Your Edge Isn't the Model Anymore, with Dr. Anthony Annunziata

Speaker A: Hey everyone, my name is Elio Narciso. Um, I'm the co founder and CEO at ScaleStack. Welcome to the Revenue Engine Masters podcast. As you know, we typically dig into AI, go to market and all the ways in which businesses are changing, work is changing. And today's guest has really interesting background to me, like, because he started as scientist but now is like deeply involved uh, in AI at larger scale with IBM and Dr. Annunziata, he's of Italian descent. We were just chatting earlier. Uh, I'm Italian. As you know, he is currently the director of AI, uh, open innovation at IBM and also a co founder, which is interesting to me, like being founders within larger companies. So maybe we'll talk a little bit about that of the AI alliance, which is something that IBM and Meta started together and now has lots of members across the globe. But you also like a physicist with like lots of patents, which is also very interesting. It's been, I have to say, like a life goal of mine eventually to own like in my name, like a patent. So maybe I'll ask you about that too. Anthony, welcome to the podcast. Maybe like you want to tell us briefly about you. And so maybe the first question is really like, why are we here? Like how did you get here? You know, like in, like in broad strikes basically.

Speaker B: Sure. Well, uh, Elio, thanks so much for having me on. I'm excited to be here. Happy to tell you a little bit about how I got here. Uh, yeah, it's a many part path, brief. As you mentioned. I started as a physicist, I was trained as a physicist, went to grad school. I started at IBM doing research into magnetic devices. So trying to build better memory, better transistors, but based out of magnetic materials and devices.

Speaker A: Yeah.

Speaker B: And uh, you know, I did that for a while, it was great. But I was really hungry to keep going more in the direction of applying and bringing things out to the broader world to make sure that they have broader impact. And I found that very satisfying. So what, uh, what I, what I was able to do in a few steps is you know, always like important to convince management that you can do something. But once, uh, once you can do that, you know, you can have opportunities and I had several opportunities inside IBM to build a new quantum computing business, to build a new AI for science platform which we launched during the pandemic where science was on everyone's mind.

Speaker A: Yeah.

Speaker B: More recently, uh, you know, in the generative AI era. Ah, really kind of take lead of IBM's open source AI strategy and launch, uh, What I think is a really interesting organization called the AI Alliance. And um, tell you more about that in just a moment alongside another part of my mission which is very important, which is to work, uh, across our startup partners and to try to help them use IBM technologies better and to try to help, uh, try to help bring, um, bring their technologies with us to market in more places and more scale.

Speaker A: Very good, very interesting. So let's unpack that. Uh, maybe the first thing I'm curious about is that like, as a physicist and a, uh, researcher, like, you know, I think the mindset and mentality is very different than like, as you said, like, applied science. No. And so maybe the first curiosity I have is like, what parts of that like, researcher mentality is helpful today in your applied science work? And which part is not? Maybe.

Speaker B: Yeah. Uh, so first I like to remind people I'm not just a scientist, I'm a physicist.

Speaker A: Yes.

Speaker B: Physicists think about things in actually really basic terms. Right. I know it maybe sounds weird to say, oh, physicists only deal with basic things, but it's true, fundamental, basic building blocks of nature. Right? So when you're a physicist, you're taught to think very deeply and very foundationally, like, what are the really essential elements for how something works, what really matters, what doesn't? What's the simplest model you can create to explain something? Right. There's uh, something could be, you know, an atom or a transistor, or it could be something more complicated, could be a product or a business. Now the same exact, you know, approach doesn't, doesn't fully apply. But this idea of thinking and first principles and foundational ways I think has been extremely helpful in everything I've done. Trying to just drill through, like, what really matters? How do we structure thinking? How do we structure? Ah, a go to market, right? What matters when you're building a go to market. What are the fundamental things? So I'd say that that's, that type of scientific thinking, especially the physics thinking, is super helpful with everything else, first principles.

Speaker A: It's something that we use a lot in the tech industry. So what's.

Speaker B: Absolutely, absolutely. And there's a lot of physicists in the tech industry, as you know, especially in AI, and I think there's a reason for that. Now on the flip side, not everything is, uh, is transferable. One of the things I'd say that is a challenge for researchers and me included, or was, I think I've mostly gotten through it, but you can be the judge is in research you're always pushing the envelope and trying to come up with a new discovery, a new insight, and publish it and get the word out. But then you're kind of like, that's victory, right? Once you've discovered something and you've communicated it to the scientific community, then you keep going, you discover something new, right? So there's this. You're always pursuing new information, new understanding, but you're not like, building and shipping a product or completing a, you know, a deal. So there's a little bit of, uh, going further than science means you have to kind of extend your definition of done. Right. So, for example, when you're engineering and shipping a product, just demoing it is not nearly enough. Right. And you see that challenge in agents today, right? Demoing an agent is extremely easy. Shipping it to production in a reliable, trusted way, evaluated the way you. In a way that actually matters is actually a lot harder. So I'd say that challenge, going from research to more of a business mindset, is like, to extend and kind of bolster that definition of done to make sure it matches what the world expects.

Speaker A: So while you were talking, I was thinking that, um, I mean, in my home country, Italy, you know, we have, like, too much of the science and too little of the application. And so I was trying to ask myself, like, what is that? Like, you know, is different. In other words, like, is it. You become successful at what you do now because you're more entrepreneurial, you think, and that you're not satisfied with just like, the theoretical, uh, layer, but then just to see it in action in real life. And so keep pushing the envelope and keep pushing the bar so that it sort of satisfies you, uh, because you're creating something that did not exist.

Speaker B: Yeah, I. I agree. I think that's a big part of it, you know, trying to. Yeah, I mean, for me, the personal satisfaction with going further. And I think also just looking at whether it's products or businesses or people that have been successful in that way, it was inspiring. And so to me, um, yeah, it just. It just made sense as something I wanted to push farther along. Yeah, I don't. I don't know, the. The bigger question of cultural differences, organizational. That. That's a big topic. An interesting one, though. I, I think for me personally, though, it was just kind of something that just made a lot of sense.

Speaker A: Yeah, no, because if you think, I mean, like, uh, so Renaissance in Italy, right? So there was a lot of science, a lot of theory, but also a lot of, like, you know, applications. I mean, buildings, churches, like, you know, if you go to Florence, like, you know, the chapel there, like, you know, in the Duomo, is like an incredible technology that was theory only like, a few years before they decided to go and build it. And so to me, it's like we sort of lost it in Italy, by the way, over the last, like, century or, you know, maybe, I don't know, maybe the last 50 years. But I do think it's this hunger for, like, just not to think something. But then I actually want to touch it and see if it actually works, even though the theory tells me that it does. Because then once you start doing something, there is like a million problems.

Speaker B: For sure. Yeah, absolutely. In fact, like, um, you know, you don't really know if it works. It. Whatever it is.

Speaker A: If you actually do it. Yes.

Speaker B: Build it, do it, ship it. And, you know, and actually, I think the other, you know, really big realization, I mean, you as an entrepreneur, I mean, I think probably get this from very long ago, but, uh, it doesn't really work until someone else thinks it works and uses it and actually pays for it or, you know, validates it in some strong way independent of you.

Speaker A: Yes. Uh, I've also experienced that in Europe. So I was talking to someone. There was this huge, huge program that the European community put together 10 years ago. They had billions of euros because they said, we're gonna, like, start innovation, like, and push Europe, like, to become, like, more, you know, like the us Billions of euros available. And they started funding, and we got some money from my Pario company. We got some money and, you know, but then essentially the people that we're funding were mostly research and scientists, which really looked and remained like, crazy. What I mean by that is that what you just said, that, uh, like, someone else needs to follow you. Someone else needs to believe exactly where you believe. Otherwise you stay as like, you know, the researcher level that, like.

Speaker B: Yeah.

Speaker A: I mean, in his or her mind, like, this is true, but it's not true for most people. And there is no new future that has been created. I don't know, maybe I'm poo pooing researchers, but, like, this is not what I think. I think, like, you do need some validation.

Speaker B: I'd say that, um, you know, the, the currency or transaction or validation for researchers is just kind of different. Right. There's a whole economy around research. And I don't mean like a money economy, just like publish, citation, the whole citation economy, citations as a measure of worth. Right. Like, there's a whole kind of economy around it, but it's not the same as uh, the broader actual economy.

Speaker A: You played that game too, right? To have like a hundred.

Speaker B: Yeah, for sure.

Speaker A: Are you still playing it or. No, not anymore.

Speaker B: Uh, not in a while. You know, I did, I was part of a, uh, let's see, when was the last paper I was a co author on? Uh, about a year and a half ago. So not like distant past, but it's been a little while since I've been in the thick of research or filing.

Speaker A: And is the patents, uh, like similar to that game, to that type of economy, you think?

Speaker B: Somewhat. Patents are funny. Right. And by the way, the field that you're working in plays a big role in whether you can get a lot of patents or not. Right. So I happen to be in fields that were met, amendable to like very novel structures and processes that we could patent. If you're a software engineer or an AI researcher, it's actually harder to patent kind of AI methods and concepts. Those would be. So not to go too deep, but like short primer on patents, like. And you can imagine because a lot of the, the, you know, the concept of patents and the rules and the patent agency in the US emerged from like a time period where things were very physical. Right. So physical structures are much more straightforward to patent. So if you have like a physical thing. Yeah, but software or data structure or data, it's. That's a lot harder. That tends to fall under like a method patent. But methods are a little murky. They're hard to kind of patent. Right. So that's why it's one reason I think you don't see as many patents in software, uh, and AI, just because, um, it's kind of less of a fit to like traditionally how we understand and how the law of use patents.

Speaker A: Right, yeah, very interesting. So I think that this is relevant to like, uh, to what we're discussing here because, you know, like, how do you generate innovation? Like it is like to, you know, try to convince someone to follow you. No. And then what I'm curious about is that you've spent most of, you know, the last 15 or so years at IBM. And so then there is the other angle, which is innovation in a large company. And I've had some taste of that because I told you, like I spent some time at Amazon. And so maybe another question is, okay, you need to have a certain entrepreneurial mindset to transfer like science and research into like applied stuff. Then like within a larger company, I mean people tend to think that larger companies don't innovate as much as smaller companies. And so what have you found to be true in order to succeed as an entrepreneur within a larger company?

Speaker B: Yeah. Let me start by saying there is a small part of my resume. It doesn't always pop up. I actually did co. Found a company in grad school with a couple of other friends.

Speaker A: Okay.

Speaker B: We made it not too far. We actually were in an early class of Y Combinator back in 2000. Uh, 2008.

Speaker A: Nice. Okay.

Speaker B: And it was a, um, it was a wear, it was a, it was a wearable sleep analytics product. Um, but didn't get that. But uh, kind of long, long fun story for why I didn't get that far. You know, the net was like, I've always had that kind of itch. I went to IBM as a, I kind of every, every few years was like, you know, I gotta, I gotta start something new. Is it still gonna be at IBM? I, I'd ask that all the time. And uh, IBM keeps giving me really interesting opportunities to start new things. And if you could do that right inside a big company, there are some advantages. Like you can leverage the scale, you can leverage the connections, the go to market apparatus like is very well formed. So if you can manage to innovate in a big company and take it forward, there are some advantages a big company has. There are disadvantages too. Right. Um, it's complex, lots of decision makers. Although in aggregate there's lots of resources. Getting resources to do something new is hard. It's very hard actually. You need to really make a convincing case and have people trust you and prove yourself. But if you can do that, I would say that fundamental innovation does still happen in a lot of big companies. IBM has a deep tradition in research and in our development, um, of coming up with a lot of new innovative things. I think the challenge is not kind of the kernel of innovation. The challenge is how can you scale it up inside of a big org where there's a lot of complexity.

Speaker A: Yeah. And so now you are like in charge of open source, open source AI at IBM, which is also like, well, maybe actually you can explain what your job is more precisely. But it also is not just creating innovation within IBM, but then convincing other companies. Right. To adopt open source AI. Um, and so first, maybe if you can explain what is the job actually today, how big is your team and what is the charter essentially? And then maybe what it takes to convince others about what you guys are trying to do.

Speaker B: Sure. Yeah. So today I actually have two roles. One is the open source, uh, Ecosystem and strategy development role. The other is actually working with our partners to uh, help support and build out, go to market. Uh, that's joint with IBM. Okay. Mostly on ISV partners. On the open source side, what I do, you know, fundamentally IBM is actually a uh, business that's been built on open source for a long time. And even if you look at our recent large acquisitions, Red Hat, Confluent Hashi. Open uh, source is a big component of those businesses. Right? Yeah. And it's not because it's just a good thing or charitable and all that. That's true. It's a good, that's nice. But it's actually, it's a very good way to base a business.

Speaker A: Better way to build.

Speaker B: Yeah. Better way to build. Why? Why is it a better way to build? Uh, well, you're pulling together innovation from the open community.

Speaker A: Many people, yes.

Speaker B: Much easier ways to access, to get started, to contribute. So you have innovation, you have validation, you have transparency. Right. As you know. Well, lots of companies still value that. Now in the AI era there's a whole like hierarchy of priorities and sometimes that falls by the wayside because we just need to keep moving and innovating. And I'll get to that in a moment. But I would say that many organizations still in the AI era care a lot about whether fundamentally the core technology is open source, whether it's transparent, whether it's accessible, auditable, uh, I mean all of that. Right. Part of my job is to kind of work across IBM and develop, you know, kind of the next generation of strategies, specifically on AI models, eval, uh, data structures, tools, all of that. And then a big part of it is to work out in the community and engage partners and contribute to open source projects and try to support them that we uh, in particular ones that we think are gonna be important for the future. Right. So you know, new methods to evaluate AI, new methods to tune models, you know, that, that kind of thing.

Speaker A: And I mean AI itself, I mean we just had like uh, Elon Musk bring into court some admin because I mean OpenAI, just like the name was born as an OpenAI.

Speaker B: That's right. It was a, yeah. Nonprofit organization, open source focused. Right, right.

Speaker A: And so, but then it's not now.

Speaker B: Right.

Speaker A: So it's a closed model. And so like what do you think? Uh, you know, how is it going to play out in Europe? So like they will coexist like closed models and open source? Or you think that like open source models will eventually win?

Speaker B: So I think the regime we're in now is everybody like, okay, let's start from the user side and then we could get to the developer, the tech company side. Look, I think everything's moving so quickly that people just are going to grab whatever is best, slash most convenient and run with it. And you know, well managed platform services tied to frontier models that are proprietary just, it works to do that for a while.

Speaker A: Yeah.

Speaker B: Now I think you already see that, you know, fundamental progress in models has slowed down. At the same time, you know, the cost of progress on the margin is high, it's getting higher, uh, for lots of reasons. Size, scale, amount of data needed, etc. Um, you know, and you already see kind of the effects of that. Right. OpenAI, anthropic, Google, others aggressively moved up stack. They have more comprehensive offerings. Right. I mean I think if you look at quad code, quad desktop. Right. Cowork versus just the RAW model, like arguably the key reason Anthropic has been so successful is because they've built this really good product experience right around the model which has a lot of sophistication. So I think you already see the market, uh, moving past the model itself as the key differentiation point.

Speaker A: But it doesn't matter that much.

Speaker B: Yes, it matters.

Speaker A: I mean, it matters. I mean models are like.

Speaker B: Yeah, so what else I'll extrapolate and say what that means is other factors, like once, now that, you know, it's, it's becoming the case that you can get really good models in lots of places, both open source or proprietary. Other factors, traditional factors cost, the ability to deploy on your own, uh, your own infrastructure, the ability to tune in your own way and so on. They're going to factor more and I do think that will lead to open models being more and more popular. And I think you already see signals of that.

Speaker A: Yeah, yeah, yeah, that's exactly how we see. And that's what I meant. Like, you know, it doesn't matter. Meaning that like it becomes commoditized. I mean in many ways they are like phenomenal models. But like, you know, it's just like.

Speaker B: Yeah. And there's a whole next generation of model architecture that is coming on the scene. World models based on for example Yann LeCun's Jepa architecture. Architecture, that's cutting edge stuff that is going to be differentiated for a while. The model there is going to I think rapidly progress in capability, but that's a very different architectural approach. So LLMs, transformer based, I think I won't say they're commodity yet, but you know they're on the way. And so that.

Speaker A: What about developing your own model? Like so small language models like you know, do you think that companies should start thinking about that?

Speaker B: I think there's steps before getting all the way to that. You know, I think taking you know, the right open weight model and you know, fine tuning it or if you're proficient with rl, you know, having an RL pipeline that you know, you can run with your, your environment in mind, that's probably the better way to, to kind of create a custom model for now.

Speaker A: Yeah, very good. And then let's go back to the question about like enterprise adoption. No, so like you know, you started stuff inside of IBM like you know, we're talking about like uh, AI, but so how do you convince maybe other companies that are not, you know, building just like ABM or Meta or OpenAI, like the AI itself that this is the new world? What you've seen like changing the mentality on uh, the culture of people in adoption that switches from oh, uh, we're checking and piloting to actual adoption. What's been things that you've seen across customers that need to happen.

Speaker B: So I think the first is kind of a team or leadership or cultural kind of aspect, which is people that um, I think there needs to be a willingness to embrace the full challenge. Right. Which is to acknowledge that um, enterprise application of AI is going to need more than just prompting, you know, a cloud hosted model. Right. An eval strategy that's not just an academic benchmark will need to be developed like real integration with the workflow and thinking how that's going to work needs to happen. So I think it's first just have a, you know, a good understanding that you're, you're jumping into something that has, it's going to need a fair bit, you know, fair bit of focus and some work. Right. Not just kind of a demo and okay, now it works because it doesn't many times I think, um, technically I would say that a well thought out eval strategy, like definition of success. And what I mean is not just kind of the academic benchmarks. Okay, do some QA question answer pairs and oh, it looks good, golden, you know, we're good, good to go. I mean like, you know, comprehensively, like first, like specifically to your use cases. How you're really going to measure, you know, in development and then in, in production, are you going to monitor, are you going to ensure that um, you know, in production when you get lots of, lots of prompts and lots of things you haven't maybe tested explicitly. How are you going to envelope and make sure that the system is working? Admittedly it's still not going to be fully deterministic. So I think like the right eval strategy grounded in the use cases, not just kind of one time at development that, that I, I see as a really key thing. Um, um. And then I'd say, you know, after that. Yeah, I, I think just a, ah, kind of smart understanding progression of where to start and how to scale up in steps. Like I think internal pilots and internal use cases have been a, a good pattern of success that I've seen. So start internally, start with your internal user base and you know, if you're a good sized company that has pretty good scale so you can learn a lot that way.

Speaker A: So you're saying back office. Start from the back office.

Speaker B: It doesn't have to be back office, but uh, yeah, just internal users first I've seen is uh, kind of a good step to prove out something beyond like a demo but not quite live external production.

Speaker A: Yeah. What about a topic that is dear to my heart which is like data preparedness? Have you seen that as like you know, also a uh, differentiating factors across like successful big time.

Speaker B: Yeah, I should have led with that actually. Keep going. There's a lot of. But yeah, we could, I could, we could talk all day about that. Uh, yeah. So fundamentally, if you go back to the argument that you know, the model itself is commoditizing, then what's the differentiator? Right. Okay. So for a company, a tech company, you're going to build a product experience, developer experience, SDK, there's a lot of stuff. But if you're, you know, a company that just wants to use AI, it's going to be your data, your knowledge. Right. So that actually means not only is data prep and you know, context preparation a knowledge platform. Right. The ability. Yeah, like not only is that important, it's almost like before long like the only differentiator. So totally important. I think that, yeah, I could go on. Uh, you know, m. Personally I think that depending on the domain, like well structured, semantically driven, ontologically driven layers are going to be increasingly important. Um, yeah, we could.

Speaker A: Yeah, no, that's a failure that we see. I think. Yeah, you're totally right. Like the culture, the commitment, like the connection. I mean you talked about integration with existing workflows. We 100% believe that back office are great ways to begin. I mean maybe internal processes in your languages, but because I think that there is so much that we still need to learn that, like, starting from like frontline stuff. It's very hard. But you know, there is so much inefficiencies in everything that we do. Particularly like, you know, and what is the first thing that AI has created? Like tremendous efficiency coding. Right. And so no developer right now is writing code. No good developer right now is writing code because, I mean, it was so efficient to begin with. And now applying AI is like, you know, made the whole thing possible at, uh, completely different speed and efficiency. Sales and marketing, which is sort of like the space where Skill stack operates. It's still very inefficient. And so starting from internal processes and back office, it is what we also believe is a great opportunity. And then, yeah, data preparedness, because we see it everywhere. They're like, you know, you launch a pilot, oh, it's failed. Like, no, it's the model actually. The model is like irrelevant. That's what I mean. We can only do these things because model are so sophisticated. But once you have them, you don't have good data.

Speaker B: That's right. That's right. And you can, uh, even through a pilot stage, you can mislead yourself to think you've got something that really is good in scales. But mostly you're relying on the model and then that'll break down, you know, as you try to scale it up into an out. Um, yeah, I'd say like also when, yeah, when you look, when you look at. Yeah, if you look at data. Right. How. How do people use data? Okay, so. So for a couple years we've been talking about, you know, rag retrieval, semantic search and retrieval that kind of. And that's fine and still works okay for basics, but doesn't, you know, you know, it doesn't always work and certainly doesn't always work when you have, you know, large data, when you have high accuracy requirements and so on. So I do think that, you know, ways to better structure, you know, context data, uh, to better update. Right. Context structure is going to, it's going to be even just more and more important.

Speaker A: And then, I mean, there is a whole topic within this which is like, you know, companies, I think are like moving fast, trying to adopt AI, uh, in pilots in larger deployments. I think that a lot of the security concerns that we had until like a moment ago are now gone, but will reemerge very soon because your employees are, um, sending potentially private, very sensitive data to these models. And so we are starting to see, particularly in Europe, which is always very concerned about privacy. I mean, Think about gdpr. We're starting to see some European customers that are telling us like, you know, we'd like to do this on um, prem. And so do you see that? Is this like something that IBM is thinking about or you personally?

Speaker B: It's a big part of our focus in business actually.

Speaker A: Even in the U.S. even in the U.S. okay.

Speaker B: In the U.S. yeah. I mean certainly, you know, Europe and some other places internationally are like even more focused on it. But even in the US regulated industries, I mean, certainly government. But I think there's a growing realization that in the era of AI it may be even more important because now it's not just a vendor lock in and kind of, um, you know, kind of kind of cost story now. It's like there's a true like privacy beyond just personal privacy, like privacy in this sense that like, you know, it's so easy for AI systems to learn from the prompts, from the data you send them, from the patterns of workflow. Right. They're very smart. So the big model companies, they are learning everything possible. So if we're not careful, some um, companies know this. You're going to transfer your proprietary differentiation to a tech company very quickly. So I think that one is.

Speaker A: And so that means also open source or small. I mean like it's all interconnected. No, because if you believe that you have also to think, you know. Well, okay, if I want to do on prem, I cannot do OpenAI or anthropology. Um, I need to. Right, Sorry, say it again.

Speaker B: I said or you'll be sparse and careful about what you use large scale cloud hosted models to do. Right? Yeah, I mean you may still want to.

Speaker A: Yeah.

Speaker B: Use them for some things.

Speaker A: So distribute like not just one model, distribute like the workflows and the use cases across models. Is that what you.

Speaker B: Yeah, I mean that's one tack you could take. Uh, I think the simplest is though, you know, determine a set of use cases that you could reasonably support with, you know, kind of a smaller mid sized model and then deploy locally. Right. In a stack that, that, you know, that can support it, you know, inference and you know, deploying agents and all of that. And you know, I think it's, it's the case with most enterprise use cases that a small model in a reasonably scalable stack is good. It'll actually serve the use case need because you know, if you look at uh, you know, one of the large proprietary models, you know, they're trying to be general, everything.

Speaker A: Right.

Speaker B: You can ask them what to cook for dinner, you can ask them to analyze your data. You can, you know most enterprise use cases are more focused than that. Right. And so if you could be more focused then you can do something that's lighter. So a smaller model, the con, you know, the right engineered context, the right eval strategy and you can serve that locally on reasonable hardware. So yeah, we can, we can do that. Today IBM has a, you know, kind of growing business and supporting those sorts of use cases across, you know, IBM Red Hat, you know, new acquisition like we're very excited about. Confluent was just joined to add uh, you know, lots of streaming data which you can do a lot with in the AI context now.

Speaker A: Yeah, yeah, super interesting. And then like uh, yeah, the differentiation of the use case. I think it's also very important and like to help companies like okay, there are certain things that maybe is not like as critical, you know to you know, be protective about your data because maybe is mostly data that is outside of the company. It's often like our use cases. Right. So like you know, okay, I want to prioritize sales and marketing efforts or even like HR workflows when I do candidate recruiting mapping. And so uh, that's mostly data that it's outside of the company for that,

Speaker B: you know, uh, Internet based market research, you know, public domain things. There's a lot of kind of what I mentioned before. Yeah, sure. Use whatever model for that stuff you'd like. It's public domain mostly anyway.

Speaker A: But you know, if data.

Speaker B: Yeah, yeah. If it's something that is about your proprietary manufacturing process or, or hey something deep in the architecture of your back end for your systems like. Yeah, maybe use a local. A local model for that.

Speaker A: Yeah. And can you give an example for the audience about like what's a local model?

Speaker B: Sure, yeah.

Speaker A: So if everybody is familiar with like the entropic of this world.

Speaker B: Yeah, yeah, yeah. So I mean if you look out today you have Google's uh, Gemma models are very good. Mistral in Europe, you talked about Europe, they're making lots of progr with open weight models. Certainly still from Meta. IBM of course has our Granite series of models which are small models. The whole strategy there is not to try to compete at the super large scale generalist model. Yeah. Kind of small to medium scale models that are good fits for enterprise use cases and there's others as well.

Speaker A: And so for instance for the IBM models like what are they good at? Ah, like what type of use cases that you found like a lot of success within companies the are concerned, let's say about like Having their data protected.

Speaker B: Yeah, yeah. So, um, so Granite has a family of small models that are text and text and image. Right. So multimodal. And if you think about a lot of what you know, there's a whole lot of enterprise use cases that are somehow ingest documents, you know, and then do stuff with that knowledge. Right. So Q and a, generate, etc. So you know, one of the things Granite has focused on is the ability through you know, vision language models and then you know, kind of pure language processing to be able to ingest, structure, summarize, generate from, understand business documents.

Speaker A: Yeah. And so is that like the charter of your team? Is that like what you're helping customers with or you're helping like ISVs, like you know, be more fluent in.

Speaker B: Yeah, so I primarily I focus on two things. So first, work in the open source community with partners to drive fundamental advancement in new open source projects and AI. So that's the AI Alliance's mission. We have 205 uh, partner companies in 29 countries doing that. We collaborate on um, many, many projects and topics. There's a new one I'll tell you more about in just a moment called Tapestry, which is a very ambitious project led by Yann McCun and it is uh, we are going to build a globally sourced uh, foundation model. So we're going to bring data together from across many different organizations globally, do it in a, in a privacy and ownership preserving way and train a model that could be better than anyone else could train. Because this way of consortium training is what we're calling it could access much more information, much more data to train than any other single organization because we have this distributed ownership structure. The goal there is to build a foundation model that is beyond current frontier capability, but is an open based model and one that is held by a consortium, right. By the alliance as a neutral entity so that anybody can take and use it and customize it and build on it. Um, and therefore can you enable their AI model sovereignty that way?

Speaker A: What's a good metaphor of stuff that has happened in the past? Because I mean like in cloud that's not really what happened. Right. So in cloud you have like Google, Google aws, Microsoft. I mean I always try to look for other examples, metaphors, things that have happened in the past. I mean like for instance, I don't know, I started my career mobile technology. Really. And so like, you know, I remember when we were telling customers, oh, there is going to be more phones, which is really like, you know, connected devices than people and everybody was like, Ah, uh, you guys are crazy. And then like that's what we have now. Right. And so now we're saying like, oh, uh, there's going to be more agents and workers by like a large amount.

Speaker B: I'm sure. There's already way more agents than tech workers, you think? I think so. I mean I think every developer has multiple agents basically running at any given. I mean I have over there. Yeah.

Speaker A: And so, yeah, what is a good metaphor for something like this, like the open model to succeed?

Speaker B: I think there's a pretty good metaphor in Linux and I think there's a decent metaphor in Kubernetes. You mentioned the, you know, the cloud era. Uh, so let's maybe just spend a minute on those. So Linux, everybody knows Linux, right? Open source operating system, it runs most of the business and server world at this point wasn't always that way. Windows had a very strong position back in the day. IBM and others, IBM played a leading role here to uh, to promote Linux and to build that as a, as a true, robust open source operating system that businesses could rely on. And now, I mean it's, it's everywhere. Right? I mean all the clouds are proprietary. Clouds are running Linux basically.

Speaker A: Yeah, yeah, yeah.

Speaker B: So Linux as this fundamental open platform to build on is a decent analogy actually with. To where.

Speaker A: That's great. I love it.

Speaker B: Yes. Yeah.

Speaker A: Um, you, you also brought up almost this invisible layer. No, because now nobody thinks about that. But back in the day, I mean like where people were making these choices, it was very much talked about. But now it's like, you know, a given.

Speaker B: No, yeah, definitely. Yeah, yeah. It's not controversial at all. It's controversial not to use Linux.

Speaker A: Yeah, yeah, yeah. What about Kubernetes? Because I also lived a little bit of that and that was super interesting also as an example.

Speaker B: Yeah, so super interesting, right. You know, the big clouds, you know, on the growth path led by AWS back um, you know, 10 or so years ago and um. Is that going to be the only model? Well no, because you know the ecosystem came together by Google. Red Hat had a big role there. I mean IBM had a significant role there and others and got behind Kubernetes as a standard and you know, as a, as a platform.

Speaker A: Started at Google, right? I think Google was, I think so, yes.

Speaker B: Internal project at Google was open sourced very early on and a number of people that kind of knew that, you know, public cloud only deployment is not going to work for everybody jumped on that bandwagon. It's not just a very strong and popular platform. It's growing, and it's growing even more in the AI Era in some ways, because kind of like what we've been talking about. Right. I think people realize that, you know, managing your own AI is not just a cost and privacy concern. It's like a fundamental preservation of your business value concern.

Speaker A: Yeah. Anthony, this was super interesting. I recognize that I may not have followed strictly the guy that I said to, but we went into many interesting areas. And so I want to thank you for being here and sharing all these ideas and knowledge with me and with people. So thank you very much. And definitely, let's follow up.

Speaker B: Uh, that'd be great.

Speaker A: All right. Anthony, thank you so much.

Speaker B: Elio, good to meet you. Thank you very much for having me on.

Speaker A: Thank you.