industry. We have the six sigma thing from the 80s and 90s. And you can apply that as well. And you can look okay at 99.38% we're going to be in a level where we say, okay, the agents don't get everything right, but my humans don't get everything right either. I can write the benchmarks and if I give um um to a system um a standard operating procedure and I just give them standard operating procedure I don't do any programming in a traditional sense uh what you see here is deepc goes up to 98% and as O4 mini goes roughly up to 99.3%. It's very close in the likelihood for easy task today without a lot of scaffolding and you can get at a higher level with more scaffolding and you can actually start managing these things like you would manage other manufacturing processes.
So all the internal business processes where people have to interact with a lot of uh systems they can now be automated. And a last uh thing that I want to share if you want to see an agent in a action a very simple one just go to charge EBT UC3 model upload any picture and ask it where is that what you see here is a picture they didn't it didn't get any other information and told correctly it was land canal it was something like 500 meters off so that's quite impressive going through the sinking process uh and I'm leaving at at that thank you thank you Patrick Maybe maybe I'll just stick here if there's a question from the audience for 30 seconds.
Um any question there is one maybe Patrick you or two is it for Patrick or for fu maybe you come to me because then the camera can see you in the live stream. Okay. So uh large linguish models or uh foundation models they have a lot of uh uh data for training right yes so they collected a lot of information from our past but we won't collect it in the past but we want to use them in the future to solve the problem. Do you see this contradiction between the past and the future? Thank you. Yes. I mean um what what these what these agentic systems need is essentially the past inputs the past outputs of the process ideally if the process changed the process description and that's kind of the data source that you want to use on top of a model and we start out with just a baseline model um you measure these these metrics and at only at a later stage you go into can I use that data to fine-tune
these models but it's like you know you don't start tuning car before you have ever driven it, right? You first drive it, you figure out where it goes wrong and then you need that data from your past interactions to actually um make that car even faster. And also I can add right as I mentioned in the beginning like llama 4 you can have with the open weight you can easily do fine tuning and also you can have even 10 million of the context right uh you can use the past data do the fine tuning on top of general purpose multimodel I think for the industry multimodel uh AI model is essential on the other hand you have have the real world data the short-term data in the con as context in context learning this kind of combination ation make you take advantage of the your past experience and also get the real world uh information.
Yeah, thank you. Thank you so much Patrick. Thank you too and I'm hoping you I'm hoping you stay stick a bit around and there is another announcement. Yeah, next year I will check the answers to the questions from you. That means she's also back next year on stage. We got you there.