Limitless: An AI Podcast - THIS WEEK IN AI: Nvidia Acquires HuggingFace, Leopold vs SEC, New Waymo Chip

Episode Date: August 28, 2026

We cover NVIDIA’s record earnings, strong GPU demand, and its position in the AI chip market, along with reported moves involving Hugging Face and open source AI. We also discuss Z.ai’s new model, Waymo’s autonomous vehicle hardware updates, SEC scrutiny of Leopold Aschenbrenner, and recent developments in robotics and Apple’s upcoming event.------🔒 Check Out Our Sponsor: LEDGER AGENT STACK 🔒https://developers.ledger.com/docs/ai-tools/overview/?utm_source=Audio&utm_medium=Podcasts&utm_campaign=Limitless------🌌 LIMITLESS HQ ⬇️NEWSLETTER:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------TIMESTAMPS0:00 NVIDIA’s Record Quarter5:05 Hugging Face Joins NVIDIA9:26 Cloud Margins and the Chip Play11:36 GLM 5.3 Flash Emerges15:40 Waymo’s New Hardware Leap19:48 Tesla’s CyberCab Challenge23:24 Leopold Under SEC Scrutiny25:44 Anthropic Usage Looks Flat29:27 Agents Need Better Security30:20 Robotics Learns From Video34:07 Apple’s Big September Event36:36 Sam Altman’s Rare Watches------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠Josh works with Anthropic as a contractor. All views expressed are his own and do not represent Anthropic, its leadership, or its affiliates. Nothing in this episode is investment advice.

The very same company that just got hacked by a rogue open AI model just got acquired for $13 billion by none other than Nvidia. Jensen Huang believed so strongly in owning open source models in America and he's made his biggest purchase yet. But that's not without cause. Invidia just reported their earnings and they had a record core. They earned almost $100 billion in just three months and the company is growing 100% year-over-year. This is unprecedented for the most valuable company in the world worth.

over $5 trillion, and every single bank has raised their price targets for the invidia stock. But that's not all in this week's random. We had a new secret model that revealed to be GPU's new GLM 5.3 Flash, which competes at the level of Fable and Open AIGPT 5.6, but at a fraction of the cost. We've got Leopold updates, unfortunately, of the bad kind, where he's being probed by the SEC, where they've subpoenaed a bunch of banks for accusations of insider trading. And we finally have an update on robots from the worlds of Waymo, Autonomous Driving, and Tesla, all that are more on this week's roundup. To start with the Nvidia quarterly revenue, I want to give you a sense of how outrageous

this company is. NVIDIA's quarterly revenue now exceeds the annual revenue of roughly 480 of the S&P 500. That is a crazy statistic how big it is. Invita is printing about just over a billion dollars in cash every single day that it operates. So the day you're listening to this, and V-Denra is printing a billion dollars of revenue. And that number is going to go up. They're projecting huge amounts of growth over the next quarter. And they're guiding now for 70% year-over-year growth in 2028 relative to the expectations, which were 44%. So this earnings report was about as good as you can guess. And Wall Street's kind of rewarding them now. I mean, the stock is up. It was down after trading, but it is now up 7% early market. We're looking on screen

now at all of the banks that have changed their price target after reading this earnings report, pretty much every single one of them has raised it significantly higher with the highest being $420 per share. And to grow that much as a $5 trillion company is a meaningful amount of revenue. And I think this is very high signal that the market's still in a really strong spot. And Viti is basically saying like, hey, guys, we know we have a lot of competition. We know everyone's trying to build their own thing. In fact, Jensen was on CNBC yesterday and he was asked about Open AIs jalapeno chip. I don't know if you saw the C-Js. It was very cool.

Because jalapeno is a really great model. And they are a really great chip. And they came out and said, hey, this is actually more efficient than the Nvidia chips. And Jensen was like, dude, we're the kings here. We've been doing this for 33 years. You've been doing this for three months. We'll talk to you when you've actually produced a chip, when you've put it into data centers and when you've scaled this thing.

He does not seem concerned. The market really liked that. And now it seems like, I mean, Nvidia is really off the races. And there's no end in sight. Everyone wants the chips. They are sold out for as much as they can. make them. They're basically operating fully at capacity, and they have a very clear path to dominating more and more of the industry. They're growing bigger by the day. They now are

working with five of the largest banks in the world to help fund the build out of more data centers using Nvidia GPUs. And it's this unbelievable business. The entire world of AI is centered around this one company. And their earnings report is saying, it's going to be like this for a while. So buckle up because we are up only for the foreseeable future. I just want to show everyone a chart which basically shows Nvidia's revenue breakdown over the last, you know, a year or so. And it's just absolutely exponential and crazy.

Now, I want to give a bit of background as to why this is happening with Nvidia, especially recently, like, why do they continue to earn more money even though everyone's accusing them of, you know, building up this massive bubble? Well, there's a few things that are happening. The most obvious one being is

they've just raised the prices of all their GPUs by 15%. Now, granted, a lot of that is because the cost of memory has absolutely sought. So, you know, those memory stocks have also been very volatile recently. NVIDIA has to pay the most for memory to build their own GPU. So that's a main reason why they're projecting more revenue in the future that may not necessarily correlate with the profit side of things. The second thing is, you know, your point around Halapeno, which is Open AI's new custom AI chip, which they're building themselves. Typically, they buy from NVIDia, but now they're building their own selves.

Some people see this as a potential threat, you know, like Google's TPUs are to NVIDIA's kind of moat. And Jensen's response is simply, we own the moat for generalized GPUs, for generalized AI chips. That means any AI lab, whether you are a neolab, so a smaller AI lab, or a large hyperscaler, or whether you are anthropical open AI, can use Nvidia GPUs for whatever type of model. And by the way, it's not just LLAPs. It's financial models. It's fine-tuned models. It's models that can help make scientific breakthroughs.

Jensen owns that entire mode. When you compare it to what open air is built, Halapeno, that's not. known as an ASIC. It's a very specialized thing. They have to prove that out at scale. We haven't even seen these things in operations. We're going to see it by the end of the year. So Nvidia's response to this is we still very much control 70 to 85% of the entire market, and we aren't going anywhere. Now, the second bit of news on Nvidia that I'm most excited about is this acquisition, Josh, the $13 billion acquisition of this open source platform called Hugging Face. Now, for those of you who don't know, Hugging Face is basically this platform

which hosts open source AI model. So typically when you are an AI lab, and you can be any AI lab, if you launch an open model, you can deposit it or upload it onto the Hugging Face platform. It's where everyone goes to check out the latest open source model.

And you can do a few things when you're on this platform. You can upload your own datasets to fine tune different open models to kind of create your own version. A very good example I like to use is for Jipu. You know, all their GLM models.

They have about, I think it's like 10 official models, but they are collectively around 155,000 GLM models that exist on Hugging Face. And some are financially tuned for financial trading. Some are financially tuned for research and all these other different niche use cases. And you can go up, sign up on account and get access to it. Now, I ask myself, why on earth is Jensen paying 50x the amount of money that HuggingFace earns to acquire this open source model? Like, doesn't he just care about anthropic and open air?

Well, the answer is simple. Jensen cares deeply about open source because he believes in a world where it's not just going to be Claude or GPT running the world. It'll be hundreds of hundreds of very niche and also very broad models that run the world. And people can collaboratively use different models at different times. Agents run the entire world. Why this is a bull case for Nvidia and Jensen in particular is, guess what, they all need to run on his GPUs. So he's been investing heavily. He invested in Poolside, I believe, last week for $9 billion to acquire $100,000.

of their employees so that he can build new open models. He's investing tens of billions of dollars in Nematron, which is a video's own open source model over the next couple of years. And now he's made the third and biggest purchase of his entire kind of like focus for open source. So this is a very specific strategy that I think the banks are rewarding and that's why they're raising their price targets as well. It's amazing to see all these acquisitions happening. We have open router. Now we have Hugging Face going. Hugging Face is essentially the GitHub for open source AI. It's where everyone goes to build these things. It seems like there's three reasons. they would do this, and there's three reasons why NVIDIA has kind of been very adamant about moving

to open source. The first being what we just discussed, OpenAI is building their own custom chips. Google is building their own custom chips. Amazon's building their own custom chips. Everyone is trying to build alternatives to the NVIDIA GPUs. It's very difficult to do, but they're working on it. So open source is a hedge in a way against its own best customers, where it allows them to find more use cases, more need and demand for tokens that can be generated in the case that they lose some of that market share to these at the large companies. The second is, like, open weights seems to very much be where the volume is going. We're starting to see the trend somewhat ossify in the sense that open weights models and open source models are producing

a lot more tokens because they are cheaper to use. They're mostly effective enough to get the job done for most people. And a lot less tokens are going towards frontier intelligence to some extent. That frontier intelligence is much higher profit margins. It's going to be worth a tremendous amount of money for the people that use it, but the amount of use cases that require frontier intelligence are going down because the quality of the open source models are going up so quickly. So therefore, why would you not lean into that? And if you are a GPU provider who gets paid per token generated and your performance per watt to generate tokens and how many token you can generate on a per second basis, open weights

is definitely where you want to be. You want to be able to cater towards the open source community that is going to be generating a tremendous amount of this inference. And then the final part is just like the moat is continuing to move up the stack. So like we have the GPUs. Then Nvidia is going to kind of optimize these models built for Kuta, which works by default with their GPUs. And there will be a one-click deployment onto the Nvidia cloud.

And there's just this huge vertical integration that they get by being supportive of open source and building tools for the open source community to very easily deploy on their GPUs. And that seems like the strategy. And that is a strategy that is going to perform very well. Because as we know, the cloud-based businesses in the. these companies have insane profit margins. They print cash. And Nvidia's like right in the middle of that. So really exciting news from Nvidia this week. Can I put my tin fall hound for a second? Yeah. What are you guys? You reminded me of something just now. So Nvidia actually doesn't have

a really good cloud business at all, which seems remarkable for the guys that actually create the GPUs that they then sell to Amazon and Google. They're on it. Don't count them out. They make a bank of money. Well, do you know about the margins for these cloud providers? So like, okay, typically when AWS has their cloud system. Like, forget about AI, right? And Google has, what is it? Google Cloud GCP, they typically make around 80% margins. But with AI Cloud, they make only 40 to 50%

because guess who's taking the bulk of the profit? It's Jensen, right? And so they kind of want to be independent. They want to build their own chips. But Jensen hasn't been focused on building his own cloud provider. They had a service called DGX, and it kind of didn't make much money at all. My conspiracy theory is this.

he's purchasing Hugging Face, which one of their mainly used products is this thing called the inference service. Well, basically, not only can you go on and download these open models, and there's actually 3 million of them, but you can inference them there on the platform by just kind of clicking a button,

and that gets routed to different neoclows. So my whole thing is Jensen is not only providing the chips, but he's buying a company where he can effectively churn into his own cloud business that can compete with AWS and GCP. It's a really genius thing because he just focuses on open source models, but he can build that out into private models in the future if he so chooses to do it. The second thing is, remember last week when we spoke about Nvidia raising $500 billion to basically back a lot of the purchases that end up going back to himself?

Well, a question that we had on that show was, what happens if he can't sell the chips? Well, guess what? If he's purchased Hugging Face, he could just sell. the chips to Hugging Face to serve that same cloud business. So there is like this kind of circular virtuous loop, but it makes sense from a business sense. And I think that that's how he might be trying to play it. It is a very small business model. I like seeing Nvidia expand beyond just making bleeding as GPUs. He's like a very small business man, obviously. And Jensen, we trust

man. The dude, he's got to figure it out. He's been projecting these crazy revenue numbers for years now. And he just like happens to be right every time. So at some point, you got to trust the guy that's been doing this for 33 years and say, like, all right. Maybe he knows he's talking about here. We have some other stuff. Yeah, this is very important. Talk to us. We were right, first of all.

We recorded an episode, got published yesterday. If you're listening to this, we promise you a follow-up. Here's the follow-up. There's a mystery model that gave 100 trillion tokens of availability out to the public totally for free. They said, here, come and use it, do whatever you want with it. It's all on us for a full week. We're not going to tell you who it is, though.

We discovered yesterday who it was, and it is z.a.i. It's the people that brought you GLM. And it's their new model 5.3, Flash. This looks like a normal model. This is all fine until I was reading through this, I realized that all of those tens to hundreds of trillions of tokens that were generated were not only generated for free, but they were generated entirely on a chip cluster that runs on Chinese silicon, which is a first of its kind. I don't think we've ever had an instance in which Chinese silicon has produced this amount of tokens and this amount of intelligence. Because as the benchmarks come out, we've started to discover that it's about Opus Tier, a level, which is very good. This isn't frontier, but it's 15 cents per million tokens, and it understands video, audio, and text. So this is a really impressive model out of China. And on the back of our robot Olympics episode, EJS that we just recorded talking about how dominant their robots are, now I'm seeing that they are actually getting silicon to work

at scale with pretty quick inference, and a lot of it. I'm like, oh, wow, okay, China's like, they're making moves this week. Yeah, it's concerning a little bit. And I'll tell you why. recently America, at least for the last like four months, has imposed export bans on China, meaning that they don't allow Nvidia to sell chips to China, which, by the way, just to kind of reference the quarterly earnings from Nvidia, all that money was made by not even selling a single chip to China.

Very impressive, because China was like, you know, a major profit revenue for them. But anyway, for the Chinese Silicon, right, they're like looking at this, they're looking at this export ban, they're like, damn, we can't train frontier models without Nvidia chips. I guess we need to build our own.

own. And whilst they haven't built a good enough chip that can be used to train their own AI models, they still use various different techniques to kind of circumvent that, such as distillation, they can build an amazing chip to inference the model, which, by the way, is where all the money is going to be made in AI by a vast order of magnitude versus pre-training. So when you look at the Chinese chip companies about six months ago, they were two and a half years behind. So can someone please explain to me how six months later, they're now here running the top stack of one of the top open source models, which happens to compete very well with U.S. Frontier models purely on Chinese Silicon. Now, if you're wondering, who's the company that built these AI chips? Look no further

than Huawei. They released a bunch of different things. I think it's called, actually, I'm not going to try and make this up because I can't remember the name of the exact chip, but the Costa became live or got announced about a month ago. And now it's at full scale. production. Now, partly, they're being forced to use these chips because the Chinese state government has mandated them to just not rely on Nvidia chips for anything. And they replied and said, we can't do it for training, but we can do it for inference. So let us do it. And this is the first real example. And I don't know if you used the model when it was in the Ox Alpha stage. But, okay, so you saw, like, it was pretty good, right? But you would get a response really quickly.

It was super cheap. It didn't cost you anything. And so it functions very similarly to an video inference stack. And yeah, it is scary, man. Like, what happens when they own the chip infrastructure layer and the robotics hardware manufacturing layer? Like, what else is left? It's just not a break ers. They're making a lot of progress in a way that's like definitely a little unnerving. I'm like, okay, this is new. And I mean, this is kind of the expectation, right? Is you, like, I know Huawei is banned in the United States, a lot of the devices. You can't really get them. They're not sold anywhere. The inverse is happening. And what does a country do that doesn't have access to chips? Well, they have to go and build their own. And now they're going to continue to build their own and scale them.

it's going to be really interesting to see how this plays out. They're not the only ones building chips, though, because we have a new company that's building chips. And this one I'm far more excited about Google. Look at this beauty. Is that what you're telling me? Is this Waymo? Yes, Waymo is looking good, man.

So Waymo has this new vehicle, first of all, Ohio. And then even more importantly, they have a new chip architecture that they're introducing into this new vehicle design. And it seems like it's a fairly large improvement. I know you were going through the blog post, so maybe you could walk us through. But my understanding is that Waymo is about to get a serious upgrade.

And not in the sense that it's going to cost less, because these cars, I think the number it was like $260,000 per Jaguar Waymo. Like, it's still very expensive. But the chip that handles it is now 20 times more powerful than it was over the last eight years. And that seems to make a meaningful difference in terms of the quality of your ride that you can actually go out and get today. If you live in one of the areas, you can go and ride a Waymo fully autonomously.

I've done it. I think you've done it. Incredible company that has a really meaningful upgrade to the hardware today. Yeah. So a few things with this new chip that they just created. in prior Waymo generations, so in prior vehicles, because they have like so many sensors, how many are there, like 15 to 20 sensors, it's like cameras, proximity sensors. Bulky ones, man, it looks crazy.

I'm sorry, I have to say it. I respect Google a lot, one of the ugliest autonomous cars out there. The new ones are cool. The new ones are cool, but mainly they've been horrific. The point is, the car has to ingest a lot of data, translate that data into some kind of an action, which is stay right, stay a left, break, you know, get this person safely from A to B as quickly as you can.

Now, in order to process all that information, you need a really powerful AI chip. Waymo didn't have that. So they kind of used a hybrid of like a chip and a cloud service. Now, that's kind of dangerous because what if the cloud service kind of breaks connection, you end up being in a whole pot of mess.

Now, with this new chip, it brings all the compute on board the actual device, the device in this case, being a car, which I think is really cool. So the entire system, Josh, is being run locally on device in the single car. So when in the trunk of the car, I love that visual that they show. Exactly. Yeah, yeah, wait, let me get the visual up. Yeah, that's why they are showing the trunk. If it wasn't clear for the last one, it's like, yeah, all of that compute lives sitting underneath the trunk there. It's so, so cool. Pretty cool. Yeah. Yeah, and it's a single chip or a cluster of these different chips. And so it takes the input data from the

sensors, it processes everything through this one chip, which by the way, is 20x more powerful and is 5x cheaper. They lessen the cost. So do you remember in our last Waymo episode, Josh, we were like, why do each of these cars costs like 100K? It's because like more 260 or 280. Oh, my God, 260K, yes. Now it costs 25K a car to pop out because they were able to create the silicon chip and

reduce the cost down and run everything. This new car costs 25 grand. Look at this. Wait one second. Where's the breaking news? Yeah, over here. So it used to cost 100 to 125K. Come to the street, Wall Street, OK, this is an important distinction. We need to make. Talk to me. Talk to me. This is, okay. Po, tell me. Tell me. No, no. So this is for the hardware. This is not for the vehicle. Okay, that's where I was getting confused. So the hardware part is like the sensor suite that sits on top of the vehicle. So I guess it was like 125K for the last one on top of a $100,000 vehicle. This one's $25,000 on top of an assumed lower-house vehicle. So, okay, that's an important

decision. Yeah, sorry. Just to be clear, we're talking about the chip, the chip hardware here, like all of that, which kind of like, you know, makes the car sexy and do the cool thing, right? And so the point is, with this new chip, they've been able to achieve a few things. Number one, it runs locally on the car, which means it's cheaper to run. It's way more effective, so it's quicker, it's way more reliable, and it's incredibly efficient. So it costs Google less. it costs way more less, which is why they've been able to kind of reduce the cost by like 3 to 4x, which is very impressive for any successive generation. I just think it's a really cool bit of hardware device.

You know, we like hardware on this show. And I'm happy Waymo's taking this path forward. It looks cool. It looks like a fun science experiment. Like the car itself looks interesting. The new Ohio, it's fun, it's different. It looks like Disney World. This is a really cool promo video showing it.

Oh my God. Wow. Yeah, it's like a unique design. I mean, granted, if you were to design a fully autonomous car, I'm not sure why there's two seats still facing forward in the way they are with the steering. Wait, hang on, wait. Did he turn around in this? Let's see, Josh.

I'm not sure. Oh, no, no. He's the shit just, he's turning his body around. So we mentioned the cost of these sensors is about $25,000 for the suite, the hardware, the sensors. You know what else is $25,000 each ass? The entire cyber cab from Tesla, the whole thing, including the vehicle. It will be $25,000. 25K per vehicle.

Per vehicle, with the whole sensor stacks. So what you'll notice with the waymos is they have this. this huge amount of spinning LiDAR sensors and very complicated sensor stack. The Robotaxi that we're seeing on screen here, the cyber cab from Tesla, has, I think it's eight cameras, very similar to the ones that are sitting in your smartphone right now. So basically, if you imagine the cameras on your smartphone, they are placed into a vehicle and put in the right place, that combined with the cost of the actual vehicle is probably going to fall around $25,000 in terms of cost of it sold.

And it works. and it is being deployed starting September 3rd at scale. So what we see here is actual production ready level cybercabs that are going to be going around, select a few cities. In the announcement post, could we pull that up so I can read it off? I'm not sure exactly which ones is going. I know there is one in Austin. I know there's one in Dallas.

So yeah, Austin, Dallas, Houston, Miami, Orlando, Tampa. If you live in any of those places, you can go and get in a cyber cab right now. And I think this is going to be a really unique and interesting inflection point for the autonomous vehicle world in general because we have Waymo that's making really good progress. And Waymo is right now much more accessible. They're doing many more miles per day. And people have gotten comfortable with it. If you live in Los Angeles, if you live in San Francisco, having Waymo's go around, it's just a part of your life. And you don't even think twice about it now. The Cybercab has an opportunity to really change that in this sense that the cost per mile on the

cyber cab will be significantly less than that of the Waymo. When you look at the Waymo vehicle, it has like these big bumpers that's really big and it's not a very efficient car. The cybercab has has much more efficiency, which allows it to decrease the cost per mile. And it's going to be really interesting to see how they're able to scale that, because Waymo's notoriously are geofenced. They work in very specific locations. I've been using a Tesla fully self-driving for years now, and with no oversight, they work anywhere. So as soon as regulation comes online, we may see a world in which like 12 months from now, cybercabs are just roaming around everywhere. And we have the opportunity to get in them, drive them around. And the idea is that you'll never need to own a car, because the cost per

mile will be so low that it will offset any expenses incurred by actually buying the physical vehicle. So really cool time for autonomous vehicles, autonomous cars in general. Keep an eye out for Wimo, keep an eye out for the Tesla event that's happening next week. You know where I'm not going to be keeping my eye out in New York City? Because they like to just fan all innovation. New York City sucks, man. It's so brutal. It's so cool, but it sucks. Like bring us the new tech early on for goodness sake, please. Now, I want to have a call out just very quickly on this particular topic Josh. Producer Luke,

we love him. He is in one of these places. He's in Austin. Guess how many times he rode in a robot taxi? Go on, guess. Well, I think once maybe. I think it's zero, dude.

I think it's zero. He may have done one, but it wasn't a cyber cab. I don't trust it. That image was AI generated. I need to see you, Luke. I know you'll listen to this because you're editing this right now.

Like, I need to see you in a cyber cab, dude. It would be awesome. Anyway, moving on. We'll get a fact checker in here. Who is our favorite investor in the world, Josh? Oh, you mean the 25-year-old

who was fired by Open AI, raised a billion dollars, and then had 45, and then lost it all, and then Citadel bought it all out, and now he's being investigated by the SEC. Wow, actually a very good, succinct summary. That's a very chat chit-esque, Josh. Very impressive. Yes, that is actually it. That guy. Wait, but what's his name again? Yeah, Leopold Ashtonbrana. Oh, yeah, that one. That one. Yeah, yeah, that's right. Okay, listen, like, I'm...

My hands are up. For those of you are about to, like, write nasty comments, right? We love Leopold Ashroomer. I still think his thesis isn't intact. He just, you know, don't trade with Lerpold. leverage kids. But he is now being investigated by the SEC, or rather specifically, the SEC has subpoenaed a bunch of banks which facilitated situational awareness funds, which is Liverpool's fund, trades out of suspicion that potentially there was inside trading. Now, this is something that we have kind of briefly referenced and spoken about on the show. Ajit Balwit, who is now officially Liverpool, Ashon Breeder's wife, is the chief of self-ed Anthropic. And there's a few other types of

things going on there. But effectively, the question is, has Leopold been kind of getting information and trading on that information when he should have kind of like publicly disclosed it? We don't know, but this is just another kind of pitfall into this whole probe where, you know, Leopold's been taking hits, hit after hit recently. He recently got back into the game, actually. He's like got a new open book. He made a $400 million private investment in some other company. Goodness knows where he got $400 million from, but he's doing the thing. And technically the fund is still up. But yeah, just curious to see. why this all ends up. I don't know if you have any thoughts on this, Josh. Yeah, it's mostly just like

people don't like Leopold. And what does it take to open an investigation? You file a complaint. Someone says, okay, we've acknowledged your complaint. Now there's an open investigation. And then you can report the investigation. It could be nothing. It could be something. Chances are, it's probably nothing. And I think Leopold is just down. People want him to be down even more. And chances are he's going to be just fine. This will kind of blow over and there won't really be much that comes from it. In Leopold, we trust, man. I'm still rooting for the guy. I'm excited to see the new 13 to see what on earth is going to be on the comeback portfolio after the entire public one got wiped out. So we'll keep our eyes peeled, as we always do on Leopold. I'm also seeing

some charts here, Ejez, on Fable 5 and Anthropic and like enterprise sales and businesses. What is, what are we showing here? I just think there's some chart smithing going on here. So like, what are we got? A few things going on here. So the Financial Times released a chart and it shows the usage of Anthropics' different models over time. And I don't know if I can open this up, Josh. I don't know whether it makes it a little tricky for you. But we'll move it. There we go. Yeah. But if you can have a look over here, you'll notice that Fable growth kind of like accelerated in mid-June to the kind of the end of July and has kind of been flat. Now, number one, I don't know where they've got these numbers from because Anthropics are not a private company, so I don't

think these are necessarily reliable. But the point they're making is it has stagnated. And instead, people or companies, rather, are choosing to use some of their other cloud models. Now, there are many reasons why companies would do this, whether it's cost, whether it's efficiency, whether it's just kind of like preference or taste of like the model's personality. But it's interesting to see that a frontier model isn't being taken up as much as they can. And the point that they compared this to is that Open AI's models have seen increased usage. Now, my comments on this is simple, which is open AI is just like has a smaller user base than athropy. Now, it's still granted very large, but obviously any.

hike in usage, which they're particularly seen with Codex specifically, which, by the way, is a fantastic product, is going to result in a larger uptake or percentage increase, right? Whereas with Fable, maybe if the stagnation is the thing, maybe people are just, you know, factoring in the cost, Fable is notably, you know, more expensive than a lot of other models, but it is also operating at the frontier better than any other model. So I can see why maybe there's a hesitancy, you know, you've got all these new open source models, JLM 5.3 flash, and you're thinking, okay, maybe if I can get 80% of my work done using a cheap model, I'll use FAML for the other 20%.

But the point is the pie is growing bigger and bigger. And I think this is not other than a bit of a hit piece. Well, and it looks like the chart is reflecting the pie growing bigger, right? It's like the numbers are going up into the right. The business spending in total based on this chart looks like it's going up. And maybe this is kind of the phenomenon that we described earlier where there is just a lot of demand for lower cost tokens, but the overall demand is continuing to increase. And perhaps that is a conclusion we can draw from this.

I'm confused what the other actually is. Like, are they talking about, like, haiku or something? Because they have sonnet, opus, fable. What else is there? I'm not quite sure what's going on here. I don't know. But, hey, everyone's having, I mean, the models are great. On a day-to-day basis, I'm still using it.

I'm getting a lot of usage out of it. And I'm kind of enjoying the time there. So it's, yeah, we'll see. I think is kind of like where we can leave this one. It's like, okay, well, I'm sorry for enterprises. But, like, hey, we're cooking. Like, for $20 a month, $100 a month, $200 a month, like, the tokens are flowing. I think there's another reason, though, right?

Which is like, did you see this whole debate around zero data retention? I started yawning when I read that phrase, Josh, because I was like, oh, my, what do you mean by this? But like, yeah, oh, sounds fun. And another big thing for these enterprises is they want to use the latest frontier models. And to be honest, when you're an enterprise, you don't care about spending more money if it means you're going to make even more money. So, like, I'll spend money on the most expensive model. But a big compliance thing that they face is this thing called zero data retention, where if you have a company, a software company which you kind of lease their tools or whatever, and they retain proprietary

information for their own usage, which they could technically use to build better models, which is what Anthropics has to do right now, but I believe they're in the process of evolving that potentially. Who knows? This is just rumors at this point. I believe, you know, this is something that will just kind of flip that switch and suddenly we'll look back on that chart and suddenly it spiked up growth. So I just think this is a nothing burger. ZDR is very common. I don't think Open AI is enforcing it right now, which might be the reason why they're like the surveillance. are waiting faster over the last couple of months, but just something to be aware of. Okay, so maybe the chart is nothing burger, but the thing that is real is the use of tokens,

and the use of tokens are generated mostly by agents. And if you are working with agents, we have just the people for you. Their name is Ledger. Ledger is offering a way to build with your AI agents because if you're worried about security, you should be, because agents, as we know, are not very secure at the moment. They're getting very smart very quickly. Ledger has an agent stack that fixes this. It's a series of open source tools that you can go and use in order to make sure that the agent is giving you the expected outcome that you want. It does this through a three-step process where the agents propose, the humans approve, and the ledger signers enforce. And this is all open source. It works everywhere that you work with your AI agents. It works in codex. It works in

cloud code. It works in cursor. So if you are interested, the link is in the description down below. Thank you very much to Ledger for sponsoring this episode. To nothing burgers in a row, we got to turn this around, man. Leopold. All right, dude. Anthropic, what do we got? This is not a nothing burger. This is a something burger. What if I gave you continual learning robotics model, huh? Dude, it's a big week for the robots. First we had the Robot Olympics, then we have full self-driving vehicles, now we have robotic learning. And this is one of many announcements that I saw this week. There was another one that I thought was super cool where you can actually go and get paid

like $15 an hour to perform tasks in your house. And as long as you do that with the camera, that allows you to train the robots, you will be paid as if you are a full-time employee. It's like kind of cool, very fun. This is another version of that, it seems like, where they're just basically throwing a ton of data at these models. and emerging from that data are physical capabilities that robots are learning from scratch, and they're doing so in a way that's fairly efficient.

What I like about this example that we're seeing on screen right now, too, is it's not, this isn't purely humanoid. This is more of a narrow purpose, narrow, very specific type of robot that they're working on. So it seems like this is a kind of new paradigm where robots are getting enough data to start meaningfully progress and get good at these tasks. And we're seeing so many unique and creative ways of collecting this data. remember we did one, was it DoorDash or something, where they were paying dashers to strap cameras to their heads?

It was DoorDash. That way they can go and collect data for what it looks like to deliver an item to a person's house. And then taking all that data, you get this emerging properties. And the Antig, the Pokemon Go company as well. That's right. That's right. Yeah.

So there's all of these really creative ways of paying humans to collect data to then automate the tasks that they're doing. And I think it's really cool to see the actual progress come from this in terms of the ability of these robots to do more and more interesting tasks that. that we as humans do, and maybe we don't want to. And I think a lot of people probably default to like, oh, this is bad. We're paying human beings to offload the responsibilities to robots. But the reality is, is like, how many people actually love doing this stuff? Like, how many people love to clean or love to cook or love to take the trash out or love to like deliver food?

There's perhaps much better options on the other side of automating that. And it results in much lower costs for everyone who is using those services. So this seems like a win, even though on the surface I could see why it's scary. And this is a really cool. example of that in practice. I mean, this is a pretty sweet video. Yeah, I mean, just to kind of explain what's going on here, they made a breakthrough in robotics models. Now, typically, the problem that you have with robotics models is they're really hard to train. In order to get the data to teach it how to do really complex stuff, you need to get those videos, like you mentioned, Josh, and you need to kind of like teach it a million different things that hard code it into a model. It is incredibly complex and

hard to do, which is why we haven't seen a cool robot in your house as of recently. The breakthrough these guys made is something called in context learning, which means you can feed this robot. One video of you making pancakes and flipping the pancake and it'll be able to figure it out from that 30 second video and do it itself. But the craziest part is, Josh, let's say I filmed you making a pancake and you flipped it and it fell on the floor, right? The robot's seen this, right? So presumably you would think it would do the same. It doesn't do that. It knows when you've made a mistake and it can take it further and start making multiple pancakes at the same time and then stacking them on top of each other. It has this intuitive feel. So an example that

they show on this video here, I think they already passed it, is it uses a kind of dustband and brush to kind of sweep a block onto the pan, right? And then they replaced the handle with a banana. And it was like, what would it do in this case? It realized that it was a banana. It realized it wasn't a brush, but it still did the same thing and brush the block onto the pan. So, So it has an intuitive understanding of what it needs to do and is willing to use any and every kind of tool to do the thing, which is the breakthrough in this robotics model. And I think it's going to get us much quicker to having a robot in our home that can intuitively just do things without us needing to hard code it or tell it what to do.

Very, very so close on the edge of the robot revolution is a good way of describing it. We had the Olympic Games that we had a full episode about earlier this week. So if you haven't watched, I go check that out. The opening ceremony is a little bit militant for China, which is scary. So perhaps like, you know, take that one with the grain of salt. We have a few rapid fire ones at the end of this episode. My personal favorite here. We have a date for the Apple event.

It's coming. It is September 9th at 1 p. Eastern time. Is it Siri AI? Dude, we are finally, no, it's not going to happen just yet. But basically what this one is is the hardware. So there's two types of events. There's the developer conference, then there's the actual hardware event.

developer conference released or unveiled Siri. The release of Siri will come shortly after this hardware event, and it will compare with the new hardware devices. The new hardware devices are expected to be a new iPhone 18 Pro, a foldable device, which I'm incredibly excited to get my hands on. And more importantly, this is the first event with Apple's new CEO in charge. Tim Cook isn't going to be on stage here. It's John Turnus, who's running it.

There's a new guy in charge. He's a hardware guy. This is all the new hardware we're going to be getting, specifically built for Apple intelligence. And man, we've been waiting for this for, what, three years to get actual intelligence inside these phones. So very excited to see how that's going to play out, what the hardware's going to look like, how the software will integrate with it.

And probably a month from now, we will all have brand new iPhones with Siri intelligence inside of them. And that is a very exciting thing. So fingers crossed. I mean, the chip thing is cool, right? I saw the M5 Pro and the M6 kind of get announced. And the price points of those are. Yeah, new Mac Mini, New Mac Studio.

Kind of up there, right? So it's like, I think it's like 900 bucks for the M6 or... Oh, and then 1,500 for the M5? Have you heard the high end of the studio? Go on. Wait, wait, let me guess. Let me guess. Let me guess.

You're going to be wrong. $3,000? No. I think it's like $18,500. The Bras. If you want a maxed out, Mac Studio with like the M6 Ultra and all the RAM that you can get, it's crazy. I'm going into debt. I'm going into debt.

But hey, if you got the money and you want to run some local inference, there's your cost of entry. All these people talking about free models. I just buy an inventory. video GPU at this point, dude. Like, what are we doing here? That's crazy. Free model, this, free model that. Like, just how about you just pay for like a subscription or like an API key? I'm like, let the big guys serve your inference because it's getting real expensive to do it locally. By the way, blame the memory manufacturers to this. That's why the price is hiking across everything right now. And they got great margins too. I was looking at any videos. I'm like, man,

these people are charging way too much money for this. They're driving up all the prices. Last topic, I thought this was really funny. E.J. as you put it in, it's like, I'm a watch guy. Yeah, so basically, breaking news, guys, arguably the biggest topic of this entire episode. Sam Altman just ordered seven ultra rare custom Swiss watches from this company called Van Gua. I'm definitely butchering that name with a branded opening a logo on the doll. No commentary, right? First, let's observe. Look at this masterpiece. For those who like wear Apple watches, I need you to just take that thing off your wrist for a second

and just appreciate the mastery that is on your screen right now. This looks so cool. Can we take a second? Because I just so happen to have an Apple watch in my hand. Please. It appears like the band on my Apple watch looks oddly similar

to the band on that watch. Absolutely. And that is because the memory component, the equivalent of this, is that complex cognitive function that is happening inside the thing. Who cares about the plastic strap?

You can literally, you see this little attachment over here? How many tokens per second? Does that generate? Oh, God. Not enough. Not enough. But inscribed on the back

of each of these watches is if, if AGI. Deploy. Which is written by a programming language Python. So the point is,

I think this is like, you know, Sam kind of like giving a peace offering before potential, rumored, IPO that I think is coming very soon and he wants to give it to is ranking order of people

that haven't left open air over the last two weeks. Sorry for the shade. I've just seen a lot of executives leave recently. But it's a beautiful watch. Nice. People have been discussing what the secondary values

value of this might be. And people are estimating this would be in the order of $500K to a million. I personally think it would be higher because there's only seven of them. And one of them will be won by Sam's on Sam's wrist. So that's definitely not pretty strong. This is kind of like

what it's like the push presence. Like when your like wife has a child, it's like when your executives produce AI. If my fiance is saying please please turn off now. Thank you. I think this makes sense, right? It's like when your executives produce AI. Like hey guys, thanks, like here's a million dollar watch type thing.

Because I was listening to Sam Altman, or I was reading the time publication of Sam Alman and what he was saying. And they fully expect to have AGI internally by the end of this year. And for those that are not familiar, we are about to enter September, which means we are close to the end of this year. So maybe perhaps he's getting these custom made in anticipation of their internal rollout of AGI, whatever they believe that to be. So that's just an interesting thing worth noting, worth following.

Lovely watches. Like, hey, the limitless guys, I mean, I'd take a half a million dollars. I'd take a million, whatever it's going to cost, that'd be cool. But that's the round of. This is a long one today. We're like 40 minutes into this thing. So if you made it this far, man, thanks for coming along with us on this journey. It's been a good one. A lot of stuff happening this week. The things are picking themselves up. The roundup's getting longer. There's more stuff to talk about. I say it's a big robotics week this week. A lot of robots. A lot of China talk this week. Good week for the Chinese. Well, there's good for us. We don't know TBD, but like good week for the Chinese. Good week for Midian Jensen.

It's good for limelis as well. We got a lot to talk about. So much content. So, well, we're putting out. We got so much content. much to talk about. What was the other thing? That was like the open-rower acquisition that might have been last week's news. Who knows? What is track, man? We are covering anything and everything that is happening in AI and Frontier Tech. Whatever just makes, you know, Josh and I pretty excited. And we're doing that four times a week. So if you haven't listened to any of our episodes this week, please check it out. We also have a newsletter that goes out to, I think it's like 110,000 of you that are reading our stuff every single week, twice a week. We do an essay. And we also do the

weekly roundup, which if you don't want to listen to our beautiful, soothing voices telling you about AI, you can just go read the quick summary. Digest, we are covering you everywhere on Spotify, Apple, as well as YouTube. If you aren't subscribed, if you haven't left, such as a comment, if you haven't given us a rating, what are you doing? Please do all of that. Send it, share it with a friend. It helps us out massively. Josh, have I forgotten anything? I mean, you can, and I know this is kind of rough for me to say, because I haven't been posting much, but on screen, our handles on X are showing, and you can go and follow us there for more information on the day-to-day. You just as I know you're always,

You've done off stuff about updates. So that's just like a fun little call out if you've made it this far. And you know, you're interested in hearing a little bit more. I'll work to post more on that platform. But that's it. Like you can go, feel free to like, you're dismissed. Go touch grass. Go enjoy the weekend.

Get off your phone for a little bit. You're now fully up to date and all of the things. AI frontier technology. You're in the loop. You can go out to your like weekend barbecue. I mean, close to the end of the weekend. We're almost there. I don't want to think about that.

But yeah, that's pretty much everything. Thank you so much. I mean, share with your friends. If you got some, that would be interesting. and yeah, enjoy the weekend. See you guys next week.