# Local AI Will Take Over Within One Year | Ian Weinberg

Markdown copy of https://realalpha.app/v/PJlTf2ACCuY/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Ian Weinberg says AMD’s Ryzen AI Halo makes complex local AI practical, with local models handling many tasks while frontier models cover the rest.

- Show: MTS
- Hosts: [Sophia Dew](https://realalpha.app/people/sophia-dew/)
- Guest: [Ian Weinberg](https://realalpha.app/people/ian-weinberg/) is an AMD representative.
- Published: Oct 31, 2026
- Length: 25 min

## The best minutes

- [9:58](https://realalpha.app/v/PJlTf2ACCuY/?t=598) [Shared memory lets local models exceed 128GB workloads](https://realalpha.app/s/shared-memory-lets-local-models-exceed-128gb-workloads/) (5 min)
- [14:43](https://realalpha.app/v/PJlTf2ACCuY/?t=883) [Local agents can turn a prompt into a Blender scene](https://realalpha.app/s/local-agents-can-turn-a-prompt-into-a-blender-scene/) (3 min)
- [18:49](https://realalpha.app/v/PJlTf2ACCuY/?t=1129) [Local AI can work offline and route hard tasks to frontier models](https://realalpha.app/s/local-ai-can-work-offline-and-route-hard-tasks-to-frontier/) (4 min)

## Also worth watching

- [7:57](https://realalpha.app/v/PJlTf2ACCuY/?t=477) [AMD says its software stack scales from desk to rack](https://realalpha.app/s/amd-says-its-software-stack-scales-from-desk-to-rack/) (2 min)

Skip the first eight minutes of hardware setup and developer-center walkthrough, the sponsor reads, and the closing ads.

## Summary

Ian Weinberg works at AMD and presented its Ryzen AI Halo Developer Box. He says he came up in the security world and describes himself as a lab geek who makes demos, including local AI workflows.

Weinberg says the Ryzen AI Halo’s 192GB of shared memory lets developers run models with up to 300 billion parameters locally, and he demonstrates a workload that goes beyond 128GB. He describes preinstalled tools and an open ROCm stack intended to make the box usable by both beginners and experienced developers. His demos explore music generation and multi-agent workflows, including a Blender task that was still running at the end of the segment. Weinberg says local AI can work alongside frontier models: in an AMD demo, a semantic router sent 70% of general office reasoning tasks locally and 30% to frontier models.

### Minute by minute

- [0:35](https://realalpha.app/v/PJlTf2ACCuY/?t=35) **192GB enables bigger models.** Weinberg says the Ryzen AI Halo uses a Ryzen AI Max Plus Pro 495 processor with 192GB of shared memory, which developers can divide between the CPU and GPU. He says up to 160GB can be reserved for GPU VRAM, enabling models with 300 billion parameters on a box less than two inches tall.
- [3:01](https://realalpha.app/v/PJlTf2ACCuY/?t=181) **Preinstalled tools shorten setup.** Weinberg says the standard x86 box supports Windows or Linux and runs natively, without ARM emulation. Its Ryzen AI Developer Center includes preinstalled models, applications and playbooks for beginners and advanced developers; AMD describes the goal as getting to a first token in minutes.
- [7:57](https://realalpha.app/v/PJlTf2ACCuY/?t=477) **ROCm skills scale across systems.** Weinberg says the same ROCm stack runs from AMD’s smallest hardware to its Helios systems. He argues that developers can learn at desk scale and carry those skills to rack-scale systems, and imagines the box being used in university classrooms.
- [9:06](https://realalpha.app/v/PJlTf2ACCuY/?t=546) **CPU, GPU and NPU split the work.** Weinberg frames the platform around fitting models into memory, using the CPU to orchestrate tool calls and encode/decode work, and using the GPU and NPU to run models. He demonstrates prompting a model to revise a song and then sending the result through ComfyUI for music generation.
- [12:18](https://realalpha.app/v/PJlTf2ACCuY/?t=738) **The demo exceeds 128GB.** Weinberg moves the music workflow into ComfyUI with another model. During decoding, the task manager shows 128GB of dedicated GPU memory and 24GB of system memory in use; he says the workload exceeds what a 128GB system can hold.
- [14:51](https://realalpha.app/v/PJlTf2ACCuY/?t=891) **Agents link models to Blender.** Weinberg runs Goose with a 35-billion-parameter model and connects it to other applications through MCP. The agent is tasked with generating a toy robot image, passing it to a 3D modeler and putting it into Blender; the task is still running when the segment wraps.
- [18:49](https://realalpha.app/v/PJlTf2ACCuY/?t=1129) **Offline AI serves varied users.** Weinberg says the box works without an internet connection, including for coding on an airplane without Starlink. He sees uses among developers exploring local AI and small and medium-sized businesses that want to sandbox workloads and provide local access.
- [20:20](https://realalpha.app/v/PJlTf2ACCuY/?t=1220) **Interest has outgrown skepticism.** Weinberg describes a shift in the questions he hears: skepticism about local AI two years earlier, questions about how to get it last year, and now requests for more systems. Sophia Dew says interest in owning data and running AI locally has also grown globally, and mentions concerns among some of her European friends about losing access to frontier models.
- [22:14](https://realalpha.app/v/PJlTf2ACCuY/?t=1334) **Local routing complements frontier models.** Weinberg argues that local AI and frontier models work together rather than as alternatives. He says AMD’s configurable semantic router sent 70% of general office reasoning tasks to local compute in one demo, while the other 30% needed frontier models; Dew notes that local handling could affect users’ token costs.

## Transcript

MTS, hosted by Sophia Dew, with Ian Weinberg.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:00](https://realalpha.app/v/PJlTf2ACCuY/?t=0) Two years ago, if we were talking about local AI, there'd be a lot of skeptics coming up and saying, I don't believe this. Last year, there was engagement asking, how do I get this? Now the question is, how do I get more of these? So the barriers to interest are gone. Hello, everyone. Welcome back to MTS Today. We are joined live in the studio with Ian Weinberg from AMD for a very special demo. I know you guys unpacked earlier the big Ryzen AI halo, and we're going to actually test it live on air with this cool demo. And Ian's going to talk us through what it is and everything about it. So welcome to the show.

[0:37](https://realalpha.app/v/PJlTf2ACCuY/?t=37) Thank you. Hi, Sophia. Yeah, so today I am here to talk to you about the Ryzen AI Halo Developer Box. Am I open by telling you a little bit about what it is? Please, please do. Okay, first of all, first thing I want to do is put one of them into your hands. And we're going to talk first about what's in it, but then I also want to take a second to talk about what went into making it. So inside of it is our next generation Ryzen AI Max Plus Pro 495 processor. Now, from here on in, I'm going to call that 495 because that's a lot to say. And really, the headline feature of the 495 is 192 gigs of shared RAM.

[1:18](https://realalpha.app/v/PJlTf2ACCuY/?t=78) So why that's important is that can be shared between the CPU and the GPU. And when you look into the world of developers today doing local AI, what they run out of very often is memory. it's not the compute side of it it's the memory side of it so of that 192 you can reserve 160 of it for dedicated gpu vram and that unlocks 300 billion parameter models wow 300 billion parameter models that you can run using these like glm 5.3 flash can be run directly on this box okay that's really impressive and so i was asking him earlier because i for some context we're breaking fourth wall today because we're doing a demo so we're breaking fourth wall thank you to the audience here who's watching so i would i had the pleasure of being at the amd conference back in july and so advancing ai we saw the halo rack system actually being unveiled and i was asking about this and apparently this would not have been possible back in july so the demo i'm going to show you exceeds so back in july when we were on the 395 version that had 128 and actually you know one of the things i wanted to mention about this is take a look at the industrial design And when you look closer, you'll see that every one of these little pieces here is an AMD logo.

[2:32](https://realalpha.app/v/PJlTf2ACCuY/?t=152) Wow, that's beautiful. If you have the close-up version here, I'm not sure what the camera is, but you can see the very close-up AMD really just on here. It's like you're in a purple. It's just hundreds and hundreds of AMD logos. But it's a beautiful piece. And the thing that we mentioned there is that 300 billion parameter model runs on a box that is basically the size of a book, less than two inches tall. And inside of it is a suite of tools that are meant to help developers of all abilities. Wow. So when you turn it on, the first thing that you'll see, I don't know if we, do you want to jump right into demo?

[3:05](https://realalpha.app/v/PJlTf2ACCuY/?t=185) Yeah, let's just jump in the demo because I think what's pretty interesting is oftentimes, sometimes you can see the hardware. But then when you plug it in or try to install it, you have no idea how it should actually run. And what's really unique about hardware and software is there's certain ways you can actually run the hardware to optimize it to make it as efficient. But often they're separated. So, I mean, for a little bit of context, I came from a developer background. I was really obsessed with open source models. I would try to figure out what hardware I needed at home to actually run my own open source models.

[3:32](https://realalpha.app/v/PJlTf2ACCuY/?t=212) And you kind of have to, like, you're, like, Googling different software, installing different things, and then testing it. They're always separated, right? This, pretty interesting because it looks like as you actually install it at your computer, as you plug it in, you start to see something that just pops up. And you get to understand how to run these things with its ultimate capacity. because you can run something on it without using all of its capacity, right? Depending on how that software is designed. So maybe that's a good intro to your demo. You threw so many balls to me that I want to talk about all of them.

[4:03](https://realalpha.app/v/PJlTf2ACCuY/?t=243) So the first thing that I didn't say before that I'd love to is that as developers, you become accustomed to a development environment. There's something you love to do. This is a standards-based x86 platform. If you like your version of Linux, install it. You like Windows, install it. You're not paying any cost for having to emulate something in ARM, 100% native, and totally open to go the way that you do. But if you take it the way that we build it, We offer it in both Windows and a version of Linux. You can get it either way. And after you go through the out-of-box experience, this is the screen that you will see.

[4:37](https://realalpha.app/v/PJlTf2ACCuY/?t=277) Now, not with the task manager on the right-hand side. That's, of course, me setting up for the demo. But in the middle, both the Linux and the Windows versions have the Ryzen AI Developer Center. And inside that Developer Center, there are a number of playbooks and pre-installed elements that are directly on the box when you buy it. So the message we put out is minutes to first token, right? But just all you've got to do is turn it on. Minutes to first token, that's a good. So it's like, hey, how many minutes to your first token? 36 seconds, not even a full minute, something like that, right?

[5:06](https://realalpha.app/v/PJlTf2ACCuY/?t=306) Totally. And you know what the best part is? That's the greatest challenge on Earth. We would love to see people go faster and faster and tell us what we did, right? But inside each one of these, you can then go choose whether it's for the Ryzen AI Halo, what kind of system memory you have. In this case, it's the 192. 192 system memory. That's a good number. One night, 300 billion parameters. I'm going to keep saying that over and over. By the way, something I'm going to say as we go through this entire thing is I'm going to say sit, orchestrate, and stream. I'm going to say those things over and over again.

[5:36](https://realalpha.app/v/PJlTf2ACCuY/?t=336) And we're going to get to there in a couple of minutes. But I want to set that up to start. You'll see that for some of these playbooks, it says pre-installed. So if you were interested in trying. Pre-installed. Pre-installed. Okay. Okay. Models, applications, all of that directly on system right away. Wow. So if you're a very beginner developer, you want to just start from nowhere, we can help you. But if you are someone who is more advanced and you want to look at only the kinds of things that are for intermediate or advanced people, we have those as well. Because while we want to make sure that we are comforting and bringing the beginner developers along with us, we also want to make sure that this is an open source platform based on Rock'em, based on all of the open source software stack that you're used to, Python, Llama CPP, all of that is supported so you can make it do the things that you want it to do.

[6:27](https://realalpha.app/v/PJlTf2ACCuY/?t=387) Yeah. Well, this is exciting. I mean, let's talk a little bit about Rock'em. I think traditionally, this is a little bit before Vibe coding and before it got as easy to program. Rock'em is a little trickier. It's a little harder, especially if you weren't necessarily taught right off the bat, especially for some developers who were trying to experience and experiment with AMD hardware. But I know I talked specifically to Anush about this and all the latest updates in the last year. So this is maybe much, much, much easier. Things have really, really changed. But what I think is really exciting are the things that you can just do straight out of the box.

[7:01](https://realalpha.app/v/PJlTf2ACCuY/?t=421) So you don't really have to know exactly what's happening because you have these, it looks like it's pre-installed. Can we just look at one of the beginner ones? Because I think this is... You know, we're going to do that. Okay, cool. The couple of applications that I'm showing are pre-installed applications, but I've made them do the thing that I want them to do. Okay, perfect. I want to shout out our sponsor, Lovable. You know the app you've been meaning to build? The internal tool, side project, or product you'd otherwise lose a weekend to? With Lovable, you get software you can ship today.

[7:28](https://realalpha.app/v/PJlTf2ACCuY/?t=448) That includes an editable code base you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now, onto the episode. So the interesting thing about Rokam is I'm sure what a noosh.

[8:01](https://realalpha.app/v/PJlTf2ACCuY/?t=481) **Ian Weinberg:** Great simplification, great report. And that feedback is coming directly from the developer community. So I can stand on other people's backs and say that. But the really neat thing about our stack is the same Rockham stack that is installed on the smallest hardware that we have is the same Rockham stack that goes all the way up to Helios.

[8:18](https://realalpha.app/v/PJlTf2ACCuY/?t=498) really fascinating so if someone's coming into the amd ecosystem and you're learning at desk scale going to rack scale it's just it's just a group of skills that you're developing today you're doing local fine tuning laura q laura kind of stuff or you're doing pre-training up in the instinct stack your learning is one yeah curriculum i could see a version where you're a freshman in university and you get one of these and your syllabus and you're early i don't maybe not Exactly. I don't know. You're going to borrow it with your classroom and you understand how this works. So in theory, that same skill set now when you graduate and you're now operating like rack scale systems, you have that muscle and you can do the same thing there. By the way, that sounds like the coolest path for university. I know that is really cool. That would be awesome. If you're a university and you want to get some of these for your classroom, let us know. 100%. That'd be great. Yeah. So what I can do now is I can actually talk through this. So what we're going to do is a bit of a capability demo and a capacity demo. Cool. Because when I talked about those three things this platform represents a balance that we have built and that balance is one is fit fit

[9:21](https://realalpha.app/v/PJlTf2ACCuY/?t=561) means does it fit into the memory signature the second one when you're talking about orchestration is as we're transitioning into the agentic world and agents are just a whole ton of agents that are moving very quickly the cpu becomes very important because it's doing tool calls it's doing encode decode it's doing a whole bunch of stuff that is traditionally traditionally historically ai would be done gpu but it's being done cpu so that's where you get into the orchestrate stream which is the third component is the gpu and also the npu because small models run really well in npus now so we built this box with that balance in mind all of those capabilities so the thing that i'm going to show you today is um is think of creativity think about when you wake up in the morning and you're like am i a creative person and i will say i am not i'm artistic creative human being but i'm a very musically oriented person that's very cool very musically oriented so what i've been exploring is how do i start to create some music via prompting so there's a couple of things that are here and i recently went oh this is really cool okay so i know i get what you're saying when you're saying you wanted to design it something for something

[10:36](https://realalpha.app/v/PJlTf2ACCuY/?t=636) you specifically wanted to do okay something i wanted to do and by the way the next demo after this when we continue to talk it's going to be a self-running agentic demo that will happen right afterwards but this one is taking advantage of the quen 3.8 flash next so with that particular model it's um it's a hybrid model so it's got n grams and it's got traditional parameters and it's about 120 billion parameters plus another 56 or so um of of those n grams the n grams run in the cpe so that's something to understand but this is the kind of thing where you know i want to make me i want to make a song and it made a song for me and i don't love it so let's turn around and say hmm i like this but make it more like dance hall robin with a slower beat so effectively what i want it to do is think through how what kind of song is that i'm looking to build change the lyrics build it up however it's going to do and then i'm going to feed that into comfy ui and i'm very large model is running at you'll see it's running remind me the size of this model so this is a it's a hundred and seventy six I think total of the parameters they 120 million a billion of them are the per the traditional parameters and the rest of it is engrams which would just be

[11:55](https://realalpha.app/v/PJlTf2ACCuY/?t=715) look up on the CPUs yeah did you jump around so what you'll see here is this Loaded up, 98 gigs of memory. This is a cool task manager so you can see in real time. Isn't it fun? It's happening. But this is not the whole story because right now it's got 98 gigs of dedicated GPU plus 22 gigs of memory. So right now, add those two things together and where'd we get? We're like at 120, right? Yeah. Roughly in that range. Now, that would be great if that was all that we were doing. But if that was all that we were doing, then I could do that in a 128 gig system. Oh, fascinating. And are we going to demo and show what that might be?

[12:30](https://realalpha.app/v/PJlTf2ACCuY/?t=750) We're going to go past 128 gigs now. Oh. We're going to drop this into ComfyUI, and I'm going to have ComfyUI using another model make a song for me. And when it does that, what you will see happening is, so in this case, we use A-step 1.5. A-step 1.5. Yes. So from there, what I'll do is I'll have that kick off. And really, it's when you get to the decode that things get interesting. So you'll watch that it will start to build up. 105 and as we're moving through the prompting step and the end code yeah and give it a couple of seconds and what it's doing now is this coming up i don't know what the audio is going to work on this but if it does i could play it for you at the end yeah i think we have to share audio and but it's okay we can we'll hear it at the end we'll we'll get there towards the end so as it continues going through you're seeing more and more stuff being loaded into memory as it gets to the decode stage.

[13:26](https://realalpha.app/v/PJlTf2ACCuY/?t=806) Uh-oh, here we go. That's going to be real hard to do 128 gigs of dedicated GPU. Can you tell me what's happening right now in the task manager? Yeah, so right now in the task manager, you'll see dedicated GPU memory has gone to 128 gigs. But you have to add the memory that's being used by the system to that.

[13:47](https://realalpha.app/v/PJlTf2ACCuY/?t=827) **Ian Weinberg:** So right now, 128 plus 24, we are outside of the envelope of a traditional 128-bit system.

[13:53](https://realalpha.app/v/PJlTf2ACCuY/?t=833) Yeah, I'm just trying to imagine running this on my MacBook. So I'll let the other guys sell their platforms, but what I will say right now is the only platform I could do this on is this one. Wow, yeah. Even the other competitors. So it gives you an idea of what this unlocks. It unlocks the ability to go to much more complex models, to do more complex reasoning, to be able to take on tasks that traditionally would not have been thought of as being local. Yeah, that's impressive. I think one of the things that I'm particularly excited about is the connection between the software and the hardware. I haven't particularly seen that before. I love playing around with these different hardwares, but you don't see it straight out of the box. You're often having to figure that out on your own, and then you don't know what's actually optimized for the hardware that you're trying to plug into. So this is really fascinating to see.

[14:39](https://realalpha.app/v/PJlTf2ACCuY/?t=879) Let's rock them all the way through. Yeah, let's rock them all the way. So that's the other thing to know here. So this is just a couple of the applications that would be preloaded. Just for fun in the background while we're talking, I'm just going to have something else running. Totally. Just try to see what's actually happening here. Because the other thing is unlocking the creativity. Because what we want to make sure that happens here is that developers are able to use whatever tools they're looking to use. And you'll note that I'm using a different tool here, one called Goose, which is another agentic harness.

[15:06](https://realalpha.app/v/PJlTf2ACCuY/?t=906) and what i've done is i've taken goose and i've asked it to connect in the background to a number of other applications via mcp and other ways to create stuff for us so really what it's going to

[15:19](https://realalpha.app/v/PJlTf2ACCuY/?t=919) **Ian Weinberg:** do here is i've asked it to make a toy robot and put it into a blender scene wow blender

[15:26](https://realalpha.app/v/PJlTf2ACCuY/?t=926) on an open model you would not really think that's possible yeah and this is a 35 billion parameter model behind it connecting in the background to blender via mcp and as we sit here having a conversation oh that's actually on me so you know what i didn't do right i didn't turn on lemonade but now i did so now that i've turned on lemonade it will go um what it will do then is produce the image then it will hand it using flux it will then hand it off to a 3d modeler and then it will punch it through to blender now by the way this is just something we're playing around with right now so it's not an end state but i wanted you to see how multi-agents are running and as they're running what they do to fill up memory as well yeah that's pretty cool and i saw you had some different models you can run nan tasks and have them all linked and do different workflows all together you can do parallel agents at the same time doing these things together yeah and so trying to figure out how you can run all like the different harnesses for running all these different models at the same time while optimizing your hardware isn't that straightforward yeah no and you know i think that the that's the what how can i say this that's the

[16:34](https://realalpha.app/v/PJlTf2ACCuY/?t=994) virtuous work of geeks right that's that's what we care about right so the the neat thing about that is that that's the fun thing about an organization like amd and how we're committed into the open source community because very often those answers built off the work of geeks right like all of us right so say we all but i'm sorry that was a battle star galactic no i'm just kidding anyone's there but the idea behind it is that you contribute yeah if you bring it in you're contributing to rock them right that's true the world as it goes today yeah and then you have different packages that someone can then another developer can install and just even just the discovery of it you don't have to go to github or follow your favorite developers and see how they're using it you get it right off the bat absolutely pretty interesting yeah so anyway this is just going to run in the background you'll see how we'll do that but this was effectively one of those standard kind of internal coding tasks.

[17:25](https://realalpha.app/v/PJlTf2ACCuY/?t=1045) One of the things that I think is so beautiful about platforms like this, platforms like ours, is that it extends the capability of the developer community even into people who didn't traditionally study coding. Right, so a world where I came up in a non-coding world where I now produce applications is so fulfilling. We'll continue monitoring right after this message from our sponsors. 11 Labs, AI that communicates at human level across every channel and modality. 11labs.io slash mts Scale your startup on Neon Millions of developers and startups have already chosen Neon for their back end Start on the free plan or get up to $100,000 in credits for your startup at neon.com slash mts Special thanks to our sponsor Kong, the AI connectivity platform Connect APIs, LLMs, agents, and systems with serious security and governance konghq.com Finally prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line.

[18:26](https://realalpha.app/v/PJlTf2ACCuY/?t=1106) MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee, billing and monetization infrastructure for the AI era. Chargebee.com. Yeah. I am interested to see, though, who do you think your target customer is? Are you seeing enterprises who want this for sovereign AI? Are you seeing just the indie developers and hackers who just want to experiment with different hardware? What are you seeing?

[19:06](https://realalpha.app/v/PJlTf2ACCuY/?t=1146) Yes. I'm so glad you asked that question. I actually came up in the security world. When you say sovereign AI, this is right up my alley, right? The idea that you can unplug.

[19:15](https://realalpha.app/v/PJlTf2ACCuY/?t=1155) **Ian Weinberg:** Right now, if I unplugged the internet, which is Wi-Fi here, I didn't have to unplug it in that manner, it would function exactly the same way.

[19:23](https://realalpha.app/v/PJlTf2ACCuY/?t=1163) You can vibe code on an airplane. You can 100% vibe code on the airplane. I have done this. With no star link. That's the important piece. No star link, yes. That's exactly the target. So the idea here is community of the interested. So from the developers who are going to get in to try to figure out the edges of local AI to many small and medium-sized businesses that are trying to answer the question, what What if I do this locally? What if I sandbox it off and I just have people access via this? We're seeing all of those use cases exploding. And the questions around sovereign AI are just really interesting.

[20:00](https://realalpha.app/v/PJlTf2ACCuY/?t=1200) questions. It's just such an interesting time to be alive in that space. So yes to all. And the message here is that if you are looking for a standalone platform that has high capability, that has a real good balance of compute across CPU, GPU, and NPU, and has that memory signature, then we take the constraints off about being able to use very complex models. Yeah, no, that's pretty exciting. If you had to guess a little bit in terms of the market size, is where are we right now on how this market is growing? Oh, okay. So I'm a lab geek who makes demos. So you're asking the wrong person. But what I will say is this.

[20:41](https://realalpha.app/v/PJlTf2ACCuY/?t=1241) The level of interest, if I measure it the way that I can measure it. Yeah, anecdotally, like what you're hearing from the developers themselves kind of thing. Yeah, for sure. So I was at Advancing AI. But if I measure CES last year to CES this past year to what we imagine was CES two years ago this past year to the next year, the levels of question and skepticism. So two years ago, if we were talking about local AI, there'd be a lot of skeptics coming up and saying, I don't believe this. Last year, there was engagement asking, how do I get this? Now the question is, how do I get more of these? So the barriers of interest, the barriers to interest are gone. And what I'm hearing now is just people doing amazing stuff with local AI? No, I mean, I can't agree more in the past year. I would actually argue in the past few months, not even the past year, like from July to now, the conversations around local AI, the conversations around wanting to own your data, own your sovereignty, just whatever it might be, you're seeing a huge influx of it. I'm actually seeing it globally, a huge rise, not just in the United States. I would actually argue maybe in the United States, of course, it's a conversation,

[21:48](https://realalpha.app/v/PJlTf2ACCuY/?t=1308) but amongst the everyday customer, they're not worried that they'll maybe lose access to some frontier models very much my european friends are kind of worried and so they want all these things at home just in the case that one day they lose access to these models or they're ultimately restricted they want to make sure that they own their intelligence and that no matter what they always have intelligence they can just plug into that that is a global political interesting question but but but what i will say this is i think what you've pointed at is balance yeah and the the really neat thing about using local AI and using frontier models because by the way it's not it's not or it's and it's 100% and right there's a whole bunch of stuff it's not or it's and that's a great way to word it it is and right and in our AMD lemonade software there is a semantic router built into it too that you can configure and it can make decisions I can tell you a few of the demos that we built at the advancing AI period when And we pointed complex reasoning tasks on it.

[22:46](https://realalpha.app/v/PJlTf2ACCuY/?t=1366) In this case, it was a general office tasks. It diverged 70% of them locally. Yeah, yeah. But those 30% really did need to go to Frontier. 100%. But that is an important distinction that 70% of it could be done locally. So imagine what happens then when all you have to do is your local compute can pay for 70% of your tasks. Super interesting what that might mean for that user, especially when it comes to their token costs. I know we're kind of at time. And what's the latest on these demos? Because I want to make sure you get a chance to show it off and close it out before we actually have to go.

[23:19](https://realalpha.app/v/PJlTf2ACCuY/?t=1399) So it's actually still working on it. All right. Because the Blender task takes some time. Of course, it's a Blender task. But we'll have to show it afterwards. We'll give a little sneak peek. But, Ian, I did want to thank you so much for coming during this launch, getting to actually see how it works. I think it's one thing to talk about it and say, hey, here's how it works, what we actually do. There's another thing to actually show it and do it live and show us how it works. And so that's been really exciting to see. I was telling Ian right before we went live, I love doing demos. This was my whole background.

[23:47](https://realalpha.app/v/PJlTf2ACCuY/?t=1427) This is how I actually got into media. I was just a good geek who loved showing off all the cool things you could build. And so that was my early entryway. So thank you for coming back on the show and showing how it's done. Thank you, Ian. And we'll be back with MTS. A huge thanks to MTS sponsors. Blitzy, autonomous software development for enterprise code bases. Ship 5x faster. Blitzy.com. AdQuick, make your brand a billboard. Out-of-home advertising as easy to scale as digital. AdQuick.com. Arena, measuring AI performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech.

[24:27](https://realalpha.app/v/PJlTf2ACCuY/?t=1467) The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech. Visit GetVCX.com for more info. That's GetVCX.com. Carefully consider the investment material before investing, including objectives, risks, charges, and expenses. This and other information can be found in the Innovations Fund prospectus at getvcx.com.

[25:02](https://realalpha.app/v/PJlTf2ACCuY/?t=1502) This is a paid sponsorship.