 Hello. So I know this talk is, the speakers are listed Miley and myself, but she had some international traveling problem, so it'd be just me, and you have to bear with me now. So the talk is, I changed the title slightly according to the return of the open core, because open core used to be something people talk about, and then people stopped talking about it. You know, then people, you know, then, I think with the new generation of, say, people finding new, trying to find business models in cloud native, and also very interestingly, open source, large language model, people have started to talk about open core all over again. So that's, would be the topic of my talk. And so my name is Michael Yueh, and I'm the founder of the CSF project called Wasm Edge, so we are one of the, you know, server-side web assembly runtimes, and I'm going to tell you more how that relates to the idea of open core and the large language model and things like that. So, well, many years ago, I used to work at Red Hat, you know, that was, in fact, I started in a company called Jboss, and a couple years later, Red Hat put us. And around that time, it was a time that the first generation of open source business models had been figured out. So I remember in early days of Jboss, and also in early days of Red Hat, we were selling CDs, right? We were selling software installers, and, you know, we would go on site to do consulting for people, right? You know, that's, people who use our open source software, we go there and do consulting. We all thought that's a lifestyle business. We thought that's, you know, we contribute to open source for fun or whatever, right? You know, and never thought it would really be a scalable business. But, you know, then, you know, I think in the early 2000s, Red Hat figured out the business model of open source. And that's, so, you know, I took several pictures from this blog post, I think, and Peter Levine did it, you know, a couple years ago. That's before the current problem with, you know, open source models, open source licenses in cloud software and also in large language model. At that time, you know, he talked about, you know, the commercialization of open source. I think around the time of 2018, 2019, you know, right before the pandemic, that time, right? So he talked about three generation of, three phases of open source commercialization. The first is, you know, from the 70s all the way to 90s is mostly just the people do it for the love, right? You know, for the interest, you know, whatever, right? You know, that's, and then from the 1990s to 2000, that's the era of, say, Red Hat and a generation of companies, my sequel and companies like that. So what they have found out is there's a business model that comes with support and services. You know, I remember very vividly, you know, that's when we first started, you know, why did we know when there's a real business to be had is that when people start to call us and ask us, first, who are you and second, how do I pay you, right? You know, and the reason they are not asking us to go on site for consulting anymore is because they have figured out how to use our software. However, they need someone to take responsibility when the software fails, right? You know, so at that time, you know, we used to joke, that's our business model is really the insurance business model, right? You know, is that we don't really provide a lot of services, but we provide insurance, you know, in case something goes wrong, right? You know, that is a really good business model. And that's what made, say, Red Hat is a multi-million dollar company, the name, my sequel and all that. But in the years after that, we have found there's very few companies can replicate this business model because, you know, I think someone wrote a really famous article that says, you know, the Red Hat over XYZ is Red Hat. You know, that's meaning that there's no opportunity to use this model in a niche market. And the reason for that is that when people pay for insurance, it needs to be mission critical software. You know, it has to be something, if it goes down, then, you know, the CIO may lose his or her job. You know, the CEO may have problems. You know, so that's why they pay for insurance, right? So the software that fits that bill is only the operating system and database. Nothing else fits that bill. So that means it's really difficult to commercialize software that outside of the scope of, say, you know, the open source database and the open source operating system. So, but open source moves on. You know, that's even though, you know, it's the first generation of a business model is discovered and then, you know, successfully commercialize those people, those companies when IPO and made a lot of people rich. Then we can't replicate that anymore. So there's no more red hat after red hat. Then there's open source 2.0 where a lot of companies figured out that they can provide the cloud offering of their software, right? You know, so I have some kind of infrastructure software and enough people use it. I provide a cloud hosted version. And the cloud hosted version have some benefits, you know, mostly it's tooling and easy to use and easy to get started, right? So that's where the SaaS business model comes on, right? And so the next slide is also from the same article. You know, so at the time it was mostly focused on, say, service and support is the first generation and SaaS is the second. And he had the insight to put open core in the middle. However, the companies he put in the open core is really what we would consider today. We would consider SaaS companies, right? Confluence is a Kafka version of SaaS. And you know, Elastic is the Elastic version of, you know, is the SaaS version of Elastic Search, right? So although, you know, even in 2008, 2019, there's distinction of those three business models that support open core and SaaS. Really, the ones that made it are support and SaaS. And there's a long footnote in the article about why open core is not considered to be closure, right? Because open core is a mixture of open source and closed source software. So you want to distribute those two together. And that's typically give the community a bad vibe. And people don't know which part of the software you're going to close next, right? So and the part that is closed source that you have to pay is on some critical paths. So people tend to stay away from those. So that's why, you know, for a very long time, I think support and SaaS are the dominant business models. However, things changed, you know, in those past five years, I think we started seeing that from MongoDB first and Elastic, those are the big ones. And then we can, we see HashiCorp do that. We see LinkerD, which is a CNCF graduate project, right? You know, so they all moved away not only just from the SaaS model. You know, that's they, the way they change their license is they become not open source at all. You know, they become source available licenses, which when people look back at this, and they think it's a lot worse than, than open core, you know, because with open core, at least you know, there are some part is open source and some part is closed source. There are clear, clear distinction. There's this expectation from the get go for the community, right? But for, for successful projects, and well, I have deep synthesis for them, by the way, you know, because I'm also a software developer, I also try to monetize, you know, open source software, you know, it's, it's, it's very disheartening to see that AWS with scale or Azure, you know, any cloud provider with scale can offer the same service that you offer based on your software, much cheaper than you do, because they don't need to incur the original R&D cost, and they have much larger scale, and they have a much lower customer acquisition cost, because people are already paying them for other things, right? So, you know, so it becomes really difficult to compete, and those, and those players come up with ways to change their software, very specifically, not allowing cloud providers to compete with their SaaS offerings, right? And by doing that, they made themselves not open source anymore. They made themselves a source available. Source available, I remember, is one of the terms Microsoft used, you know, when they hated open source, right? You know, they say Windows is a source available for government client customers, so that, you know, you don't have to worry about security issues, because you can always look at it. But, you know, you can't really contribute to it either, right? You know, so that's, you know, the, the, the state of cloud, cloud native at this moment, you know, that's, you know, a lot of those companies started from open source, and they grow big, they become SaaS, they made money, they were an IPO, they made 100 million dollars, you know, even billion dollars, but then they find out they can't compete with AWS and Azure on those fronts, and they have to change their source code license. And that goes back to Peter Leving's article. I think this all started because of this graph, right? Because of so many open source projects, REC found it, so they have to look for return, you know, so that's very different from the open source that, you know, say, before 1997, right? You know, where there's virtually no money in there, because nobody knows how to make money from that. But, you know, post 2008, there's, there's, with success of Red Hat and the companies like, you know, SpringShots, MySQL, and, you know, things like that, there's lots of, you know, funding, which is great for open source development, but also fast forward 10 years from that, we now see the effect of all this, you know, a profit motive that's, that's, that's, that come back to HONUS, right? You know, as software community. So people start to think, you know, maybe open, like I just said, maybe open core has benefits, right? You know, so it's first, it's compared with things, where things are 25 years ago, it's a lot easier to understand open core now, because back then, it's difficult to understand, you know, what, you know, what's, you know, what has a mix or combination of features that are open and closed, but today we are entering premium products all the time, and we have sort of come to expect what is, you know, say a consumer product, what would be free and what would be costing money, right? You know, so I think the general rule of thumb is really anything that to do with collaboration would cost money. Anything that to do with reporting to the boss would cost money, right? But all the basic features that for individual users would be free. So I think the new generation of open core models would be followed the same rules. So you would see open source software that's, that they would charge, you know, the particular pieces of it, like monitoring or, you know, high availability or communication, you know, forming a cluster and, you know, things of that nature would become more and more closed. However, the basic functionalities for individual developers to start on their own laptop and on their own would be, would stay open. So, yeah, that's, so I think looking back, a very clear benefit of open core really is that it sets the expectation, right? You know, it's to say, I wrote this software, I want to make money from it, right? You know, that's, so, but I also want a community. It's not to say I, let me develop a community first and have, you know, community product fed and product community fed and then change the license later, right? You know, so it would be, I always want to make money from this software. So I always kept something that I saw large enterprise would need and perhaps, you know, a regular developer wouldn't need, that's as close as us, right? So, we're talking about, you know, all those, while interesting, sounds like the battle of the last era, you know, it's, you know, we were talking about things like 20 years ago and, you know, that's the business model that was 20 years ago and, you know, so, but why did we propose this talk? Because all those I thought was fairly well known, you know, that's because we see a new opportunity that's, is tries to redefine open source all over again. That is the large language model, you know. So, if you look at the, the so-called open source large language model has been making a lot of noise since OpenAI came out and, you know, we all know it's not open, it's closed. That's, you know, that's, so they had the whole rational argument why they need to be closed and then, you know, the traditionally not very open source friendly Facebook meta came out and opened and released the Lama 2 model and build a huge ecosystem around it, right, you know. But if you look at what they have released, they have released the model weights under some kind of open license, okay, and they did not release the data to train it, did not release the program to train it, so you cannot reproduce it. The, you know, I thought that's what we call closed source software, right? You know, that's the source code and the way to reproduce the binary is not available. I just give you the binary for free, right? You know, that's, so I always thought this open source large language model is a misnomer. It's that, it's, you know, it's very much open source only in name, you know, because what they released is actually a binary format of a model file. It's not something that, that individual developers can, can reproduce. So if we look into a little detail about those, you know, like the Lama 2 from Meta AI, you know, they have, they release the whole series of models, however, in their licensing terms, they, they said commercial use is only available for companies with 700 million MAUs, you know, which essentially excluded all their, you know, large competitors, right? That phrase probably doesn't affect any of us, you know, but it's, it does affect, but strictly it is not, it's what made it not open source by the OSS definition anymore, right? So that's, but even then, you know, even they have this course, I think they still are the most generous, you know, software developers, you know, because the, because they release the whole thing and they release the model architecture, they release code around it. So because of that, there is a huge ecosystem of it. If you look at just this one person, you know, he, you know, quantized a lot of, you know, fine-tuned Lama 2 models into the GGUF format, which is a inference format, right? Under his repository, he has almost 4,000 models, right? You know, so those are, there's a tremendous amount of, you know, fine-tuned and derivative models that come from Lama 2, you know, so that's open source, although I wouldn't call it open source, again, I wouldn't call it open source, but, you know, it's, that effort has to be super successful in building an ecosystem and achieve the goal of someone who want, who may want to open source stuff, right? And then there's company like the Mistro AI, which is based here in France, right? You know, and they released model ways under Apache 2.0, which is a lot better, which is in a way better than what Facebook did, you know, that's, you know, Facebook, okay. So what they did here is that they only released it for their smaller models, like the 7B models or the MOE models, the eight experts, you know, times the 7B, you know, so that's, so they're larger models, like the Mistro Media and Mistro Large are still SAS only, and they don't allow other people to compete with that, right? You know, so it is also a partial open source strategy. And then there's the most recently released software called GROC, that's from Elon Musk's, you know, x.com, you know, that's, he also made it so that technically he released everything under Apache 2.0, under open source license, but really he doesn't release anything because the model is really huge and it's infeasible for any individual developer to run, and it doesn't have any of, it doesn't have fine tuning, you know, so it doesn't follow conversations, so even if you can't run it, you can't converse with it. And to figure out how to fine tune it, I think it's going to take a major effort, however, because he has the model which released, so I think someone's going to figure it out fairly soon, right? So this is the current status of open source models. Well, you know, there's, I think we're running out of time, but there are other things we can get into, you know, that's people, people also asking, you know, why do you need those open source? Why can't you just use the web APIs, and you know, things like that? There are lots of reasons why you can't, why you shouldn't, or you probably, it's very, it's not okay to use the web APIs for, you know, for our own use, we have fine all those, you know, central hosted models constantly refusing to answer certain questions, right? You know, because it deem it politically correct or it deem it harmful, you know, whatever it is, right? You know, so, you know, there's great benefit from cost, from alignment, from safety, from everything to have your own model that's built for your own use case, right? You know, that's based on, some base model and fine to yourself, right? So what do we envision that OpenCore could play a huge role in the future of LIMs? I think we are getting really close with those open source releases from Meta, from Mistra, and from X.com. You know, first, I think it should be, we should release complete open models with the training data and all that, but the OpenCore component comes with the fine tuned stuff, you do not have to be released, right? Because fine tuned, you could use your proprietary data, but you could use your own proprietary data. And as I discussed, most of the open source model use cases are such that the model itself are not huge, however, it's fine tuned for specific purposes. So for most of the commercial applications, you could, in fact, you know, that's, you know, make money from it by, you know, having the fine tuned component as the open core, as the core, but the base model as the open, right? And then, from that point on, so this is the second to the last slide. So, you know, we can also envision a world where the model runtime is also open core, meaning that the model runtime works on most of the hardware that you can think of. However, there are data data center hardware. So for instance, the Google's TPU or AWS, you know, Neurochip, and you know, things like that, a lot of those are hardware devices that has NDA attached to it. So you can't, even if you write a driver for it, you can't really release that open source. So you could build a model runtime that is, that target many different hardware devices and do it this way, you know, that's having to release that open core as well. So it works on most people's device, but if you really want to go enterprise, or if you really want to work for, offer commercial services around it, you could have the, the closed source component attached to it, right? So, well, so that's, so I talked a lot about the landscape and all that. And you know, that's, so I'm at the end of the talk here. So this is a shameless plug of our, our software, which is, you know, the same CFS was a major project. So what we do is that we follow this open core model, we build a build a runtime that runs on the Mac using the GPU on the Mac runs on NVIDIA runs on AMD runs on NPUs. However, it has proprietary plugging components that allows it to, to, to run efficiently on, you know, proprietary hardware's doing model inference so that, you know, you can write a Rust application and compile it to Wasm and then have it run everywhere, everywhere of any devices that you may have. Yeah. So that's it. Thank you.