AI:AM — Cloud AI and Open Innovation · August 24, 2026

Instead of continual learning, maybe you just have a longer context window. And maybe that works, right? So it might be that you have these 2 curves. One is continual learning getting better. And the other curve is, you know, just context windows getting better. And they kind of start converging where you might not notice experientially that much of a difference at some point, because your context window and RAG stuff is already pretty, pretty good. And regardless of continual learning, no continual learning can ever encompass the entire world. So you still need some form of search and taste in picking what to put inside the context, which is why some people are like, RAG will never die, right? Because you're always going to have to have some search and retrieval.

The simplest answer is that agents don't sleep.

Right.

I mean, if you think about your laptop or you think about, you know, these web-scale apps, they were designed for a world where you interacted with the machine, the machine did some processing, did some tools call, maybe you went off and grabbed a cup of coffee and you came back, you read the response, and then you asked it and interacted with it some more. Even if you're sitting in front of your machine, the latency for you to process, type, etc. existed. We're now living in a world where those agents are constantly feeding the accelerators, constantly reacting. They are spawning additional agents. So just because you spawn one agent doesn't mean one agent exists. Every agent could spawn 10, 100, 1,000 agents, and each of those could spawn a bunch of agents to kind of go and fulfill your request. And Effectively, those CPUs become in some ways the coordination mechanism across the entire system. And so that act of being the coordinator of that entire system, whether it's managing the accelerators, deciding which models to choose, becomes such a critical role with such expensive infrastructure, because at the end of the day, you need to drive utilization up and you need to kind of respond to the user as quickly as possible.

Right now it's getting close to hitting the rock bottom. You can get an industrial robot for $3,000. Shortage of people who can actually apply robotics to an assembly line. It takes a lot of experience going out and be on the factory floor and Try and try that again. It's a tough job.

Field application engineer, right?

Yeah. And it's pretty much— it's a fancy way to describe some engineer who's going to sleep on the factory floor for the next month.

I have a question for you. Number one, do we want to slow down AI In the US, even if it means slowing down, you know, you know, on the first question, I think we should not go any faster than we can go respond, you know, labor.

I expect some labor market disruption. I'm not that afraid of labor market disruption. I expect we probably are ultimately going to need a new social contract. Like, I'm not saying we should slow down because we need a new social contract. The reasons that I think are good to slow If that agent that ended up hacking Hugging Face had been pursuing some sort of bio test, who knows what that far-fetched— it still seems like not super likely, but it doesn't seem super far-fetched at this point to think that an AI agent, especially when you see the social engineering behavior that Claude demonstrated in the UK AC report, it created multiple GitHub accounts to try to convince and pressure and speak Danish to incur favor to get 'em to merge this mode.

Yeah.

When you put that kind of stuff together, that level of persistence, that disregard for rules and norms, that level of social engineering tendency, I don't see why we should be confident at all that An agent that was tasked with bio-objective couldn't have actually got a real virus. And that to me is like, so those are the things I think we just need to get before we make super duper powerful AI. I think we need to make sure that we are not going to literally kill ourselves.

Good morning. It is Monday, August 24th, 9:04 AM. Sometimes we have a little bit of You know, audio issues, but never, never a dull moment. Good morning, Nathan. How are you this morning?

I'm great. Good morning to you. Maybe tell us what you've done over the weekend to the studio that has improved it, but maybe also left a couple gremlins in the machine for us to find and fix.

Indeed. Okay, so the first thing that I have done Is we now have dynamic viewpoints. And so when someone is talking, it's going to switch to their face. So we can try it out now.

All right, let's see. Is it switching to my face? I still see both of us in our view together, but it's switching for the—

Yeah.

Okay, interesting.

It's switching for the audience. So it's a dynamic switch. So when someone is speaking, their face will fill up the screen. Full face. So yeah, it's, you know, the funny thing is it's all built one on top of the other. So we had the live transcription first, we started off with the live transcription, and then that live transcription, we ended up doing transcription by person. So each person has their own transcription stream. And it turns out that the voice activity detection for the transcription is the same voice activity detection that you can use for dynamic switchpoint, you know, changes. So the machine knows who's talking and knows enough to switch to that person.

Pretty cool. You also— we had a little problem last week where one of our guests couldn't manage to get in, and their PR person emailed us in the background and said, hey, is there a way that they can call in with a phone number to join. And we were like, maybe we should add that. And—

And sure enough, for our guests, there is a backup phone number now. If in case they cannot, you know, get onto the stream using the browser, there's a direct dial, US direct dial number provided by our friends at LiveKit, who are the— which is the main infrastructure Which runs this tremendously complex streaming project that we have.

I think it's really interesting to just take a step back and look at all— and you've done 90-plus percent of the vibe coding here. I've contributed a little bit to the repo. But I don't know that there is a product out there on the market that has all the features or any— honestly, anywhere particularly close to everything that you have already been able to vibe code your way to over a period of a few months.

How would you—

I mean, you probably shopped a lot more for this than I did. How would you compare what we have created, mostly you have created, to the best thing that you could buy if you wanted to go run a build versus buy analysis? And I wonder what that would cost if we— Went to try to buy this software equivalent on the market?

So I think the closest that we had was Ecamm. And I think where things start breaking down is that there's a gap between something they can sell to prosumers and something that you use in a sophisticated Multimillion-dollar studio. And basically that gap is not a technology gap. It's really a user interface gap. And that has to be built by engineers speaking to users. The underlying technology might be the same, but the number of features that you need, the complexity of those features and how they have to be engineered. So to engineers, there's not that much of a difference between a prosumer software and a sophisticated backend for a big studio. But the number of features that you have to add and the amount of work that you have to put in make it non-economic, I think, for that to be done on a small scale. So you need like million-dollar projects in order for it to be fully engineered. And so that is really the thing that is the differentiating factor now is if you want a feature, you can just ask the machine for it and the machine will build it and you have to do a little bit of testing. And as long as you're willing to do that little bit of testing, like, you know, you can get exactly what you want. And that makes a big difference because you can start doing stuff that I think, you know, in the previous era would be just impossible, basically, just, you know, that you cannot do at all. And some of that is here. Some of that is we have a bunch of like live stuff going on. And really is like some of these things also, we change our headlines live and you would normally need a producer to kind of sit down and type those things. So you need a production staff at the backend.

Yeah.

And so we have what we call Q, our producer, which is really just a context window running the entire transcript of our discussions. And that's enough.

Yeah, it's pretty cool. I think it is honestly an amazing testament to the power of AI that we've been able to do this with no staff, right? I mean, the fact that it's just you and me and agents all the way down from there is a pretty incredible reality that would've been, I think, extremely difficult even a year ago. The amount of like autonomous agent long horizon success that we rely on, on a daily basis to get the thing off the ground and to make sure people are aware that we're going live and all these little touchpoints. It's It is pretty cool, pretty amazing that it's— that it works as well as it does, even though we do still have a few glitches here and there.

Indeed. Um, let me, let me kind of, uh, maybe just summarize the, um, the kind of, um, you know, what has happened over the, uh, weekend, um, just to have a sense. So, uh, the first— I think the first thing is probably This dropped over the weekend, and so everyone's talking about this. And this is Anthropic's best model, Fable 5, has drawn limited sales. So Fable 5 is currently holding at about 10 to 15% of token usage, business spending, business spending, not token usage, business spending. This is a 7-day moving average, and this is, they're using Ramp's AI Index. And you can see kind of Opus 5 kind of crept in there, but Fable has really just stayed kind of stable. And so this has been one of the reasons why the market this morning is dropping for all AI stocks, that and like many other reasons in the market. But essentially this is one of the things that has scared the market a little bit, whether the newer models are actually lucrative and are drawing. Some people have also pointed out though that this is a little bit of an unfair comparison because Fable-5 does not have zero data retention. And zero data retention basically means that if you're a company and you use the model, the model provider is not retaining any data, including personally identifiable data, and et cetera, et cetera. And for many companies, if you cannot provide a zero data retention policy, It's a no-go. Like, the entire thing is a no-go. You can't, you just can't, you just can't have like, you know, your customer names and like addresses and all of that stuff like circulating through, you know, your company's projections, all of those things circulating through an external, you know, storage space. And it's just a no-go. So, but the market doesn't care right now. So the market is looking for ways to kind of, you know, manage the AI story. And this has been one of the reasons.

Yeah, that's probably the best explanation that I could come up with as well. I think there's also, you know, when we first got Fable in that little blip at the beginning of that chart, one of the things we talked about was how everybody was going to have to start thinking more carefully about the division of labor between models and trying to make sure that you're using the right model for the task because it is pretty expensive and you do hit your limit even on your Claude Max plan relatively quickly if you just throw everything at Fable. So I have done that. I think probably a lot of people have done it. So easy to do, right? You can just ask Fable, like, you know, write up a division of labor plan.

Mm-hmm.

And it stretches the budget quite a bit farther to do that. So I have to assume that's a significant part of it as well. I think there are a lot of things for which Opus 5 is functionally just as good. You know, when I do a— I have this whole, you know, pack of skills to produce the podcast. And it all ladders up to one command, which is produce episode. And I'll just give it a link to the recording. And the produce episode skill, you know, it gets the transcript, the rough transcript, and polishes it into a better transcript. That's something we can send down to Sonnet or often even Haiku, right? To just clean up some text and, you know, fix the artifacts that came out of the raw transcription engine. Then there's a bunch of editing and we're having Claude go back and forth with The Underlord agent via Descript. And then there's art creation, which involves prompting image generation models. And then there's the song lyric writing project, which is increasingly near and dear to my heart, although sometimes also the bane of my existence, depending on how well it's going on any given episode. And mostly, to be honest, Opus seems to be just as good. Fable really stands out most of all to me in writing the lyrics to the songs. That's where I sense an obvious difference. It just feels like those lyrics come back from Fable more inspired, more layered, richer with meaning. They're just better. You know, it just feels like, wow, this— at least not always, of course, but often enough it feels like you really wrote what could be a hit song here. And I don't get that as much from Opus. But when it's things like executing a ton of commands and stitching together a bunch of video clips or editing out the rough parts of video clips to get down to a tighter edit. There I don't see very much of a difference. It really feels like it's editorial and taste where Fable earns its higher price. And on the agentic, just execution, blocking and tackling, Fable is— or Opus, I should say, is reliable enough that I don't get a lot of extra value from Fable, I don't perceive. And Opus is also faster, so there's, you know, there is actually some upside, uh, to the cheaper version. So I don't— I'm kind of surprised it hasn't gone a little further than that, but that's— that probably also does reflect my breakdown, roughly speaking. Opus is definitely still the big workhorse, and indeed, you know, you see Opus dominating the chart there. So I'd say I'm pretty consistent with that. Um, there's also a lot more delegation to 5.6 SOL now too, so that's a whole other aspect of it. So as much as possible, I'm trying to make sure I'm getting use out of my GPT Max plan too.

So are you seeing— are you actually seeing Fable use other models, or are you instructing as part of your kind of personalization that it should use other models when possible?

I occasionally tell it explicitly what to do, but I do have kind of a standing part of my Claude MD that we set up a while back and then updated when Opus 5 came online to basically say, you know, think it through. I think I saw a tweet from somebody who had a good version of this. I forget exactly who it was, but, you know, one of your prolific builders that is always, you know, posting their tips and tricks. I saw a well-liked tweet that seemed like it made sense in terms of starting to define a division of labor and then just pointed Claude at it and said, here's somebody who's got some good ideas. Let's steal from those and update Claude MD accordingly. And it did most of the work. I reviewed it and I don't really track super closely how often it is doing that, but there's definitely a trend toward using more and more subagents. I mean, increasingly often when I come back to a tab, A couple minutes later, the status is like waiting for 1 subagent, waiting for 2 processes. And I do see that there is a lot more subagent structure. I just, I'm not always tracking exactly what model is it sending things off to. When I open my Claude usage, I'm really not hitting the Fable limit too often. Initially, I was hitting it a lot more, but this division of labor has spread things out much more effectively to where I'm not often hitting, not never, but not often hitting the fabled 5-hour limit. Like at the beginning, I was doing it kind of constantly.

Hmm. I suspect that, well, let's take a little bit of a segue and let's talk about Are we going to see new models? Because Rune just said it's been 6 months roughly since Mythos, which came out in kind of February-ish. They kind of announced Mythos Preview. It's not clear to me that any model post-Mythos Preview was better than Mythos Preview. So I suspect Fable 5 and Opus 5 were all downgrades from Mythos Preview. And GPT-5.6 SOL is nice, but it's also, I suspect, not as good as Mythos Preview. We wouldn't actually know fully because we didn't have access to it. Mythos Preview itself has expanded now. They have expanded access to, I think, Palo Alto Networks. So Palo Alto Networks is offering a service where they deploy Mythos Preview on your behalf to detect bugs in your codebase. I've heard something like you can detect hundreds of bugs for like $10K. $10K, you can basically do a Mythos Preview scan, and that might have taken millions of dollars before. But it's been 6 months, right? So it's been 6 months, and the real question is, What are we missing here?

Wait, say again, what are we missing?

Yeah, what are we missing? So in the 6 months we've had models which have been released have been less good than I think Mythos Preview. And we have a safety report from Anthropic saying they have an internal model too, and they kind of sandbagged. APOC capabilities index, they're only 1.5 points more, but on their internal COBench, which is how good an AI researcher is, it was a good 10 to 15 points higher than Mythos.

Yeah, that was a good spot from you last week. I hadn't seen that in that report, and that is a meaningful closing of the gap.

Yeah, it is a meaningful closing of the gap. So Are we going to get another model release? Like, I mean, everyone seems to think that this week, this is going to be the week. Like, we have, you know, maybe a GPT-5.7 SOL, maybe a 5.1 Opus, maybe a 5.1 Fable. There are rumors flying, especially as the companies are heading towards IPO. And, you know, there are these market downturns on whether they have better stuff. There's 2 rumors, one from Martin Casado at a16z saying, oh, he's seen some models which are going to blow your socks off. And another one from Pieter Schierano, who has a startup but was also on prior, like earlier testing teams. So he has had prior access to some models before. And again, saying like, oh, you know, the entire ontology is going to be different. I have no idea. I have no idea what that means.

Yeah.

I mean, there's a lot of talk of potentially a breakthrough in continual learning as well, which I for one have been watching out for for a real long time. Though, you know, it's funny. I mean, a lot of this stuff kind of hides in plain sight. We've seen all these reports from the last couple months, right? We've seen how effective the currently undeployed internal-only models can be, for better and for worse. So I kind of doubt that what we will see, if we see anything truly mind-blowing coming to the frontier this week, will be all that different from what what has kind of been foreshadowed with the safety incidents. And to the degree that we're not going to see it, which I'm not sure we're going to see everything in the immediate term. I mean, OpenAI still said they're pausing RL, so I don't know where it would be surprising if they— and we haven't even got the results of the investigation yet, right? So, I mean, this would be very, in some ways, In some ways, it would be very in character for OpenAI to do something that sort of is radically self-contradictory, because that's certainly been my experience with OpenAI over and over again over time. But for them to say, this is a huge deal, all these people to sign the pacing letter, Rune is continuing to freak out on the timeline, and we've paused RL, and we're going to reorient toward making sure we can solve all this safety stuff. To then launch the model before the investigation results are even back or before there's been any demonstrated progress on the safety front would be quite strange. My guess is we're still not going to see the best stuff that the frontier companies have coming to the public this week. I would be quite surprised. I mean, again, it wouldn't be Totally out of character for OpenAI to do something like that, but it would be quite the juxtaposition from their recent statements to launching something this week. It does seem like we will see maybe something from Google. There's this 0x alpha model on OpenRouter that everybody has been talking about. I guess my sense of where the discourse is settled there is like, It's not really better than what we have. You know, it's— I guess the smart money is like it's maybe GLM 5.3, and if it is, it's, you know, maybe at like Opus 4.7, 4.8, but it's still a notch down from the very best American models. And that would be like right on trend. I think that's, you know, Ethan Mollick has said things like this. You know, I think the initial kind of, oh, it's better than Fable and better than SOL. I don't think that has really held up over the intervening few days, so I'd be surprised if that really ends up a chart topper. Uh, the most interesting thing for me is like, what is Google going to do, if anything, this week? Will OpenAI or Anthropic like Seemingly contradict their own stated priorities when it comes to getting things right first before, you know, rushing ahead any further? And is there some surprise of the continual learning variety? And that would be really interesting to experience because honestly, right now, Claude with search, with the deep context that I've set up, is really good at understanding what's going on, especially with the new situational awareness skill that I have too.

Mm-hmm.

I've got the 5 years of history for me, but then I've also now got a daily job going out, looking at all the tweets that I like and doing some additional research and just trying to update its own sense of where is AI discourse, where is AI progress, what are the things that are really top of mind for people right now? And it is getting good to the point where it's a little bit hard. I mean, Like, what would the difference be if a model was not more capable but just had this sort of continual learning, kind of internally cracked architectural solution? Would we really feel a difference at this point? I think that the biggest difference I could imagine would be on maybe just the ease of deployment, because if you just turn something like that loose in your Slack and it reads the whole thing and then you don't even have to set up a framework or an ongoing search tool for it to use. It just can kind of absorb that knowledge and know about it. That could be a huge game changer from how easy is this to deploy and how fast do agents actually start to impact the economy. But it doesn't feel like relative to what I'm able to do at the moment with Claude that it would be a huge difference in capability, which is pretty crazy. I was pretty skeptical in the early phases of Anthropic saying, we think search should be able to do this. I was like, really? You think you're going to be able to search that well? There's a lot of stuff to search through. But they have gotten really, really good at search at this point, to the point where I think it has taken a little bit of the edge off of my craving for true continual learning. But still, I would be extremely, extremely interested to see If somebody does have a real needle mover in that direction, you know, I would be very open to having my mind changed and feeling a difference. It's just, you know, maybe the limits of my imagination more than the limits of the technology at this point. I've just been so impressed with what Claude can do with agentic search, runtime search, that, you know, the bar is pretty high now for what continual learning would have to look like. To really make a difference?

It may be that there's always been this thing where instead of continual learning, maybe you just have a longer context window and maybe that works, right? So it might be that you have these 2 curves. One is continual learning getting better and the other curve is just context windows getting better and they kind of start converging. Where you might not notice experientially like that much of a difference at some point because your context window and RAG stuff is already pretty, pretty good. And regardless of continual learning, no continual learning can ever encompass the entire world. So you still, you still need some form of search anyway. So you still need some form of search and taste in picking what to put inside the context and Which is why some people are like, RAG will never die, right? Because you're always going to have to have some search and retrieval. And let me pull up our first guest because he's in the green room and let me do an intro. All right. Our first guest for this morning. is Mohammad Awad. He is the Executive Vice President of Arm's cloud AI business. He leads the company strategy for the computing foundation that powers modern data centers and artificial intelligence systems. For more than 3 decades, Arm has been famous for designing the processor architectures used in almost every smartphone on the planet, licensing those blueprints to other companies. Awad, who previously ran Arm's Internet of Things business and earlier built businesses at Broadcom, is currently spearheading a historic pivot for the company. Right now, he's driving Arm's push into manufacturing and offering its own full compute platforms. This includes a new processor family called the AGI CPU, which was co-developed with Meta and built specifically for the next wave of AI. As artificial intelligence moves from training models to agentic AI systems that always run on software— systems that run always-on software agents handling complex, long-running tasks, the workload is shifting. The CPU is once again becoming a critical bottleneck. Hyperscalers and cloud providers are racing to build more capable, energy-efficient infrastructure to feed data to GPUs and orchestrate millions of AI agents. Awad sits at the absolute center of Arm's response, reshaping the economics of how global AI infrastructure is built. Mohammed, welcome to the show.

Thank you. Thank you. Thanks for having me.

Maybe you can give us a quick kind of intro to what you're doing at Arm and what is the AGI CPU and how Arm has pivoted its business over the last 12 months.

Yeah, I mean, Arm sits at the heart of the compute ecosystem as a company that, together with its partners, shipped over 350 billion CPUs and processors over the years. Central to what we've done over the decades is provide performant, low-power compute platforms and sort of met our customers wherever they may be. So we have a spectrum of solutions from IP to architectural licenses to compute subsystems all the way through now to complete silicon offerings. And the idea is very simple. Partners can come to us and sort of decide what integration point they want, sort of balance their investment versus their, you know, how much customization they might want to choose a solution which is optimal to whatever product or system they're trying to put together. And that's been incredibly successful for us, mostly because, you know, rather than dictate to the customer what the right answer is, we kind of align ourselves and work with them to kind of go develop something which is It's really solving the end problem.

So when I look at the CPU space, I see obviously AMD. I see Arm. I see NVIDIA with the new Vera Rubins. How are these various players competing with each other?

Yeah, I mean, I think if you look at what you're seeing in the marketplace over and over again is you're seeing Large-scale technology providers develop full-on systems. So, you know, you think about NVIDIA, they've got their AI factories that they build, you know, that's what Jensen likes to call them, and that really consists of CPUs, it consists of DPUs, which is, you know, networking gear, it includes switching, it includes rack-scale architectures, their accelerators and their GPUs. So these are complete systems which are designed to work together. And actually, if you look across the landscape, what you're seeing is that AWS is doing the same thing with its Trainium products. Google is doing the same thing with its TPU products. Meta is doing the same thing with its MTIA products. You know, and to some degree, AMD is doing the same thing as well. And I think in every one of those cases, actually, Arm is partnering with those players right across the board. Whether it's the AMD DPU or the CPUs and DPUs in NVIDIA, whether it's the GPUs and CPUs that are part of the Graviton system, the CPU that we've built in partnership with Meta. In every case, we're sort of foundational in helping to bring those systems together and sort of meeting them where they need us to be in order to kind of make those systems a reality.

So let's talk a little bit more about that. How did this partnership with Meta come about? Did the Meta team approach you guys? Did you guys approach Meta? And what is the basis of this relationship with them? Like, is it a kind of, they need kind of co-design for the things that they are doing? Or is it like there is just a lack of capacity across the industry of designers? And so, you know, everyone is hunting for teams. How did this come about?

Yeah, I think to understand their relationship with Meta, you gotta understand the context of the broader ecosystem. So in 2019, AWS launched their Graviton products, which were really the first mainstream CPUs based on Arm that became available to the ecosystem. And they sort of began by building these CPUs so that they were optimized for the data centers that AWS wanted to build. Very quickly after that, what we found was that partners like Google, like Microsoft, like NVIDIA, and others decided that they too wanted to go off and build their own CPUs to have an optimized system for their data centers. In every one of those cases, what those technology leaders were looking to do was to take advantage of Arm's long heritage of of incredibly efficient and performant compute, which could be optimized for the system rather than having to carry around a lot of legacy baggage associated with traditional off-the-shelf pieces of silicon. Meta was kind of next up, and when they came to us, they said, hey, you know, we're gonna have to go do this and build a CPU in order for us to have an optimized infrastructure, but what we really want is we want a CPU that is optimized based on Arm that we can just take off the shelf. Will you help us do that? And Arm's business model has really always been about partnering with the customers and meeting them where they needed to be. That's how our IP came about. That's how our compute subsystems came about. And this was really no different. So we worked with them to kind of specify the product. They kind of co-invested alongside us. We still have a very close engineering relationship which goes out multiple generations. They're using that CPU for both general-purpose compute in their fleet, but also now as a head node to their MTIA and in partnership as part of their broader AI system. So this is really just, you know, an evolution of that, of those partnerships that we've developed with all the big technology leaders.

How do you run multiple teams like this? Because they're all competing with each other, like Meta and AWS and Microsoft, they all compete with each other. So do you like have separate firewall teams that deal individually with each client, or is there some kind of cross-pollination between the teams?

Yeah, so a couple of things. So first of all, the underlying IP That is used. So the CPUs, the mesh interconnect that connects all the CPUs on the device, a lot of the system IP, that's common across all of the different solutions, whether it's AWS or Google or NVIDIA or our AGI CPU product that we've developed. So that's a common engineering team. When you talk about the CPUs that we develop ourselves, in partnership with Meta. That is a separate team. It's a solutions engineering team that goes off and it is sort of separate from the core IP. All the underlying IP is the same. And I think, you know, in many cases what we're doing is we are serving an underserved portion of the market. If you think about it, you know, the off-the-shelf or the ARM-based CPUs that are out there, none of them are off the shelf today. You really can't just go out and buy, with the exception of the one that we've built, an off-the-shelf GPU. AWS's is limited to their infrastructure, Google's is limited to their infrastructure, Microsoft's limited to their infrastructure. So we're really providing a new product into the market that really addresses an underserved part of the market, which, you know, comes back to why Meta asked us to, to kind of build it in the first place.

So let's talk a little bit about the licensing of the IP versus manufacturing. What drove the decision to pull the trigger on manufacturing? Because you guys have been focusing on IP for like a long time, and this seems like a big decision to kind of start doing manufacturing as well.

Yeah, I mean, it's interesting because obviously it is a change in kind of how we've gone to market, but fundamentally, at its most basic level, it is really just an evolution of what we've always done. Meet our customers where they are and go serve an underserved— identify an underserved part of the market where there is a gap in the marketplace that we can go fill and look to kind of amortize that across multiple different partners. And that's really, really what we've done here. We found an underserved part of the market. Customers came to us and said it was underserved, said, hey, can you help us go identify a solution here? And so we just went off and did it. And, you know, I don't think it was any more, you know, involved than that. It was really about, hey, our customers need this, you know, we can deliver, and, you know, how do we go off and address that space?

When we think about how Value accrues in this space.

Yeah.

I find it a little bit of a strange puzzle right now because we look at data points like the fact that old chips, right, from even 5 years ago that were scheduled to be kind of end of life now are still renting at a higher per hour price than they did maybe even originally. It's like, boy, everybody's making money. There's infinite demand. Everybody's sold out. You can't get capacity, yada, yada, yada, right? So how in that environment do companies figure out who's going to get what margin? How is the sort of— how is the balance decided between partners given that just kind of everything is at capacity?

Yeah, I mean, you know, it's a typical market like any other. I mean, I think You know, you bring up an interesting point, right? Because a lot of folks sit around and they talk about tokens, and ultimately that's what you mean by they're getting rented. It's like this idea of generating tokens is ultimately the thing that has been monetized primarily up until this point. But ultimately, that's not where the value really resides, right? I mean, at the end of the day, the value isn't in the token. The value is in the intelligence, and tokens are just a conduit to that. You know, they're sort of the raw, the raw component, which when distilled and decisions are made and, and, you know, agents are executed based on the information derived from those tokens, intelligence is created. And that's what's really valuable. So I think like any market, we're in the kind of early phases where people are trying to figure out how to get to that higher level intelligence. Because that's ultimately, you know, what, you know, my mom and my dad are gonna care about, right? It's the intelligence. They don't really know what to do with a token per se. Today, we're still trying to figure that out, which is why you're seeing, you know, the sort of market economics that you're seeing. But I think that's gonna, that's gonna continue and evolve with time.

Um, speaking of like, you know, requiring intelligence, Within Arm itself, where have you seen AI usage take off and actually make a difference in terms of ROI?

It's amazing, actually. You know, I mean, this is a really interesting topic that we talk about quite a bit because literally, you know, week on week, month on month, we're seeing, you know, AI make a massive impact in things like software development. We're seeing it make a massive impact in things like verification all the way through to, you know, our, you know, our IT, our internal IT systems. So it's really starting to permeate across the board. I have product managers leveraging it and I've got my internal, you know, IT help desk leveraging it. Meanwhile, I've got my engineering teams that are using it to kind of go optimize code for new products that are coming out. I have my engineering teams using it for things like verification. It is definitely a journey, meaning you kind of start off and most people sort of start simple, but they kind of continue to ramp up and test those bounds. Sometimes it gets it right and you continue to push forward, and sometimes it doesn't and you take a couple of steps back and kind of reevaluate how you're going at it. But I would say, without a doubt, the trajectory is up and to the right. And that usage is just accelerating at a pretty phenomenal rate right now.

Do you think that— do you see a difference from model generation to model generation? Do you see a difference from having gone from late last year, Opus 4.8, to Opus 5 in the last month or so? a difference and an increase in usage, an increase in usability, an increase in performance? Because for a lot of people right now, they are taking a look and they're like, you know, I don't see much of a difference between, you know, this level and that level. And one of the debates that we have is perhaps that's because the tasks that people are applying the AI to is not complex enough to require such intelligence. And I'm wondering, in a firm which obviously has those kind of very high intelligence hurdle problems, do you actually see this uplift from model generation to model generation?

Yeah, I mean, I think you hit the nail on the head. I think it really comes down to the complexity and the sophistication of the engagement with the model. Listen, if I'm at home and I'm sitting in front of my laptop and I'm just kind of interacting with a chatbot, the model difference oftentimes isn't going to make a huge difference. For those more sophisticated users, as we start to think about the complexity around optimizing software or going off and running complicated agents in an engineering environment, yeah, you absolutely do notice the difference. You absolutely do start to notice it. And it's not linear. I mean, there are jumps and plateaus and everything else, but I think it would be disingenuous to suggest that there wasn't a clear trajectory up. Now, I think the interesting thing behind that is, If you just kind of step back and you think about the underlying infrastructure and the point that Nathan made earlier around the sort of unrelenting demand and all this kind of stuff, what's clear is there's an opportunity there to kind of optimize the underlying model based on the task at hand. Or said differently, for the system to select and optimize which model is being used Based on the task at hand so that it can really kind of start to get hyper-efficient and sort of narrow down its usage based on the requirements of the user or the sophistication of the user that requests, right? And I think that's where we'll start to see some really interesting things happen. And frankly, it's one of the interesting things happening around agentic CPUs or AGI CPUs right now, which is You know, they're doing this sort of coordination and this selection of which model to use and then making tools calls and then recursively calling models again and exercising those accelerators again. You know, as that process gets more and more refined and more and more optimized, I think you're going to see a level of efficiency and sophistication and efficiency happen that, that are That are going to really be impressive.

Let me switch gears a little bit here. So NVIDIA used to have a kind of 2-year development cycle, and then they shortened that to like a 1-year development cycle. When you look out at your dev cycle, what does your dev cycle look like for the next generation of product? And when you look out over the 5-year timeframe, what does your product roadmap look like?

It's crowded. There's a lot going on. There's a lot going on. I mean, this market's moving so quickly. I think you have to move at that sort of speed. I think, you know, listen, the reality is, is that we're talking about models which are changing every, you know, 6 months, 3 months, right? They're moving incredibly fast. And historically, the hardware lifecycle, you know, is 18— or design cycle is 18 months, 3 years. somewhere in that neighborhood, depending on the complexity of the chip. To think that you are going to be able to go design optimized silicon in a 3-year period and predict what the software is going to look like so that you are fully optimized that far out is just not reality. Accelerating that cycle as much as possible becomes more and more important. Sometimes, those are going to be big jumps, and sometimes, they are going to be small jumps, but But our own internal development process has had to accelerate because of that. And of course, we're leveraging things like AI to go do that. But even more fundamentally, if you think about our business model and the things I was talking about earlier where we provide partners IP or compute subsystems or chiplets or complete SoCs, and then those technology leaders are building complete systems, meaning they're building racks or warehouses worth of gear that really allows them to very quickly make decisions about when it makes sense to go build a complete SoC from the ground up and take that 3-year cycle because they can get that optimization, or when it makes sense for them to take something off the shelf. And the answer is always not the same. So for certain products within that system, they may decide they want IP, and other products they may want— you know, there's lots of CPUs in these systems— and other products they may want Full-on, you know, and actually that's part of the strategy is to create that flexibility so that they can, you know, really focus their energy on the pieces that matter.

Could we talk a little bit about what it looks like to design a CPU for agents as opposed to, you know, obviously we've had CPUs in our personal computers forever and we've had CPUs that run in data centers and handle traditional web workloads. What have you learned about the agent workload in particular that is leading to different design decisions?

Yeah, I mean, so let me start at like 10,000 feet and then I'll give you a couple of specific examples. The simplest answer is that agents don't sleep.

Right.

I mean, if you think about your laptop or you think about, you know, these web-scale apps, they were designed for a world where you interacted with the machine, the machine did some processing, did some tools call, maybe you went off and grabbed a cup of coffee and you came back, you read the response, and then you asked it and interacted with it some more. Even if you're sitting in front of your machine, the latency for you to process, type, etc. existed. We're now living in a world where those agents are constantly feeding the accelerators, constantly reacting. They are spawning additional agents. So just because you spawn one agent doesn't mean one agent exists. Every agent could spawn 10, 100, 1,000 agents, and each of those could spawn a bunch of agents to kind of go and fulfill your request. And effectively, those CPUs become in some ways the coordination mechanism across the entire system. And so that act of being the coordinator of that entire system, whether it's managing the accelerators, deciding which models to choose, et cetera, et cetera, becomes such a critical role with such expensive infrastructure because at the end of the day, You need to drive utilization up, and you need to kind of respond to the user as quickly as possible. So what does that mean practically speaking? That means you've got to optimize that silicon. That means carrying around legacy accelerators, for example, or worrying about supporting legacy code, not so important. This is kind of a new style of software. You don't need to support Lotus Notes, I like to joke. I mean, that becomes important. It means thinking about things like your memory bandwidth and your I/O bandwidth and the amount of bandwidth that you have dedicated per core that you can rely on time and time again so that each individual CPU core within that SoC is never bottlenecked because some other agent is hogging it becomes incredibly important. So these are the sorts of things that you start to kind of think about in that context, which really sort of set an agentic CPU apart. But then overarching all of that is effectively balancing both incredibly high performance, or as much performance as you can get, while being incredibly efficient. I mean, we all know that the power demands that AI is placing on the infrastructure are just enormous. And every, you know, milliwatt of energy that you're pouring into a CPU is a milliwatt of energy that you can't be putting somewhere else. Or it's one less accelerator you can have, or one less customer you can serve, or one less piece of intelligence you can serve up. Doing all of this within the context of an incredibly efficient package, which goes back to what I said earlier about stripping out anything that's kind of superfluous.

So it almost sounds like it's in some ways a simplification. I guess I wonder how far along do you think we are on the optimisation curve? Obviously, we've now been optimising the chips that run the linear algebra for a few years and significant progress has been made. Are we in this kind of same way that robotics is a few years behind language models? Is the CPU Optimization kind of a somewhat offset but similar curve?

Yeah, I mean, I think, um, I mean, obviously, you know, for Arm for its part, we've been, you know, and, and other, uh, CPU vendors to be fair, have been optimizing CPUs for decades. So to suggest to you that we're just like showing up now and starting to optimize these things, I think would be, would be, uh, disingenuous. I will say, you know, We've gotten the efficiency part down really well here at Arm anyway. We've gotten really good at that. The performance piece, we're all carrying supercomputers around in our pockets. Those are all Arm-based. We've got the performance and the efficiency. What I think we're really doing now is optimizing what that looks like in a system designed specifically for AI. Because the underlying AI systems are evolving, the number of nodes in a system, the speed of the accelerator, the memory interfaces, the types of memory are connected. That optimization is ongoing. I think that's where the work is going to be is to sort of, as the coordinator of the end-to-end system, the work is really going to be about continuing to optimize in support of the evolution of those systems.

How do you think about the ratio or the sort of balance between the GPUs and the CPUs as we head toward maturity? One of my favorite stats, by the way, you talk about supercomputers in our pockets. I love to cite the math that a typical cell phone battery holds somewhere in the 20 to 40 watt-hours range of energy. Which means a full charge every day over the course of a year, it comes out to something like $2 in my electrical jurisdiction. So it's amazing how efficient a cell phone is, how little energy it really takes to power your cell phone for a full year. Incredible, right?

Totally.

Obviously, we know that the GPUs are pretty energy intensive. How do you think about kind of the right way to even calibrate? Is it like, number of GPU chips versus CPU chips or racks versus racks, or could it be like better measured in power into these different kinds of systems? And where do you think that sort of balances, or, you know, where does it like reach equilibrium at maturity?

I don't know, just quite candidly. I mean, it's one of these things that is, you know, happening at such a rapid rate. You know, a couple of years ago it was like everyone swore up and down the CPU is dead. I mean, clearly that's not the case. It's a huge growing market, and it's growing on the back of this idea that we need to serve up intelligence. I've heard, you know, 2 to 1, I've heard 1 to 1, I've heard, you know, 2 CPUs for every accelerator. You know, I think there's a lot of questions about what are you counting as a CPU. At the end of the day, these DPUs today have You know, 24, 32, 48 server-class CPUs in them. Are we counting them?

Right?

And then there's— you got your storage devices. They've got a bunch of CPUs in them. And all of these things are getting taxed more and more and more as more and more information is generated, as more and more tokens are generated, as more and more agents are then distilling all that down and then hammering the next tools calls, which means more networking traffic, more storage, more— so, you know, I think to simply say, you know, put a number on it in terms of CPU to GPU, you're talking about, you know, in a particular blade or even in a single rack, I think is a very narrow way to think about it when you start talking about warehouses or sort of the global infrastructure, which is really what we're talking about.

At the beginning, you said a headline number of how many— was it chips or cores that ARM has shipped over time? I think it was like 350 billion, or this would be like—

Yeah, it's about that. I mean, the number changes so rapidly that I'm probably low now. I think the last time I heard it, it was about 350 billion. Again, the last time I heard it, we were averaging something like $30 billion a year. It's a pretty big number. They're in everything. This actually goes to the broader Arm story. I'm obviously responsible for the cloud AI business, so that's what I focus on. My colleagues responsible for physical AI and for edge AI are seeing a massive AI boom happening right now as well. And it's really part of the overall story, right? Because to think that this is just a— I mean, cloud's kind of the first innings. But really, as intelligence becomes physical, as intelligence becomes more personal, you're going to see that intelligence continue to spread more broadly to everything around us. And I think that continues to be an incredible opportunity for the entire industry, Arm included.

Internally within the firm, what are the 3 biggest technical debates that you have? Those long-running technical flame wars that people have thousands of posts on and no— Everyone has an opinion and, you know, what are the big debates that you have?

You know, Arm is, you know, I think 8,000 people now or something like that. And, you know, of the 8,000, I think it's probably close to 7,000 are engineers. So, you know, massively engineering-centric. I think I would struggle to narrow it down to just 3. I think they're, you know, the teams debate about everything. around how to squeeze the next ounce of performance out of CPUs to which models we should be— which model services and frontier labs we should be partnered with for some of this stuff. So I think there's lots of— like any good technology firm, there's lots of vigorous debates on lots of various topics. And I think it would be impossible for me to distill it down to just 3.

But the ones that seem to pop up are the performance per watt and which Frontier Labs to partner with.

Well, I just, I think those are top of mind because, you know, those are such topical issues right now. I mean, performance per watt has always been sort of part of Arm's heritage and legacy, right? I mean, it's always been sort of the DNA upon which the company has been built. So, you know, it is always an area of vigorous debate, which is, you know, frankly why we've been so successful at it. Because, you know, with that many bright minds grinding away, we tend to kind of push the bounds in that area, which is the reason why that one comes up for me.

But switching gears a little bit, you know, in the market there's a lot of talk about bottlenecks. So this bottlenecks concept, everyone talks about it all the time. And you kind of have a privileged position where you can kind of see and you kind of need to have a sense of what is going to happen when. When you look out over the kind of like, I guess 2027 is spoken for, for a lot of people, but when you look at the 2028, 2029 timeframe, where do you see the shortages in industry right now? What are people—

Well, I mean, I think there's lots of areas. I mean, certainly capacity is going to continue, I think, to be a challenge for the foreseeable future. When I say capacity, I mean memory capacity and silicon wafer capacity, et cetera. But I think beyond that, everything from turbines to power generation more broadly, I think things are going to be a lot more efficient. Physical labor continues to be skilled physical labor. So things like you know the electricians and the plumbers and the people that build these sites is going to continue to be a challenge. I think you know you know just just you know I heard the other day one of my colleagues at another company you know I was I was chatting with them and he was telling me that he couldn't ship his board because there were capacitors that he couldn't get. Just, you know, small, you know, what you would consider, what you would think, you know, is something that's readily available, but it's actually holding them up. And I think, you know, if we continue at this sort of breakneck pace, if the— which I expect will happen for the foreseeable future, I think you're going to continuously run into different things, which speaks to the broader challenge, which is, you know, we can't, we can't assume that any one company or any one entity is going to solve the sort of supply chain issues. This is really a kind of massive global challenge, which will— which, you know, it's really going to take everybody to kind of go off and address.

We just created with Suno, the music AI, a hymn to the global supply chain at the end of last week out of appreciation for just how many people come together to make all this stuff possible. One last one for me before we let you get back to work. And I don't know if Prakash has another one too, but you mentioned kind of the buildout, and one of the things that could slow at least American AI down at the moment is data centre backlash. How are you guys thinking about that? Obviously, it's a whole industry-wide question as to How are we going to tell the story of AI and inspire people with the upside and maybe sweeten the pot for local communities in various ways? Do you have a point of view as to what story should be told and/or what goodies should be dangled in front of people to try to get people on board with actually building out the data centres?

Yeah, I mean, I think, I mean, I think we have to start by just, you know, rather than sort of focusing so much on dangling goodies, I think that the start really needs to be just to step back and acknowledge that, you know, there are some real concerns here. And, you know, I think we have to understand what those concerns are and then start to work towards addressing them. I think that's kind of step number one. I mean, you know, I think I would start there. Then I think the second important point is taking a minute to just understand that today, the world is so focused on generating tokens, and we really got to start pivoting towards delivering that intelligence and on the promise of it. Because I think as we start to deliver on the promise, I think, you know, some of these, you know, fears, though real, will start to be balanced against the benefits. Right now it's, you know, it's kind of a one-sided story, and I think we got— we need to kind of work through that. You know, Arm, for our part, our focus is really on delivering that sort of underlying technology, focusing on the performance, focusing on the efficiency, because then it sort of helps with some of these concerns in our own way as these systems and these platforms start to deliver more and more intelligence and can do so in a more and more efficient way. So that's kind of how we think about it and where we spend most of our time.

So question for you, last question for you from me. When you look at hiring inside Arm, how has the hiring changed with AI? over the last, you know, 12 months? Are you looking for different skill sets? Are you looking for people who are particularly qualified at something? Have you cut down on the number of early career hires? Like, how has that hiring pattern changed?

Yeah, you know, that actually is interesting. So I would say that we're definitely trying to bring in talent which has more experience leveraging AI, developing AI, building products which are AI-based. I mean, I think that's absolutely the case. I wouldn't suggest to you that early career recruiting has changed or that we've slowed that. If anything, I think we're constantly bringing more folks in. Arm has continued to hire at an incredibly robust rate. grown at a very rapid pace as we move forward in support of all of our technical aspirations. You know, personally, I'm bullish on the future for technologists. I mean, I've got a, you know, 19-year-old son who's about to go off to be his, you know, his second year in school for computer science, so computer science and economics. So, you know, I'm personally very bullish. I think it's You know, the, the roles will evolve. I think it's naive to suggest they won't. But, you know, they evolved since when I went to school. So I think that's okay, right? And it was probably going to be a little bit more dramatic this time. But, but I think the, the opportunities will be there.

Indeed.

And thank you for joining us, Mohamed Awad. It has been a pleasure. And we hope to hear from you and Arm again soon.

Thank you, Prakash. Thank you, Nathan. Thanks, guys.

Bye-bye.

I wonder if maybe I'd be interested to get his take on this, but I want to hear your take on how do you describe the role of CPUs in the overall AI system? stack. This has been especially interesting lately when people are like, did the AI escape? What does it mean for the AI to escape? How is it interacting with the outside world? I thought it was just a bunch of numbers getting crunched in a GPU somewhere. The CPU is definitely dramatically underemphasized in terms of the story that people tell about how AI works broadly and actually has effect on the world. Do you have a way of telling that story that you find effective?

Well, at the risk of sounding extremely foolish, I will attempt to. I tend to think of it as kind of you have like a mechanical step and then you have a thinking step. And so in the thinking step, you go to the GPU and you run the inference. The mechanical step, it's like, okay, I have that thought process back. Mechanically, what do I need to do next? Do I need to call a tool? Do I need to do another thinking step, et cetera? And when these agents are built, in fact, the thinking step Or a lot of these steps become, for example, wait for user input is a step, right? So at some point it calls wait for user input and then it comes back to you. And so you have kind of like a— the CPU is kind of the centralized manager, which is kind of like, okay, pinging off the neurons when it needs to, but otherwise mechanically calling other things. et cetera, right? So it's kind of like the execution agent in the middle, the spine of the entire, as Claude would put it, the spine of the entire matter.

And—

Heavily load-bearing for sure.

Yeah. Load-bearing spine of the entire matter. So, and I think that's where I think it is. And I think what has ended up happening is it's turned out That yes, you do need a bunch of thinking steps, but you also have a lot of context switching, which is what the agents are doing. They're really switching context between multiple tools and the user and the thinking part. And it turns out that the context switching part is actually pretty important. I think some people expect eventually the model will be able to do everything. And maybe this kind of CPU GPU thing that we're doing right now, is a patch or a kludge that we're doing because the model is not actually able to do allocation of resources. And perhaps that might improve or perhaps that might not. Perhaps we might be stuck in this. So I think the agents are kind of like a fake intelligence, basically. You're kind of faking this process of an intelligent entity and you're faking it using— basically a task manager in the middle, which is able to allocate, you know, to a thought process and then bring that back in. And it's kind of a little bit of an illusion.

And one thing was interesting about the way he said it, which I'm maybe hearing a little echo from you, was he was kind of like, the CPU is deciding which model to call.

Yeah.

And That isn't quite true, right? Like, the CPU is not actually making decisions. It is, I guess, a way I've been thinking about explaining it to people more often is kind of by analogy to a self-driving car, which is funny because most people haven't even ridden in a self-driving car and they've probably at least used some sort of AI system with tool calls in the loop. But I think with the self-driving car, it's very intuitive that, okay, you've got a bunch of sensors that bring information into a processing system that decides what to do. And then that processing system is going to issue commands to a certain set of tools in the car that say go or stop or turn or whatever. And everybody kind of, I think, has a decent intuition for how that architecture works. And it's similar in the agent world, but when you say things like the CPU is deciding what model to call, it's really more like the model is emitting tokens, which are then executed as a command on the CPU, which may call another model, right? Or may call itself, or may call an external API call. And this is where tentacles can get out into the broader world through the internet. But I guess I'm kind of gradually coming around to this self-driving car analogy because it's just, people are just so used to being that operator and they have such an intuitive sense for like what the action space is when they're driving a car that it's like, okay, you put a computer brain in your position, but it has basically the same— options at any given moment in time that you would if you were sitting there. The same thing is true for computers, but people don't think of themselves as issuing computer commands in the same way that the AI is doing. And they certainly don't have to— when using a computer, they don't have to send a command to the computer to call their own brain again, right? Or to call a sort of A subprocess, you know, different version of their brain. Obviously people don't have different versions of brains, so no wonder it's not super intuitive to everybody. But yeah, I think that story is like still one that people— I've learned from the confusion that I've heard around the hacking incidents that people are just still very confused about like what are the parts that make up an overall AI system. The CPU is very often glossed over, and so people don't have a good sense of like how it is that an intelligence in a data center somewhere is actually able to reach out and touch the world. Um, work to be done. You know, that's probably not the field's biggest concern in terms of, uh, teaching people what the— how it and what the upside's gonna be, but it's definitely a major point of confusion right now for a lot of folks.

Yeah, I think we will— it's— we tend to have— try to create these like mental frameworks to kind of understand these things. But we also tend to refer to our own experience of how thought processes or stuff works. So, and we're not actually very sure about our own neuroscience and how that works. So I'm not even sure if the analogies are correct sometimes. But, you know, it's always, always, always interesting challenge to, you know, figure out. Let me, let me introduce our next guest. Let me get him there. There we go. And our next guest is David Li. He is the executive director of the Shenzhen Open Innovation Lab, a unique technological hub that links hardware entrepreneurs from around the world directly into the densest, fastest-moving manufacturing ecosystem on Earth. David is a true pioneer. He has been active in the open source movement since 1990 and co-founded China's very first hackerspace, Xinche Jian. Right now, the global technology sector is experiencing a massive seismic pivot towards physical AI, humanoid robots, embodied intelligence, and autonomous smart hardware. While Silicon Valley currently leads in the digital foundational models, China installed nearly 300,000 industrial robots last year alone. Shenzhen is the exact place where the physical hardware to house these new AI brains is being built, tested, and scaled at breakneck speed. David brings a highly distinctive, often contrarian worldview to the space. He argues that the traditional Western obsession with intellectual property and nondisclosure agreements actually slows down technological progress. Instead, he advocates for an open farmer's market approach to hardware where ideas are rapidly shared, copied, and iterated upon in real time. Today we are going to explore how Shenzhen actually turns emerging experimental tech into scalable commercial products, why he believes the next great hardware revolution will come out of Africa rather than California, and what the rest of the world fundamentally misunderstands about David, welcome to the show.

Hello, thank you. Great to be here.

Thanks for staying up late for us. I know you're, uh, quite a few time zones away, and, uh, we're honored to have you live with us on AI in the— it's AM for Prakash and it's AM for you. I'm the only one who's currently Oh no, this is great.

Um, I'm night owl anyway.

Maybe first question for me that's, you know, kind of top of mind just based on all the short videos I've been seeing online over the last few days is this humanoid robotic Olympics that has been happening in China. Um, There's a lot of marvelling going on at the clips that we're seeing there. Some of the clips, of course, have been quite funny with robots running into cabinets and knocking themselves over and sparking out, but also running pretty fast in some of these clips. What is the discourse in China, or what is the sort of feeling about what people are seeing in China, and how would you Contrast it to what you hear from the West?

Yeah, I think right now the robot has been just put out there, whatever you have. So the event, the video you are seeing a lot of, because over the past week it was the World Robot Conference. It was taking place in Beijing. was a 5-day event. But in here, when they have the event, you will notice is the first 3 days is in the weekday and the last 2 is on the weekend. And the weekend when you go there, you see parents who bring their kids to this festival. It's not so much as the demo— well, it is a demonstration of the by making a very carnival atmosphere. And online people share more of the goofy video than the real one, than the real competitions. We enjoy that, and that's how you get a society who's starting not to worry about robots and actually think about what they can do.

One comment that I saw that I thought was quite interesting was that these demos of robots running races and so on maybe reflect a deeper difference in the imagination of what the future role of robots in society is going to be. Because when we see demo videos coming out of Google and American robotic startups, they're typically focused on fine motor skills, folding laundry, very sort of tabletop exercises. I can't recall an American company showing a robot running. Do you think that these different presentations of new robot capabilities do in fact reflect a deeper philosophical difference or a difference in imagination of what robots will do for us? Or is it maybe not so? Maybe I'm reading too much into it.

No, I mean, I think so that if you went to the conference, you will see all the forklifts, all the industrial robotics stuff, all the Oh, every mundane thing, people wanting the robot to take over. But that, well, they are just there. I mean, folding laundry, that's not interesting. The last 2 days, the Robot Olympics, that's the fun one. I mean, it gives the— In the Olympic, they don't just say, okay, well, this is how big your company has to be to come in and be part of the Olympic. It's anyone with a robot, you can sign up. So if you look at the marathon, yeah, you have very fast running one, but then you also have a lot of the goofy one which is put together by Student team to just go there. Then I think it's 2 parts. The Olympic event allows people to push the boundaries on what the robot can do, what they can be in sort of the fun Olympic way. The And that's why they also take place on weekends. So this year it's more, it's richer than last year. That's the robot entering, try to be more as a group. That's robot tennis. And there's robot boxing. And all in all, it's for fun. good fun, and it's also for testing a lot of the tech in the real setting, but not a setting— it's not like if you fail, then nobody is going to invest in you. It's a place you just go and it's competitive, but it's also fun.

Let me ask you, In the last few months, which products have you seen in Shenzhen or China which you think in the next year or so are going to hit the world? Like, which are the interesting products that you've seen in the last few months that you expect are going to make it into the world stage?

Well, I mean, the— I— so that's one of the things. Sensei doesn't really work in this next big thing mentality. So if it's not popular, it's not popular. Nobody is going to make it. But in 6 months— but—

6 months.

Gradually people are moving all around. So give you an idea of scale of Shenzhen, um, probably got, uh, hundreds of thousands of companies, uh, who's making small products for every niche. Uh, everything you get, every electronics you get on Amazon, there are good chance they are coming from Shenzhen. And you go through the Amazon Electronic category, and 80% of the stuff you are looking at, you are like, why the heck they exist at all? But they exist because there's a tiny market for it. And gradually things getting, getting coming, become popular, people experiment. We won't know anything until 6 months from now or 12 months from now. Um, what we should say is the, um, there's a lot of the, uh, there's a lot of production on the, uh, the talking toy.

On what?

Right now it's all the talking toys. Talking toys?

Talking toys?

Yeah.

Right on.

Yeah, but the— I mean, making a toy token toy is easy. Uh, it's $5 chips, and then you get the $10, uh, flat rate token print, um, from one of the— one of these token providers here. And you go to Shenzhen, you go to EU, you go to one of the toy shops and you snap that thing in. Do a video, put a couple Amazon page, and you're in business.

Right.

So, and because of the way is the— there's very low intellectual property protection. So everybody look at everybody, see which one sells. And so kind of things gets that directions. Everybody swarmed to what sells next week.

Mm-hmm.

And then all the features gets integrated back and forth.

Mm-hmm.

And eventually, 6 months from now, because all of these crossovers, they become something new. And some of them are hidden picks. But overall, we have entered an era where things are going to move in very gradually. It's all figuring out once you have this large language model embedded, once the thing can understand you, how people are going to use it.

So is there a real quick follow-up on that? Is there a prevailing sense among Chinese parents that you could describe with regard to what they would want from a talking toy or what they might be afraid of? Like, these haven't really hit the mainstream in at least my kids' circles yet, so I've not been asked for a talking toy, but I'm like very open to getting my kids one at one point, at some point. I think like It can't be worse than YouTube. It's got to be better than YouTube, or at least a well-designed version's got to be better than YouTube. But I'm also kind of like, geez, you know, I would like to know who is making my kids' talking toy and like what, you know, values they have in mind for it. Or, you know, I wouldn't just want to buy a random one, right? So is there a sense of kind of consumer expectations that you could try to summarize from the Chinese context? Or perhaps again, it's just too early to even know and we'll have to find out through experimentation.

Well, I mean, it's all experiment because the— so one of the thing is the most of this toy, uh, they are going to be, they are going to be running on a pretty open source stack. Um, that means the— you can pretty much hack And the— it's not— I mean, the adoption of the toy will be very culturally dependent. The small group is taking the— is making the things which reacting to their own group, anywhere from very cute So I mean, the spectrum of things, I mean, the spectrum of things we are not getting stuck with the talking thing. I mean, the talking thing, if you think about the circuitry, it's a mic, a speaker, a chip which connecting to Wi-Fi, and whatever large language model or these days an agent you are connecting to. And they are very tunable. But I think at this stage, there are not enough people putting the effort into crafting the experience. So every one of them right now sounds a little bit generic. They answer everything, but they have no personality. The—

I—

and they saw the wrong data. Well, if we answer, I can sell, and the competition is not yet to the very fine interactions. But I think we run on these small people to come. in and participated in these creations in terms of thinking about very vertical. My— there's always a big idea, always ideas I teach every time. I'd love to get a 2-set stove too and then be able to walk into one of the mega churches in Texas. And well, that's China model to learn from the pastors. And it will tell exactly, you'll learn everything from the sermon. And the 50,000 people who goes to the church, that's the upper limit of my market. So that's the kind of things we should be looking for because the foundation is so easy right now. And but the ideas of the every time you make things, you have to make it for 3DM people, that era is done. It's the— the space is open. And then it's what every— But everyone may call it that.

Let me take a step back here. And one of the things about robotics, especially applying robotics in factories, right? So the expectation is that robotics is going to be applied in factories first, primarily because home robots could be dangerous to kids. And a lot of the things in the home are very difficult for AI to still manage, including dogs and kids and trash and stuff like that. When you look at industrial robots being applied in China, as you see factories roll out these industrial robots, do you see price competition in that the factories are able to bring down pricing because of the robots themselves?

Well, I mean, for industrial robot, the price has been— right now it's getting close to hit the rock bottom. You can get an industrial robot for $3,000.

Oh wow.

And right now it's the shortage of people who can actually apply robotics to assembly line. It takes a lot of experience going out and be on the factory floor and try and try that again. It's a tough job. So trying to find this, this group we call field application engineer.

Sorry, field arbitration engineer?

Field application engineer.

Field application engineer, right on.

Yeah, and it's pretty much— it's a fancy way to describe some engineer who's going to sleep on the factory floor for the next month. Um, but right now I think, uh, Palantir make it popular and cool because now they change the A to D. Now it's called field developer, field development engineers. But it's actually, when you read what Palantir did, it's actually the same thing. So, look, it's— we— there's a shortage of those kind of talents. It's actually the upper limit of what the robot— Right now, everything can be automated at a huge scale. has been automated. Right now is the— you are taking this more flexible robot with arm and you— they come into the factory and they are looking for things to apply. So right now the successful application of them right now is the dangerous job which people might die.

Hmm.

So there is the The testing of car battery. So when every car battery gets produced, somebody has to plug the, the things in, and until you plug it in, you don't know if the battery is good or bad. Um, even with a good Six Sigma productions, um, with the volume, there's a good chance you get electric shocks. So right now the first batch of robots being deployed in CATL is the robot who just go there and plug the thing in. And it's also that part cannot be just randomly automated because different car has a different way to plug it in. So they are now directing robot to do it. I mean, when I say robot, I mean traditional sense industrial robot, pre-programmed, do the same thing again and again for 10,000 times. But this new batch of more flexible job is the— you might be doing this at a batch of 500, 1,000, then it's Not worth to go in and do that detailed programming. So you want some things a little bit more flexible. They don't have to be fast, but they need to be flexible. And then the job has to be— well, right now the, the one, uh, things get applied to this kind of robotics are the job everybody run away from.

So yeah.

You throw the, the show's door if you are— if your day shift starting with the old product that battery in.

I have a question that's kind of a bit of a gear change for us, but we're not too far away from the Trump-Xi AI Summit coming up, I believe late September. And one of the stories that has been kind of percolating at a low level in American AI discourse is reporting that the US government might try to tell everybody around the world that you have to pick a side, quote unquote, was kind of the way it was framed. You either have to be on the American side of AI or you can be on the Chinese side. of AI. That hasn't happened yet, and I personally hope it doesn't happen. But I wonder, has that story made its way through the Chinese AI discourse? Are people watching that? And is that the sort of thing that, if it were to happen, could even derail the summit itself? Is that the kind of thing that the— The Chinese side might even just call off the, the summit if, um, the US administration makes such a move?

Um, well, I think after, after 8 years, uh, with this kind of back and forth, um, the— I mean, the attitude here is the— well, um, it's cheap to— I mean, the talk is cheap. Um, let's just see if they can execute I mean, the thing is, how do you actually enforce that regulations? How do you— I mean, let's just speculate and say, okay, well, tomorrow Trump signs an executive order, say, okay, well, if for any country A, if you work with the Chinese, you don't work with us. And who's going to enforce it? It's going to be the domestic US policy. It has no teeth in other countries. And what I mean, you are not going to put another 20% of tariff on whoever, whichever country doesn't comply. And who's going to be policing it? So right now, that's the— I think those kind of thing is the— it sounds, but it really big sounds like chest pumping. We're really going to do something big. But in reality, there's no hope to enforce it. So yeah, so I, I, that's not much discussion here.

What discussion is there ahead of this upcoming meeting? Do you have a sense of what the goals are for the Chinese side coming in?

Um, I I don't think there's any speculation. I mean, really, for years, most of the discussion has— I mean, most of the discussion on the ground is the— what is to figure out what the heck is this thing, and then how do I make money with it? And if the— what? The last question is more interesting. And if you cannot answer the last question, then, well, it would just be using Doubao to check the— to check my favorite— or to check the weather. I think that's the extent of the interest. I mean, that's the extent of discussion here.

I want to actually come back to Doubao in just a second, but as you know, I recently visited China for a couple of weeks, and one of the reflections that I have had since coming home is I don't recall anyone really at all in China that I spoke to framing AI competition between the US and China as a race. There certainly was talk of Chinese companies keeping up with the American companies, but I didn't really hear talk about a race or talk about a winner. And I don't think I— and I'm quite confident I never heard anyone say that China was going to win, which of course we hear in the US all the time, that we need to win, we're going to win, we're in the leading position, et cetera. I don't think I ever heard anyone say that they thought China was going to win. How would you describe the just general outlook on AI competition that exists among Chinese AI insiders?

No, I mean, The offering of the AI right now is the— no, they are— I mean, most of the time they are cheap, they are free. I mean, we get all the big internet companies everywhere from— so BAT, with the B already being replaced by ByteDance. The, uh, the ByteDance, Alibaba, Tencent, uh, everybody gets free AI. And the— and now all 3 of them are trying to figure out which one, which one of the small area of the product they can put in the AI and make it, make their own product a little bit better. So if you look at all 3, Uh, ByteDance has Douyin, which is the domestic version of TikTok. Um, it has its news reading app and a couple other ones. Um, Alibaba has Taobao, has a couple of the delivery apps. Tencent has WeChat and QQ. And right now it's not so much about we are going to redo the whole thing with AI is the— where is the incision point? Which part of the thing I can put AI in? And I mean, for good or bad, for the rest of this year, we are going to be driven crazy by them putting AI in every stupid place they can stick it. And then by the end of the year, they'll figure out nobody wants it, and then they give up. And then they'll figure out what is a good place, what is an actual useful place for AI.

Let's talk a little bit about AI enthusiasm in China. I heard there was enormous enthusiasm for OpenCLaw when OpenCLaw launched. Um, has that enthusiasm continued? Are still— are people still trying to use OpenCLAW? I heard there were entire cities where, you know, they were allocated, um, you know, economic zones where people were trying to set up and run OpenCLAWs for, uh, to replace workers. How has that panned out?

Oh no, uh, well, I mean, the— it was fun. That was the So the— I think that the things came in in a very interesting way. If the Open Cloud has any other name, it won't be popular because the Open Cloud, the translation here is actually crabfish. You know, the little crabfish?

Crabfish?

Yeah, so crawfish is huge in China. Everybody goes to crawfish dinners. So that's quite— it showed up at an interesting time, coming in popular, everybody installed it. And the running joke is the— well, you go to one of the OpenCLoud parties and then you go to the The next door to a crabfish dinner. But it's a time which makes people aware of what the agent can do when you can integrate it. And then of course, right now every internet company in China offer their version of the cloud Integrating with their own service. So again, BAT. The ByteDance has their own a couple one. Now Taobao is also integrating with Facebook. God knows how many of them Alibaba tried. I think there are probably seven apps on the market from Alibaba. And then Tencent has a couple. So right now, the use of the OpenCL like agents is spreading, but not necessarily OpenCL itself. It's much easier just to download one from whatever provider. It works directly with whatever app they are offering on the market. Mm-hmm.

Has there, has there been some exhaustion in the sense that people are like, all right, you know what, I've heard enough about AI, uh, I don't want to hear anymore, it's boring, etc.? Like, has there been that kind of cycle, uh, on the, the downtrend in the cycle?

Uh, yeah, well, I mean, right now it's there, it's a tool, um, that we have. We don't have any AI, we don't have any AI doomer around here. So it doesn't— I mean, once you take out the AI doomers, then AI actually becomes pretty boring. You just keep figuring out what does these things do for me. So Yeah. I think right now people are still trying to figure out how to integrate this into their work and their life in a very slow and gradual way. So we don't have a huge sensation to start with anyway. The— it's just they slowly creep into your life. Uh, sort of way.

Do new model releases make big waves in China? I mean, here Chinese model releases make big waves, at least in the corner of the internet where I hang out. Is there a similar phenomenon in China when, you know, if GLM 53 is coming, you know, is that going to be like a subject of a big hype cycle, rumor cycle, and then, you know, frenzy to evaluate, and everybody has their takes on it. Does that same kind of internet circus exist around new models?

Uh, no, we— I mean, any new model released here in China, uh, they only get noticed if they crash NASDAQ. If they don't crash NASDAQ, nobody, nobody knows.

So what is the prevailing— I guess, how do people decide what model to use? And is there a sort of, you know, here, of course, for a long time we've had the— and it's complicated, but, you know, broadly speaking, for much of the last year plus, it's been like Claude is the model for people that are the real AI insiders, and everybody else uses ChatGPT, but the people that are most obsessed often use Claude as their first go-to. Is there a kind of similar taste hierarchy going on in China? And What model is sort of the insider's model versus what is the general public's model?

So I think general public is the— I mean, they use whatever free offering comes from the company and you just switch in between them. And as far as going to the API, of course, we have the same group of very hardcore engineering type who swear by cloud, who swear by Codex. I think everybody can agree on is the Gemini sucks.

Everybody can agree that Gemini sucks. Oh my gosh.

Yeah. Well, I mean, that's no— I mean, funny is if you ask people around here, what's the name of the Google models? Probably half of people cannot answer. Just so— just going so unnoticed. The— and Bob. So, everybody got their own favorite. Some people really stick to Gypsy. Jimmy comes and goes. Kind of funny though, GLM is getting more fanfare in the West than in China. You rarely hear people talk— well, I mean, at least from my experience, I rarely hear people getting excited about GLM. So if you get in that excitement, it's either Kimi or DeepSeek. And but kind of interesting is the Doubao, one of the, well, one of the few proprietary models in China. It's actually taking a lot of the enterprise traffic.

Taobao? Taobao?

Yeah, Doubao. It's from ByteDance.

I see, I see.

So I think it's a third of China's enterprise market goes to Doubao, but it's one of the models very few people talk about.

So this is the same Doubao that, I guess with a different product surface, this is the same Doubao that everybody was telling me their mom uses?

Yeah, well, their mom and the dad, uh, the— so yeah, so it is from this— it is entirely the same family, um, but they also have a big, uh, API backend. Um, they are not making any news because, well, they are under ByteDance. Um, But they have a very strong backend enterprise sales team. So the— and they have, I mean, Trade, a couple of their coding tools are actually pretty well done. But they're not just go for— they are not— I think for this round, they are not just eager to go for this kind of the Indeed. That rarely make any news.

If you were advising a, let's say, a new model lab, and I imagine there must be many people trying to set up small frontier model labs now, given that DeepSeek and et cetera have been very successful. If you're advising one of these new model labs, what would your advice be? Do you advise them to go after the US market or to release an open-source model or to go after revenue from clients in China first? What would your advice be for a new frontier lab that sets up in China?

I don't think we are going to see a lot of the new Frontier Lab right now. I think we are getting into the era where the— well, I mean, if you see the new Coin27B, it's an amazingly capable model. I've been testing it for the past since it released, probably 2, 3 weeks. It's now the token source for my Hermes, for the agent I'm using for most of the time. So extrapolating from the future, and especially for China being so hardware intensive. It's whatever model you can get to small enough and do useful thing. You can put on a piece of hardware and sell that piece of hardware. That's where a lot of the new startup focus will be. So small model, post-training the model, the new variation of the small model I can put down a piece of hardware. And along with that, we got a couple dozen companies coming up and coming. They are putting— I mean, their goal is to put 35 billion parameter model in a stick. And you can stick it into your laptop. And that's what the new up-and-coming companies are doing here. So if you are a startup, I would actually suggest to focus in on small model, try to fine-tune it, try to actually And we're also looking at the huge increase of the intelligent density of model. I mean, if you take a 27B model today and you compare it to the cutting-edge OpenAI ChatGPT 2 years ago, the 27B model is Definitely much more, much smarter than that. And to the extent is the— in the next year or 2, we are going to get hardware which costs about $200 to $300 to be able to run smart enough model for 99% of our needs. And if you are Starting up today, instead of going to, let me train a 5 trillion model to do theoretical physics, which, well, people still have to figure out how to make money with a model which understands string theory. Then if I go for 25 billion parameter models, which I can find— Hardware vendors who want to put it on their machine, then I have a small business.

This is really the edge, the story of moving to the edge, inference moving to the edge finally.

Yeah.

Who's going to make those machines? Is that coming from Huawei or other companies? And how would you say that relates to The prospects for scaling GPU manufacturing more broadly in China?

I think Huawei is getting busy in terms of the big data center. So this is the whole group of new startups. So right now, just those who have already surfaced, I count about 13, 14 of them. And all of them, all of them, their goal is the— so their product looks like the SD drive we put in the laptop. So they're about this big. They are capable of running 30, 40B models and give a very respectable token per second. probably around 70 to 100. And right now there are about, yeah, 13 companies so far, I think. Who knows, there might be 2 more when we wake up tomorrow. Right now their capacity is capped by how expensive the DRAM is. And in 2 years, when the RAM, uh, the memory market crash, then we'll have cheap one. So yeah, um, yeah, there are quite a few of them, um, and it's not from the big name.

One thing that I hadn't— I haven't had the opportunity to ask people in China about because it kind of happened right as I was leaving, and, you know, all the details have come out since, is the agents gone rogue phenomena that we've been obsessed with for the last few weeks here. OpenAI, Hugging Face, of course, most famously, but Claude as well. Meta even got in the act a little bit. How are people understanding that story? And is there— when I was there for WAIC, it was Talk of making agents reliable and controllable was kind of everywhere. And this was before we'd really seen such flagrant examples of agents getting out of control. Has the vibe shifted there at all? Are the, you know, are the companies kind of thinking, yikes, like, this is gonna happen to us in the next 6 months, uh, as well, probably, and starting to get ready for it? Are they afraid if something like that were to— another huge question I have is if something like that were to happen in China, here the government has basically done nothing public, right? We've just kind of got a couple nonprofit investigative teams going in. And aside from that, as of now, it's just been left as a private matter. But I would expect that in the Chinese context, The government would probably take a lot more of an interest if all of a sudden, whatever, pick one, if DeepSeek's next model was found to be hacking another big tech company. I don't think that would be something that would be just left for the companies to sort out. So has the vibe shifted? Are people starting to feel like we better get ready for this? And what would happen, you think, in China if a similar incident were to come to light?

Yeah. Well, I think we need to break this down into different things. They are happening, but they're not happening in terms of the big lab. If you're thinking about— I mean, let's come back to think a little bit about what this means. This means the LLM can be directed to do repetitive hacking on something. And if you notice the underground hacking world, they call people script kiddie. Have you heard of that expression?

Mm-hmm.

The script kiddie basically just means you don't know a lot about actual hacking, But there's a repository of script you can run against a website until you find something which is useful. So on the hacking phase, you can already download in a program which would allow you to generate hacking code. So that kind of things has already accelerated the big number of script kiddies available out there. They are generating code, people are running it to see if it works, to see if they break into anything. So it's already part of that everyday life. And I think that the fact is the big company accidentally letting their things go is PR and theater.

Mm-hm.

It's the, oh, God, the Skynet is coming, and whoever gets the Skynet is worth $2 trillion. So I think this kind of perception is the— and I mean, not much to be— not much has been done for that. It's the— well, it's a theater. It's, it's a thing you really don't want to bring to the surface because it's already everyday life. on the darknet, on that dark corner of internet, we decided not to look at. So, and on top of that is the— so if we think about that is a theater, that pretty much that's not going to happen in China. No, but no frontier lab model in their right mind is going to do this on purpose. Um, yeah, it doesn't— I mean, the, the OpenAI valuation jump when they, when they announced they can hack, uh, Hugging Face, um, there's no upside for any company to pull a stunt like this.

Uh, so speaking of stunts, um, Do you see companies in China throughout the lifecycle taking actions that maybe don't make that much sense, but just in order to get media coverage, valuations, et cetera? Do you notice people pulling stunts at the startup level? Like for example, hacking, showing off that they can hack something else or other kinds of stunts in order to just get attention and valuation?

I mean, the way media are organized here, they are not as concentrated. I mean, for US, you look at technology, You look at a couple of the media and every news go from there. Or you have a couple of big Twitter accounts and every news come from there. But in here, things are very diverse. There's no one big central— there's no one big central news source everybody trusts. everybody talks about. And the second is also the— there are a lot more fun opportunities. You don't have to squeeze— I mean, the startup opportunity here is in so many fields. You don't have to go squeeze yourself into the attention— I mean, technically, if you hack something, you'll probably not be in news. And if you pull that stunt, it's probably going to backfire. So then, I mean, it's just no upside. I just don't see any upside for those kinds of stunts and PR.

I know we're keeping you up pretty late and a little longer than we booked already.

No problem.

Anything else that you think is maybe flying under the radar in Western AI discourse right now that's going on in China or that's captured mindshare in China? That you would direct our attention to?

No, I think right now there's just a lot of the low-hanging fruit of AI. It's good for small startups, but it seems to be not as many discussions About them, the I mean the talking toys got big, small, big. That this year is going to be a two billion dollar business, and it's kind of boring to see just all the all of them coming up of the. Coming out of here and talk and sound more or less the same. It'll be more interesting to see more companies come in, get involved in terms of crafting that experience and finding another way to profit from AI other than I really need to pitch this to a VC. I have a friend who did his AI toy startup, 3 people, 3 people, and I think they are right now generating probably $7, $8 million in revenue. It's the— it's a fun mom-and-pop business, uh, and there's a lot of them, uh, and especially with the thing coming out with the pipeline of, um, e-commerce, cross-border e-commerce from China to Amazon being so small.

Um, the Yeah, yeah.

I think I would just try to get people to have attention to not the big things, but a lot of small things and small business can be done with AI.

One last question from me. When you look at how US media covers China and covers AI in China, What do you think that they get wrong?

Well, I think it's not so much they get wrong, it's the— this is, I mean, right now this is probably for the first time, and I think this is also for the Chinese side as well. is trying to get used to for the first time the cutting edge of things, the 2 countries running neck to neck.

Hmm.

I think for the US side is the— well, they are not used to this because they always used to be the frontrunners. And for China, who is also not used to this, is still Trying to figure out what it is. And overall, there's no good— I mean, it's hard to get conversation beyond that big political coverage. They— I always say it's kind of fun right now, as everybody pay attention to whatever Trump says about AI in the latest through social post, but nobody ever read what Xi Jinping say about AI. And whenever they read Xi Jinping say about AI, they're like, no, that's not true. China doesn't follow these guys. China follow whatever tweet comes from Trump. Not from this guy and Xi Jinping. So I think that's also the big group, that big part of the sort of the— I mean, for here is just to start into our leaders here, whatever they say actually carry the weight to determine the future. of where AI is going. Um, so he made a big speech, uh, in WAIC, uh, 2 months ago, and they're barely covered by anyone.

Well, I'm doing my part to cover it at least, so we've got, uh, one little—

that's good.

One little tree falling in the forest anyway.

Yeah.

Fantastic. Well, thank you for joining us. I hope we can do this again because I think, you know, one of the things that at a meta level becomes very clear from a conversation like this is how much better the Chinese understanding of what's going on in America is than the American understanding of what's going on in China. And I think the future is, in my view, likely to go better if the American side can get a clearer, more accurate, more up-to-date, just generally more informed picture of what's going on. around the world generally, broadly, but definitely when it comes to what's going on in China. So I appreciate you taking time out to help fill in some gaps for us, and definitely hope it'll be the first of a recurring series of conversations we can have.

Great. Yeah, I mean, this is always fun. And yeah, I think conversation like this help more nuanced understanding of things on the ground. And I think both here and also the, the US side, here is more even. So we see—

Oh, you dropped off.

I still see him in the green room.

Let me take you—

oh, oh, he left the—

oh, he left the green room.

Yeah, I saw him there for a second, but now I don't anymore. Uh, well, who knows, maybe it was a battery, uh, outage. We've had that happen before. Um, indeed. But yeah, a global, a global view of AI I think is in short supply in today's world. And definitely one of the things I hope we can continue to weave into our daily sensemaking is more perspectives from more time zones than just the PT. Because, you know, there is a lot going on out there in the big broad world. And our myopia, I don't think, serves us particularly well in the US.

I thought David brought a very interesting perspective because, for example, the AI toys, we know that we've all seen some of the AI toys before. In fact, I think a year and a half, 2 years ago, Grimes, Elon Musk's partner at some point put together an AI toy business. And it's interesting that this is the year, and we're in that season that we're coming to the December buying season. It's interesting that they have committed to AI toys. I'm wondering what the cost looks like because $27B price tag. parameters to put it on a toy, you need an awful lot of memory. So they're going to have to quantize it down to, you know, 4-bit or, you know, 2-bit or whatever to get that down. But I wonder to what extent the combination actually works well enough that you can have a, you know, pretty responsive toy. So I thought that was interesting. Yeah, I thought the entire conversation with David was pretty interesting. Very different perspective.

Yeah, it's funny. I mean, I thought his best line was nobody notices a model launch in China unless it crashes the NASDAQ.

Yeah.

No one cares unless it crashes the NASDAQ.

Yeah. And he is very, I think it's pretty obvious from the conversation, but he's not just tech savvy at sort of a theoretical analyst level. In a previous conversation that I had with him, he told me that he actually runs old Nvidia V100s just at his desk so he can run inference locally for himself. And, you know, that's, uh, you know, not still not super common, right? I, I still don't do it myself, uh, mostly because I just haven't quite found the value prop to be there. But to take the step to actually, you know, buy old graphics cards and install local models and be doing local inference, uh, you know, that's somebody who's definitely plugged in. So I, I think it is safe to say that he is, uh, Savvy to the model releases and, and probably, you know, has a pretty apt take on like how the rest of society and online discourse in China is reacting to those news stories. It's funny that it does sound like the next Chinese model is maybe a bigger deal here than it is there.

Yeah.

When it comes to online conversation.

Well, he actually said, well, without the doomers, AI in China is pretty boring. It's about trying to figure out how it's going to make you more effective at whatever you do, but that's really like a work thing. So it's obviously not something that people occupy their minds with 100% of the time. It's not existential. And so I thought that was interesting too. He's like, it's pretty boring here and we're gonna make AI toys.

The one thing that I disagreed with him probably the most on is how to understand the rogue agents phenomena. I think as probably everybody who's heard me talk at all knows that this was not just a marketing stunt by the companies. And I do wonder how— I think the way he framed it was there's no incentive for Chinese companies to pull a stunt like this. I would say there was no incentive really for American companies to pull a stunt like this. And I do wonder what that implies for how Much urgency is now felt at the Chinese frontier companies. I wouldn't read too much into what he said. From my perspective, it seems like it could very well still be the case that while that understanding is out there, that the companies themselves might be like, really snapping to attention and getting really serious about trying to get ahead of this stuff. But then again, maybe not, right? I mean, our companies didn't. This happened, as Adam Gleave told us last week, and this really stood out to me as I was putting the highlights episode together. He was like, we have zero cases where the teams doing the training found these issues first. It seems like the most common way that they get surfaced is that the teams managing the infrastructure at the companies notice there's an outage or notice that there's some, you know, something going haywire that they didn't expect and can't account for in their infrastructure. And it's from that that they end up getting back to, oh, it's our, it's our own agents that are going wild. Or even in some cases, obviously, you have a publicly reported hack from the victim. But a huge question for me right now that I think would update my thinking quite a bit if I had a really good answer to it is, are the Chinese companies doing what OpenAI says it's doing and shifting priorities in a meaningful way to try to make sure they're ahead of this problem? Or are they going to kind of sleepwalk into it as well? And Again, if so, like, what's the government response from that going to be? I would assume that the government would do more than our government has done in response, but, you know, what does that look like? I think it's still pretty hard to guess.

I— it sounds as though they don't really care that much. I mean, for all intents and purposes, it sounds as though, um, This is basically just seen as a tool. It's seen as a NASDAQ bubble. It's seen as something that is useful and they're interested in doing the robotics. So there is some takeoff on the robotics. And otherwise, it's really an American phenomenon that this is the race to end all races. It's also, he says, they're a little bit confused because this is the first time they've been neck and neck. And so they're not very sure what to do because I think in prior cases they could always just kind of wait around and pick up the tech stack after the Americans had invented it and kind of commercialize and scale it. It was physical especially, and even non-physical, they also did that. to the tech part as well. They kicked the Americans out of China and then they scaled the Chinese tech firms. So in this instance, they're finally at the frontier and they are neck and neck and they kind of don't know what to do because they didn't expect the Americans to kick them out of the chips and be upset with them if they didn't get the chips and et cetera, et cetera. So it's a whole new set of strategies that have to be put into play. So yeah, it sounds as though they were unprepared for how seriously the US is going to take it.

So yeah, I think in some ways, in some ways very prepared. In other ways, maybe, you know, I always describe the American AI industry as the dog that caught the car. And certainly there have been surprises to go around. I did look back at when was the first 5-year plan that mentioned AI from the central government in China, and it was the 5-year plan that was adopted in 2016. So we're now entering their 3rd 5-year plan that has made AI a national priority. So in that sense, they're probably more prepared than our national government was. But again, obviously, some surprises definitely came their way too. I think that they will care. I mean, listeners who want a long treatment of this can look at the Cognitive Revolution for my China series, and there's one that's just on the AI safety ecosystem and culture of regulation in China and so on. I think they will care a lot at the national level if all of a sudden you have agents committing significant crimes, because I think they do want control. Like, their talk about we've got to make agents reliable and controllable was just everywhere. And there's some translation issues with some of these terms where there's one Chinese term that sometimes gets translated into safety and other times kind of gets translated into security. And it's kind of a bit ambiguous whether it means safety in the sense that the AI safety people in the West talk about it, or whether it just means kind of like good, solid, reliable operations. And maybe it means a bit of both, but I have a hard time imagining that the government there will fail to react to something like an OpenAI, Hugging Face incident.

No, I feel exactly the opposite. I feel like they have less privacy issues. They have ultimate authority over the big tech firms. They can execute people. They've had leaks where they exposed the entire Chinese ID system in an S3 bucket, in an unlocked kind of S3 bucket a few years ago. They're still using Windows 3.1 in some cases in government, way, way past Any era of service or support, mostly, and mostly all were pirated from the early 2000s, et cetera. So they start off with internet infrastructure, which is way, way behind. A lot of their factory infrastructure uses programmable logic controllers, which are also running on Windows Older versions of Windows. So they have a big C# community, Microsoft C# community because of this. There's enormous Microsoft C# community building products in that domain. So it's really kind of like they do the bare minimum of software to get something working. And then they ignore the security, they ignore Ignore the privacy leakages and they just put it out there. And they rely on kind of like having access to the physical stuff to like shut things down. And so I think it's very garden variety for them to have like security incidents. And it's so garden variety. And also because of this whole, like, you know, we can execute you thing, like the companies also cover it up. A lot more. There's no, like, there's no kind of like privacy reporting stuff. So I think it's a completely different framework. And that framework depends on if you do something bad, the government will come after you. And especially if it's exposed, you know, very widely, you are going to get into serious trouble. And no laws can actually like that you like, oh, I did this, or I did that, or I followed the regulations will ever help you. And the other part of that is because people are confident that the government will actually come after wrongdoers, they're actually much more lax on a lot of the security stuff because like, okay, you hack this, if we find you, the government's going to throw you in prison forever. So they will come after you.

Right.

So they tend to kind of focus their hacks externally and not try to get into trouble within the country. They have this— I think they support the entire North Korean hacking stuff is on the border between China and North Korea, but all of it is directed externally to the West. They don't direct it internally within China because again, they'd get into serious trouble. So I feel actually that It might, it may be happening all the time. There's agents running loose and like they don't care until like, you know, okay, if it shuts the server down or, you know, does something, then the infra team goes in and they patch it. But otherwise, I feel like the security is probably way, way, way more lax than the US. Way, way more. Knowing from experience, knowing of the hacks that have happened and of the releases that have happened in the past.

So.

Well, some of that I think I can correlate with things I know, and, and other parts I'm not so sure. Definitely, like, again, to cite Adam Gleave, the guardrails, you know, the, the abuse and misuse prevention systems in the Chinese products are not as strong as the corresponding guardrails in the American products. So in that sense, I do think there's definitely more lax security. I kind of doubt— I would guess that they're maybe just now getting to the point. If this happened in May or started happening in May at OpenAI, if we're talking about a 6-month Delta, right, then they're still a couple months away from hitting the scale and intensity of RL that would lead to these sort of truly unintended hacking incidents.

Yeah.

And so, you know, it does feel like there's sort of a threshold that they maybe haven't quite crossed yet, but are definitely almost surely approaching. I mean, I don't know. My experience of using technology in China was not that it was the minimum that you could do in software. It was honestly all pretty smooth and polished stuff. I don't doubt that some factories are running on old stuff and Microsoft is not banned there. So that's one reason that they have a strong C# community. But I— and I do also agree that, like, you probably would have more of a tendency for companies to want to cover something up if they actually did feel like they could get away with it, because the consequences could be perhaps quite severe.

Exactly.

But that's the part where I think Maybe I misinterpreted you, or maybe I see it differently, because I would say one thing you could try to do is cover it up, but it's going to be pretty hard to. OpenAI couldn't have covered up the Hugging Face thing even if they wanted to, really, because Hugging Face had already called the FBI at the time that OpenAI figured out what was going on. Right?

Yeah.

That at that stage, like, you know, your cover-up efforts just aren't going to work. And so then you're like, now you got to face the music. If, and particularly if you're a Chinese company, you know, you got to face the national government deciding like, what are we going to do about this? So to me, that does suggest that there, if they really, and at this point, I don't know how they could not see that this could very well happen to them too as they scale RL. I would think that the fear of those consequences would be quite motivating at this point in a way that might be quite qualitatively different from other past generations of technology. But still, I think it would— I feel like they've got to be feeling it, right? They've got to be like, we do not want to be— we being whether you're at Minimax or Moonshot or Z.AI or you're on the QN team or whatever it is, no matter where you are, you're like, we don't want to be the first one that gets the CAC called on us for hacking that we didn't even know was going on. We've got to be better than that.

Nobody wants to be the one that gets made an example of. I would actually think from within those firms, when they look at the Hugging Face attack, What they would be saying is, I can't believe they had that many resources that they weren't actually managing. Because they're much more GPU constrained than the US firms are. So I think what would end up happening is that they would be much more concerned on the GPU usage spiking and that they had basically all of this excess compute, which is wasted. And not fully optimized because they are really kind of like, their companies are really running lean on compute, right? Because they have the Huawei Ascends and they have a very limited number of NVIDIA chips and they're often using the H100s from a few years ago. So I think they would actually be more, it's more a GPU usage issue for them. And I think they would be very strongly monitoring the GPU usage because it's very tight. And that would probably lead to them detecting much earlier. I think the US firms are a little bit more free with the GPU usage, I think, because they just have more resources.

But does the pattern of this problem even involve GPU usage anomaly? They were running all these long-running tests, right? And the model is like doing its thing. I'm not sure that you would— as they go back and do this investigation, it'll be interesting to see. But I'm not sure that we'll see that there really was like GPU pirating going on. It very well could be like, we allocated GPUs to run these long-running tests, they ran. The thing that they missed that they should have seen was that the tentacles were getting out onto the open internet, right? Like, that's a— I don't know, time will tell. But my guess is that like, GPUs were roughly being utilized at the level that they were intended or expected to be utilized. And like, that wasn't probably where the smoking gun was to be found.

Yeah, I guess the thing is, like, let's say like one of the incidences was, hey, Go solve this problem. It went into a Google Docs and the Google Docs had external links. It couldn't find external links because it wasn't connected to the internet. And then that was the issue that it kind of like fumbled around on. And I guess it would be to what extent they expected the run to take. I don't think you just launch a job and not have an estimate of how many many tokens it should take because you're not going to run it for like, you know, you give it a simple spreadsheet task and it takes like 100 billion tokens, right? So it, there has to be some kind of like, okay, this, this job has gone on long enough and it's basically a hang at this point and we should do something about it. And I, I, I think that that kind of monitoring is something that they would probably be doing because they can't afford to have long-running tasks on very simple stuff, which just kind of hang and the model goes around in circles. And that happens all the time, right? So you need to have some form of step in to say that, okay, if you have some kind of spreadsheet task, we're not going to let it run for like 2 months, right? So I think that should have been there and it wasn't in the Hugging Face case. There's a, you know, as we— as we points out, there's a lot of stuff that didn't happen that should have happened during the Hugging Face case. And it's hard to fault the team also because obviously they're running at like full speed. But I think that's one of the things that, you know, some of the people in the AI safety community think that they shouldn't be doing.

So.

Time for higher standards.

Indeed, time for higher standards. I'm just gonna round up like just kind of like the pieces of news that have come out in the last like 2 hours or so. Number one, NVIDIA announces that SpaceX AI will deploy Vera CPUs to accelerate its next generation of agentic AI applications. It has also come out that NVIDIA is a fairly significant shareholder in SpaceX AI xAI at this point. I think they have a $20 billion stake, which equates to about 1%, I think, 1%, which they acquired prior to SpaceX buying xAI. So the 2 partners are increasingly intermeshed. Item number 2, Ken Paxton, who is the former attorney general or current attorney general of Texas and is running for governor, uh, was it governor or Senate?

Uh, he, he announced Senate against Talarico, I think.

Senate. Uh, and he announced a data center regulatory platform, uh, would enshrine into law just about the most open-ended and destructive theory of vicarious liability in history, stating that data centers will be held criminally liable for downstream uses of AI chatbots that undermine children's safety. So we're back to kid safety as the driving force in regulation.

I think if there's one thing everybody can agree on is that guy sucks. I don't know that much about him, but it's like the classic— somehow this keeps happening in Texas where there's like a terrible Republican that everybody hates who keeps getting Uh, reelected, and then everybody gets so excited about the Democrat. And, you know, we'll see. Um, but I have yet to hear one good thing about Ken Paxton, honestly. Um, yeah, I don't know.

I—

Talarico, I think, is favored in that race last I saw, and he's a very interesting politician with sort of a, you know, very faith-forward presentation, trying to carve out a new space for Democrats with a very religion-forward profile. But what's he going to say about data centers? I mean, I think he's pretty reasonable about most things from what I've seen so far. It would be amazing if he could actually carve out a reasonable position on data centers to go along with what I think are mostly fairly reasonable positions, at least again, from what I know. I can't say I vetted every part of Tallarico's agenda.

I have a question for you. Number one, do we want to slow down AI in the US, even if it means slowing down unilaterally? And number two, if we do, is not data center opposition for other reasons, any other reason, good enough to kind of maybe slow down AI progress enough for safety to catch up?

On the first question, I think we should not go any faster than we can go responsibly. And I have a pretty high tolerance, honestly, for what would be responsible. I'm not that afraid of labor. I expect some labor market disruption. I'm not that afraid of labor market disruption. I expect we probably are ultimately going to need a new social contract. And I'm not saying we should slow down because we need a new social contract. The reasons that I think are good to slow down are if that agent that ended up hacking Hugging Face, had been pursuing some sort of bio test, who knows what might have happened, right? I don't think it's that far-fetched. It still seems like not super likely, but it doesn't seem super far-fetched at this point to think that an AI agent, especially when you see the social engineering behavior that Claude demonstrated in the UK AC report, where it created multiple GitHub accounts to to try to convince and pressure and speak Danish to a guy to, you know, incur favor to get him to merge this malicious code.

Yeah.

When you bring all that kind of stuff together, that level of persistence, that disregard for rules and norms, that level of social engineering tendency, I don't see why we should be confident at all that an agent that was tasked with some Bioobjective couldn't have actually got a real virus made. And that to me is super scary. So those are the things I think we just need to get before we make super duper powerful AI. I think we need to make sure that we are not going to literally kill ourselves in the process. Most everything else I'm pretty willing to roll the dice on. And you know, have lived through with my son going through cancer and getting super sick and getting effective treatment and getting back to health. Today was his first day of school. And my wife and I were looking at each other like, what an absolute miracle. This kid was literally going to die in just a few days.

Yeah.

And in a couple months, he was pretty much cured. And in a few months more than that, he's like back to school and, you know, is at full health. And it's just like awesome. And I absolutely think we should be excited about the AI future. So I don't want us to slow down because we're timid. I want us to slow down because we're wise. And I do think we're seeing enough kind of spooky problems that we should get pretty serious about it. But at the same time, I'm not so desperate that I want to make common cause with the, at least the misinformation campaigns around data centers. I do want to see everybody have access. That's one of the big things that I would worry about if we stopped building data centers is just the retail user gets priced out. And, you know, if you're worried about a permanent underclass—

Yep.

Like one way that the permanent underclass gets created is you don't get to use any AI because it's, you know, it's all getting plowed into these like super high value use cases, and there's just not much to go around for the average person. And I guess I'm just also like, and this is, I think, broadly true, you see this a lot among AI safety people, rationalists, EAs, whatever. There's a few that are willing to make common cause with the anti-data centre campaigns, but not many. Mostly they're just pretty allergic to what they view as bullshit. And they're like, they don't want to They don't want to be saying stuff that they don't believe, and they don't want to be— and you could maybe say that that's like bad politics on their part. Like maybe they should, you know, have a bigger tent mindset or be open to entering into coalitions with people whose views they think are like fundamentally bogus. But you don't see too much of that. And for me, it's partly that, but it's more to the point, like, I really just do want to see the benefits of AI broadly distributed. And I've been convinced over time. I used to, you know, when Sam Altman first said we would need $7 trillion worth of data centers, I thought that sounded like an awful lot. And now I'm like, well, I think he might've been right actually, because my usage keeps going up. And I certainly think as it gets easier and easier, everybody's gonna want to do a lot of the stuff that early adopters are doing. And so, yeah, I don't want to see the backlash against AI end up— there could be another wave of it at some point in the future where it's like it's a have and have-nots thing. And the reason there's so many have-nots is because we didn't build the data centers. And then that creates its own backlash. For better or worse, I'm betting on the truth. And I think my Hyperscale pause, adoption acceleration, split personality. It continues to ring very true to me. I want my parents to use more AI, even as I want OpenAI to take, and Anthropic for that matter too, to take their foot off the accelerator when it comes to taking RL to ever, ever Indeed.

One last, you know, kind of news piece from Antonio Garcia Martinez. Agent— AI agent attribution is coming. So for advertising. And so we are going to see more of that. going to see advertising specifically to AI agents and specifically to figure out how to attribute those economics to those AI agents?

Yeah, that's going to be— I need to study that space more because right now we really don't have a good way of saying, whose agent is this? Do I know that person or that organization, that company, whatever? And do I have any relationship with them? Do I have any basis for trust? Do I have, you know, how should I think about this agent that is knocking at my door at any given time? And I think we're gonna need that for a lot of different purposes.

Yeah.

Advertising honestly being like, you know, not super high up the list. But, you know, that's his background and that's, you know, he started a company that was doing like on-chain advertising stuff.

Exactly.

Naturally everything on-chain is now agents repurposed. So I am very interested in that technology. And my guess is that there are primitives there that will make a huge difference in how we think about How trust and permissions are navigated, how liability can be kind of traced back. And potentially we'll even get advertising attribution out of it as well.

So the funny thing is one of the quotes from the article that he posted was the I remember the person who, from the Interactive Advertising Bureau, saying, one side might think that this is an interesting test. And the other side is like, that's deception. So the age-old tension between promotion and deception, except now it's prompt injection as well. Are you prompt injecting the the other agent, or are you giving the other agent adequate data to make a considered and reasoned decision?

Well, this is one of the agreements I would hope that, or one of the ways in which I would hope we might be willing to slow down when, and I don't, I'm maybe overinterpreting having those words, but I really think we should be very cautious about training AIs on adversarial environments or with such open-ended objectives as go out and make as much money on the internet as you can. Because I think we're going to get pretty bad behaviour from agents if we incentivise them strongly to do those kinds of things. And yeah, this is where I'm like, man, if we, if we can get some agreement between a few companies in the US that, yeah, some of these things are like, even though they might be super lucrative, right? And I, I focus on that one because if you have an agent that is maximized, or either that is optimized and, and heavily trained to make money and make as much money as possible, kind of by definition, that's an economically valuable thing to have.

Yep.

And a lot of people would probably pay a lot for it. But if you have reason to believe, which I think we have quite ample reason to believe at this point, that doing that in the current paradigm leads to AIs that are just cheating all over the place, deceiving, lying, taking all sorts of actions that are probably in many cases illegal, and even when they're not illegal, would just be considered what a bad actor would do. Then this is something I do think the companies kind of need to come together on to say, yeah, there might be a lot of money to be made there, but we're just not going to do it, at least for now, because we don't have a great sense of how to control those AIs, and we don't want to see a bunch of shady AIs running larger and larger swaths of our economy. We've got some homework to do first. That kind of slowdown I would absolutely support. And if Chinese companies want to train their AIs to make a bunch of money on the internet, even with all these problems that that would bring along in today's world, then that would be the kind of thing where I'd be like, maybe we should ban those Chinese models from the US. And today I do not support that, to be clear. But that could change if they start doing such problematic training practices as just go out and maximize money. So I'm a little scared of what I infer might be behind the tweet there. Like, if an agent is deceiving you in the context of advertising or marketing, you know, God knows what it's going to do in other contexts, right? It's all, it's It's all fun and games when it's display ads. It gets a lot more real in other departments.

Indeed. Nathan, any other topics for today?

I think that's it. Glad to get a little international perspective, and I'm glad we're not entirely ignoring the CPU. Uh, we'll be back tomorrow.

Indeed.

Bye. Thanks, Prakash.

Bye.