This Man Visited the Best AI Labs in China. Here's What He Saw | Matt White

Speaker A: AI isn't always the best solution for everything, right? One of the things that I saw early when LLMs came out for fraud detection, everyone's like, I'm going to use an LLM. Well no you don't. You have perfectly good classical machine learning models that are able to identify potentially like fraudulent transactions, illegal devices on your network or these kinds of things with higher accuracy. How much money would you trust your agent to invest on your behalf right now? People are like, I don't want an agent to touch my money. We've seen instances of chatbots or customer service agents that provide refunds that they shouldn't have given back. And so a lot of these things have to be built in a deterministic way through code as opposed to relying on the model to follow your intent every time.

Speaker B: The share of Chinese downloads of Chinese open source models for the first time is higher than the Western ones. What does it tell you?

Speaker A: I think what we're seeing with the Chinese models is that there's.

Speaker B: Hi Matt, welcome to finoverse podcast.

Speaker A: Thank you, thank you for having me.

Speaker B: For those who don't know, uh, what's your role, uh, as the AI CTO at Linux Foundation? Can you walk us through what your responsibilities are, uh, what your day to day activities are? What are you working on?

Speaker A: Yeah, so a lot of my work is very community focused and so making sure that we have a healthy open source AI, uh, ecosystem that there's a lot of really great projects that need to grow and get eyes on them. Um, projects that we've brought into the foundations like vlm, Pytorch, um, deepspeed and others and Ray as well. And then some of my work is very focused on education so we're working on some training and certification especially around Pytorch at the moment and uh, also ensuring that we have a very strong technical direction for AI in open source. Um, so some of my prior work is on the model Openness framework which we wrote uh, a paper on a few years ago and then the Open MDW license which is um, something that is very important in the open model space. Um, and then just sort of getting out there speaking to the community. Uh, yeah, MCP is another one. So we recently stood up the Agentic AI foundation, uh, which is MCP, Goose and AgentsMD. And so we're building a very healthy ecosystem around Agentic AI now as well. And so my job is to be a good steward to the community to help grow this very vibrant open source AI ecosystem which sort of encompasses a lot of different things. Right. It can encompass anything from open data sets and open models to all of this great open source software like through the stack. And at the Linux foundation we have that represented through a, uh, many of different foundations. Right. And so like many workloads rely on Kubernetes and Linux and others that are represented through other foundations as well. And so part of my job is to ensure that we have the ability to democratize access to tools, to democratize access to education and help really just keep turning the wheel on the open source AI ecosystem.

Speaker B: So before joining Linux and you were also leading um, a tech side of the uh, Ptorch previously, but then before that you spent almost like 25 years in enterprise tech. Was it IBM?

Speaker A: Yeah, telco, yeah.

Speaker B: Uh, how does this experience now helps you to kind of look at sort of the new, kind of within the context of the new hype of AI?

Speaker A: Sure, sure, yeah. I've spent most of my career in industry. Um, I had a little sidebar for a while working at cable labs and doing research there and then went back into industry and then working for another nonprofit again. Um, I think it's helped me to understand both sides of this. The industry enterprise needs the consumption, adoption, applied AI. Uh, and then on the foundation side is growing the ecosystem, growing access to tools, helping to create a flourishing culture of open source AI. And so it sort of uniquely positions me to have both of those views. Um, the other thing I think that is pretty important from my perspective is to understand what AI researchers have to go through like, like what developers have to go through so that can solve their problems. Right, Help them solve problems.

Speaker B: We'll get back to um, your work at Linux, but I'm curious, uh, first of all, you just came back from your long trip in mainland China and you visited pretty much every single major AI lab, Deep Seq, uh, Moonshot, um,

Speaker A: Zi,

Speaker B: um and Quen and other teams. So what was um, your kind of the most surprising moment apart from things that everyone knows? Um, what surprised you the most?

Speaker A: Yeah, I think from the, I've come uh, to China a few times over the last couple of years and what is probably the most surprising and from like a uh, American perspective is that we don't really know what's happening in China. Like we see a lot of great work being published in open Science and Open Source and released, but we don't know generally where they're going within the research. And so something that was very eye opening for me was when I came to China and started to meet with these Labs is that they're very open about the work they're doing, they're willing to share all the great innovations they're working towards, um, not necessarily share their product roadmap, because everybody's got a commercial position, but very, um, open about the work they're doing. And I think one of the most enlightening parts of that journey for me was to see how passionate they are around the work they're doing and uh, how excited they are to be able to publish their work and get their model out there and see how it benchmarks against particularly a lot of the closed API models. And so, um, this is something that I think was very interesting for me was to see that they're not holding back and they're being very upfront about, uh, the work they're working on.

Speaker B: And from your observations through these meetings, how is it different from, say, U.S. counterparts?

Speaker A: Yeah, in the U.S. folks will talk a little bit off the record about what they're working on, but not necessarily, um, invite you to the lab and then have these kinds of, like, broad disclosures. Um, the U.S. there's a lot of movement between the U.S. labs, like researchers moving between those labs. And, uh, a lot of that is. There's not a lot of secrets, I mean, because there's a lot of movement, especially in the Bay Area, uh, but in China, the labs are more geographically dispersed. And so they're in Beijing, they're in Shanghai, they're in Shenzhen, um, they're also in Hangzhou. And so you don't have that same movement of people, uh, that we do in the US Labs in the Bay Area. And so, um, I think the slightly different sort of cultures, I would also say. But the. I think the culture lab culture in China, like, is fairly sticky. Like, you're, you work for like a, a company, like, let's say Moonshot, which has like a very, like, you know, rock and roll themed office. And, um, you know, young, young, young folks. And so, um. But in the US Maybe it's Moonshot

Speaker B: is Kimi, is it?

Speaker A: It is, yeah. Yeah, yeah. And so, yeah, it's. It's definitely like different and different culture. But there's also, I think, something that people don't really recognize is that a lot of these researchers were, many of them were Western educated, especially the founders. Um, you know, UC Berkeley, cmu, Stanford, and they, uh, were, you know, did their education in the US and then they came back to China. Um, something that's happening more now is that a lot of entrepreneurs, a lot of researchers are uh, coming back to China to work. Right. And then not staying in the US to work. And so I think where historically they were spending, you know, they wanted to work in the US and I think China's done a good job of you know, encouraging a culture of innovation and startups. And so um, and of course like, you know there's a Bay Area here too, right. That uh, you know, there's a lot of like, you know, folks that are for, in this space as well.

Speaker B: So, so there's actually very few people realize that. But if you look at the world map, there are two massive epicenters of AI within the Bay Area. And the next one is just um, few minutes away from here in Shenzhen. Yeah, um, but also Hangzhou and Beijing obviously. But um, in your sort of observation, conversation with a lot of Chinese researchers and um, how do you find that ah, um, deserve the what in the west called the ChatGPT moment and in China is called the Deep Seq moment. Right. Sputnik moment. Um, uh, what's the take on answer of deep sea? They still have like dominant position within China or you see it as more like um, evenly distributed.

Speaker A: Now I think Deepseek is viewed as the, as sort of the tip of the spear. The innovators, the doing more with less. Um, they have put out a lot of great innovations that have been adopted by a lot of the other labs, including in the U.S. um, and so I still think they're seen through that light. Uh, there's a deep respect uh, in the Chinese labs for what DeepSeq has done and what they're working on. And so um, I can really appreciate that coming from an open source background and um, looking at what they're putting out into the world. Their research papers are very detailed. You can reproduce the research they're doing. Um, and that's greatly important for the advancement of AI across the board. And so we want to make sure that we have the disclosures so that you can go and replicate the research. Um, we've seen sort of a trend that I'm not extremely enthusiastic about, which is that the research papers have shrunk and then become technical reports. And so there's not enough disclosure in a lot of these technical reports to be able to go and replicate the research. And um, that doesn't really advance the cause for AI or um, depending on your position, how you think we're going to get to AGI. Uh, it certainly doesn't help us move the dial if everybody's research is closed.

Speaker B: Just when you were in China I think MIT and Hugging Face issued a uh, new report where the share of Chinese uh, downloads of Chinese open source model is for the first time is higher than the western ones. What does it tell you? Are we in a situation where can we say that sort of developers globally, um, the mind share of the Chinese um, open models is prevailing in the western ones.

Speaker A: I think what we're seeing with the Chinese models is that there's better performing small language models and medium sized language models and that is driving more of the downloads and those models are highly performant. Um, the llama family is not currently there's no other additional llama models. And so uh, Qiwen or Kwen as people call it, um, is the most widely deployed model and uh, even in enterprises in the US and so I think what we're seeing is there's a huge number of hobbyists and researchers, developers that want to download these models and use them locally with Ollama. Uh, and then we have the enterprises that want to benefit from cost uh, optimization and so it's more appealing and more attractive to use some of these smaller models or medium sized models uh, as opposed to you know using the APIs and paying a higher per token price.

Speaker B: Yeah, I think that the latest numbers of um, deep seq version 4 which is like 50 times cheaper to certain models of the um, uh commercial APIs, similar capabilities kind um of triggered the price war across the mainland Chinese. Uh but interestingly um, you also met the Minimax founders and they just released the first time the closed model. Have they explained anyway what's the logic behind?

Speaker A: They didn't go into detail about it but I think this is going to be a trend in China. And so um, the perception was that anything above a trillion M parameters was so expensive to train that you would need to move it behind an API because it would be harder to deploy but also to realize commercial gains. Right. Um, and so where the US is being some open model, some APIs and now much higher on the API side, uh, I do think we're going to see that same model in China going forward. I think that the labs, some of them have done IPOs earlier this year and some are eyeing IPOs and the

Speaker B: pressure to kind of demonstrate that yes,

Speaker A: yes, they need to and revenue costs money to train these models. Right. And so they have to have a commercial, they have investors like they have to realize some sort of gains and so I think they will shift to APIs. I'm sure all of them at some point will have uh, commercial APIs and then help to subsidize some of the cost of the research and training.

Speaker B: So in my personal experience with Extra Minimax founder we're developing house of Solution AI for events and um, as a voice based um agent and we were struggling with finding a right Cantonese model for the Hong Kong audience. And then I messaged that I happened to know the founder and message can anyone maybe help with the Minimax? She immediately set up the group, brought engineers, we sorted it out and there was a contrast. Um, we used one company in the US um for the sort of stack on voice AI and then suddenly in the middle of the event they changed the API. It took us so long time to figure out where the problem was. It was like no heads up, nothing. So I messaged again to the founder in LinkedIn saying look, um, no response, nothing. The reason I'm mentioning that is I feel like in my kind of perception the new has so much demand um on all the surf AI stuff and um, on the other hand with Chinese um, sort of customer first approach and willingness to do as much as possible is minimum cost and then proactively help. I feel like maybe it's just the wrong example but that was just our goal. One of the experiences made me think um, okay, maybe I should reconsider. Right?

Speaker A: Yeah. I think people, some people and organizations are going to want to have more control over the model and the life cycle of it. They may not want the model to be upgraded in the background. Obviously if you have an API they can upgrade the model minor release, maybe break something um, or make changes the API. Whereas if you have more control over that and you're running your open models you have the ability to set your own terms and set your own upgrade path. And so um, I think that's enticing to a lot of folks. Right?

Speaker B: Yeah. I feel just overall kind of the attitude um towards any kind of the aggressiveness. Ah and in a way in a positive sense of like let's do everything for customer, we can do it just to gain market share because they are not um spoiled by overall surf over demand. Sure. Like you see that in, in the, in the US companies. But going back to your trip, you also mentioned um, in your um blog that you also had a lunch with some of the dipsic. Um uh one of the, the senior staff um who was this company from there back in the days when they were high flyer, um from from what your sort of conversations uh with with them. What is their approach or internally how they approach their work might be different or Interesting that we can learn from any insights you can share. Sure.

Speaker A: The uh, I think what struck me the most about deep seek and meeting folks there was the humility. Like they're very humble people. I don't think it's really you know, in some organizations it's about the, the individual and so being able to like, you know, show what you can do and be first author on a paper and that sort of thing. I think the, the impression from Deepseek is like, it's very like communal. Like it's. We're all working together, um, you know, towards solving real problems. Right. And so I, I think the you know, from my experience is that the humility in that, in that office and how people are really um, motivated to solve problems and um, not as concerned about maybe you know, personal growth and you know these um, you know, six figure, seven figure salaries and things like this is really just doing a great job working with your colleagues to do something to help solve a problem, um, especially in these resource constrained environments that we're working in here.

Speaker B: Your trip also coincided with another interesting uh, event that happened back in Washington when White House, um, released the memo on distillation attacks on the US AI models. Was there any sort of reaction that you've seen within your Chinese counterparts and your conversations on that?

Speaker A: Yeah, I think what I've seen kind of across the entire AI ecosystem is that um, distillation is something that a lot of labs do. Right. And so the, there's one perspective you can take of like hey, we're all working towards the same goal, right? This AGI destination and everybody wants to leverage the best of everything that's out there. Right. And so um, it happens that a lot of folks are data constrained. They don't have the quality data that they need to be able to kind of move m. You know, move along and improve their models. And so I do see that you know, distillation is done quite often and by a lot of different organizations, um, both Chinese and us. Uh, and so it's not something that I think is a unique, something unique to China. Um, and so it's, it's definitely interesting because we, we talk about um, openness and trying to move ahead and make some progress. And so I think it's important that we recognize that um, m Much of the original data is copywritten and so models are trained on that, pre trained on that data. And ultimately um, the utopian world is where data is a commodity and, and you can use that commodity to help improve your models. And I would Especially if you're working with like smaller models and you want to do domain specific training on those models. Having the ability to distill from a teacher models, a model that's got um, you know, can produce high quality data for you is super important, right? Because you don't always have, you may have the SMEs in house in your enterprise, but you don't necessarily have all of that data needed to be able to enhance your model and make it good for your application. And so I still see distillation as being something that's beneficial to the community. Um, but again there's terms of use and things that some of the providers have in place and um, it's always good to abide by those if you're uh, a commercial user.

Speaker B: M. I think in your substack article you were saying that there's a narrative in the west that China is copy pasting everything that west does. Um, and then another narrative is that the export control slows down the development of Chinese tech and Chinese AI. And you believe that both are wrong. Can you walk us through why?

Speaker A: Sure, yeah. Uh, so the, the Chinese labs are innovating, right? They're creating a lot of great innovations and those innovations are used for learning. They're being applied in actual both US Open models and um, on the API side. And so these innovations are cross border, right? Like this is just part of everybody that's working on AI research, working together and sharing what they're doing. And I think that's extremely beneficial across the board. The U.S. you should be doing it in Europe, Canada, China, everybody producing great open science, right? Um, the export controls probably haven't had intended um, effects. And so when you're forced to innovate because you don't have access to resources, you're going to innovate. And so I think what's happened in China is that the labs have figured, okay, we need to innovate, we need to get more out of the, more mileage out of what we've got. Um, and so that has happened, right? And I think that's great because um, we can apply the same principles and those same um, discoveries in what we do everywhere, universally, right? And across the world. In AI research, the uh, sort of the restricted access to hardware, um, is something that again has kind of forced China to uh, force Beijing to start to invest in their own silicon. Right? And so we're seeing more development in that area, a lot of new startups in the chips, uh, manufacturing space. Uh, and so, and you know, China is Uniquely positioned for um, rare earth minerals and the ability to build a very vertical, um, stack of uh, everything from raw materials through to production, manufacturing and deployment.

Speaker B: And so in order to continue this question, you also named uh, four specific Chinese contributions that Western labs are now building on. Dipsic, uh, grid po, the moonshots work on um, Muon Optimizer by Dense Verl Minimax, Scaling, lighting and attention. I wonder which one of these four will never happen, if there will be no export control existing at all.

Speaker A: I almost think that it's hard to say what innovations will be tied to export controls. Um, I like the example of the Muon Optimizer because it actually, the original work didn't originate in China, it originated in the US and so there's a paper published on that work. So the very initial work on that and then it was sort of built on by subsequent labs. And so that kind of um, iterative work is why we're able to make the progress we're able to make today. And I would love to see more of that open science being embraced because you create something, somebody else improves upon it and then improves upon it. And we've seen this with, this is how LLMs have evolved, right? Like the transformer architecture has been modified, there's been more optimizations, different flavors of attention and all of these sort of innovations that help us get to higher performant uh, systems. Right. And so I think this is very important and we're seeing the same sort of things happen higher up in the stack as well. So not just on the model AI research side, um, but also in, higher up in the stack and the frameworks, the agentic frameworks, all of these other harnesses and scaffolding and things that are being released out in the open as well. And so, um, obviously openclaw, hugely popular in China and now Hermes Agent, very. You know I think Hermes Agent, I think almost everybody I've spoken to at the the conference here is like talking to me about that. And so um, these open innovations, open source innovations are the ones that get the most eyes, they invite the most community participation and invite developers to contribute. And so I think uh, the same parallels we have on the research side we also have in the open source software development side.

Speaker B: Yeah, the open claw is crazy, is massive in China. I feel like it's now getting a little bit, I feel like it's a, and reached the peak and now it's just going a little bit plateauing and going down there. A whole new service offered how to remove it from your computer Ah, now that everyone install it. Um, but interestingly your um, stand on open source AI models and if you would be um, if, if there is one um, founder among the audience say building uh, AI, um and has like I don't raise 2 million US dollars. Sure. And has 18 months which path you would recommend him, where he should build it on. On Claude, I don't know. ChatGPT, a deep seq.

Speaker A: Yeah.

Speaker B: What's your take?

Speaker A: Well with $2 million I would strongly encourage them not to start an AI

Speaker B: lab because that's not going to maybe trading AI, but application.

Speaker A: I think there's so much opportunity for innovation and you know for startups and developing products you really have to find your niche, your vertical and like are you going to do something in finance, are you going to build something in a uh, particular domain or are you going to provide some sort of generalized like you know, professional services or these sort of things. And so um, I don't think I could say like you know everybody really loves cloud code. I mean uh, it's an excellent tool. Um, the decision around using running open models in local infrastructure or in your self hosted infrastructure, um, or using even the you know, companies like Fireworks or together to you know, these inference providers, um, or just using the you know, APIs, the paid APIs. I think that's a decision that you have to make based on your needs as an organization and so if you've got a startup and um, it generally you can do a lot more with less when you're using the open models. Um, and that creates a very nice uh, you know, very frictionless approach to you know, maintaining low spend on tokens and then being able to invest in things like marketing and like you know, growth of your teams and so forth. So um, but again the overhead, the administrative overhead is much lower when using an API and so I see you know both of these being very good options and it just depends on the startup and what you're trying to achieve.

Speaker B: OpenAI just discontinued their uh, model fine tuning services so you practically don't have a choice now but fine tune the open source one if you want to have something very cost efficient and apply it for a very specific repetitive task. Right.

Speaker A: Yeah and I think it'd be really fantastic to see more the fine tune open models being shared and having a very good ecosystem like Hugging Face and modelscope both have a lot of these models online but also trying to figure out which one is the best one for your application and a sea of open models is challenging sometimes. Um, but having A very trusted ecosystem of fine tuned models that you know their origination, you know what model they were originally built on, you know that you're not um, it's not being re released under a permissive license when it was a restricted license, restrictive license to begin with. These are sort of important insights that you need to have before you start commercializing your product around a fine tuned model that perhaps wasn't licensed the way that you thought it was supposed to be.

Speaker B: And I think one of these examples like uh, say meta, right With Llama model where licensing is pretty restrictive in a way for the applications um, in the view of the open source or framework and specifically a framework in relation to AI models that you promote. Where do you see the standard who is currently close to achieve reach? Uh the standard that the bar that you're raising with this um framework in what we can call as a true open source model today, is it Deep Seq, is it Llama, is it Quinn, Is it um Mistral?

Speaker A: Yeah I think so. This is kind of a two parter. So for one any of the models that are using permissive licenses which they're all using software licenses today so using Apache 2.0 or MIT, those models have the most permissiveness, right? So any of the Deep SEQ models, any of the Mistral as well, um, and anything that's kind of coming out of like um, AI2 or the LM360 initiative out of MBZUAI, these are all um, models that you can kind of do use for any purpose right? So research, you can build uh, your applications on it, you can create a startup around it, these sort of things. The community license are the ones that have the restrictions. So there may be um, utilization trigger limits and you just have to be discerning but any of these kinds of community licenses have um, a condition in there that would trigger you to have to go negotiate a new license right? Or um, non commercial use or things like this research only. So it's important to read the license, it's important to just drop it into your favorite uh, um agent or into ChatGPT and you know get the details. But um, so that's one part of it. The other part is that uh, we haven't been open source software licenses were designed for software, right? And uh, you know a couple of years ago we embarked to create a license at the Linux foundation that was directed towards models. And so three, four maybe four years ago now we started working on this um, model openness framework and we wanted to identify what constitutes an Open model and what constitutes um, something that's higher than that, which would be like Open science. You're releasing not just the model and its associated weights but you're also releasing um, all of the data sets, all of the uh, training code, all of the recipes, benchmarks, all of these other accessories that were needed to help get you to the final product. Um, and we've seen a few. AI2 has been a good steward of that approach of open science and um, allowing the community to be able to replicate the work that they've done. The um and we kind of aspire to this open science and this is really great for education, it's great for reproducibility, transparency and research purposes. The minimum viable product for what you need to be able to um, build a business on or study uh, to a lesser extent would be just the model and its weights. Right. And so we had defined that as an open model and at the time, and it's still being used widely now is this open weights term open weights. Although it implies openness, it's really being attributed to the restrictive license models. So those models that have conditions in place, what we call open model is those that have permissive license and so you have no restrictions on what you can do with it. Um, in practice open science is amazing. Um but when there's commercial interests involved people sort of default to the open model. Um, not everybody's going to be able to reproduce a trillion parameter model um, and have the resources to do that. And so I think it's important that we uh, at least have these open models out there that are using permissive licenses and that the entire community has access to them. People can learn how to work with this technology, they can build on the technology. I think that's extremely important.

Speaker B: Which of the current projects under Linux foundation brings the most attention of community and contributors? Where do you see the most sort of the hottest ones?

Speaker A: So there's definitely a lot of excitement around the Agentic AI foundation, uh, mcp. There's a lot of people building MCP servers and integrating MCP into their organizations. Linux foundation is uniquely positioned with all of these foundations that kind of work at different layers of the stack. And so up in the application layer here we've got the um, Agentic AI foundation and the projects that they're working on. And it's very early for us, right? We're five months, six months into establishing that foundation. We have coming close to 200 members involved now and so there's a lot of excitement about Agent Ki. Um, there's projects that are a lot more seasoned like Kubernetes, but are absolutely integral for all of the AI workloads and all the cloud workloads. And so um, obviously the Linux kernel is extremely important and very widely deployed. Um, and then we have the Pytorch foundation which has projects that are in the AI infrastructure space. So how to build the models, training and fine tuning and then deploying models in production and providing inference. So the foundations all kind of play a role in this stack.

Speaker B: When I told my engineering team I'm meeting AICTO of Linux, the first question they had is is it going to be AI introduced or embedded in a kernel level in Linux system? Is this any kind of plan that you have in mind or any kind of idea of that could be. Or uh, what will be the, where the AI can play, what kind of role AI can play within this fundamental.

Speaker A: I can't really speak to the kernel development aspect um, because I'm not involved in developing the kernel Linux kernel. But I would say that AI as a tool is being leveraged in many places. Right. And so to help optimize code and for some of the projects we work on, a lot of folks are asking now about when they want to use claude code or another um, agent, like how they make contributions. Uh, a lot of the projects have policies now around whether an agent can contribute or not. We're seeing a lot of the projects get a little bit overwhelmed by the number of PRs and how people that have very little knowledge or coding background are able to issue a PR and overwhelm maintainers a little bit. And then we're seeing AI being used to help mitigate that, to help review PRs and consolidate them. Um, so I think we're kind of in this storming space right now and I think we'll get to the norming a little bit more once we figure it all out. But um, definitely the agent coding, the coding agents are very pervasive now, especially in open source. And so um, it's kind of interesting to see how all the different core maintainers are kind of handling this and um, trying to reconcile the influx of

Speaker B: uh, PRs in regards of where the things are going now with the um, surf AI becoming the interface, uh, ultimately or even conversational AI, uh, do you really need the current interface of the os, uh, as we used to it, uh, with button settings, the command lines and all of this if you can just speak to AI directly if it's embedded into serving the root.

Speaker A: Even though we have the capability of using voice models now, like voice to voice. Um, you generally don't see people using that as much as they're using on their phone, like using the keyboard and so the QWERTY keyboard. Um, and so I think there's still a place to play for the UIs.

Speaker B: They were traumatized by Siri probably.

Speaker A: Yeah, maybe. But uh, yeah there's still, there's a lot of folks that are just not interested in the voice interfaces and they like also, you know, if you're in private, in public spaces, do you really want to be advertised?

Speaker B: Maybe just conversational, not necessarily voice. It could be through uh, the typing. But the idea is basically that you have a layer of interface that currently is um, redefining the way we look at the software in general.

Speaker A: Yeah, I think there's definitely a shift in how people interface with systems. Um, and using AI, there's also, it depends on who it is, right? The consumer versus the engineer, the researcher, the developer. Um, obviously those people that are more technical want deeper access. Uh, we're seeing even more agents use CLIS and um, integrating against APIs directly. And so as the models become more agentic and more capable, I think we'll start to see um, maybe a different set of interfaces for the agents versus the human interfaces. But right now we're training a lot of the models on the programmatic interfaces and the, the human interfaces.

Speaker B: You mentioned agentic AI, MCP and um, now that access to APIs, you also talk about the safety and the shift from the sort of safety uh, of the LLMs towards the safety of agent. Can you share a little bit more on what's your way of thinking towards this problem?

Speaker A: Yeah, so safety is obviously something that is important to everybody. It's important to downstream users, to consumers, to those that are building the systems. One of the things that's really important is model safety is one aspect. But as we build systems on top of the models, agentic systems and other forms of systems to come is taking into account safety. And so a safety first approach in engineering and safety by design, uh, is something I always encourage because you got to think about how people are going to use the technologies that you build. Right? And so we want to build those responsibly. Um, at the same time it presents a great opportunity for those building startups to create startups that are built around safety and security and be able to put in guardrails and harnesses that uh, will enforce the safety of that system. And so there's a lot that you inherit in agent AI from The model, a great portion of it. But you're also building programmatic scaffolding or programmatic, um, the agent is codified. Right. And so being able to also uh, enforce safety at that level is also very important.

Speaker B: When you look at the edge and swarms or multiple agents, how do you, how do you imply this safety? I mean when you don't really have that much of control, unlike when you're just working with one particular model.

Speaker A: Yeah. So for agents that are multi agent systems, that has to be built into the framework. Right. And so the framework and the harness help to contain. And so there's a lot of different architectures in terms of multi agent systems. Right. Like sports spawning sub agents or multi agent coordination, these sort of things. For the multi agent systems they have to be built into the protocol. Right. So enforcing safety and security at the protocol layer, but also within the framework itself that's running the agent. And so safety has to be viewed at each one of these levels the same way we do it on the web. Right. For instance, there's enforcement in the operating system, there's enforcement in, in the actual program itself. There's enforcement on the network layer, on the interfaces. And so um, often security and safety are kind of like a last thought when you're working hard at innovation and also experimenting with things and working at the tip of the spear. But it is something that needs to be addressed. And so it's very important that we um, look at these considerations not just from security on the network or security or privacy issues with the actual applications themselves, but also downstream effects. Right. Effects on society, effects on the environment, these sort of things.

Speaker B: The one question I wanted to ask you is about the LLMs. Uh, around a year ago you wrote that the reasoning is not the kind of reasoning that with human reason, it's kind of sort of like a fake reasoning. The models can't reason. Did anything change since that, um, one year later, now that we see these capabilities of reasoning and significantly improved?

Speaker A: Yeah. So I think the, for me, nothing's changed because it's not reasoning in the same sense that humans reason. Right. And um, when we're talking about like to next token prediction, we're mimicking human reasoning through strictly text. Right. And so through human examples being able to go and replicate how you reason. But we don't generally reason out in tokens or written aspect.

Speaker B: Right.

Speaker A: We reason through our brains. Right. And so um, I think it's important to recognize that the reasoning that we're seeing with LLMs is not the same degree of Reasoning as at human level. But the reasoning that we do have in LLMs is no less useful. Right? It's no less useful. It doesn't have to be the same type of reasoning as the human brain. Right. But the same effect. Is it useful for applications? Most certainly. Um, is it always right? No. Um, are humans always right? No. And so I think we have to look at this from the lens of I think the personification or like believing that we're interfacing with a human when we're interfacing with an LLM or chatbot is a little bit harmful. Uh, why? I think we've seen cases of folks that believe that they're interfacing with a human or they're um, falling into a delusion that perhaps what they're thinking is the right path. And uh, we've seen some court cases where um, it's been implied that particular chatbots have encouraged people to do things that uh, they shouldn't be doing. And I think it's important for us to realize that we're working with an innate system and that we're basically not interfacing with something that is thinking the same way we do or reasoning through things. They, they've, it's pattern matched data and um, sometimes it's just going to echo back our own, our own thoughts to us. And so it's important for people to recognize that these are you know, systems. They're not, they're not people.

Speaker B: I just want to say that I really like your metaphor of uh, LLM reasoning as with um, sort of parrots who can replicate the human phrase or even a sentence without actual understanding them.

Speaker A: Right. Yeah. I can't be credited with the stochastic parrot, uh, uh, analogy. But uh, definitely again when we're training LLMs, we're pattern matching. Right. We're intentionally um, especially through fine tuning and reinforcement, learning to make sure that the intent is followed with an LLM.

Speaker B: Right.

Speaker A: And so um, you throw a bunch of context at it but you also want to express your intent and, and ideally you get back exactly the kinds of things you would expect. Right. And I think we've improved over the last few years in our data mix and the way we approach training models that has yielded much better results.

Speaker B: Yeah, there are obvious limitations of large language models. Uh, what's your take on uh, Yann Lecun's new venture with building a so called world model?

Speaker A: Yeah, so Yann Lecun has this JEPA architecture a different way of um, looking at how not necessarily sequential training but um, I think it's great. I think for me my interest in world models is more associated with robotics, um, with embodied AI. I think if, if you come from the argument of like we to achieve artificial and general intelligence or super intelligence or whatever it is you're aiming for, um, that you need to have all the same ingredients like the multi sensory, like you know, physics informed all of these aspects, then the world models make sense.

Speaker B: Right.

Speaker A: And there's differing accounts of what a world model is. Some say it's okay, well it's video and images and te, um, others say well you need to be able to create 3D spaces in there, take um, real time construction. Yeah. And so um, there's different approaches taken for the world models. If you are a roboticist, you care about action, you care about the ability to actually uh, perceive the world but also predict outcomes. Right. And okay, what are my, you know, what's the next move for my robot? Right as I'm going to grab this particular object. And so you're, you're triggering action and you care about you know, real time sensory input, um, and being able to you know, activate actuators and these sort of things. So it just depends on where you're coming from. A world model I think means something different. Uh, and you know, this idea that like even gameplay or like generating games or like um, choose your own adventure style movies, those can all come out of the world model as well. And so I'm very curious to see how things uh, evolve. I mean diffusion models and transformers are sort of the architectures of choice. Unless um, you're using the JEPA architecture.

Speaker B: And you mentioned robotics. You also met a, um, founder of um, Unitree and other robotics companies in China. You also hinted um, that this industry partially going into bubble without revealing what kind of bubble is that. Can you share with us what's your observations in that sense?

Speaker A: Yeah, I mean there are a lot of robotics companies worldwide, but a lot of them in China. Um, there's probably over 150 or so humanoid robotics companies. Um, there's a, a lot of competition uh, for a market that is still small and, but growing. Um, and so the, the robotics companies that are able to solve a lot of the like real problems in robotics, like the human hand dexterity problem. And um, they're, they're still like the autonomy piece. Like a lot of robotic uh, systems right now are um, teleoperated or they're well orchestrated, they're programmed ahead of time. And so the autonomy thing hasn't been solved. And a Lot of people are working on that. Um, what I do see is that the Chinese robotic startups are really good at the mechanics, the physics, um, building the actuators, bring all the components together. The US robotics companies are more focused on the brain, the AI behind it, the multimodal, not multimodal, but M multimodal systems that are behind robotics. Um, but it's possible that there's going to be too many, there's oversaturation. And that generally happens in any industry was as folks rush in. Um, but the us, the Chinese robotics companies, it is a priority for Beijing and so they've invested in being able to grow that ecosystem. Um, and so there are going to be winners and losers, I think, like in any market. And so we'll kind of probably see that in the coming years, like who's going to rise up and who's going to kind of not be able to compete at that, at that top level.

Speaker B: One thing is interesting to hear your opinion as someone who is an expert, who spent years, decades in enterprise tech, now that you're looking at the applications, um, and you're working within this frontier sort of level with all these AI companies. Now, from your observation, what is the most common mistakes that enterprise make when they try to apply AI? The way, how they look at, and strategically finding, uh, the way to implement AI into the workflows, or looking at how kind of business models they can build or uh, automate internally. What's your, uh, take on that?

Speaker A: I don't think there's a challenge with adoption of AI. I think is successful deployment of AI in enterprises and that often, I mean, there's a multitude of factors that lead to poor outcomes. Um, one of those is the I know how I can build this better mentality. And so instead of the building, instead of buying, you're building and you don't have the talent in house, you're spending a lot of time experimenting, burning through cycles, burning through capital to, um, try to create something novel which may already have the core components out in the ecosystem. Um, one of the things we want to do in open source is provide those foundational components so that organizations can build on them instead of having to replicate them entirely. Right. And when we're all kind of running right now, especially on the agentic AI side, um, it's attractive sometimes to try and to build it yourself. Uh, instead of using a partner that may have already built a great framework, and even if that framework is an open source, they could provide professional services around that and help guide you. And so I think it's important to look at partners. Um, don't think of it as a journey you have to take on your own. Um, and obviously there's consultants that will um, happily take your money to uh, provide you some guidance. But I think it's the uh, what I've seen the most is the hey, we can build this better or we can build this and then it doesn't perform at sort of the expected outcomes.

Speaker B: M. Do you have any specific examples of like industries or like use cases that you see this failing, this kind of approach?

Speaker A: Yeah, originally, without calling out any particular organization. Um, you know a few years ago it was the, the, the chat bots. Right. And so um, how can we build a better chat bot on our infrastructure? Oh well we have the data, so let's go pre train a model. Well that's not a good idea. Right. The um, there's a lot of great you know, pre trained models out there. You can work from that base and then fine tune it yourself. Um, but we were seeing training models

Speaker B: is generally a bad idea if you're

Speaker A: not, if you're not a lab then you know, don't focus in on that area. Um, and we saw a lot of companies that were building their own models for different domains, different industries. Um, they put one model out and then they stopped. Right. And so um, and there's a lot of organizations I'm aware of that tried to pre train their own and fell short. Right. Burned through a lot of capital to kind of come up empty handed. So um, I think build on the blocks, building blocks that are out there. There's a lot of great stuff in the open. We definitely want to see more of that. We want to see more models out there that are open and permissively licensed. Um, and so I think it's very important that we have this really vibrant ecosystem of projects that are out there that solve the different problems in not just developing AI but also in applying it.

Speaker B: The one example that I know of, I heard of is Revolut, which uh, released their own model which is not as they claimed as the language model but trained on their um, numbers basically uh, the data which they apply internally. But I guess they probably first didn't really have much intuitive uh, to choose from. Uh, the existing projects, uh plus sensitivity around the big data and privacy I think is also somewhat maybe pushing them into that direction. But uh, we'll see so far it was first release, we'll see it will be the second in terms of um, besides the model, um, besides we build it better internally rather than use existing ones. What are the other mistakes on management, uh, business. When companies approach AI nowadays it's also

Speaker A: really about identifying the low hanging fruit. When you're looking at use cases to explore, don't reach for the top shelf and try and pull down the most complex use cases with the highest risk. You definitely want to do a risk assessment on the use cases and so ensuring that the use cases that you're applying are ones that um, are achievable and help you actually realize optimizations in automations. Um, if it doesn't and you've just built something out that demonstrates hey, we have the capability of doing this but it doesn't actually have any tangible benefits back to the organization then it certainly wasn't an endeavor worth taking. Right. And so you want to make sure that you're working on use cases that are going to be achievable. You're not going to spend you know, six to 12 months trying to build something and just to come up empty handed. Right. Um, and obviously do a risk assessment because a lot of people are very excited about the prospect of and actually one of the topics that came up yesterday was around like you know, how much, how much money would you trust your agent to invest on your behalf? And I think as we go on m through time and they become more reliable, you're going to trust it more and more to invest on your behalf. But uh, right now people are like I don't want an agent to touch my money. Right. And uh, we've seen instances of these chatbots or customer um, service agents that provide refunds that they shouldn't have given back, um, that have been jailbroken and um, again always going back to safety and security and privacy like make sure you have a very strong posture there, make sure you have strong guardrails in place because you don't want to do things like exfiltrate data or personal information. Um, and so a lot of these things have to be built in a deterministic way through code uh, as opposed to relying on the model to follow your intent every time.

Speaker B: Yeah, I just realized the guardrails that you prompt model with not necessarily the best way to do it. Having a sensor uh, hard coded in between the user and LLM and sort of edge cases might be a way to look at that. In regards to agenda ki. Do you see that realistically? Do you see any successful examples where autonomous agents, ah, fully replacing humans, um, kind of day to day job.

Speaker A: Yeah. So I would say there's sort of different views of what autonomy is with agents. Um, I've seen it as extreme as, like this, these like fully sovereign agents that exist in the ether on their own and they're, you know, going out there doing things and um, that at the end of the day, uh, you want an agent to be exactly what the definition of an agent is, right. Is to work on your behalf to achieve some sort of outcomes. Um, where I'm seeing probably the most sort of energy focused right now, I guess in two spaces. One is on the building blocks, right? So I go back to these concepts of building blocks where uh, you have these open source building blocks and you can go and build on them. Whether that's a protocol like an MCP or um, some sort of harness that's generally purpose or um, these other building components that you can use to build your agents. Right. And so skills and kind of just taking all these things together and creating something out of it. Um, I think that's where I'm seeing a lot of focus on people building protocols, standards, uh, specs, like the, the frameworks that are needed and then all of these, you know, reusable components. On the skills side, the other area where I'm seeing a lot of, a lot, you know, that's these are all. It's very experimental right now. Right. Like all of these things are very experimental. The, on the adoption side, it's again, it's the simpler use cases that have the, the most benefits. Right. And so you have to kind of cut through a lot of the um, hype and you know, people talking about like they just made like a trillion dollars off of like their, you know, their financial like agent that went on the, you know, started buying up stocks

Speaker B: or I fired my team and Now I have 50 agents.

Speaker A: Yeah, yeah, I have no one person company now. I had 500 employees and laid them all off. And I've got like 12 agents that, you know, that's, that's the kind of thing you have to cut through. Right. And so, um, you want to be realistic. There's a lot of aspirational sentiment around things like, oh, I can, you know, do my one person company and enable myself. And that's great. Like it's great that people can um, you know, build something that they may not have had the capability of building before, that they can automate a lot of the things that they do. Um, where we don't generally see some, you know, an agent handling everything for the entire workflow. There's an agent in this part step of the workflow. And this step of the workflow. And this step of the workflow. And so, um, as an example, like doing your taxes, right? Like if you don't want to be sorting through receipts and trying to tabulate them, you can run all your, you know, scan them OR run your PDFs and run those through an agent and have the agent go and tabulate them and create a spreadsheet and then total it and then kind of give you all these pieces and you go and validate it and then, you know, before you file your taxes. Um, and so there's a lot of, there are a lot of nuanced use cases that are out there. But the, the one thing is that you always want to validate, uh, you know, and make sure that you're especially, you're building an application, you want to benchmark it. Often you have to develop your own benchmarks because you need to make sure that it's working for your use cases. And so before you ship something, it's important or put into production is make sure you have the right benchmarks in place.

Speaker B: When it comes to AI adoption, um, within the business companies, um, of different size, what would be the best sort of strategy where you should start? If someone is running a business now today, there are different approaches, right? Some bosses saying, oh, everyone must use AI or if you're not, we'll give you a bonus and I will find you. If you don't use, um, what will be the, from your point of view, what will be the first step, the second step, third step, where we should start?

Speaker A: Yeah, So I think one of the things is important to recognize is that you don't need to work with AI directly to realize benefits. And so AI is finding its way into the features of the products that you use. Right. And so whether you, you don't necessarily need to interface with ChatGPT or Gemini, you can um, like Adobe strategy of like integrating behind the scenes, putting more generative features in, putting more agentic features in their products. These are, it's, it's a great way to help up uh, the productivity of businesses by um, adding features into their existing applications. It's the lowest lift for anyone that's adopting because they don't necessarily have to work with the technology directly. If you are in an industry where

Speaker B: you, on the customer surf level, you're.

Speaker A: Yeah, yeah, exactly, yeah. And so even if you're an enterprise, talk to your suppliers, talk to those product suppliers and say, hey, like you know, we want this feature, right? And have them built it in so that you don't necessarily have to compensate for that by building something yourself, right? Because um, you're offloading that feature development to your third party supplier, right? Whoever's supplying you those applications, that's the easiest lift, right? Especially if you're not an AI native organization, if you don't have skilled people in place.

Speaker B: So talk to your vendor, ask them to introduce feature built with AI.

Speaker A: Yes. And this is like if you've got a 100 person company, 50 person company, you don't need to go out and build it all yourself, right? You should be leaning on your vendors to help you. If you're a larger organization, I always say talk to a consultant. Um, don't try and don't assume that you can learn it all on your own on day one. Um, if you're a Silicon Valley startup, you're probably very focused on building it yourself. Um, but if you're building for some other purpose and you're not an AI first company or an agentic first company, then again you want to leverage the tools that are out there. Um, and it's really important to, and this is why open source is so important, right? Is having access to all of these tools, all of these components to start building your product or your service or enhancing your own internal systems. Right? If we talk about the enterprise and integrating with a lot of legacy systems and back office systems, um, being able to work with your own data, uh, now there's a lot of tooling that's in place that in a classical SQL world now you can bring that into your language model and um, there's just a lot of really great like products out there now a lot of them are commercial products, um, but they, you really have to solve for what it is, your, your use cases, right? And so don't I, I wouldn't go searching for a solution for a problem you haven't first identified. Right? Uh, and so again it boils back to the, you know, first principles, like what are the, what are the use cases that are most important to your organization? What are going to, you're going to realize the most optimizations through automation. Um, how are you going to realize cost savings, all of these things, right? And so you have to be very focused on the problem you're trying to solve.

Speaker B: But if you ask your vendor to do it for you, that means that all your competitors will have access to this as well. So you're not having this edge anymore.

Speaker A: Um, it's definitely, if it's differentiation versus productivity, like internal productivity is one thing. It doesn't always make sense perhaps to push that to the vendor, but often when it's a differentiating factor, then that's where you want to invest. Right. If you're going to differentiate yourself from your competitor, um, if your only differentiation is that you are able to process claims faster, uh, if you're in insurance, um, that's not really a moat. Right. That helps you move quicker and maybe provide downstream better customer service. But ultimately figure out what the differentiating factors are and see if AI is a good solution for driving that out. Uh, AI isn't always the best solution for everything. Right. And certainly, um, generative models are not the best solutions for everything either. Right. One of the things that I saw early when LLMs came out was that um, for fraud detection, everyone's like, I need to use an LLM. Well no you don't. You have perfectly good classical machine learning models that are able to do um, you know, identify potentially like fraudulent transactions or fraudulent device or like, you know, illegal devices on your network or these kinds of things, um, with higher accuracy and reliability. And so you don't necessarily need to shift to an LLM because it's the latest and greatest thing. You can also use a lot of your classical machine learning models.

Speaker B: I think now a lot of insurers facing a massive problem where the whole thing was claim insurance. Claiming is collapsing because everyone is using LLMs now to claim insurance or getting credit or anything like that. And then you mentioned, uh, for differentiation you should invest for kind of internal optimization. You probably should better off working with existing vendors. And I agree, it's like, because internal optimization is sort of limited, the upside is limited by default. How much you can is like 100% there. Right. But um, building a value, new value proposition, changing your business model, differentiate that surf. Upside is unlimited in a way.

Speaker A: Right.

Speaker B: So and then your returns, uh, also you look at the returns, right? Yeah.

Speaker A: And that's for the area that it's better to look at. Uh, innovating and building something yourself or at least partnering with someone that can, with an organization, a supplier, provider that can help you innovate in that space. Right. Versus everyone's going to be doing this anyways. Like internal optimizations. Uh, is that really where we want to invest all of our money?

Speaker B: Everyone is building solutions for that. Every, every big AI company now focused in their investment into that. So yeah, why would you. I mean you will be out competed by them anyways. You don't have that much resource.

Speaker A: Right, exactly.

Speaker B: Yeah. And you mentioned AI first companies. If you're not AI first company. You shouldn't probably do that. But can you build an AI first serve a spin off of your, what would be the best strategy? Uh, would it be like creating a lab or a spin off or anything like that that you can just build from scratch and imply? And what does it mean to be AI first company?

Speaker A: Yeah, so if you're a well established company and you want to be an AI first company, there's really two approaches, right? There's the embedded uh, skill sets in the organization where you may have, in each team you have somebody that's very AI savvy, either on the application development side or potentially lower in the stack. Um, or you have these sort of um, I hate to use the word like tiger teams, that's a very like 2000s word. But uh, that have these teams in place that can uh, work with all of your other departments to help get them onboarded, educated and leveraging AI and to various extents, right, because we're talking about enterprises, we're talking about startups, we're talking about a lot of different organizations that have um, varying needs for AI and then varying capabilities to actually build that capability in house. And so if you're a startup, AI first is definitely the way to go. You want to compete in this space. If you're not working with AI, you're going to be left behind. Um, but if you're running logistics, let's say um, shipping logistics, lots of applications of AI to be used and maybe it makes the best sense for you to create a small team that's going to help. Look at these use cases and what are the tools that we need to build? What can we buy, uh, to help with these particular use cases and optimize. And so um, there's a lot of low hanging fruit like the sort of customer service chatbots, um, more nuanced and complex things like when you talk about startups that are coming in the space of replacing call centers, right, and they're using voice agents and they're building an entire service around this uh, that doesn't require you know, call center with 200 people in it. And so that kind of company is going to be uh, an AI first company, right?

Speaker B: Thinking of the customer service now that after the introduction of MCP, your agents or your cloud, uh, open uh, ChatGPT can go and make certain action on behalf of you via MCP natively. So the question would be then do you really need then a customer service team to service you? If you can just connect directly and let your Agent, talk directly to the agent of the company that you want to buy something from. Do you see that is going to happen within the next two, three, four years? Do we really need this, uh, websites and overall, um, like I don't know, booking flight, uh, or hotel. Do you really need that in the next few years? If you can just ask, you know, your choice of your LLM provider. Book me a ticket.

Speaker A: Right, Yeah, I think that's what everybody aspires to have is to have their own personal agent that can act on their behalf, that knows their preferences, that can, you know, take initiative on their behalf, understand their intent, um, and be able to navigate that uncertainty in the commercial space or whatever systems they're interacting with, uh, to act responsibly on your behalf. Obviously human in the loop is still important. Uh, if I tell my agent I want them to go and buy me this particular couch when it becomes available below a certain price, um, that's a very simple use case. Um, but there may be nuanced things in there that you as a human would recognize. Um, hey, like the, you know, the trend is actually going downward and I know that I'm informed by other sources of information that the price is going to continue to go down. And so it's, you know, it has hit the threshold to buy. But maybe I just want to wait like you talk about this with stocks and you know, cryptocurrencies and all these other things. So there's um, there's things that we may be better suited to do. I think when it comes to making purchases. Again, how much you're going to trust are you going to put inside your agent? Um, but definitely I would prefer to have an agent go and you know, buy my tickets and organize my itinerary and uh, make sure everything is in place and, but to do that in a way that there's no surprises, um, and it recognizes my preferences and knows how to apply those. So I think that's the direction we're moving in the next few years. I think that we'll see that more and more materialize. There are evolving specs and standards that's coming out now that will help with agentic interfaces, um, to help standardize agentic interfaces. And so when you have your online store instead of uh, having it in a human interface where you go through and select different products, that sort of thing, um, that that interface is maybe more text driven and better suited for an agent to interact with. And so um, we're seeing a lot of that. We're also seeing a Lot of innovation in microtransactions and like payment rails and all these systems and agentic AI. Um, at the Agentic AI foundation we kicked off day one with like seven different working groups that are focused on everything from regulatory issues to, to um, working with um security and privacy and reliability and observability. All of these areas that need to be solved for in agentic AI. Um and out of that we'll see evolving standards and specs come out of that and um, solutions and ideas that can be applied by the community to help build this very strong ecosystem, build the networks, building this um, space where we can endow agents with a certain amount of responsibility and go and act on our behalf. And, and I think everybody's very excited about this idea particularly in China about like the opc, like the, the one person company where they sit back and do nothing and all their agents are out there like making that money. Um, you know certainly anybody would be very excited about that. Um, and, and so uh, you know how do we get to that kind of economy, the evolving like agentic economy? There's ideas of these self sovereign agents and uh, agents that are working on your behalf, potentially doing things on the blockchain or doing things through commerce, um, traditional commerce. And so um, it's all very exciting and I think it provides a lot of opportunity for innovation. I think that's why we're seeing so many startups that are very much agentic focused. Um and certainly the word agent is probably overused in many cases. But um, I think there's an opportunity here for innovators and people to solve real problems.

Speaker B: Interesting. And when it comes to yourself, when it comes to Linux and I'm not sure if you're coding yourself but your engineers, are they allowed to use AI to which extent how do you govern that internally and what would be the use case that uh they normally use AI?

Speaker A: We have a lot of projects, the Linux foundation, many, many projects and so each of those projects has their own policies around what they're using AI for. But I would say that the majority of projects use coding agents to help um, augment their abilities. Uh, I myself code as well and so um, my go to is cloud code like many people. Uh and I appreciate that it's a very useful tool. Um, can I ship the code that it produces? No, I definitely need to um, tack pieces together and run my unit tests and things like this. Um, but it's becoming a lot more proficient at doing um, all the unit testing and all of the um, you know pieces that need to be done to productionalize systems. So um, but yeah we, we see different projects have different policies on um, AI. And obviously you know, AI is being used extensively in the community for open source projects. So a lot of people that wouldn't be able to code previously have created a PR because they ran it against the code base and tried to find something that they could fix or optimize. Um, which has inundated a lot of the maintainers with PRs that they have to go through and review and sort out. And then on the tail end of that we have a lot of folks that are implementing AI to actually go and review the PRs, um, and consolidate changes and validate whether they're viable PRs. So um, I would say everybody's using AI to some degree because you almost have to use AI to counter the effects of AI. So it's very interesting.

Speaker B: I think engineering teams having a little bit of a hard time now that business uh, folks uh, come into them with something, caught it with clot caught and say look, um, it works for me. Why it doesn't work uh, uh, uh with you guys. I feel like the consensus is actually where uh, the distance between the engineering teams and business teams are now uh, shortening because business teams can actually wipe code, um, MVPs of the prototype of the product where they can apply the business logic and see and adjust and adopt much faster than it will be going through every iteration through the engineering team. While engineering team can benefit tremendously by seeing the business logic implemented, albeit maybe the code, it doesn't really matter in that case.

Speaker A: Yeah, the software development lifecycle is getting truncated. Right. And so um, being able to go from requirements to like POC is a little easier now. Um, but generating code doesn't mean it's optimized. Generating code doesn't mean it's secure. And so there's still a lot of things that the developers need to do.

Speaker B: You need production scale, the security, all this stuff that you probably won't be able to do this uh, with your cloud code. And if you're just studying if the last question, if you're a founder today, if you have US$1 million and where you would invest in right now, what will be your focus on doing right now in this space?

Speaker A: So that's a pretty wide open question because people have to figure out where they're going to differentiate themselves in the market. Right.

Speaker B: Where do you see opportunity in maybe sort of what segment or yeah industry or where they see the low hanging fruit right now.

Speaker A: Because I favor from like an open source lens, I obviously favor general purpose systems that solve problems. I'm um, very you know, bullish on agents. I, I would, you know they do um, they aren't great at like long horizon tasks right now. Uh, there's a lot of other problems to be solved in multi agent systems and safety and security. I would like to see what I'm not seeing, ah, enough of I think are companies that are focused on security and privacy for agentic systems. So there are startups out there, um, but for general purpose applications, uh, not seeing as much. So I think I see folks that are like oh heavily regulated industry, they're working with finance to like create scaffolding and guardrails for systems that are very specific. But how do you apply that more generally? Right. Like how can we create general purpose safety and security systems like we do with deterministic systems, um, how do we apply that and the stuff stochastic space and so uh, it would be great to see people that are working on ideas there um, and especially around improving reliability. Right. So in agentic AI the failure modes of LLMs are inherited, right? So like hallucinations, these are inherited into agents. And so we have um, the ability with agents to add ah, more deterministic layers and create tighter context in the harness and be able to become more focused. But um, at the same time you're not going to solve everything through the model. You have to actually take a systems building approach and you're going to build a system, you're not going to focus entirely on the model. And I think that's how we scale the capabilities of today's LLMs is by building systems around them, not necessarily focusing on the model itself.

Speaker B: Fantastic. Thank you Matt. It was great to have you.

Speaker A: Appreciate it. Thanks.

Speaker B: Thank. You.