Masters of Data, BI & AI
Podcast fΓΌr Unternehmer und unternehmerischen Erfolg im Data, BI & KI Umfeld
Masters of Data, BI & AI
Open Source Doesn't Mean Nobody's Accountable | Michael Berns & Rocky Khan - #39
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
π The most powerful computers in the world run on open source. The internet runs on open source. And yet most companies still think "open source" means "cheap license fee" instead of strategic infrastructure.
Michael Berns spent seven years as Director AI at PwC Germany, recognised as a Leader in AI Services by Gartner and Forrester. Then he left, to become Head of AI at the Eclipse Foundation β the Brussels-based, vendor-neutral home for over 440 open source projects, including the IDE most developers have used at some point without ever thinking about who's behind it.
In the current episode of Masters of Data, BI & AI, he debunks the biggest myths about open source AI, tells the real story behind his move, and explains what European digital sovereignty actually looks like once you get past the slogan.
A few things that stuck with me:
On the three biggest myths about open source AI:
It doesn't mean free β many projects come with SLAs, quality gates, and years of paid development behind them. It doesn't mean less safe β an open codebase is inspectable by anyone, which is a stronger security posture than a black box, not a weaker one. And it doesn't mean nobody's accountable β when a project sits inside a real foundation, committers are named individuals, and Eclipse just introduced voluntary identity verification for contributors.
On why he actually left PwC:
No corporate spin here β becoming a father, Brexit, and what he candidly calls "a late midlife crisis" pushed him to rethink what he wanted his work to mean. He'd spent years volunteering on the side for causes like the UN World Food Programme. Eclipse Foundation offered him the chance to make that his actual job, not a side project.
On PanEval, the story Rocky specifically wanted to hear:
Michael brought Chinese BAAI technology into the European open source ecosystem to build PanEval β an EU AI Act-aligned, open source AI evaluation platform. Not a press release. A running project, and a concrete example of what building optionality actually looks like instead of just talking about sovereignty.
On the real data problem, even now:
Despite everything AI can do today, 60-70% of real implementation work is still spent finding, cleaning, and governing data. Department heads will tell you their data is perfect β until someone actually checks. His advice for consultants and vendors alike: do the full assessment first, manage expectations early, and get buy-in from the people who actually hold the data, not just the C-suite.
On where regulation could help or hurt open source AI:
The current OSI definition of open source AI is strict enough to disqualify roughly 99% of models on the market. Michael expects the field to move toward a graded system instead β matching transparency requirements to the sensitivity of the use case, healthcare needing far more disclosure than a low-stakes internal tool.
What's your take β does true open source AI require full training data transparency, or is a graded system the more realistic path forward?
LinkedIn Profile
LinkedIn Profile of Michael Berns: https://www.linkedin.com/in/aithoughtleader/
LinkedIn Profile of Rocky Khan: https://www.linkedin.com/in/khanrocky/
Business Book
The Five Dysfunctions of a Team by Patrick Lencioni β Michael still returns to it years later. His reasoning: the more senior you get, the more your real problems are people problems, not technical ones.
The most powerful computers in the world run on open source. The internet runs on open source, and yet when it comes to AI, most companies still think open source means cheap, license fee, not strategic infrastructure. Today's guest is building the infrastructure that changes that framing. He's head of AI at the Eclipse Foundation, the vendor neutral home for 440 open source projects from the IDE. You might be using right now to the AI evaluation platform that will shape how Europe governs AI. I'm Rocky, I'm host of Masters of Data VI AI, and every week I talk with thought leaders about the impact of data at AI. Good morning, Michael. Happy Friday.
SPEAKER_01Of course. Thank you for having me.
SPEAKER_00No problem at all. When I ran your CV, I saw you have been working for a long time, quite a long time, at one of the renowned Big Four vendors, PWC. And then you became head of AI. Can you tell us a little more why did you change? Why does someone leave one of the big four?
SPEAKER_02This was a very personal uh journey, uh Rocky. You have to remember that um I had some 16 years working uh globally out of the UK. So being based in London, but actually being based in London to fly globally from Heathrow or um in Europe from City Airport. So, you know, uh it took a while to to kind of settle there, but uh becoming a father and and and Brexit meant that I was thinking about what to do next. And I applied uh at a number of places uh in in Europe, plus uh in Asia, and as it happens, uh P2C pressed the button first. So then uh my journey was clear to help them to get ahead in terms of AI um and their positioning in the market and in the German market and and and so on. So I I uh enjoyed my role there, but for some family reasons and and some wider changes in my life, you could call it a late uh midlife crisis if you want, but um I uh I took some time to take take uh stock and to really think what to do with my life, how to have more purpose, or how to um kind of apply myself, my skills and knowledge to to things that make a difference. And for me personally, when when uh Eclipse Foundation approached me and asked whether I want to uh contribute in terms of the AI policy and in terms of the AI ecosystem in Europe and help Europe to get ahead, that was an exciting uh challenge for me. And while I've never worked for a foundation as such directly, I've often spent some of my free time on uh volunteering for good causes. Some of them are more in the um fintech for good, some of them were more on the AI for good. So I worked, for example, with uh United Nations World Food Program and a number of other initiatives over the years. So I've always done some kind of foundation charity as on the sideline in my free time, and that gave me an opportunity to um really do this as my main um uh job and and and role. So it's in in a way, you could say it brings me closer to my my purpose, has more meaning for me than just doing another AI implementation or project. Um, but that is a personal choice that sometimes in your life you you want to make.
SPEAKER_00Thank you very much for your very straightforward and uh honest words, Michael. Can you tell us more what does head of AI mean in a non-profit foundation context as opposed to it to a tech company?
SPEAKER_02Sure. So um unlike my previous role where I was, for example, um often supporting uh CDO functions and and CIO functions and and so on as uh director for AI at PwC, this one is a slightly different different. There's an element that is more outward-facing in terms of ecosystem development. So that means I work um with uh AI startups on a daily basis. I I look at um what are the propositions um that they're putting forward? Is there a chance to collaborate openly with them? Um what are kind of the the major themes going on in the market? But I'm also working with our internal um AI governance um uh kind of committee where we're looking to take the the whole of Eclipse Foundation forward in terms of our own uh journey um towards uh a higher um AI maturity. So you could say it's kind of a a two uh uh a role split in two, um outward facing and to some extent also inward facing. And and so it's not so much just building functionality inside as more collaborating, um working with you know the startups and also some of the the corporate members in that space um to build something meaningful um in the open source arena.
SPEAKER_00I'm confident that you have seen now many, many clients in your lifetime since you have been working at PWC now at the Eclipse Foundation. You now mentioned that you are also working with startups. Let us clarify a little bit about the myths about AI. So, what would you say? What do most people, even when people who are working in AI, still get wrong about open source AI?
SPEAKER_02So, I mean, for for me on the open source, um often the discussion is uh oh, this is uh free. Um, and of course, um it is correct that um some some work in the open source arena is uh provided for free, but there's often multiple tiers, and you often get some common um like uh SLAs on top or some kind of quality gaze, and therefore it becomes a product. Um I think what what you also have seen recently is the accusation that open source is is uh less safe uh than some of the proprietary. Um I would also divide that Murph by saying that the fact that it's open, that anyone could inspect the code base and run their own analysis on it, um, that it's open to all the vulnerability scans, to all the um mechanisms that will provide a safe implementation, makes it much more transparent than um a lot of the uh propriety solutions. And that means that um, well, in in this very dynamic world in AI where you have releases like Mythos or um others uh being classified as dangerous, um nobody will be absolutely safe under any circumstances, but at least by using a transparent solution for some of your key domains um can give you um, let's say uh a better shield than um other paths. And then the the last myth maybe to uh explain is um that if it's open, then um nobody's responsible, right? Um it's that might be true for some small open source project that is hosted um by uh a small uh player somewhere um where you can just use it and and so on. Um when a project comes to a foundation like us, then you have um fixed roles, the people that are committers are uh named. Um we've just last couple of days introduced new roles for voluntary um kind of um verifying the identities of of contributors and so on. So really going a path of uh transparency and responsibility. And that means that while you don't have you know your service agreements and so on, so you haven't got a formal route to complain, you still have people that you can uh notify, that you can relate to, that you can um kind of make responsible for for everything that happens based on the projects that you've been using.
SPEAKER_00Thank you very much, Michael, for clarifying up those myths. I would like to dig deeper a little bit about Eclipse. Yes. So I'm pretty sure many people know Eclipse now from IDE. So, what is the uh Eclipse Foundation now as an institution? Can you can you tell us a little bit more about that?
SPEAKER_02Sure. Um, I mean, as you say, the IDE is what most people associate with us. Um but the foundation is uh a different thing, right? So when it originally was um set up, I think 2003, it was a spun uh spun out from IBM where it was decided to release the IDE code um as an open source project. Um today we are a Brussels headquartered European organization. Yes, we still have uh a strong presence in in Canada, um, but then most of our employees are European-based uh today. And um, when you look at the address, you will find it's very close to the European Commission and um all of the government men there. So, in terms of the political and government alignment, the the statement is very clear. Um, and also yes, we started with the IDE, but today we've got more than 400 uh open source uh projects. Some of them, you know, uh OpenVSX, for example, they get 300 million downloads um uh a month. Uh and others, like the uh Jakarta EE, um, are now pretty much the biggest uh Java-based open source um place um uh in the world. So you you might just know for the IDE uh and from your development work. I also used it when I was at Morgan Stanley to design some um um ETLs uh with um uh Java back in the day, 2008-2009. Um but it was very clear once I had the first kind of discussions with them that the organization has very much moved on into many other fields. And you could also say the same about the industries they're in, from the uh SDB for um the car automotive industry to some of the IoT or security things that we're doing. It's a pretty broad um portfolio, and the AI piece that I'm working on is kind of kind of really cutting across a number of those industry kind of groups.
SPEAKER_00Thank you. Thank you very much for that, Michael. Um you said IoT, uh you there are more than 400 projects, and very often, Michael, I see almost everywhere there is this Pareto Pareto distribution. Yes, do you also see within the 400 projects any specific industry where you say, Rocky, this is really dominating? Um, that would be interested, uh interesting to understand.
SPEAKER_02I mean, based on the DNA, of course, anything with Java that is kind of uh used to be our bread and butter, but it's very clear that we've branched out from there. So we really have projects across Python, Rust, and many different uh languages as well. But you know, based on the 20 years or so in business, there's a lot of the historic stuff we've done in Java, and that might be still dominating. But you can tell by STV, for example, the automotive ones I mentioned, we've got the majority of the German car manufacturers, and there's some international ones in Japan and elsewhere, um, where this is a big initiative to build like um the software for um operating and and and so on. Um and that is something that where a lot of the Europeans are interested in in terms of this being the next step. Um so you could say that there's some some definitely some uh distribution across some major industries that are based on Europe, right? Um in terms of what we see here in Germany, in terms of the heart of Europe, is very much automotive trim. Um, but then if if I have discussions sometimes with the Middle East or elsewhere, it could be quite different uh topics in the IT or in the cyberspace.
SPEAKER_00Um could you tell us a little bit more about the use cases? Yeah, so do you see any use cases which are reborn or which are always uh favorite? I mean, you worked at PwC for I guess seven, seven years now at Eclipse. What what are very specific common or favorite use cases what you always face?
SPEAKER_02So, I mean, personally, you know, with this uh 13 years or so in applied AI, the use cases that I used to see back then were of course very different than what we see today. So even back then we would call it AI, but when you would be honest, there was uh RegEx and there was some um ML on some occasions, and then eventually uh some um neural networks and deep learning and so on. Today you've got your um gen AI and agentic um coding and so on. So the topics have moved on. Um the ability and uh the desire of companies to do AI have also moved on, right? The early discussions, even with American firms some more than 10 years ago, were more does that thing work? Do we really have to do this? Do we have to put our data here? And and then over time, you you know, I worked on a lot of these uh target operating models or infrastructure projects where you spend one, two years just um getting the data in a central place and um building the infrastructure so you can then build uh powerful and performant uh models on top of it. I feel a lot of those discussions are now maybe not fully done, but the the basic infrastructure around this is there. Maybe the skill gaps and some other things have resolved. So the focus has shifted. But in terms of what the favorite use cases are, the cluster has changed. I mean, I used to work a lot in financial services, as you said. Um I've had maybe around uh 22, 23 years in financial services, mainly with some kind of uh adventures in healthcare and telecoms and so on on the side. But um, from that perspective, um, the first movers back then were not doing AI because of, oh, we must do innovation or we have a spare innovation budget. It was really the regulatory fines. And we're talking about some of the American banks where the fines were in uh 200 million, 1 billion uh dollars, and that eventually included European institutions as well. And when you have that kind of level fine, you start looking outside of your firm. You're starting to look at startups, you're looking at fintechs, you're looking at joint ventures, you're looking at getting vendors in for things that may take you two or three years to develop yourself, right? So then you you do innovation via that program and you try to solve your own problems by getting more insights, more control. And maybe it was in some of these early things where yes, I was competing with uh Palantir and some other um big players, but it was really trying to give the banks some insights that would prevent uh another regulatory fines in the multi uh millions, if not billions, um space to give them more control. And then over time, just kind of protecting them in terms of regulatory and compliance use cases moved on to. So I've got I've harvested all this data now. What else could I do with it? Could I build uh market insights that gives me some ideas about optimizing my trading? In the early days, I built solutions that were um harvesting uh Twitter back then uh near real time um uh for insights on commodities and and other things. And um so you could build use cases that were using alternative data and satellite data. But then moving on today, where a lot of these compliance protective use cases are built, it's more now enabling people and in some kind of way also making their own resources uh more efficient. So the banks have really shifted their focus in terms of what use whose use cases they're doing. But that's also a a sign of the times in terms of the Gen AI um uh availability, where back in the day I would spend days um just to uh work with my team to build an initial model and then test it over weeks and weeks with a client and provide they can they provide you know feedback back in an Excel sheet in terms of supervised learning. We then go for another loop and another loop. So the POCs tended to be really, really long. Um today you start at a very different level in terms of Gen AI models. It's more about okay, given given what the client use case uh looks like in terms of what um kind of how fast does it need to respond in terms of latency, how how smart does the solution have to be? It's more you've got more in your toolbox and you can build up those solutions very quick. And then jumping to the other part of your question, I mean, a lot of the discussions I've had at events I was visiting over the last couple of months, whether it was our own uh OCX uh in Brussels or uh Gosim in Paris or some other smaller events locally. Um, a lot of this is on the um agentic um side of things, but it's more, I would say the um agentic hype was pretty much last year. This year is a lot about the security discussions. And you could say that you know the release of Mutos um contributed to that, but I would say there was awareness, particularly of clients in Germany. Um, I would really love to try all this uh agentic stuff, but I I don't want to expose my systems to it. So I create some Docker container or I do it in a sandbox, but I can't do it for real. And last year, my discussions with prospects were still, yeah, yeah, we are running some of those POCs, but they're all kind of interested in how do we bring that to reality um in terms of in a secure way. And and so some of the propositions that we've been working on are very much in the agentic harness in the um agentic security space to make sure that you know um a number of the bigger players can collaborate and then kind of come up with secure standards uh together rather than one dominating players and and and so on. Is that different than you know, um given your podcast, what you've been seeing in the discussions with with others? Is it also the your focus in terms of the agentic security?
SPEAKER_00Oh yeah, thank you very much. There's a lot of uh content. I I will I would also come back to your to your point, but um uh I want to understand also one one thing you said um that you were also partly competing with Palantir, right? Michael, in one of the projects. Can you yeah?
SPEAKER_02Um that was in my early days back in back in London when I was with uh Digital Reasoning and uh Palantir was uh like us, uh uh originally funded by the um CIA uh venture fund together with Peter Tier. So we were um seeing them close up at a number of organizations.
SPEAKER_00And Palantir is extremely well connected with uh the governmental, with with all the agencies in the world. Uh I just want to understand with such a big player, Palantir, how were you competing with them? I mean, uh what were things like what were your USPs or differentiators? I'd like to understand more.
SPEAKER_02Just to clarify, I'm I'm certainly not uh competing with them uh anymore at the moment. And um, I don't think they're involved in any open source projects that we host, would be interested to see them. Um, but I haven't had a discussion with them uh for some years. But going back to that situation, I mean, um Palantir with the Gotham and Metro offering, they were basically um doing this uh fast forward um deployments of their engineers. So Some of the larger French banks, as well as some of the Swiss ones, were using them to gain a lot of insights that they wouldn't have otherwise. The problem was back then that it wasn't a product offering, but more an expensive data scientist solution offering with like 10, 20 people on the ground. And that meant that even those organizations that gained a lot of insights and value from this were paying around 5 to 10 million a month, not per year, which is in terms of return of investments. Once you are beyond the initial insights, not always easy to justify. I feel Palencia have their point in terms of defense and in other areas. But back in the day they came under a lot of pressure in terms of the financial services arena and other things.
SPEAKER_00Thank you very much, Michael. I would like to also understand when it comes to those projects with AI, there are always a lot of problems people, clients, clients face. Most of the common problems that I have here now on my podcasts is Rocky Data Foundation. We come up and we see there is a lot of mess, messy data. So this is like I would say maybe 80 or 90 percent of the problem. What would you already nodding? Okay, I'm not saying nothing new, but I was thinking maybe there are other things, other issues. I mean, you you you have a great overview over over 400 uh projects, I guess. So is there anything else what you would add?
SPEAKER_02I I think we we're gonna separate this a little bit. Um back in 2020, I know it's quite a while ago, I did uh a study with about 150 or so um CDOs and and chief of AI and so on uh across the DAC region, and we we covered some obstacles in terms of deployment. These obstacles have shifted. So my former colleagues at PWC have published more on this, but what I would say is back then it was uh around the skill gap, uh around data, around um systems, around buy-in, those typical themes. And I think, and as a consultant, that hits you harder than others, right? Where the client comes to you, can you build me the system? And back in the day, sometimes was it was like, Can you build me an AI system? In the recent years, it was more can you build me my own LLM? And um, you have to take a step back and kind of really um think, is AI even the right solution for the problem that they are having? Can they describe some more? Um, is Gen AI the the right hammer for for this nail? Um is it does it make sense to build uh this kind of things? But over time you learn how to to manage the expectations and to kind of make them aware that you know not every use case is suitable for this. But as you as you correctly mentioned, the data is often the problem. And I've seen that in early days of AI as well, where um you you work with the department heads, and of course they're telling you uh the data is perfect, it's very complete, um, good quality, and and so on, because that's what they are being told without having done a very thorough analysis based on what you know the different units, sometimes global units, uh, had several global uh rollouts uh giving them, and sometimes they don't even have access. And then when you um work on the vendor side and you try to scope that project, over time you learn that you get burned if you don't do a full assessment on your side. Maybe that's not required, um, that is not necessarily paid for, but if it's about managing expectations later, that will save you. Um, so over time you learn, gotta do the first week of full assessment uh first, and then give your comments in terms of does this project make sense? Yeah, that's maybe not what everybody wants to hear, but that is what you you probably have to do, and then also give some um indicative um success criteria that you might be able to reach based on the data. Now, that discussion has changed a little bit. Given Gen AI, you started a higher um base level, as we just heard earlier. Um but it the the problem remains the same, right? You could um download public data sets from Kaggle, from wherever you want to and and build something with it, but that doesn't mean USP. And that means you still rely on the bank or on the um corporate solution to um the the the company to provide you with something. And apart from the the quality earlier, it's often that they give you some test data, which then turns out not to be representative of the the whole uh set. So you fine-tune too much, you have to do a lot of iterations to get it then work, and then at the end, everybody is disappointed if you only reach maybe 70% accuracy. I think there's more potential to do this better now in terms of Gen AI, but it all depends on what are you trying to launch uh here. Is it something internally? Are you providing a system that was previously not there and you're all running it maybe on small local models that doesn't leak anything uh to anyone outside the firm? Then that system might still provide value, but you're probably not going to see like an 80% or so. And then it's kind of depending on the use case, might be a mix of um like uh on-prem, hybrid, and and and so on solution, um, with also a mix of public data with with your own enrichment or a reg solution that prefers your own one. It it really depends. But as you say, I would say the the data is now the main hurdle that compute you can solve if you're willing to in terms of cloud and other things. Um so the data rather than the skills or um the buy-in should be the main problem now. If you look at all the statistics, I mean all the CEOs are seeing Gen AI as absolutely critical, or I would say the far majority of this, and the rest would still recognize that this is a factor they have to deal with. Um otherwise they might not exist in the current form um after some five years or so. So everybody has to deal with it. The data problem probably needs to be solved. And um, if you've done the implementation, then you know that that is the maybe uh unsexy part, uh, but it's the reality on the ground where maybe in some cases you spend 60-70% of the work getting the data, finding the data, cleaning the data, uh managing the data, governing the data. So it's it's still all about data until we get to the cool new stuff and the Gen AI things.
SPEAKER_00Thank you very much. I think I have um seen in the last two decades more than I don't know, 100 innumerable studies that most of the time, 80% of the time or even more, uh the time gets wasted only what you just said with data crunching, ETL prepping, etc. Yeah, that's wrapping up, etc. So this is still uh it's another still, I don't want to say it's funny, but it's uh yeah weird that it's still still still a current problem. Yeah, it's not fully solved.
SPEAKER_02I mean no, it's not fully solved. If you were honest as a vendor or as a consultant, sometimes you would really love to go and say, you guys sort out your problems and you come back when when you're done and and you have your things ready. But of course, that's not what you're gonna say, and then you're gonna say, I can help you to sort it. But sometimes it's it's really this this journey where you are exploring an organization where you kind of go by the org chart and try to see who's responsible, but you also try to find data and and you kind of really need the internal buy-in of the different levels to to guide you to that data. So um the critical mistakes sometimes people make in the early stages of the project, if they're just communicating to C-level and senior management, is if you don't get the buy-in from the admin people and the data administrators and uh so on, you probably never find the critical pieces because then it's a very long journey, and before you can prove any value, you might be out. So how you communicate the whole project at the different levels and how you involve people and make them feel part of this thing and kind of take fears away and and making their work life more interesting going forward. That is all for me really key success factors overall. So it's not only about you know, oh, we we need to fill this gap, it's also how how do you incentivize and motivate people to show you where else could should you be looking for that data, right?
SPEAKER_00Now let's yeah.
SPEAKER_02You could compare it to you know the the search for the holy grail, you will never find it, but um the the data is distributed across the whole organization, and nobody has all the pieces of that possible.
SPEAKER_00Thanks for for your insights here, uh Michael. I'd like to shift now the the conversation to a different perspective, this different topic. Um you are engaged with uh G7 under the French presidency and with the UN AI advisory body. Well what are the conversations typically there?
SPEAKER_02So I mean I can't talk in in detail, of course, about um uh specific conversations, but um it's it's uh from from from our side, I mean, given how close we are in terms of um Europe and and um the scenery there, um of course there is a lot of interest overall and in the digital sovereignty piece, given the way that the world is uh shaping up, uh, right? And then um I think the um kind of what you could perceive as a threat by uh frontier models and kind of the open question um for Europe and others, do we kind of demand to have access up front, or should we um uh I think the uh Trump just released uh some some guidance on this um a short time ago where yes, the US government also would like to have access, but they decided that they cannot uh impose or force it. Um, Esther would like to have it. So I think Europe and others, they have to find their own uh way of dealing with the new reality that we are in the middle between East and West. We could do more on AI, but we are probably short on the energy. Um, because particularly in Germany, we've decided not to do nuclear. Um so it's kind of the question: how many more data centers and um AI factories and things can be reasonably build unless we somehow find a way to speed up our energy infrastructure, whether that's both alternatives and others, and the power grid. Um and and the same question is probably for for yes, rest of Europe, but then it's all about you know the sovereignties, uh the discussions about the um the uh processes and and the chips, all the key pieces and in the puzzle to really be somewhat more independent. And there's a lot of um those those questions to to solve. And when it comes to United Nations, I mean I was um on a panel with um uh two representatives from United Nations about two weeks ago, um it was about open source and open data and open models. I I personally was surprised how much interest the UN uh takes in uh open source. Um they're now in about a month's time hosting their own uh open source um week uh in New York. Um so you can see um open source is is definitely not a niche topic, and it's it's very clear, clearly connected as a key building block to digital sovereignty, or let's say a a fair way of companies of collaborating um with each other. Companies often that might be competitive in another scenario, right? But there's always enough incentives to say when you look back at the DVD or USB, or um, you could argue that MP3 by Fraunhofer was initially also open sourced, and so on. You have enough examples in history where there is um better, a better case for doing something together around the table where everybody has the same influence and and and vote, and where the standards that will benefit them all should not be dominated by by one player, but you want to take things on board from others. I compare it a little bit with you know the situations I've often seen in in boardrooms where sometimes you have these um boards where essentially the the bias of higher people that are like yourself have led to what is an echo chamber, but the ones that you want to really go forward with and that are more sustainable are the ones where you allow different voices to be heard, maybe some concerns that you wouldn't have known about otherwise, or some industry or uh some um um thoughts cutting across. And I think you build a more robust uh solution if you allow for those other voices to be heard.
SPEAKER_00Thanks, Michael. I I I have a follow-up question on on that. Where do you think regulation is most likely to help open source AI and where uh could it maybe damage it?
SPEAKER_02I feel we are still um on on on that front on on the the the earlier days, right? I mean the definition by the OSI in terms of open source AI. Um the way it is cut right now. I think it was defined about a year and a half ago. Um I would say it it rules out like 99% of models because they're basically um uh not uh compliant with the full open source definition in terms of um sharing um the weights, the uh rules or methods, and the the training data, right? So I think where we probably will move on to is a more graded system where we we look at the different components uh separately and we match that possibly with the use cases. So let's say, for example, in in healthcare where you have some of the most private data in in the world, you probably want to be using a very uh transparent model uh where you have um uh open weights, but you also know what the training data set was, you have some um full transparency in the outcome and and so on. So I think we will be getting to a more um graded system rather than a uh black and white, this is open source AI or this is not. Um but I also feel that there's more awareness of you know uh what this transparency should mean. And for me personally, I mean I've been following all these um legal cases in terms of the um people, sorry, the some of the model builders being sued for uh copyrights and and so on. I feel uh a lot of those legal cases are uh are not done yet or they will be um decided over the coming years. But at some point in time, hopefully we'll get to a place where the value of data is somewhat clear. So if everybody wants to use Reddit or YouTube, Wikipedia or whatever it is, there should be a a cost or a value put to this, and then you wouldn't have a problem disclosing some of this, right? That might um kind of shift the uh balance a little bit, but I think that will help people wanting to contribute their data over time, and maybe that's very ambitious, but I I think we will get to a place where yes, you can build some uh European um open source models by Mistral or others where all of the sources are named and and disclosed, and we can still achieve um uh a very performant outcome for a certain set. But I also see um in the near future there will be more of these uh smaller models that are more um specific to a industry or or subset, um, where we will have hopefully a bit more disclosure in terms of what went in and what are the expectations there.
SPEAKER_00And I'd like to understand more if um listener now who has heard this podcast, listen to this podcast, wants to engage with Eclipse tomorrow or next week. Can you give us a tip where should they actually start?
SPEAKER_02Sure, sure. So I mean if you are active in in building something, so let's say in a startup or you're doing this as a hobby, then of course your heaven, you could just be look at some of the projects we are doing in in this space. If you are in AI, um look into the eclipse.org AI uh projects. There's a number of them. Um I already mentioned um the Open VSX, which is used by a lot of the bycoding and other tools, but there's um Elmos, uh, there's Ethia AI, there's um Edge, there's a number of them that you can uh pick from, and then see whether this is something that you would actually be interested in contributing or use in your organization. Equally, I mean, even if you're in a startup and you're building some um proprietary tech, what you often see is that um kind of sourcing some submodule or a component of your solution to open source to kind of test it out in the masses to see if it gets traction, if it attracts people to your paid-for solution, that can also be a good way, right? You have a component that you could think needs wider distribution or should be validated by others in the market and where you would love to collaborate with others, then consider um talking to me or to the foundation in terms of whether you want to um open uh sources or whether there's an avenue for others to collaborate with you. So there's a number of things. And I mean, with me personally, I'm I'm pretty big on LinkedIn. You can find me under AI Thought Leader. Um happy to connect there and and and discuss open source AI um and and look forward to some of the engagement on there.
SPEAKER_00Thank you very much, Michael. And um, normally you should end with the next step. Now we started with the next step. My last question to towards you, and it's my favorite question. I ask it to all my guests. Which business book inspired you most?
SPEAKER_02Uh I've I've I've uh read uh so many, but um I I would say that uh the ones that were the most helpful over the years, and you know, in the executive MBA London Business School, we got to read like I don't know, 60 or more. But the ones that I I still use uh even today is like the the five dysfunctions of a team by Patrick Lanzioni, um Table Group and so on, um, where it's all about psychology and team dynamics. I feel that still got value even is uh is older, or um have a nice conflict, right? Because the more senior you get in your career, the more you will be dealing with people problems and not technical or any other problems, and the better you want to understand and motivate and incentivize your team, the more you will feel you're really working towards a purpose.
SPEAKER_00Thank you very much also for the recommendations. Yes, it's uh most of the time it's always about communication or interacting with people. This is also my perspective. But Michael, thank you very much for your perspectives for this open conversation. I wish you all the best and of course, a nice sunny weekend.
SPEAKER_01To you as well, Rocky. It's been a pleasure being on here.