#240 Manos Koukoumidis: Why The Future of AI is Open-Source

0:00For something to be open source, it needs to be open data, open code, and open models or open weights. This means that it should be possible for somebody to have the ingredients and the recipe to be able to reproduce the model. Yeah, it's good if you're giving them the model, the weights, but also they should be able to reproduce it. I think this is a very, very good definition. However, what we say is that we want to have open code, open data, open code, open models. And what we say open collaboration, what this means by this, which may be a little bit less clear, is that something that is open source by the OSI definition, it's not sufficient.

0:33If, yeah, the code, everything gets there, but it's too hard for anybody else to reproduce and build upon. Because then it is the whole purpose of open source being a community effort where it's easy for people to build upon and extend somebody else's work. AI coding assistants like GitHub Copilot, Google Gemini Code Assist, and Amazon Q Developer have quickly become essential tools for developers. They generate code with remarkable efficiency, significantly boosting developer productivity. But the widespread use of AI-generated code brings its own set of challenges. Bugs, vulnerabilities, and suboptimal code can inadvertently enter production, leading to issues with maintainability, stability, even costly outages.

1:25How can organizations minimize disruption and risk while maximizing productivity and innovation with these AI tools? With Sonar, the code quality and code security leader, organizations can amplify developer productivity in concert with AI assistants. They can improve developer experience with streamlined workflows and prevent quality and security issues in all code from reaching production. Sonar's AI Code Assurance feature found in their Sonar Cube solution offers a thorough validation process for AI-generated code, including the first in its industry to automatically detect and review AI-generated code from GitHub Copilot.

2:17The AI code assurance workflow helps to build trust in AI-generated code, assuring companies that proper due diligence has been performed and code is ready for production. Their AI code fix feature speeds up productivity and issue resolution with instant AI-generated fixes for issues discovered by their code analysis that developers can review and apply directly within their workflow. Sonar is helping redefine the software development lifecycle by use of AI and AI-agentic systems with tools that streamline the verification of AI code and provide guided remediation of issues, supercharging developers to build better applications faster.

3:11Join over 7 million developers from organizations like the DoD, Microsoft, NASA, and MasterCard who use Sonar. Visit sonarsource.com slash ionai to request a free demo and see how SonarCube can help your team with machine-generated software development. That's sonarsource, S-O-N-A-R-S-O-U-R-C-E dot com slash ionai. That's E-Y-E-O-N-A-I, all run together. and request a free demo. sonarsource.com slash IonAI. Very nice to be here, Craig. I'm Manos Kukumiris. I'm the CEO of UMI. Until about nine months ago, I was at Google Cloud. I was supporting the science and engineering for all the natural language AI services.

4:11I built also some of the multimodal ones at Cloud AI at Google. And about three months before ZGPT was coming out, I was bootstrapping the efforts to productionize CloudPalm and drove this effort almost until the general availability launch in May of 23, before it was called DeepMine and sorry called Gemini and moved to DeepMine. Before that, I spent some time at a startup. I was at Mera working on commercial AI. Also before that at Microsoft, where I was building something like ZadGBT in 2016, a multimodal open-ended chatbot, even built what now we call embedding retrieval and RAG back then. And before that, I was doing a PhD.

4:49I spent my first three years at Princeton and the last two at MIT. And before that, I was doing my undergrad in Greece. Okay. And UMI, what does UMI mean? UMI stands for Open Universal Machine Intelligence. We wanted to pick a name for the effort that, for one, clearly denotes that it's open. We want to make sure that this is very clear. And also we use the term appropriately, unlike perhaps what's happening in other cases. And we didn't want to say AI because we think it's a little bit overused term, because where everything started was machine learning, machine intelligence. So that's why we call this open universal machine intelligence.

5:36So universal means is we wanted something that open machine intelligence that is democratized, that people can use everywhere and anywhere. And that's why we end up with the name. Yeah. And UMI is a platform, right? So, I mean, I was interested in talking to you because open source, there's been a long running debate in the artificial intelligence world, certainly in the generative AI space about open source versus closed source. There were a lot of hand-waving in the beginning about, my God, what if this stuff is open source? Then all the bad actors in the world are going to have the most powerful AIs.

6:26People like yourself and Jan LeCun argued that open source is the way to go, and that if anything, having more eyes on the model creates a safer model. and since certainly I mean there's been a growth in open source you know starting with Llama by Meta and then Mistral with its models and then DeepSeek sort of came along and woke people up if they weren't paying attention to China that open source is a big part of China's AI strategy so I wanted to hear from your perspective just first generally how you feel about the open source closed source debate and then my second question is open source is an overused and under defined term and what exactly does it mean?

7:31There's open weight, there's open code, there's open training data and no one seems to agree on what open source really means. So let's start with just the open source closed model debate. Where do you stand on that and how do you see it developing yeah i agree craig if that's okay with you how about we start from the definition just make sure we're on the same page about when i say about open source what do i mean about open source great once it's clearer you know when i argue that that's the way forward you know it's clear what exactly i mean by this so um as you mentioned before there's a lot of efforts quite often many of them uh being called open source when they're not really um i think the osi the open source initiative definition uh for open source ai it's pretty on point it says for something to be open source it needs to be open data open code and open models or open weights this means that it should be possible for somebody to have the ingredients and the recipe to be able to reproduce the model yeah it's good if you're giving them the model the weights but also they should be able to reproduce it i think this is a very very uh good definition however However, what we say is that we want to have open code, open data, open code, open models.

8:53And what we say open collaboration, what this means by this, which may be a little bit less clear, is that something that is open source by the OSI definition, it's not sufficient. If, yeah, the code, everything gets there, but it's too hard for anybody else to reproduce and build upon. because then it is the whole purpose of open source being a community effort where it's easy for people to build upon and extend somebody else's work. So that's what we mean by open collaboration that for one, we have the principle that any of these AI technologies that we also are building that are easy for anybody else to use in a fully record, sorry, very easy for somebody to reuse because what somebody else did is fully recordable and the platform has developer ease of use as first principle.

9:40and also to extend a bit also the collaboration aspect is that we already have and we plan to have even broader open collaboration efforts where we'd say you know what here's the goal the problem we plan to solve anybody in the community is invited to help us make it better so this can be more inclusively developed it's almost similar to what was happening in some ways with linux compared to unix where it was collaboratively advanced but again to do this it needs to be open code open data open models but also in a way that's easy for everybody else to reuse build upon and contribute So that's our definition of open source.

10:12We think it should be nothing less because anything less is doing a disservice to the community and not helping us as humanity advance this technology in the best way possible. Let's now move to the debate. You know, why is this the best way possible? Actually, before you do, open data is the one that's a little confusing to me. Open model means the source code is published. Open weights mean the weights are published. I at the models I've looked at people describe the source material for the data but no one that I've seen really uh has uh you know a database that you can look into and see exactly what data is in there and I presume because there's a lot of concern about uh copyright liability and that sort of thing.

11:07But can you talk about what open data means? And are there any models, quote-unquote, open source models today that have truly open data? Yeah, yeah, very good question. So, you know, for an open data means that again, if you use the analogy of, you know, baking a cake, you need to have the ingredients, not just the recipe or the oven, but you also need to have the ingredients themselves to build it. Now, in terms of, as you mentioned before, quite a few companies, quite often they may say we have an open source model, we can open source the model, but usually they only mean the weights, not the data, not even the code.

11:48There is very few truly open source efforts right now, and it's mostly by universities or non-profits, like for example, AI2. Everything else, even though they may say they're open source, they're not really open source, which means it's very hard for anybody else to reproduce and build upon it. Okay, and somebody like DeepSeq who claimed when they came out that they're fully open source, including the data, they described the data. Is that helpful at all? So I'm looking forward to all the announcements that DeepSeq is going to be making this week. They said it's going to be announcing one more thing every day.

12:31But as it stands right now, the models they have are like R1 and others. They are open weights. They have released some code. Like yesterday, they released some inference code that helps do inference more efficiently. But still, nobody can fully produce their work. They had a report that detailed some of the steps. But if anything, that left a lot of questions across the industry about how exactly did they do it and what exactly data did they use and how exactly did they produce the data. So yeah, nobody can still reproduce, for example, DIPC-Card 1 because they're liking the actual data. And some of the details of the algorithms that it's not clear if everything is in the report.

13:12Yeah. But you guys are promoting full open data. And what does that mean? Exactly. So one good example is an agentic model that we released just about 10 days ago on a Friday that was called CALM, C-O-A-L-M. And actually it's an agentic conversational agentic model that even built GPT-4-O in some of the Berkeley benchmarks in the leaderboard. and to give an idea for this any of the data that we curated in collaboration actually that's say that it was UIUC Emre and Gökhan and the lecturer from UIUC that was leading this effort any of the data they generated and any of the training that we also helped them perform and scale up to 70b or even 405b which I'm not sure if any other university has trained the 405b model so far all of that is open which means right now somebody can go to the UMI repository and see where we have the the pointer to hiding face where we have uploaded the data sets and it takes right now two commands for somebody to produce they can do UMI install to install the solution and then UMI train with the name of the configuration file if they want the the smaller 8b models or the 7db or the 4 of 5b that's what takes two commands and as I tell people it shouldn't be any harder than that for anybody to reproduce somebody else's work end to end with the data, the code and everything.

14:37Of course, the model itself is based on LAMA, but at least all the post-training and everything else is fully open source. Yeah. And then presumably there are, even if it's fully open source, there are licenses attached. And is there a concern that if you fully open source in that regard, open data, that people will just take the model and not license it? Yeah, so there is that concern. But at the same time, this would be almost, you know, also if I'm saying the question correctly, you know, by design, the desired benefits of open source, which is you're saying, you know what, I'm going to help contribute to the common efforts as research community, as humanity, to make this technology better.

15:29And I'm going to tell somebody else all the critical ingredients so that they can also help me make it better. I think one good example here is I was seeing this just a couple, maybe weeks ago. It was the Human Genome Project. That it was, I don't remember how many, over 100 institutions, thousands of researchers that participated. it and somebody could have said like the same thing now with frontier foundation models hey we're going to give it a black box just so that we can monetize freely uh or we're going to make it open and share the knowledge with everybody and i mean i don't know where people stand in terms of politics you know with i think it was bill clinton but regardless of politics you know he made a statement that resonated with me that said that he's i think he said it would be virtual criminal if they were to withhold all these critical discoveries that are almost about life and death just so that as opposed to advancing humanity and helping everybody advance it instead they close in a box so that some companies can monetize out of it i think the same thing for foundation models the link may be less intuitive to um to life and death but actually it's there we're using foundation models not just for technology and the industry but for all the sciences from material science to climate science to healthcare i was actually maybe not to disclose specific names here uh in um in a in a discussion with uh quite a few healthcare leaders from some of the biggest institutions we have in the us and what he was telling that they use right now these models they adapt them to detect i think uh uh polyps i think colonoscopy if i'm correct or some of these procedures.

17:10And what they said for these procedures is that soon enough it's going to be malpractice if somebody doesn't use AI, not the other way around, because AI was like five times, if I can recall, better than humans, not 20 % or 30%, but five times or something better, which means, again, and then you get almost the question about it's almost like a matter of life and death. I mean, these are critical technologies. It's a disservice to humanity if they're not open and they're not advanced. And even if you keep those open, there's still enough opportunities for many enterprises, including ourselves, to monetize.

17:44So that's what I think things should be. And we're going to also be touching on the earlier debate. Yeah. Okay. And then more generally, how do you think this is going to develop? Do you think, and then we'll talk about UMI, but do you think that open data will increase? what will happen to the closed source operators? I mean, already DeepSeek has put pressure on them with its pricing. Yeah, yeah. I think eventually things should be more open because if anybody, even closed providers, are using data that they shouldn't be using, there should be a little bit more transparency over that or make sure that everybody's being a good player.

18:30So if anything, eventually things should be more open. But yeah, for pre-training, massive amounts of data, it's the whole web, sometimes even accounting for everything may be practically hard in many ways to be very thorough. I think for post-training, a lot of the research, for example, that we recently did or we're doing with university, it's a lot easier to do. And yeah, I think now your second part of your question was, okay, what will happen in the future, right, in terms of open versus close and the pressure that open models are putting? that was actually our strong conviction about a year ago or more than that when we were starting this effort so you know what it's clear that the future of enterprise ai and oh and beyond just enterprises is going to be open the same reasons why linux became better than unix and i could elaborate a little bit on the history and the parallels that are happening with ai and And once this starts happening, which, if anything, maybe with Dipsy could accelerate a little bit more, even perhaps a little bit more than what I was expecting.

19:37And if anything, then all the pressure is going to be a lot of pressure to the closed source companies. And they're going to start feeling a lot of pressure because they, I mean, arguably you won't be able to justify the greatly inflated valuations anymore. or for them to be able to continue raising at this high valuation so they can keep executing on the same, in my opinion, bad strategy because it will become clearer and clearer to more people and their investors that, hey, this just doesn't make sense. There is a better way to do this. If anything, history has taught us that there's a better way usually for these complex technologies to be developed and they may stop funding them or be less interested.

20:19Yeah. Okay, well, let's look at UMI. So tell us about the genesis and exactly what UMI's platform does and how it's different from last time we spoke from Hugging Face or Lightning Studio is one I'm familiar with, the other platforms out there. Yeah. So the reason why we started UMI Actually, sometimes people ask me, you know, what was the one thing that made you stardew me? I say, no, actually, it was not one thing. I not just want to do the right thing by humanity, but also I'm a pragmatic person. I want to build the conviction that something makes sense before I start doing it. But in terms of the right thing, that goes a little bit back to the debate, right, that we're discussing before.

21:09It was very clear in my mind that for this, generic technology should be like electricity, should be a common utility. It's going to be powering everything. for science's industry to advance, there needs to be an impeded access, and also to advance it faster, safer, and most efficiently. This can only be done if you have all hands on deck, all eyeballs on this technology, helping to make it safer and advancing. So philosophically, it was very clear that, you know what, this is the right thing for humanity. But then increasingly, I kept thinking, okay, but this may be the right technology, but will it happen in this way, or will unfortunately go in a different path just because, you know, it's not practical for open source AI to succeed.

21:52And then it was a sequence of things that made me realize, actually, you know what, actually this could work. And not only it could work, it's actually the most likely scenario, the most likely outcome. The first one was increasingly I was realizing how, especially for post-training, and there's so many tasks, so many capabilities that these foundation models have across so many modalities that we are already working on and many more that we haven't worked on yet that is not necessarily bottlenecked by the amount of compute you have but the amount of people creativity and people that you have to go and innovate in all these different directions and across all the different tasks and modalities to advance them which means that we an open community that is not just a thousand two thousand three thousand people but orders of magnitude larger can do a much better job if they have the right tools, which is what we're helping them at 2Me to give them.

22:47At the same time, besides the fact that, again, an open community could do better than a closed community, I was increasingly realizing that the current status quo is only serving a few companies. Nobody else wants it. So as I tell people, unless you are an aspiring AI oligarch, and the handful of ones that they are, they know who they are, nobody else wants this because imagine for example the outcome where uh let's say you know a specific closers model like gemini or open ai's models dominate everybody's going to go use those specific models on the very specific cloud provider on the preferred accelerators that this cloud provider has which means other cloud providers they won't have a competitive model to serve other accelerator providers including nvidia amd all the major ones they may be out of the loop because they know that the closed model providers that may have their own preferred accelerators where because the model is closed they can more freely optimize across software and hardware which means they can offer a better experience end-to-end so anyway and besides the cloud providers and accelerator providers then i started realizing that actually it's not just them even consumer companies don't want uh ai close the ai close source ai to dominate mark zuckerberg has said it publicly recently quite a few times that even for a company of the size of meta it would be problematic if ai one day is controlled by a single company let's say open ai or google and then they are at the mercy of getting permission from them to use the ai they need for their consumer products and the ai is going to be everywhere right so this is extremely critical so given all that i was realizing that there's so many tailwinds again most of the players in the ecosystem they don't want to succeed and if you help channel the efforts across the community and across them in a more meaningful way at the end what is the most likely outcome that open source would succeed or close source that is developed in a silo by its company individually to shoulder all the economic and human costs to make it happen so for me was becoming clear again and as history has told us with Linux versus Unix, that the open source will be actually the most viable path.

24:57And I think this becomes increasingly to more people, and that's why we started UMI about a year ago. It was a combination of, you know what, this is the right thing for enterprise to succeed. This is the right thing for humanity, and at the same time, it's possible. So it just makes sense to go build it. Yeah. Well, tell us how UMI works. I mean, how does somebody use the platform? Create an oasis with Thuma, a modern design company that specializes in furniture and home goods. By stripping away everything but the essential, Thuma makes elevated beds with premium materials and intentional details.

25:37I'm in the process of reorganizing my house, and I'm giving Thuma a serious look for help in renovating and redesigning. Thuma combines the perfect balance of form, craftsmanship, and functionality. With over 17 ,000 five-star reviews, the Thuma Bed Collection is proof that simplicity is the truest form of sophistication. Using the technique of Japanese joinery, pieces are crafted from solid wood and precision cut for a silent, stable foundation. With clean lines, subtle curves, and minimalist style, the Thuma bed collection is available in four signature finishes to match any design aesthetic.

26:31Headboard upgrades are available for customization as desired. To get$100 toward your first bed purchase, go to Thuma.co.i on AI. Ion AI all run together, E-Y-E-O-N-A-I. So for$100 off your first purchase, go to Thuma.co.i on AI. That's T-H-U-M-A dot C-O slash IonAI to receive$100 off your first bed purchase. Yeah. So UMI, as we launched it about three and a half weeks ago, so it's very, very recent, it's a fully open source platform on GitHub. It provides all the capabilities that somebody needs in enterprise or academia to advanced foundation models, anything from pre-training, all the different post-training techniques, full fine-tuning, parameters and fine-tuning, RL techniques, evaluation with all the most common benchmarks and a growing list of them, data curation, auto-evaluation with LLM judges, and many more.

27:46And the idea is that all the tools that somebody needs to do research in academia, let's say you want to develop the next Dipsick RL, the next Dipsick R1, you should be able to have all the steps in a single platform so that your whole recipe can be recorded fully. It's not you combine this with that, and there's dark knowledge and scripts, and so anything should be recorded in the same platform. First of all, for convenience, you have all the tools you need. The whole recipe can be fully recorded, which means it's easy for somebody else to go to produce it, cite your paper, and continue advancing it.

28:17It shouldn't be any harder. So that was part of the goal. Now, besides providing all these capabilities, we support both text and multimodal models, over 100 open models, pretty much all the ones that matter right now that are relevant for enterprises or researchers. They are supported. From hundreds of million parameters to all the way, a 405b LAMA that somebody can train. And also as a solution, we wanted to make it fully accessible, which means that, and seamlessly accessible, which means we want to make sure that somebody, whether they're getting their computer, Let's say they have just a MacBook with a couple GPUs.

28:56Or they're getting their compute from some cloud provider, WA, GCP, Azure, Lambda, Rampo, Together, or anywhere else. Or they have a big HPC that they can still run the same experiments, of course, assuming they have the necessary compute. But all it changes is that some simple deployment configuration says, how many GPUs do you have and how many do you want to use? but it is that the same recipe should be reproducible easy by somebody regardless of where they get the compute. So that's kind of what the platform provides. Now again, all the capabilities that somebody needs to build foundation models or adapt them to specific tasks and domain, whether it's pre-training or post-training, even evaluating data curation, all text and multimodal models and also running across any platform and across any scale.

29:46which as I mentioned before I'm not sure how many universities have been able to produce models at this scale and that's why those researchers from EIUC, we had a long relation with them, they said you know what, we really need your help to scale them up because we have been using other technologies but it's just impossible for us to do this and yeah, the good thing was very easy and it's also very easy for anybody else now to produce so this is the fully open source solution, we're starting to develop also enterprise solutions on top of that happy to discuss about it as well Yeah, on building models, how many, you know, on Hugging Face, there was, I don't know how many models are now, transformer-based models.

30:30how many people are really building models from scratch i mean they they write their first line of code and then you know code for however long it takes and then train that on on data that they've collected versus how many people are taking LAMA or some other open source and tweaking it or forking it or whatever they're doing, fine-tuning. Yeah, so the vast majority of the people are getting an existing open model because to get to that point takes significant cost. It doesn't make sense for them to produce it. They get an existing open model and they continue post-training it with different techniques to continue making it better for specific research or if they're enterprise for a specific problem they're trying to solve.

31:29Very few are working on pre-training or training from scratch, actually to be more precise, to training from scratch those foundational models. But the good news is that there is huge innovation that's still required, not just on the pre-training, but across all the other stages. And as I was mentioning before, an open community, especially when it comes to post-training or finding creative ways to get existing open models after the summary like meta has incurred all these massive costs to train them and make them better across all these different domains and modalities and capabilities and with creative approaches like deep six grpo and any new ones that they're gonna invent that's something that you don't need massive compute and there's huge opportunities us as a research community as humanity to continue to advance them sure it would be great if it was easy for more people to do pre-training but even on the latter stage where it's more of a how many people how many creative minds do you have to advance this and less about the compute constraints uh so there's huge opportunities again to advance even uh on getting an existing model and curating or finding creative ways to make it better uh and actually just something to add to my previous uh the previous question i forgot to mention is that besides providing the platform itself we already have also open collaboration with many universities and we plan to launch even broader collaborations where we say you know what here's the problem that we as a community make sense to solve let's all open and collaborate and make it better and uh does the platform require uh open code open um data as well as open weights yes exactly so for all any of these collaborations that we currently have to advance this foundation models we say you know what everything needs to be open and if it's built on the only platform for one it's very easy for you to do your experiments end to end and then the record the full recipe so that anybody else can easily reuse your data your code everything and reproduce this work and then make it better yeah yeah it's it's it's um it's an absolute necessary requirement.

33:43If you want open source to thrive, that's how we need to function. Anything less will just impede the success in open source. Yeah. And you think that people will start building truly open source models if they have this platform or, yeah. Yeah. i think for pre-training it may become my hope is eventually we'll be able to do more and more open source even though now the truly open source on the pre-training states are very very limited my hope is that in the future we may do more and more there as well but for post-training and or at least using existing pre-trained models and different continue to pre-train them or post train them in different ways that is already happening there's huge appetite from researchers in academia to do this and by the way to your earlier question that's how even we started with umi this started with a sunday discussion with uh ruslan salagudino from cmu where he told me you know what uh we are i mean he didn't say that but they are the top researchers among the top researchers in academia and and then he said that uh that's the top research it's it's my assessment um they said you know even even for us it's very hard to do this type of research it takes a lot effort for my students to put together all the pieces that they need to do this type of research and then none of them has scaled and done multi-node distributed training to scale up and train bigger models but it's too hard for any student to to do this so if you build a platform we'll support you and then he introduced me also to amit al-warkar who is also from cmu and actually in his words he said uh we are an untapped resource we want to contribute but those labs don't make it easy for us or possible to be precise.

35:36And also, you know, if you had the platform we described, that would be really helpful. So yeah, please build it. We're going to help you. It was the same response for many other academics. So that's why we started to build this. So it's easy for people to do the research, build upon it, and then give the, as we say, open source, the kind of Linux moment that it was missing. Yeah. One question I've had for a long time. If you take an open model, well, let's take LLAMA as an example. It's been pre-trained, right? So are there models out there, the code out there, the source code is out there that you can pre-train from scratch?

36:20Yes. There is a few, very few examples. One of the most notable is the efforts from AI2, where, for example, for the old paper, they said, okay, here's the data that we used, and here's the code, and then the resulting model. So we need more of that. And my take with this is, you know, but such models should not be developed by almost a single organization or small collaboration should be a much more open collaboration across the community. And that's what we're trying to help do. So it's amazing to have such efforts from Mera or AI2, but there was a missing tool, a missing capability, the Linux of AI, as we call it, that was missing for the community.

37:04And that's what we hope to help provide. And as a platform, it's something that can work with the models developed by AI2, by Mera, or any new models that the community develops. Any of the best tools. So that's why we're very grateful and excited for all these efforts that all these other organizations are doing, including Mera, even though it's not fully open source, open weights, it's still a huge asset for the community. Yeah. So if someone, because I talk to people all the time who say they're building on an open source model and they're fine-tuning it for one particular use case or another, but if you're doing that, if there are biases or frankly bad data in the original training data, it's baked into the model, right?