TRANSKRIP 🔗
Thomas Wollmann: And now the governance part, so the sandboxing part, gets much more mature, especially through the push of Nvidia, like open share sandboxes and nemo claw and all these open source projects that they are pushing so that you can also easier design off the shelf sandboxes and not have to build something super custom like a few months ago.
Sebastian: Unless you're utilizing the latest OpenAI or anthropic models that basically jump out of every sandbox. No, just kidding.
Thomas Wollmann: Mm-hmm.
Sebastian: Welcome to another episode of our new season of Beyond Vibecoding partnering with Impala Search.
André: The go to tech and executive search agency in Germany.
Sebastian: In this podcast we explore the transformational change in software engineering and knowledge work in general. I am Sebastian Heidemeyer zu Erpen, CTO at Ecosia.
André: And I'm Andre, CTPO at Trusted Shops. Great to have you back. Today we are covering the perspective from a consulting agency again. We are talking to Thomas Wollmann from Mirantix Momentum. And again, I need to say some very interesting thoughts.
Sebastian: Exactly. The most interesting aspects for me actually are that he combines a strong focus on engineering concepts with the business perspective. We're talking about how team topologies shape his collaboration approach between humans and agents. we talk about DDD, and that many businesses currently are approaching Verantix Momentum with the question, will our business still be viable in 10 years? That's true. So interesting insights. Enjoy the episode. Welcome, Thomas, to our podcast episode. We're very excited to speak to you, and as usual, we'd like to introduce yourself to our listeners.
Thomas Wollmann: hi Sebastian. very pleased to be on the show today. I'm Thomas. I'm the managing director of Merantex Momentum, an AI consultancy based in Berlin. We help organizations in their AI transformation, and I'm building this organization since seven years. We have done over 200 projects with organizations especially in Europe in pharma, finance, manufacturing and the public sector, everything where transformation is painful. That's our cup of tea. Before I did that, I built an e learning startup, did a PhD in deep learning, as well as worked some time in healthcare and in automotive.
Sebastian: Well, thank you. thanks a lot. short f first short question. seven years means that it was pre GGBT. So I guess that the transformation or that the kind of work that you're doing changed a lot in recent years compared to before, right?
Thomas Wollmann: yeah, definitely. So seven years ago nobody knew about AI. It was a niche topic. where there was a lot of talk already in academia due to especially the achievements on ImageNet, but not yet in industry, especially in Europe. And this is also why we started building momentum, because we saw the potential. And we wanted to get this out as early as possible. So the first things we did were explaining executives the
Load full transcript
potential and then it also moved towards building POC, showing the value, and it was also pre in context learning, like what we do right now with like ChatGPT and so on. So building specialized models for specific tasks. Building automations of verticalized problems. And this shifted towards a more broad range application, but also more broad knowledge of our customer base as well as their needs, obviously. So now we have broad adoption, measuring the impact of AI investments. They're pushing foundation models in areas which is not text and so on. There are lots of new needs that just didn't exist at seven years ago. and the economy also wasn't ready for that yet.
Sebastian: Interesting. yeah, and
André: Absolutely.
Sebastian: measuring impact, that that's also something that would be great to cover later on because this is one of the main themes where in our domain I would say we see a lot of through like a an increase of throughput, significant increase of throughput, not so much impact. that's why this is a very important topic for us also to discuss later on. But before we get there, let's go to our first segment, the reality check. So how sorry, not the reality check, but the the the status quo. So how are you currently working? So where do you utilize agents in your day-to-day work?
Thomas Wollmann: quite a lot. I think I have a very broad range as managing director but also as a basically the technical mind of the organization. And I do slides, I have to do strategy look at data for controlling have to manage my direct reports as well as building prototypes and ad advising on experience that we accumulated in our past projects and basically from being there from day one sharing this knowledge with people who haven't been there. and this all can be augmented with agents. And you could now use like a Bazillion tools in your daily workflows, but actually I try to bundle this. So I'm a big fan of VS Code, and this
André: Okay.
Thomas Wollmann: means I try to integrate my workflows with that. If it's like note-taking to building slides to building prototypes or interacting with our internal data warehouse, which acts as our company. brain and accumulates all the experience that we built over the past years. And then this is basically my, let's say, control tower for everything I do. And in addition to that, I have some background agents running for recurring tasks. But I'm actually more a fan of triggering actions myself. than having a thing that is running all the time and is doing random stuff because then you have to really take care again of all the stuff that gets produced by these agents and nobody has time for that.
Sebastian: This I can imagine is unlike what you do with organizations, right? Because I assume there you have processes that you wanna automate and then there are triggers. So and these are not always manual triggers, but probably also automatic triggers, right? That so something happens and then an agent picks up some work.
Thomas Wollmann: yes indeed, but the same principles apply. so it's about the interfaces. So if you also think about there's this marvelous book on called Team Topologies, how to build interfaces between teams to optimize the communication and the cohesion in teams, and the same also applies with agents. If you need very high context and have like a Poorly defined API between humans and LMs, then this is actually creating more work and synchronization than you're actually benefiting from having an additional person or agent in your organization. So what we also do with our clients is actually look at what are the right interfaces so it so that you are in control and have the best interface to offload tasks to an agent. And this is I think what we tell our own employees as well as our clients.
Sebastian: great. And then maybe one or two last questions on on your setup that you described. VS Code is something we hear very rarely, if at all. Usually people are using Clot either as the the app or CLI or something similar, Codex, whatever, right? and VS Code is is very rare. And what what kind of yeah environment do you use behind it? So w what what LLMs and and I mean do do you then use skills a lot and yeah how how do you work with VS Code as your interface?
Thomas Wollmann: For me, that's mostly the UI. So underneath I'm using Claude Codex, Copilot, the usual suspects. And this also changes from month to month depending on what's the most capable model, the pricing, and so on. But I think humans are much faster if they Are really in one specific workflow for their user experience, or in this case developer experience. So if you know the OI quite well, how to interact with every component, you're very productive and you don't have to go through this learnability curve of a new tool and then just switching out one AI assistant with another is It's quite easy. so I'm a big fan of keeping this in like a common theme, and you get the advantages of all the other tools immediately. And I'm actually using a lot of MCPs to get context into my LMs, which is very helpful. I I shared before we have a company data warehouse where you can look at every aspect of the organization and leverage that when you are generating something as well as we have certain skills that we share in the organization to augment our workflow. So in the past you would maybe write like a SOP how people should do a process. A skill
André: Hmm.
Thomas Wollmann: is basically an SOP for an LM. So we basically translate our SAPs into skills and use that so it acts like what you would expect from a human worker.
Sebastian: An automated SOP if you wish.
André: If you say data if you say data
Thomas Wollmann: Yeah. The friendly word for SAP.
André: If you say data warehouse, you mean a a a knowledge space or a real data warehouse where you have like certain reportings?
Thomas Wollmann: as a database which we
André: Okay. Okay.
Thomas Wollmann: use for analytics as well as for lms and if you think about what we do we do projects and help organizations in transformation and everybody could claim that you bring a few smart people in a room who know how to build Software solve problems and you have a service company. But what is the thing that you have over time that others don't have? This is the experience you've seen experience over the years, and this is something that we conserved as data. for the next generation of projects that we do. So every project that we do compounds to a a better product for our clients. And this is something that we did in the past with humans, but we are also now doing it with our agents, so that our agents get better over time.
Sebastian: And we will get into more details of that in the main section. Last section, last question for the first segment. What is the environment your background agents are running in? Is this some anthropic environment or do you have your own environment? Yeah.
Thomas Wollmann: we have our own custom environment that's part of the company brain where we we are running custom agents.
Sebastian: Alright. Well then, thanks a lot. so that that was already quite interesting and a little bit special. And yeah, now let's go to the main section of our podcast. And yeah, let's maybe start again by or start from the beginning. What actually do you do at Mirantex Momentum? You already described it on a very high level, but maybe if you could go a little bit more. into detail on how you exactly yeah help companies adopt AI. Yeah.
Thomas Wollmann: So maybe if we start from what I already said, we see ourselves as the AI transformation partner for organizations, especially in Europe. And this means usually we start with the strategy from the boardroom and see where an organization needs hand holding, where there are gaps in the business strategy from the execution but also the vision on how to actually interact with AI and the changes that is coming to every organization. Is the business model still defensible? Is the roadmap there to deal, for example, with the upcoming IPOs of LM providers and so on. do I have my supply chain secured So that I can still provide my product and business continuity in the future. So all strategic topics where we help the organization shape their business strategy. And then we don't stop there. So we help also in the execution of the strategy. This means rolling out software with clients helping measuring. how the software is actually bringing value, which is not an easy thing. If you look decades back at how computers were introduced into organizations, I think none of the organizations back then saw immediate value of computers, because you invested in the engine of how you run your company, but The bottom line didn't change immediately. This is something where you have to look at how to actually see if your engine improves and then ensure that later you see this also on the balance sheet. As well as we build custom systems for our clients. Many of our
clients have very high RI AI use cases, which There there's n just no product on the market. So we build it for them. and this can be a new drug discovery tool. This can be a demand forecasting algorithm. Yad yad yah. There are plenty of verticalized products that are very relevant for individual companies to build for themselves. And sometimes this even goes to helping machine learning research teams progress on their research agenda. So help them in experimentation, in shaping their research agenda, and sorting out what are promising research paths and what aren't. And we Do this because we can provide the full chain from strategy consultants to actually full blown researchers because we maintain a research team of around five postdocs, which I think is a very rare thing for a consultancy that you actually get real researchers who do this full time.
Sebastian: Interesting. And you're s so if I understand correctly, then you're pretty much in the execution phase. You're looking or screening the market for existing solutions that can be rolled out in order to yeah, support the actual AI induced transition or transformation of the company. or you build custom solutions depending on if you can find something that that suits the needs of the companies, yeah.
Thomas Wollmann: Yes, indeed.
André: So your angle is business, technology and research, less the organizational transformation?
Thomas Wollmann: I think you can't decouple it. I think there's also always a socio-technical system where you introduce a new technology, and maybe if you think the technology is the solution, I think 80-90% of the solution is humans. So I think the biggest mistake that we see over and over again is that project champions who want to introduce the technology, forget that they have a lot of legacy and processes structure, which has to be changed as well. And that this is often even more work for us than solving a technical problem.
André: I can imagine.
Sebastian: Yeah, definitely. this this is the the the cultural shift that comes with it. That that's the the difficult part, right? And I'm I'm still a little bit struggling with or trying to connect the dots from you said that your focus is AI transformation and and then what you described on the strategy level sounded like you're looking at how AI will impact companies in general. So the the business models and yeah the the interactions, the collaboration. in the company is this pretty much the starting point and then then derived from that you you screen the market for for existing solutions or or build solutions.
Thomas Wollmann: I think it doesn't make sense to introduce AI just for the sake of it. So it always starts with okay, I have a organization, I want to grow it in a certain direction, or I still wanna be there in 10 years. So we start with that vision, which is usually like the first thing you do in a business strategy, and then evaluate how Does AI interact with that? And often nowadays what we see is that
company owners and executives are very afraid of if they can if they still exist in 10 years or if AI just ate up their whole business model. So this is a thing that we do very often right now. consult on the threat of AI and also the opportunities and then you hopefully s have a corridor where you can transform your organization into that you still stay relevant in the future and maybe capture a new market and then we help the organization to get there.
André: Super interesting. You this means you're starting actually pre-product strategy. Actually you building the business strategy together with the companies. Usually I could I could also imagine consultants come or join a company and then implementing the product strategy using AI technologies. In your case, it's very different. You actually start developing the business strategy together with the company.
Thomas Wollmann: I think this is a major shift in the past years. Before I think we actually did a lot of that help with a concrete problem with the product. and this is also still a part of what we still do nowadays. But what really shifted is that AI is not just Just not one tool in your organization or one little thing. Organizations make this core of their business as well. This is one of the biggest threat and also cost factors and for for an organization. So you can't just see this very isolated in one part of the organization, but have to include this into your core strategy. And this is why we are also moving more and more on advising on business strategy, because AI strategy and business strategy is not two things anymore.
Sebastian: Got it, yeah. And now now your
André: Makes sense.
Sebastian: initial statement actually connects the dots for me as well. yeah, thanks. So and and this already hints basically probably at how the conversation with your clients has changed throughout the journey of the company, right? So initially it was probably more like okay, there's this new thing, deep learning, AI, whatever you want to call it, and let's see how we can utilize it. And now it's probably much more around. Wow, this is going so fast and we cannot see where the end game is. will our company still exist in five years, right?
Thomas Wollmann: yes indeed. These are exactly the questions our clients ask themselves. and the thing is, out of the sphere, what you can do is just crank up the speed, say like, okay, let's do all agents, automate everything, transform the organization at max speed. But we still have to consider that for example, LMs, they are probabilistic machines. When can you actually trust the output? When does it do mistakes? And now you are building everything on that. At the same time, you take the humans out of the loop because you increase the velocity so much that humans don't even can't understand of what's happening. For example, if you have, let's say, software development, coding agents, and all the code they produce, It's very hard to really grasp what is the impact of the thousandline count change. And this means you have to think a lot about what are the right guardrails of these systems so that you don't derail them by increasing the velocity. And you can still be actually in the driver's seat. Because end of the day, a human is the operator who is accountable for these machines, if it's a coding agent, if it's a car or whatever device, and if something blows up, still a human is liable for the consequences. So you also as a human have to trust your AI. To do the right thing with the right guardrails. And I always compare this to a car. Not everybody understands how a car works, how exactly everything is connected, but I still operate it and I trust that the brake is breaking and it doesn't kill me while I operate it. and we have to implement the state the same guardrails for AI. in our processes. and I think this is something which we have to do alongside increasing the velocity to still be in control.
Sebastian: And how do you sorry Andre, please go first. How how do you think about guardrails? So what is the the approach you take to guardrails when you think about a new process that you've automated or you're in the process of automating? So how how would you do this? And and maybe can you share an example?
Thomas Wollmann: I think this comes at multiple layers. So I think first of all there's the big word of alignment. So actually that a model or system does what it is intended to do. And this gets more difficult if the tasks that a system can do is more open. And there are just more chances that it does something which I didn't design it for. So this means you can align the model through evaluation, through density checks in the distributions through drifts, through fine-tuning, and so on. So the model itself, so that the thing that it actually should do is more probable. But this is just a soft basically guardrail. What's really interesting if you mix it with hard guardrails, which is additional models as a judge, which is handcrafted rules and this then provides hard guarantees in my in my process. for example if I know something is Just not plausible. I can just forbid that the model is outputting that. as well as if I can build a proxy, for example, in code, which also LM can do itself, I can give feedback to the model output with an actual calculation rather than guessing what an LM is doing.
André: So deterministic stuff actually.
Thomas Wollmann: Actually, yeah, deterministic stuff. And I think a good combination of increasing the likelihood sort of probabilistic alignment as well as deterministic rules is what we need.
Sebastian: Yeah, I would say that in in our world of agentic engineering, we are talking exactly about the same thing. So also we talk about coding factories or dark factories as they are being called, and this is exactly the same, right? So you have the a mixture of code being generated with then deterministic checks that you anyways have, be it tests, be it lintering
infrastructure, and and then you add like review agents and then you have LLMs as judges on the final PR, whatever. So very similar. Yeah.
Thomas Wollmann: Yeah, I think if you go to coding, I think that's a very interesting thing because if I look back, we had this wave of model-based software development, which reminds me a lot of what is happening right now. So you try to be very precise as the specification of a software, and then you have a black box, like then some rule-based system now LMs that then implement something based on the specification. But I think people who have been in software engineering for a long time know that humans are very bad at writing perfect specifications. So I think what's also very relevant is a good loop to actually see where there are blind spots and wrong assumptions. inner specification and removing them or clarifying them for an LM loop. Otherwise we end up in the same state that the promise of generating code automatically is not fulfilled and then people start writing code by hand again.
André: The approach you just explain, is that following a certain framework, so a mental framework, s so something n so not not a technical framework. is that something you came up on your own or is that following, I don't know, some external best practices?
Thomas Wollmann: I'm thinking more like from a machine learning perspective. So here we would for example think about the counterfactuals in an explanation. So if you think of do I have examples that challenge my specification, then this is something where you can iterate on. And if you think of how humans do that, just intuitively they do the same when they are building specifications. So I think what we have to do is instead of coming up necessarily with new frameworks here, look at how the frameworks that were built for humans were done. from a statistical point of view as well as from an organizational point of view. So I'm taking a lot of ideas actually from classical domain driven design frameworks, which I think are still very relevant today, to decompose a problem into what are actually the functional non functional requirements that then can be implemented and where you can identify blind spots in the modelling of your business process.
André: No, not the first time that we hear this. So compared to VS Code, which is a rare case, this
Thomas Wollmann: Okay.
André: case we have heard I think I can at least remember two more times. So this seems to be a a good practice. A good practice.
Thomas Wollmann: Good to hear that good old stuff is not forgotten but still lives today.
Sebastian: Definitely. Also, I I liked your remark to team topologies as thinking about interfaces between humans and and agents who are becoming team members more and more as we introduce them, right?
André: Hundred percent. True.
Sebastian: So really interesting, yeah. and now you so you we were h quite high level so far, right? So talking about strategy coming from certain questions from
from the company owners or the management, then breaking this down into strategy execution. So how would that work? Do you usually look at existing processes in the company and then try to see where you can bring these processes on the next level or yeah, how how do you do this normally?
Thomas Wollmann: Hmm. So we pick up companies where they are actually are right now, and then try to help them with their most important decisions. And this can be, for example, we have a lot of grassroots initiatives in the organization. People have their coding agents, their copilot, whatever, and the organization maybe bought some licenses. And now we want to adopt a common way of working across the organization. So here we help with analyzing the organization, the processes, to understand what's actually going on there already, and then see how we can support across the board with technology. And in the end, maybe the decision is by this platform because it perfectly mirrors How you run your organization. And maybe some functional or non-functional requirements. For example, a big topic right now is AI sovereignty, to be independent from the cost hikes that are happening right now with LM providers, as well as having access to models that And yeah, no matter the geopolitics or having something that comes with European values. And if you might know, we are building our own foundation model in the SOFI consortium. So it's not like we just rely on the big players, but we also bring our own foundation model and also the expertise to to build these kind of models. and this is a big topic. Where do you want to stay independent, where you want to maybe be quicker but be dependent, how important is cost? what's the technology that mirrors my organization structure and processes? And then we have with these decisions and then implement it with our clients. For example, if you by a platform let's say Vertex or Azure Foundry, just to give two examples, then this just brings you a platform for a workload, but we haven't solved the business problem yet. So we help building ML systems and on top of that and prioritize them and so on for our clients.
Sebastian: And like picking up on the topic of sovereign AI, what is the the current stance that you take there or how do you implement sovereign AI right now with companies?
Thomas Wollmann: Hmm, I think sovereign AI is having a choice. And if we think about software, this is this is a stack, so you have many layers of the cake from hardware to basically who runs the hardware, the model on top, the data the model is trained on. as well as the the observability and so on. So many layers where you could say where do I want to have a choice? and we help organizations in picking the the right providers with the trade-offs that they are accepting to do. I think being 100% sovereign is impossible, otherwise you have to build chips on your own from sand from your backyard. And then build everything on top yourself. You always depend
on somebody or something. but doing the the trade-offs or the dependencies that you think are reasonable is something where we navigate with our clients through. and if we think about what is the biggest worry of organizations right now, it's I have data in my organization where I think this makes my organization defensible. If I'm losing this data, then I'm basically killing my own organization because somebody else can replicate the knowledge that I built up over decades. So organizations want to have inference. At partners that they trust. I think that's the biggest question that organizations ask themselves. So we help them building inference systems locally or for example in Europe, helping with routing data depending on the content as well as then also. based on how much you're willing to spend on the individual requests. and I think a thing which is not yet there, but probably will come in the future, is as well as Basically having the systems in the cloud, having the systems more and more on the edge, actually in processes itself locally, not because of the costs or being dependent on a specific provider, but in addition, you want to have them close to where it's happening because you want the performance on the shop floor. in a robot whatever. and you don't want to stream the data to a different country, as well as the old question still remains to a non-trusted organization. And I think this is something which especially with the introduction of AI in the industrial sector, which is like very prevalent in in DAC will arise more and more. How to actually achieve that without getting into the hands of non-trusted organizations and hand them basically over your own factory.
Sebastian: Yeah, makes a lot of sense. And actually I like your phrasing of sovereign AI as a question of having a choice and looking at the the different levers. That this is actually exactly how we approach AI at Ecosia also, in our product but also in our agentic engineering efforts. And now maybe last question on the on the on the actual execution before we then maybe come to the final point of impact and and then close the main section. do you do you see a pattern of like specific processes that you always or in most companies or in lot a lot of companies are trying to transform via AI or is this totally different because your clients are totally different?
Thomas Wollmann: a big chunk is actually in back office tasks where you have a lot of knowledge retrieval. And in the past you had to search for knowledge, had to Aggregate a lot of unstructured data into structured data by hand. And now with the with LMs having at hand, this is something everybody wants to introduce. And we see this across the board because no matter which product you have, which vertical you are in, everybody can pull this productivity gain. I think another big one is actually something not many people talk about, which is prescriptive AI. So that you model what your organization is doing, let's say old school as a digital twin, and then have planning method to do better decisions. And this can mean you Test if there's a better pricing strategy in your organization. Or you test, should I invest in this or that country? How should I distribute my fleet in my organization? So this combination of I'm building a replica of my organization and then testing different scenarios for better decision making is something which I think is not that mainstream. But actually is something that we see a lot in organizations and I think also where a lot of RI actually is.
André: Yeah, and it actually brings us back to the discussion or the point we had earlier where it's not about implementation, it's about business decisions. And I really like that fact. so applying AI actually earlier in the process, like really from the actually from the beginning, beginning, beginning, right? Not just putting this AI sticker on your products, but like think about that from the beginning on. Maybe sounds so obvious,
Thomas Wollmann: Yes, I think that's
André: but it's I I think it's actually it creates quite some headache because you always need to keep in mind, okay, now where we have the capabilities which haven't been there before. now can I apply now I can apply that, right? And it actually it allows a very different kind of steering and organization, I couldn't imagine.
Thomas Wollmann: yeah, 100%. I really like a phrasing of one of my senior engineers who basically said, Every day I ask myself, why should not an AI do this? And he just does something himself if he believes that an an i an AI shouldn't do this. And I think we can apply the same basically way of thinking how we approach running a business in the future. Where we in the past thought a lot about how humans can do this and maybe we how we can augment this with automation. Now I think this flipped and we have to think in how we are building businesses and let's imagine everything can be automated. Where should a human should do something themselves?
André: And I need to stress that it's actually for me the learning or the the the the takeaway of that episode is really usually in AI adoption we think f f on the second half of that transformation, right? We do not start from the very beginning on I think this is actually where it should start from. Not just like like using AI to then implement what is already there, right? Your thoughts, but like really stepping back, think about okay. what are the capabilities? How will the market transform? What should be my positioning? Where do I have advantages, given all the data, and so on and so on? And then you take it from there, but derive a business strategy, derive a product strategy, and not just like take the product strategy which is there and then using AI implementing that. I think yeah, I would not talk this down and say like this is simple but probably not as powerful as the approach you just outlined.
Thomas Wollmann: Yeah, a hundred percent.
Sebastian: Yeah.
André: Thanks.
Sebastian: As you mentioned, right, that that's also where the strengths of LLMs are put to use because now you can actually do these these different run these different scenarios based on unstructured data. Before that it was super expensive, now it's costing you a couple of minutes and obviously tokens, but it these are the the two areas I al also always look at. So number one you have Of obviously the efficiency gains. If you can automate a process, then you have efficiency gains, right? On the other hand, you have the the you can increase the top line, right? If you're thinking about launching a new market or improving pricing, whatever, right? So these are the two areas to look at. And we already touched on on the one other topic several times, and maybe before we close this episode very briefly. on impact measurement. So how how do you measure impact on your improvements that that you implement with companies?
Thomas Wollmann: So I think best is if you can measure it directly in the monetary advantage. But this is actually quite hard. It just works in very specific use cases where you really know what you're doing. For example, you have A process which creates some cost, for example, manufacturing a specific piece or running a a delivery network and now you want to basically optimize it with AI. In the end, you can look at the costs again, what you basically got invoice, and then you can compare and you see the RI. But that's
Sebastian: Yeah, the efficiency the efficiency gains. That that that's something you can measure.
Thomas Wollmann: Yes, but that's very rare. I think that's a example we can all work with, but if you look in the wild, something like that is very rare. So instead of that, as I said before in the episode, looking at the outcome is often delayed. For example, if you as in the example before, introduce a platform for better knowledge retrieval. You don't see this immediately on your your top line because there's no costs immediately. You will also not reduce a lot of headcount immediately because people will not just all quit or you fire them, or even if you do, you have a lot of side effects which cost you money in the short term. So you don't see this in a monetary value somewhere on your yeah in your finance department. However, what you can do is look at the engine. Did it improve? So did you have a shorter process cycle? Did you had less incidence? Did you basically make decisions faster or didn't had to revert them? These kind of things tell you, okay, the engine is improving. And then you should also be able to expect a return later on. And then you can hopefully see this more long term also on the balance sheet.
Sebastian: That's again throughput, right? So looking at throughput is giving you hints if your improvements
Thomas Wollmann: Yeah, indeed. But not
André: Yeah, but also
Thomas Wollmann: just throughput,
André: isolation.
Thomas Wollmann: also quality. I think that's I would
like to to stress that because just increasing
Sebastian: Fair enough. Yeah.
Thomas Wollmann: the velocity can also mean that you are producing more shitty stuff. so we also always have to look at both.
Sebastian: You need to have exactly you need to have these balancing quality metrics, definitely. Or even just focus on the quality metrics. If you if the throughput is the same but the quality is going up, then still you have a gain. absolutely. Yeah.
Thomas Wollmann: Yes, indeed.
Sebastian: All right. Cool. Thanks. With that, I would conclude the main segment of our little podcast episode and would now go to the final recurring segments. So number one is the reality check. So we usually ask if our guests have sometimes still what the fuck moments or if there are if there are topics or technologies or approaches that are really hot right now. So do you have either or maybe both?
Thomas Wollmann: I I have both. I think there's so many what the fuck moments that that I see. I think especially on long running agent tasks, these the current LM still derail sometimes randomly, like crazy, and sometimes in very funny ways, where at some point it just decides to refactor your whole code base and you just wanted to let it improve a certain I don't know section in a loop. But since you're running in a loop, at some point by chance it just yeah wants to change everything. and there's so many of these things happening where you start an LM and then in the end it just blows up and you see this either in the changes and then you can revert it but also in in your costs so you then basically in yeah run through the roof with your with your tokens by some very you would think harmless prompt but in the end it just blows up everything Over and over again, I'm getting amazed how it can derail sometimes, even if I'm very deep into that field. and what is hot right now, I think Pi-based agents I think are very interesting because the ecosystem is evolving quite fast and it's super flexible and cool. So this is a very small and flexible harness to build autonomous agents on top. So not just like a copilot, but very autonomous. And there's stuff like OpenClaw built on that. I think that's the most prominent example. And people build very amazing things with it. And I think that's that's very cool to see.
André: Would you see this also in an enterprise context? So really large scale, not just for maybe engineering and local environments, but also at scale in companies?
Thomas Wollmann: yes, we do that. It just has to be again properly guardrailed. and
André: True that, true that.
Thomas Wollmann: and this is definitely possible. You just have to define the interfaces, harden the the sandbox and then I think it's ready to use in enterprises, just not blindly.
André: Yeah, there's still the
Sebastian: All right.
André: the reason why I was asking is there is still the question, some every now and then are there companies around which are operating based on OpenClaw? and that's why I was curious to better understand if you're running Pi somewhere at scale. But yeah, thanks for thanks for sharing.
Sebastian: Yeah, and I I would say that there is a difference, right? So Pi is obviously at the center of OpenClaw, but OpenClaw comes with
André: I know.
Sebastian: a lot of integrations already that you can set up but don't have to set up, right? And you you can choose which ones you you set up. But if you want more specific, yeah, very special agents for specific tasks, right? Then you may be better off with Pi or now even Pi durable. Right, running on on Cloudflare infrastructure, that's also quite interesting right now.
Thomas Wollmann: Yeah, a hundred percent. it depends on how much you wanna customize it. If you start with Pi, then you can build your own harness and can customize it to your needs. If you just need to have something where you hand out stuff, you can go with open claw. And now the governance part, so the sand the sandboxing part, gets much more mature, especially through the push of Nvidia, like but open share sandboxes and nemo claw and all these open source projects that they are pushing so that you can also easier design off the shelf sandboxes and not have to build something super custom like a few months ago.
Sebastian: Unless you're utilizing the latest OpenAI or anthropic models that basically jump out of every sandbox. No, just kidding. all
Thomas Wollmann: Mm-hmm.
Sebastian: right. last question as usual. do you have a prediction for our listeners?
Thomas Wollmann: I think velocity will increase. I think we just started with cranking up the speed in organizations in our society. and I think to not end up in a disaster, we really have to take care to have the right systems and guard trails in place to actually live in a better world. And not destroy everything that we built in the last centuries.
Sebastian: Amen. And
André: Got rates, got rates, got rates.
Sebastian: And it yeah. But
Thomas Wollmann: Mm-hmm.
Sebastian: but but before that, I mean yes, velocity will increase and this actually makes head spinning, I I feel like makes my head spin because velocity already is crazy, right? So how long can we still cope with that? But we will have to. there's no way around it, we will have to. Yeah.
Thomas Wollmann: We will have to. Let's prepare for that.
Sebastian: exactly. Alright. Thanks Thomas. Was great
André: Thanks.
Sebastian: speaking to you. Bye bye.