The GCC Pod Ep 15: Open Source AI Has Left the Screen

Every week, my colleagues and I unpack what's really happening in the world of global capability centers. We're living in unprecedented times, and most businesses aren't ready for what's coming. Many mistake using AI for simply adopting tools like ChatGPT without actually thinking about how their organizations operate. For example, how Ashish Vaswani has built R&J model, their first open-source model described as a world-class 8 billion base plus instruct pair. Or for example, how companies are setting up their entire data value chain to acquire highly specialized expert datasets by setting up global capability centers in South Asia and South America. The real shift is happening elsewhere. Some companies are building their own models, others are embedding AI deeply into custom workflows, but many are securing long-term advantages around compute, energy, data, and most importantly, access to global talent. At the heart of it all is this: the future belongs to companies that can tap into the world's most capable, ambitious, and underutilized talent pools and turn them into a strategic advantage. In the GCC Podcast, I sit down with operators, founders, and leaders to explore how the world's most important businesses are building global capability centers, not just to cut costs, but to acquire proprietary data, develop custom AI systems, and transform their products, services, and workplaces into truly AI-native organizations. This is a podcast about how global businesses are being rebuilt from the inside out.

Muhammad Umar, it's been a minute since we spoke. So I wanted to kind of COVID this past week as well a little bit. And I think the biggest thing for me is that open source AI essentially went physical. You talk about voice, you talk about robotics and concrete. And one of the biggest open source AI contributors in the world has just closed the door. So I want to sort of break down what's happened in the last week. And I'll start with this, right? For the first time, the complete voice AI pipeline, listen, think, and respond, can now be assembled entirely from open source self-hostable components with no API dependence. And I think this is a category shift, and it gave us both the eyes, the ears, and mouth for open source voice AI. You have Cohere Transcribe for listening, Voqstral TTS. So would you say the moat of commercial voice AI companies like ElevenLabs, is that gone over No, no.

First of all, you know how passionate I am about open source AI, and I call it conscious of freedom. So Mode, I mean, if you talk to Adel and all these people and all the technologists who are working with us, Mode is currently not technology anymore. Mode is distribution. So for example, if you ask me, like, Muhammad Umar, and we are actually working on sovereign AI stacks for using these open source voice AI systems. For some critical defense infrastructure projects. So if you ask me, like, you want to build the capability as good as ElevenLabs, let's say, yes, you can build, I mean, you can build those capabilities using these open source systems under Apache 2.0 license and you can own the entire stack. It's possible. There are two main things that's important there. Number one, inference server, you need a lot of compute. For the inference. You need to serve customers, you need to serve your users at near runtime. So your inference should be very good. That's one bottleneck. And the second is in terms of if you want to harden these models based on your own data, for example, at some stage. So then some fine-tuning, but generally I think we should push towards adapting these open source AI systems for the enterprise use cases. Now it is possible. Previously, I mean, hardening these models were not that easy and there was so many missing pieces. But I think from this point onward, in terms of voice, I mean, I think if I would be a company and I need some voice command centers, I probably won't use these API systems. I would probably build my own AI stack. Because it's as— I mean, probably it would be a little bit expensive at first, but then you can build the entire stack, right? I mean, you don't need to rent these from these expensive APIs. So yeah, I mean, you can see my passion about this.

Yeah, no, definitely. Fantastic. So the next thing that I wanted to bring up that happened this past week, there's a peer-reviewed paper that has been published in the Nature Machine Intelligence, and it's introduced this thing called ROS LLM, which is a framework that integrates large language models directly with the robot operating system, ROS. And I think the ROS is already sort of the de facto standard in robotics. It's used by NASA. It's most defense primes, thousands of factories connecting LLMs to ROS. Means that any organization, it's massive.

This is huge.

Tell me a little bit about this.

So I was in the— I mean, in the NSF, I was in this event. It was a robotics event. And one of the challenges— I mean, when you talk about large language models, Yann LeCun and all those, they say these are just language and text. I mean, it has nothing to do with the physical world. But this across— what we are trying to bridge is how to bring the physics of the natural world into the context of language. And I think this is very powerful. Even if you perceive the world around us, you perceive it with the text language, but also with the physics of it. So this is, I think this is breakthrough in terms of, I mean, this is path towards the breakthrough where you leverage the existing language AI systems to even work with the robots. And these are the gaps that's— the gap between the physics of the natural world and the language. These are the gaps that these technologies are trying to bridge. And great, again, they are open source, right? You can take those systems, work with companies like us, like CodeNinja, and harden them for your enterprise use cases. You don't need the— Boston Dynamics of the world and all that stuff. Because hardware is, uh, all of a sudden the cost of hardware is going down a lot, right? Uh, and that, and, and there's one thing I want to mention. I think probably people, uh, needs to understand, like, what's the difference between— I mean, in our factories, robots was the most of our factories. I mean, if you think about Japan, they— there's lots of robotics deployment already. These, those are the robots are built on control systems. Now, how to move those systems from control systems to basically where the robots are the brain, that's a shift. And with these kind of models, it's actually far more easier, better. And you would see a lot of adoption. I was at NVIDIA GTC. I've seen every major business, industrial manufacturing business, were driving this change. So from control systems to this inference-based robots, basically robots with a face. Yeah. And anything that is these kind of systems that allows you to leverage the language models and combine with the physics of the natural world.

And I think if we keep thinking about that context, like the physical world, the physical infrastructure, another thing that I want to get your thoughts on is the fact that Meta has very recently released this thing called BoxSqueet. Which is again, very interesting Bayesian optimization for concrete. And it is yet again, another open source probabilistic model and optimization framework for designing concrete mixes. And oh yeah, very interesting. Yeah. I mean, and so this thing I have here, it says that the model is trained on a new open access dataset of over 500 strength measurements from 69 mortar and 54 concrete mixes at 5 curing ages. I mean, this is so cool to talk about, right? But I think I want to hear more about this from you.

I mean, if you see any construction business and you see how they mix concrete, I mean, you will very quickly get a guess. Like there's no, I mean, it's mostly on the estimates of a lead structural engineer. Okay, do this. There's no real measurement and thoughts that go into it. Like, so, and the US produces 19 more than 90 million metric tons of cement every year. Just think about it. And concrete is the most used material on Earth after water. 30 billion tons, I guess, or 35 billion tons for globally every year. So it's huge, but no one ever thought about like there's a better way to mix, do this mixing of concretes in the construction materials and come up with a better way to create things like cement. I think this is, a model like these are basically, I mean, just imagine how we are bringing AI into these systems. Just imagine that you are now bringing reasoning capabilities in the mixing of concretes. And again, this is open source. So anyone can, anyone in the construction business can take those models and build their own sovereign AI infrastructure stack based on these models.

Yeah. And I was also reading about this and it said that it cured 43% faster in real-world tests. I mean, this thing is open source. It's essentially transforming industries that have nothing to do with chatbots.

So one of the challenges with open source is if you think about open source community, big labs produce lots of open source models, lots of researchers do open source models. There's an institute called Allen Institute. They have done some incredible work in computer vision models and made them open source. The challenge with open source is someone has to take those models, right? And then harden them for the last mile, right? Someone has to implement the security. Someone has to increase the intelligence density at certain parts, do the fine-tuning, implementing the security guardrails and all that stuff. So that hardening part is the reason why people go to the proprietary. And when I found it, like, okay, when I found, like, I talked to so many CIOs and when I found, like, they, that's the part they're missing. I thought, okay, let's Code Ninja do that. And we started doing it and along the way dropped that entire hyper platform that now allows us to take any open source AI system and harden them for these large model. So yeah.

Right. So let's move on to the next one.

Okay.

Between March 31st and April 2nd, which is yesterday, Alibaba has released 3 AI models in rapid succession, all of them closed source. I think this is one of the most important strategic signals of the week that I wanted to share because for anyone building an open source AI, Alibaba was one of the most prolific open source AI contributors in the world. You got Qwen models and they powered thousands of applications. But I think what I really want to hear from you on is that Is this the beginning of a broader consolidation where companies close the door once they have that market position?

I mean, once you have that land grab with the developers and the companies, and then you start closing those most important models. I think if you see Mark Anderson's recent tweet, like these are the discussions even in the White House. It's not like, I mean, these I mean, we should not open everything because it gives too much power to other, I mean, decentralized systems or entrepreneurs like us. I think we should restrict, we should go against this. I mean, the upside of open source is far more than the downside of it. I mean, People give arguments, hey, in the White House they give arguments, it's not being used in developing some bio weapons and things like that. It's not that easy. I mean, when you make anything open source, the scrutiny is very, very high. You can't do these things, right? So I think we should restrict that. The idea of OpenAI, for example, was to make AI open. Yeah. They couldn't do it, right? They could— I think it's companies like us and the community and the industry and CTOs and CIOs. It's their job to make sure we keep this, keep AI open. If we don't do that, AI is extremely centralized. If you see every workplace, the best thing they know how to make best use of AI is like how to automate the labor, right? Because if you are us, if a company doesn't own the entire AI stack, there's a limit to what they can generate new ideas from. Because at the end of the day, then you have these closed models and there are certain things you can do on top of because you don't own the stack. So in our case, we built this sales system for us and you can see like because it was built through Hyper, it was all open and the kind of ideas we applied in building these Building these sales systems is not possible if you're using any closed source systems. And in that case, if you generate more ideas, then labor automation is the last thing you want because you always have more ideas than the things you can do. But if you're not owning your AI stack, if you keep using ElevenLabs API for your voice command centers, there's a limit, right? I mean, of course you can't generate more ideas. So that's the reason why I'm so passionate about open source. Why we need to make sure that every business and every workplace have their open source AI stack, sovereign AI, so they can generate more. Otherwise, you would see more and more jobs using— I mean, Reckless fired 30,000 people. You see more and more firings. As Jeff Hilton said, these big companies, so heavily funded and the only way they're going to return to their investors is that they automate lots of the jobs. And I'm not okay with that. I think humans, and we can generate far more ideas. The technology is incredible. I mean, it's so fascinating when I built this entire CRM for myself. And built and clubbed it with all the stack. I mean, I can't do what I'm currently doing without this, but the only reason I was able to implement so many ideas was because the entire system was open source. Uh, we built the entire stack in a few days. Yeah. Hardened these models for our own enterprise use cases, for our sales use cases. Uh, and then you can build any copilot on top of it if you like. Like you can plug it with any other, but, and any copilot, but at the end of the day, you want to own the stack. You can't just use these big tech AI layers and automating everything, every last labor job that you have. I am not comfortable with it. I think we can generate far more ideas. Yeah. Back to the Alibaba thing. Yes. Issue. But what I want to mention is it's not just anybody in the white— these are the discussions. It's like, I think I've just seen that tweet of how there are discussions like how to make it closed, how to make the—

Yeah, because I honestly think it's more about building internal capability to actually train but also maintain these models independently. And the way I personally look at it, the stuff I've read so far, and I'm getting into open source slowly, I think we have to treat these open source frontier models as kind of a non-renewable resource in that sense. Because if Alibaba can do it, like you said, anybody can do that.

The major change, I think, what— there are some intricacies. For example, if an enterprise wants to adapt, wants to build a software stack, it's great if they pick anything under Apache license. So it gives you more clarity and you can do a lot with them. So yeah, I mean, I think what Code Ninja is doing, there should be 100 more companies who should be doing exactly the same thing. That's the only way we can survive this massive job disruptions. Because if we help every construction business, if we help every pharmaceutical business build their own AI stacks rather than keep buying it from Refine and whatnot, then there is a limit to the ideas they can deploy. What needs to be stored in the memory? What needs to be stored in the pre-training? What needs to be stored in the after-training? That entirety is possible only if you own the AI stack. Awesome.

All right, so before we close out, I think one thing that I'm interested about and that I just want to highlight, one thing that's on the horizon for me, let's put it this way, DeepSeq v4. I think I want to keep an eye on it and maybe that's the next thing we talked about, the next thing we talk about in the future. I think expected specs, you're looking at 1 trillion parameters, multimodal model, MIT license. I think if it lands this month, it's going to be The largest open weight model that has ever been released. Am I correct?

Yes, it would be the largest if it gets released. If it gets the largest open weight model ever released and would fundamentally reimagine the competitive landscape overnight. And I think for this thing, we would do one separate podcast.

Yeah, I think I was just—

and it's multimodal. It's like it's text, it's vision, it's audio. So it's one training parameter is great. I mean, but for example, last week, I don't know. I mean, you should have mentioned this, or maybe in the next part, Prism. Well, it's another company and their core idea is to recently release their model. I think it's on, I'm finding it hard to pronounce, B-O-U-B-I-R or something like that, right? And their core idea is like it's 4 billion parameters, but it's not just that you hire the number of parameters doesn't mean higher intelligence density. So you can get a lot. There's a lot of research going on in how to increase the intelligence density with even 4 billion parameters. So they released their, was probably in the certain benchmarks, they were at 76%. Everything that every system ties back to the energy needs. So you can run that 4 billion parameters in your hardware. Or for example, if you want to build systems that requires test-time compute, like IoTs and stuff, you want to do the inference in real time there and then. You need models that can be deployed at the edge. So we need to think about it. And I think most of the gains that we get are from edge deployments. And for edge deployments, you need models that are not 1 trillion parameters, but that are probably 4 billion parameters, 2 billion parameters with a very high density for a certain use case.