OpenAI im Gespräch: Mit Christoph Winter über Agenten, Strategie und echte Enterprise Power

Jens: So, perfect. Jo, dann starten wir mal rein. Hast du also Trinken dabei?

Christoph Winter: Yeah, I it me.

Jens: professional, very good. Perfect. Hello Christoph, nice to you here. Thanks for your time. I'm forward to a great conversation with you. But for everyone who doesn't know yet. Who are you? What do you do? Which well-known company do you And does your day look like?

Christoph Winter: Hi Jens, nice to be Thank you for the invitation. I'm Christoph Winter. I'm our partnerships with strategic enterprise customers at OpenAI in Germany, Germany and What this means in reality is that it's job to work to create as much impact as with AI for their companies.

Jens: Okay, that was a very good short pitch. I think you've it twice. And now, course, OpenAI, Germany's location, totally exciting, it's currently in development. How did your rough journey How you get where you are now? What was on your way there and did you always to get where you are now?

Christoph Winter: Yeah, good question. You can plan beforehand. My background is computer and software engineering. I'm from IT security and cryptography. That's my passion field in engineering. But I university and did startups for 10 years. I founded and led product and growth for other startups. I the market and the vertical extremely exciting. I to find in the field of healthcare. But I couldn't how difficult is to I went to AWS, to Amazon Web Services and set up the Healthcare and Life Sciences team in Europe. And about Then I the vertical at AWS, about four and half years. When I OpenAI, I wanted to the and was back to start something. Story of my life. Then former colleagues who had OpenAI came to me and asked if I was in seeing if we our partnerships in Dach and the office in Munich. The opportunity was so big that I it in the shop. I've April at Open AI and a lot of fun. The team Munich is there's so much to do, so many opportunities, many good topics and so much to do, but it's a lot fun.

Jens: I think if you have an opportunity like you probably don't a long time. Especially when you've of And you said the team is like a side in Munich, to a feeling for it.

Christoph Winter: We started, we always start new offices with 5-6 people, cross-functionals, one the main functions in which with customers I don't where we currently, but will 15-16 people, but still

Jens: Exciting. So still relatively small, but for sure it bit bigger in next years. And now you just told me that you are more of a techie. How is it currently a role where you are probably a lot advice? Can you your nerd life or your nerd passion? Or is that something where you say, would actually like to craft in the future?

Christoph Winter: I think both them are not too short. One is that what moves a to problems. Also problems with customers, different industries. I am extremely curious and extremely happy And then to really That's one thing that actually me into engineering, because in my experience that's the core of what an engineer really is. On the other hand, are also many rabbit holes that you can and many super exciting topics on the engineering side, you can even if it's not always 100 % on point, the role of the customer, dig in and people out

Jens: What is your recent rabbit hole? Or one of them?

Christoph Winter: One of my favorite use cases of AI is procurement. One the areas where you the greatest passion for it. There is a lot about building autonomous negotiation agents. For example, one rabbit hole is a guardrails. If an agent is let go and there a vendor relationship...

Jens: Ooh, okay.

Christoph Winter: and negotiate it and that really an impact on how the price looks, even if humans are in loop. Then guardrails are a very essential topic in different places and a very important topic in regulated industries, where I come from. And that's a topic where I've been very deeply involved

Jens: I think that's where guardrails play a crucial role when the company supposed to survive. Very cool point. If you understand yourself, these industries that are most fascinating are probably the most ones. these regulated industries, healthcare, is that something you say you're most interested in because they biggest challenges?

Christoph Winter: Good question. I my favourite industry at the moment, because there are so many opportunities and so many exciting topics. But if I had would come from, as you say, the Pharmacy, Healthcare and Life Sciences. Why? Because I'm extremely impact driven person. The biggest question I always ask myself, even when to a new professional challenge, is... You yourself a position where you have the opportunity to an impact and do something good from your time. And that's why healthcare and life science is one of the industries that most tangible impact. Because you really say, assuming you're turning AI, drug discovery and clinical studies, which are extremely inefficient processes, make Potentially, you have an impact, for example, to a new drug on the market And that excites me extremely, to a lever on impact. That is Pharma and Life Sciences. Another a second, is Financial Services. Not because of... are many impacts, but simply because the problems with AI are so interesting. The processes that can be everything that under the umbrella of Doc.AI, basically on documents, existing processes that are currently with a lot of just automate is extremely exciting. And then can into the area where work that really didn't any fun, over long years, and where lot of people are working, to a large part and automate and really free up of potential. And the third thing is automotive and manufacturing. And that has to with the that we Germany have our automotive and manufacturing industry, but are currently in a situation where we really ask ourselves the question how we AI to this potential. And I find that extremely exciting, especially from a German perspective.

Jens: I interesting industry. I this approach to people and this coming of AI can write nice headlines and blog articles. think that's really exciting to the big picture. That means that there is still so much to do And what you mentioned in side sentence, which I very important, these jobs where nobody wants to

Christoph Winter: See next

Jens: I think we have a lot of this discussion, you probably know that AI automatically everything feel so speak, but I think work is something that we have and also the jobs that to because most cases it didn't work I think we have a lot of in your own everyday life.

Christoph Winter: Absolutely. I think with every technology, with every platform shift in technology, the question has for us, how will work change? I don't that's anything else with AI. Our experience shows that, because that's often the narrative, that only jobs are reduced. That's just something we in the data and in our experiences. Because there is still so much potential for this talent, for these people in company. The question is, maybe we the 10x what do we with the potential we can Most companies have a very good idea of where they 10 % more manpower than to invest directly in cost reduction. There is also, but we it as the big trend. We all don't know where the future will but I think it's a shift, as has other technology shifts. We humans are very creative, will a good way to deal with and turn it into a better one for all

Jens: I think also the Especially when you realize that you can much more than just saving on people and so on. I that's just the economical way to in a company and growth. And in the best case scenario you hired that good at what you do and only work but can also think

Christoph Winter: Absolutely.

Jens: And that's what I notice with myself and with the teams we work It's shift when you your skills with KI.

Christoph Winter: Totalt!

Jens: Let's jump to your employer. I would like to know when GPT-6 But don't a fun side to OpenAI is a company that generates relatively high level of attention How is it like? Working with OpenAI, especially with the main sit-down in the states, you night shifts. How are the different locations seen? Are you autonomous? And how does the topic of tempo and ownership out, with such a young startup?

Christoph Winter: Interesting question. give you brief context, OpenAI is relatively small company, contrary to the perception we have We still very hard, like a small company, like a startup, but we grow extremely strong and things move extremely fast. So tempo is certainly what defines the work here. A question about how office in Germany works, Munich works. Because we relatively small, we very connected to the states, to San Francisco, the headquarters. Many of our teams are in San Francisco and we a close connection. That means there little of a gap between us being the German team and the SYNC. And you feel like you part of the global Open AI team and also very strongly related to the headquarters. It also with downsides. There are night shifts, of lot of work But people here do it because they in the mission and not because they have to. And think we all as much gas as we can to develop the mission, which the AGI that all people with value, and to contribute

Jens: You can the tempo even if you look at the release last 6 months. It doesn't have to the huge model updates, but a lot of smaller and bigger projects. Own updates here and there. think you can already that you're giving it lot of tempo and testing it But if talking about testing and testing... Then it's something you in the enterprise sector. Or rather think that It's more your area, the enterprise focus. Get everyone who listens a little bit. Now we JGBT for free, have the plus version, they have the teams business version and then there's this omniscient enterprise. What do you and what is the enterprise version of JGBT?

Christoph Winter: Yes, absolutely. Maybe briefly as a context, So how we companies is that we form very strong partnerships with companies in different industries. That means, as before, we are small team. Our goal cannot be and will not be to with every company. But we are extremely exciting industries that are also close to our mission, where there interesting problems to where are also use cases to crack, which then also a downstream impact on other industries such as contact center, for example, partnerships in these verticals with these companies. And then to see, as I mentioned how AI and AI transformation can implemented in that it really a added value for the business. That's our job. How we do is, we offer two products for companies. One is JetGPT Enterprise. This is enterprise variant or the company variant of JetGPT that we all privately. with a few differences, we can get in And second is our API platform. And this is platform that our models, but also many capabilities from AI, like tools, can used to software development team, speak, tailored to a certain problem in company. And with these two products, we go into these partnerships and think very carefully about the challenges of the company and how we them together in the partnership to this impact. Yes, please. In end, got the question, are three fundamental differences to the private version. One is security and compliance. means that GVD Enterprise is specifically for

Jens: Okay, that means... Do you to

Christoph Winter: to ready and the highest standards of security and data privacy. What does that A few ideas. Point number one, should never on company data. That means should in wait for a model based on any data that in or out in GTP enterprise. For our private user, is an option. That means have this option to choose from. For business customers, they can't activate if they want to. So we never on this data. And customers... owned, so they always 100 % of all inputs and outputs. That's point number one. Point number two is data privacy. means highest standards on data privacy, European data residency, means the data stays in the EEA, in the EU. And the third is security and enterprise features. There's a lot about how can I my SSO and in my... as an enterprise suite, I user managers. But it's also about how encryption And we offer things like enterprise key management, where we encrypt data, for example, to provide customers, but also things like just certifications, external certifications for SOC, for ISO, for GDPR and other things. So that's all security and compliance, to to a very high standard and really enterprise ready with the product. And the last point, and it is often underestimated, is a dedicated infrastructure for our enterprise customers. That means a completely own infrastructure for JTPT Enterprise. That means an enterprise customer never over resources with a free user, for example, capacity. We the infrastructure specifically to this, but also to security and compliance. And that results for enterprise customers in higher context windows from social or some models, but also a performance that is different from what you a free plan in your private environment.

Jens: Very exciting, thanks for the roundup of the product. Very exciting. The question that me is, higher context window, that like this How high or how big?

Christoph Winter: It's four times as high as the normal models, depending on how big they are.

Jens: Okay, of if you have your own infrastructure, it is course also different. You can that with many providers, not just with you. When the USA wakes sometimes the performance is bit comfortable, that But shouldn't in enterprise area. And where do you in enterprise area? How many seats or how do you these partnerships? What size?

Christoph Winter: We don't it to the level of selling a solution, but we like to the strategic level, as I said We at the strategic level with the board of of DAX companies. Where are you, what are your challenges and how can we all products and everything that OpenAI can or with the know-how and unique value propositions that we can bring us in such a partnership. What can we do to this It in different directions depending on what is the most exciting and relevant for the customer In normal case, JGPD is always part of it. We can go deeper into this. It is very fundamental driver of how we see successful AI transformation in companies. That is very often part of it, but there still a lot more

Jens: Okay, exciting. can deeper into What do you when you talk Because from my bubble and from my contacts, often this feeling that the chat GPT because I that data flow-comfort under. Maybe in the free account differently. But what are the most common reasons why you hear that your solution is not an option, apart from the features, rather from the framework conditions?

Christoph Winter: Yes, I think you mentioned it, Jens. think there is now and then the belief that we have problems with things like security and data protection, data privacy in Europe, data residency and other issues. I have not a case where worked with a lot of companies where this really the case, where we a hard blocker, because Chattipudi Enterprise and our API platform are now such high level, which also data residency in Europe and these security and compliance features that we already mentioned before. Also for regulated industries, such as financial services, special agreements to do this. I haven't a customer in Dach where it was really a hard blocker and we couldn't But think it of the impression that maybe JetTPT is not so enterprise ready But we in our first conversations and demos and then we get and say, okay, where are the requirements, we that relatively quickly together. Where this comes is that for all of us, the pace of innovation and development in AI is a bit unusual. Now, GTPT was at the of 2022, so speak, a few years ago. And many see the development in consumer sector, use it and say,

Jens: Mm, okay.

Christoph Winter: Normally, the cycles take a little longer until an enterprise-ready product comes I think that's the fundamental difference with AI. It's a little faster and I think we all move a little faster than we're used to. In any case, the LGBT and our API platform are what these aspects are about and are definitely ready for all industries in Europe, Germany and beyond.

Jens: Okay cool. speed. How is your own feeling about speed in the AI level or in AI field? Because compared to what was before Gen.AI and before JetGBT, it was all fast but more comfortable. Is that a topic where say, sometimes everything too fast for Plus the insight you might even have from your own work.

Christoph Winter: I know the feeling absolutely. As good as I find and as many opportunities as possible, you often the feeling that it almost overwhelming in which direction and at what speed it develops. We see it when we our release cycle of models. It's not just major models that we release, but also snapshots, improvements that we have. We 20 % 22, 23, from a time, once every six months, once a year, large model is released. Meanwhile, we large model releases, multiple times, every week. And several times in any case every month. Even if these are not as large releases as GPT-5 or Sora 2, but small snapshots, there is already lot going And you have to keep And I think that's also very openly spoken for many of our business customers, often a certain challenge, that they say, We don't know what is new and for us and what not relevant and what an or not. That's we are here in Germany to partnerships and our perspective on this. I think we have a very unique perspective because we the ones who the models and can assess well. at least hope to this with customers, what the impact is and what you should focus

Jens: I think you have to get to Sometimes you feel like it's getting slower and then again. It's like a week where everyone is spinning I think a pretty crazy time. Let's go a little deeper into the models. I think that's very interesting You just said you have new models Now you know from outside the WD-5, WD-5.1. But let's about the snapshots. What is it, why do you it and how do they get rolled out?

Christoph Winter: A snapshot is a small update to an existing model. We improve our models, like GPT-5, based on feedback from customers, based things we see that work and don't work, based on the evive we set Evive is a high-level term for unit tests, as we used say in software development. We say, out, we know... for example in a certain task, such as health advice or example studying or coding, we the model to very strong and based on the prompts that our users are interested in, perform And means we set up test sets to constantly the model in different areas. And then we a new model and notice, example, we believe that one or the other area it is not strong enough and then we have do something. Our teams are on this area, doing smaller training runs to improve And then there running snapshots releases where the model is getting better. One big reason why it doesn't happen often as we would like to is because security is a huge issue for Both major releases and snapshots require all changes from the model to very detailed. assessment process, security process, or even by different teams to ensure that they do what they are even at the level of performance where are going to to and our users, but also to the possible negative impact with those don't want to have. And that's a huge topic and we lot of into

Jens: Yes, think so. when you see not only in enterprise sector, but also consumer sector, how many people it in the end, then of course you also want something good to of Do you a feeling, or can you also say how long such a security audit, or this whole process, Is that something where you about two weeks, six months? Do you your range? I that exciting.

Christoph Winter: That's a very good question, but very different depending on how the release looks like and what the findings in early phases of a security review or red teaming where are teams that try to the model to to do something that we don't want to with guardrails. Depending on what is early, it can longer. That's an interesting question. The was our open source models, we launched So two open weights models. That means not in our API, but are provided and can be used even with a very open license to do This was a topic for us, where our security team looked very closely and invested lot of because it's just open weights. And because we the guardrails around the model in our hands. If it's outside, then it's outside. And we would have loved much earlier, but we saw that we weren't 100 % sure what the security was about one the other topic. And then we decided to these models again in a few weeks, simply because we wanted to that we have in right place.

Jens: I think it's to have in that there two or three announcements like this, that it coming soon, and then it wasn't there But of as you said, you let out, you can't it if it's That's a very important point.

Christoph Winter: Exactly. happens quite often. when to security and testing. For us it's a bit too much. Of we how important it is, but we of features internally for few weeks, months. And we extremely eager to feedback from users and it accessible, to finally our families about it and it And then it a while. But think it's a... a data point on how important the whole topic of security and alignment for models.

Jens: I think that no models are 100 % safe There are also some accounts that nothing but jailbreak It's probably interesting from your side, when you see that there is a jailbreak. We didn't think For me it's the most My grandma always me Windows 10 license keys. Or think Windows 7 Ultimate was that.

Christoph Winter: Thanks

Jens: And then, unfortunately, with the story that grandma died and that was very sad, Windows 7 volume license keys were read to That was pretty cool. It was also a version from 3.5.

Christoph Winter: Yeah, that's funny.

Jens: Cool, let's talk about what is a good use or how companies use it properly. You work with big companies, many challenges aside from how do I use AI, but also many political and regulatory issues. I'm always out LGBT for everyone, so please care And now we're ready for What you that good companies that you work with are right? In other words, can say what is wrong. But where do you see the pattern?

Christoph Winter: Absolutely. That's the right question. What we long in the context is that all the companies that heavily in AI, probably somewhere between 25-30%, that's a of a global variation, measurable ROI impact or top-line, bottom-line impact in their companies. And now the big question what decides this third of all companies from all the others? And there are a few different things. One of biggest is that for successful AI transformation, is really driven we always to two consecutive motions that to run Once it's what we bottoms up. That it's about really the technology into the workforce and everyone to level that he really understands what AI can do a certain level of AI literacy and also tools available, such TGVD Enterprise, to try out, to your daily work, to make and then certain understanding of what we could a dedicated solution for and what couldn't. This is the first motion. And second is what we tops down. And that's where about very special use cases, where I say, as a company have certain problem. That has me a very big value and that can be top line or bottom line. So either it enables me to lot more revenue, new markets, more margin or to bottom line and really big costs and to save for the services and products that I offer. And that's what to bespoke solutions with our API platform and the models. What we see is The successful AI Transfusions do everything together, so to speak, and then to via both motions, where I my workforce, I the benefits from Bottoms Up, and that is essential for And get about 70-80 % of all employees that we interview when they JGPD Enterprise say that they significant time every week, so over two to four hours.

Christoph Winter: and work that you would otherwise have to 77 % say that they could that they couldn't do because they have So I that, but then above all to the education and AI literacy in the company up. What that the with the company is, I get all the benefits from productivity, but I especially the ground for my employees and every single employee who is very own and precise perspective on every single process that the company is doing, can an opinion, can AI help or not. Instead of just a generic AI use case that was by another company in my industry, I have this process where my employees their hand and say like, I have this process, I have it 80 % already via JetJBD Enterprise, but I think if we it even better, it would be an extreme... value level for us. So the feed I want to that's Bottoms Up, and then Top-Down is exactly to invest in these use cases that have this exponential return on investment and then also to see that it the organization and that the adoption is where I want And only then, when I both at the same time, the feed for and then everything right when it the Top-Down use cases, then you have these transformations where there value and where it very strongly in top and bottom line.

Jens: I think you meet somewhere in the middle. I think from both directions that's totally important. And how do you it? Because when I of chatgbt or when I enterprise software and chatgbt, I it's pretty boring for enterprise software. Do you a feeling how much then really often your enterprise customers say, hey, we use this power from AI now not somehow in a nice interface, where you can add right what the difference between what we all know, but build on Do you feel like the direction more and more building more more yourself, so that more ideas really to the day from use cases that maybe a chat gbt or a chatbot in general can't solve? Or is it really like this chat solution that is simply safe and usable for everyone already currently has the greatest added value?

Christoph Winter: It's both. Interestingly, and to context, the roof market is extremely exciting Maybe a background check, GBT has around 800 million weekly active users at the moment. Within Europe, Germany, Germany Switzerland are the European masters in adoption. They the biggest adopters. For example, Germany, age group between 18 and 24 years old, almost 100 % of use chat2bt every week. Very strong adoption and that generates, so to speak, a macroeconomic level, this ground for this transformation and for the change. Because of especially young people, general, the employees carry the topic into the companies. But it's also exciting, now one million Business user, also on GDBT Enterprise. But to to your question, this background, is interesting to the bias of big companies more exclusive top down. They say, no, we understand what the use case is that we implement and we know where the value is and let's focus implementing one or shiny objects on board level. And interesting thing is then What we see is that you both and that I have to Bottoms Up and also the feed so that these use cases will successful and go into production. If you the motions, what are the biggest risks that I as a company, a board, then it's from Bottoms Up to primary adoption. I know that has benefits, know my employees use them in their private lives. They benefits from it has added value. The question is, can I an adoption in the company context so that they have in their work? The primary risk is the adoption. Do I the right tool that my employees want and is it useful in company? In the top-down context, I a lot more risk. Do I the right use case? Does it really have the value I imagine? Is it possible to over the current models?

Christoph Winter: I it built, which is not trivial, and then get the adoption done in my company. And that means that's a big focus. And we often companies to say, yes, top-down use cases are great and let's them into a partnership, but don't forget, bottom-up, because A. fast return on investment and there is a lot of value to get now. And you will the AI literacy so that the top-down use cases can be scaled down and by your employees.

Jens: Do you this from top to bottom when it comes to a risk? That this in the roof market? Because of it's new technology, we have to about new use cases where have no idea if something is to Do you that companies in the roof space become more courageous when it to experiments?

Christoph Winter: I think they have always brave in the to We can in the adoption in Germany. That's it's so exciting. That means a lot has tried. The core of is from the down, I to these use cases into production so that they business value. The biggest challenge we see with these use cases is, it from the POC or the prototype? to something production that accepted, that finds and that then really the business value. And there are many challenges. And one of the important things that you can have in place to ensure are partnerships, for example with us. We can help much to contribute to mitigate all the pitfalls that are on the way from a successful POC or prototype in production to see that the video is also

Jens: Those would be the top 3 pitfalls.

Christoph Winter: Point one is visibility. Often it is misjudged what models can do and what cannot We often into scoping phases of partnerships. We look at what customer already in their ideas, what these use cases are that are already prioritized. By the also find, to your point, we want people to experiment, we never too few use cases. It's always Excel sheets with hundreds of use cases that are prioritized I believe in ideas, and that's a very good sign, think. for us and for the adoption in Germany, they are not There is already a lot. And then we like into a mapping where we the top use cases and them to two dimensions. First, it's really business impact. Why? Because it always difficult to the first few use cases into production. And will always be painful and I will always into problems. And I want to the first ones that are painful, that have measurable business impact. because otherwise you momentum. If I through the part of train that can be often and it into production and then the value is not there, then I don't a flywheel where I investment and buy-in for the topic. means dimension number one is the business value and that has to relatively high. The second, to back the question, is this feasibility. We used cases that so broadly scoped that the feasibility is relatively low. I'll you concrete example.

Jens: Yeah.

Christoph Winter: from a call with a customer last week. It's about, for example, accountants, book holders completely. And then the idea well, we'll an AI agent and he'll do the bookkeeping and then an email with a document in and the agent takes over and then comes out SAP in our booking system, valid booking sentence. And of course good and I'm not saying it's necessarily not feasible, but it's not as trivial as it sounds. but if I automate I to look at it I have to look at what the individual steps is it possible, do I need for parts of this process? That the feasibility of such a use case is not up because it has not done we have no indications that the whole process is by the models. That means feasibility is one the points where we challenge and say, have to down further. Or we often say, wouldn't prioritize because the perform but have to lot of work into optimizing Wait three months, three months is a very long time in AI, and then we'll we believe it will work.

Jens: Do you think that this hype about agents and agentics is overheated? that you might say, now suddenly everything is an agenda and we can automate Does that the work a bit difficult when you things? That's just not a click. Sorry, we have to little more education in

Christoph Winter: We are very positive in of leadership among customers, we are very ambitious when to agents. think what us all is when things get overhyped. Especially agents. It's a topic that last two years.

Jens: I lost the old feeling at Kai.

Christoph Winter: think so much played and so much has been which was often not possible, you to that openly, that it has already stummed That means we often ourselves in the situation to say again, watch out, let's talk about what agents are or what agent AI really means and what it is not. Let's talk for the use case we here before us. What are the changes that it possible now and that maybe half year ago? But the background we see 2025 as the year of the agents. We have a certain five-step framework, we about the development towards AGI. And we really reached the point, especially with models like GPT-5, where we this instruction following, where we hallucinations, the error rate, so down, where we special Genti capabilities, which GPT-5 was trained and also coding on the level, where we really see that very, very many, a large part of all, example, automation that we see in practice at companies, well to a large extent. That a very different situation and we often to customers and said, sorry, we wouldn't do it, we're not sure that the models can and that we to the accuracy where we want That changed in 2025 and now this this era or epoch of agentic AI really started. There are many examples, maybe a few in Germany, mention customers with whom we work. For example, DKB, for in finance sector for customer service, would be the savings fund group, Axel Springer, Bertelsmann, example, where the companies in our partnership, which we a long time. and now they are unlocked and we can work on automating lot of fun.

Jens: think if you on such scale, you have other levers that really move I think that's really, really cool. Two questions before we get the topic of the 5-step model, which maybe not everyone knows. You very often that you are also very transparent and say, yes, I wouldn't now, maybe in three months, maybe we'll have three, four, five snapshots more. What are the

Christoph Winter: Well, one category of use cases where we always put on is overly broad workflow automation. means taking a very complex task and the expectation that one agent will based on the model. We often see as being too easily estimated. I'll you concrete example. For example, customer service is one of the AI use cases that has been in production time and has had good stories for With Business Value, T-Mobile USA, example, a big partner of ours, but also Clana, example, here in Europe, with which we also this in partnership. To such system is not, I have an agent who maybe a voice interface or a voice model in it, who then over these things, but it's an orchestra of different AI agents who different tasks. And and optimizing works and it is feasible today. And you hundreds of millions of interactions, as the two partners do, autonomously with very good user experience. But I just to invest. So it's a system that very complex, I have to optimize That would say, on your question, back, it's not necessarily that there are special model capabilities that we don't necessarily like. It's the expectation of what works out of the box and what feasible. But it to in a certain way. I have evals for I have to the systems properly. I have to an agent orchestration. I have to my guardrails in place and make appropriate investments for That's where I'm always very careful with customers and say, let's see what we can and what we

Jens: I find it very, very cool when you the AI development or the model development in last few months and a few years. We also a lot of updates. think snapshots are good example of what's in the background, because the models are a certain ability that is not fancy and does any viral tweets and LinkedIn posts, which probably a brutal impact on use cases that you also see.

Christoph Winter: Yeah, absolutely.

Jens: Let's talk about the 5 famous steps on the way to AGI. For all of you who don't know or don't know what to the term AGI. What is it? What I with

Christoph Winter: Now we will come back to definitions. AGI has many definitions and I think we should not be scientific to define How we define is an AI that all tasks that really big economic value than the best person in that area. We see that as a milestone in the development because we can say that the AI can everything

Jens: Mm-hmm.

Christoph Winter: which for us at Economic Value is at a very high level. So see it as a tipping point for the development of AI and the impact it can

Jens: And you or defined It's not a law, but what the five steps look like? Where are we right now and what's next?

Christoph Winter: We are coming stage 1, was a conversational AI. It was a classic chat GPT without tools. I used to chat, to information, to input on certain topics. I asked and got answers. That was the model. Stage 2 was a reasoning model. I had models that didn't give in system one thinking, but... who who could who could tools like web search as an easy example to really useful. We always say that task of JTBT is to a useful AI super assistant. We saw a step change with these capabilities to be a useful AI super assistant. Stage 3 is a Gentic AI. This is the stage we are in right now. That means I a transition from It can questions about it tasks for me. An is, you Deep Research, which is a first-party agent in Chattupti and Chattupti Enterprise, can complex reasoning tasks for me. give the AI, I Chattupti a task, it starts, does 20 minutes and comes with an analyst grade research report. That means it really on a task for me. There will be a lot more with an agent, example, which has its own browser and can implement for me. I'm joking, I haven't booked a hair stylist appointment for because I a task in GGBT and the agent book it on the appropriate platforms for me. But that's example of what Gentic AI is. That means it can over topics for me and I give it a task, it starts, it comes back with results. The next stage, stage number four, is what a knowledge contributing AI. That means it's the first time that AI is in a phase where can new knowledge to our society that no one discovered We see this in first steps, for example in areas like mathematics, where mathematical evidence, for example through GPT-5,

Christoph Winter: being refined and improved, which no one had proven before. Or especially in biology, when it drug discovery or certain connections in biology that can in the lab. Then we see the first examples of how GPT-5 can very well. But we are still at the beginning. But that will the next phase, where it us in a few years. That we say, who... Who made the connection, principle and the new scientific discovery? And it come that it was typically the Fife or another model. What I super cool about the mathematical evidence is that when you look at and then the reports of the scientist and their experience are highlighted, then it is often a process where the scientist puts the task to the AI. For example, with a new proof, but the scientist it as input to improve That means have a nice symbiosis of, there is the human genuity, but it combines with AI, which also contributes which then a really good impact.

Jens: This whole prompting is becoming more human. We are like real people. And prompting techniques on consumer basis are becoming more irrelevant.

Christoph Winter: Yes, think so. I think if we our job well, shouldn't necessary to take a course on to solve problem, but rather it. They should me, which we also have with, for in JetGBT, is very strongly contextualized, and they should understand what I mean, even if the prompt is perhaps not good. So, Jens, I write terrible prompts and get decent results. I think that should get much better

Jens: I hope too.

Jens: Sometimes I voice is a great lever for prompting. Everyone should do

Christoph Winter: Exactly. This is stage four, this is knowledge contributing AI. And then there's stage five, the organizational AI. And we define that as a form of organization of AI. How the AI is whether it works like companies or politics or parties or whatever, today, or otherwise we don't know. It will be very strange. But as an organization... a company like OpenAI outperformed in all the aspects. That would be the last phase five in the of AGI.

Jens: Awesome. I have a question I'm about AGI. But you're the only company working on reaching AGI. Do you think about what comes next?

Christoph Winter: Yes, I But I that the world post-AGI will be so different from what we currently know that everything we imagine is nice to think and to discuss, but very hypothetical. And I think can be positive. where work and creativity and what people have and can to develop and maybe to the topic of impact, how they want and to do that are simply a matter of will completely different. And means we very focused making AGI, especially AGI that value for all people. And I think post-AGI will be a very, very interesting phase for humanity where we... where we can only speculate about what it will like.

Jens: I think just having the openness, changes will come and how it will look, I we can all not imagine But that was a wonderful final question, especially the answer from you. Thank much, Christoph, for your time, for your insights, what you do at OpenAI for the enterprise topic, but also for your opinion and us take peek into how it is to the enterprise area Thank you much for your time.

Christoph Winter: Thank you Jens. Come when you Munich, come to office. We are

Jens: I definitely Thanks for the invitation and have a wonderful day.

Christoph Winter: Thank you, you soon, bye.

Jens: Tch.