Claude Code Channels Are Coming For OpenClaw

Good morning, everyone, and welcome to The Modern Market, where every day we discuss everything to do with the modern market. But today, that is the AI market and the AI market only. Loaded up for discussion, we have Supermicrocomputer co-founder is charged in the $2.5 billion scheme to smuggle AI tech NVIDIA GPUs to China. The stock is down 17%. On that news, DoorDash introducing Tasks, which is basically DoorDash's take on the website Rent-A-Human.ai, which we talked about. I think one and a half months or so ago, when the OpenClaw madness started. And last but not least, Claude Code now works on Telegram and Discord. They introduced channels. Anthropic is shipping pretty, pretty fast ever since they came out with that dispatch feature, which we just talked about earlier this week. And we're also going to have a few separate AI deep dives to get into on the show today. It's going to be bred in the co-host seat with me. We're almost ready to get into it. But just a reminder before we do that, friends, before we get started, that nothing that we say on the show is financial advice. This market is extremely risky. We do not know anything for sure. So please do proceed with caution and exercise your own judgment at all times. With that out of the way, Brett, it is Friday, 20th of March. It is time for an AI deep dive for an AI show. I'm pretty excited. How are you feeling? That's right. Legendary is here. BCheque is out. And when, what is it? When the cat is gone. Yeah. The cat is out. The mouse will play. Exactly. We're playing. We're playing in the AI realm. You've got a hat on. Or a visor. A hat? A visor? Yeah. Got it on. So we're very in sync. It looks like, maybe it's the top of the thing, making it look like a visor. But things were in sync. We're locked in. And I'm excited to talk about this. He even grabs himself. I'm learning. These quad dudes are dropping stuff so fast. So fast. And yeah. But I'm excited. It's going to be fun. Yeah. So shall we get straight into it? Let's do the price action starting with that as per usual. That is starting with the majors first. First, Bitcoin is up on the day. It's up 1%. Just a touch under 71k. ETH is down 1% at 2.1k. What else do we have? BNB is basically flat on the day. $640 there. So is Solana still sitting at $90. So is hype sitting just a touch under $40. Quickly taking a look at the meme coin side of things. And we have Doge being up 0.5%. We have SHIB being up 5% on the day. The PUDGIs are down up 2%. SPX is up 2%. So it's a bit of a mixed market here. Some are up a couple percent, but nothing too major happening. With the exception of Pippin. Pippin is now down another 30% on the day. It's down almost 80% on the week. It's crazy to think that Pippin just a week ago or so had flipped Pudgy, had flipped Trump. And now it is down 80%. We managed to ignore everything that has been going on with the token. And now it seems to fade out of irrelevance again. With that being said, that is it on the price action. Maybe one more thing to do before we get into the agentic discussion, before we get into our AI deep dives. And that is to give a shout out to today's sponsor of the show. And today's sponsor of the show is Yeet. A crypto casino that has done more than a billion dollars in less than a year. Founded by trusted names like Mando, KBM and pro poker player Ben Lamb. We are users of the sites ourselves. I am still on my Limbo GCR arc on Yeet. And just a reminder to only play for fun what you're comfortable with. And to help there, we're going to give out $500 each week at the end of today's show for one player selected who is playing under the Modern Market affiliate link, which I'm also going to put in the comments later in the show. So do check Yeet out. They do have a number of crypto-themed games. They have the sportsbook. And stay tuned for the giveaway at the end of the show. And with that out of the way, the first thing I want to tackle on the agentic side today is payments. So what's the thinking here? On the one hand, we did speak about the X42 standard on the show as a payment standard on blockchain rails, on crypto rails. We also did speak earlier this week on the one hand about Tempo. Tempo coming out as a blockchain with their own payment integration for agents as well. Did have a look at that on how that is working. And we also did touch on Visa and their head of crypto labs coming out with Visa CLI, which is basically also made for payments using a Visa card. So there seems to be a lot of attention on how to get agents as customers for whatever payment rails companies have been building. But at the same time, someone built this very interesting dashboard, which is called X42 Everything Data AI. And it's basically looking at the payments per second, which are done by humans, which is around 29,000 payments per second. And then at agentic payments, which according to this dashboard is around 0.82 payments per second. So if I were to refresh that site and we would probably refresh that counter, we'd see one, two agentic payments, three agentic payments happening. But at the same time, globally speaking, more than 200,000 agentic payments have been happening. So it seems that for one agentic payment, there's approximately 35,000 human payments going through at the same time. And that is using X42 transactions exclusively to measure agentic payments. That's the only caveat there. So obviously other systems which are in use or are being built are not counted in there. So I wanted to get your take, Brad, seeing have you had a chance to look at Tempo and how they do the payments? How are you thinking about that payment set of things from Megha's perspective? But before we quickly do that, I've seen some YouTube comments saying that we have some audio issues with Brad. So Funky, given that you're on stage with us, quickly checking in with you if you were able to hear Brad fine so far or if you had issues as well. Jim, now I hear Brad loud and clear. Cool. Then Brad, take it away. Go for it. Okay. I'll do a little bit of a sample slow talk for the crowd. And you guys can tell me if I'm still having issues, if it's an audio thing or if it's just like cutting out entirely. So to answer your question, yes, Tempo dropped. Tempo, you didn't mention it. The actual protocol itself is MPP, Machine Payments Protocol. It was Paradigm plus Stripe. That's what Tempo is that are leading it. I actually am already crafting up a PR so that we can get Meghaeth included in there. Honestly, so like it's a framework. It's the same framework as X402. It's meant to be a little more agnostic to how things can be paid. And what I mean by that is for X402, it has to run through blockchain rails. Like the payment is like it has to be an on-chain settlement. It has to be tracked. It has to go that path. For MPP, it does not have to be blockchain. You can do a charge payment for a credit card. You could do Stripe. You could do whatever. So there's two or three other different things already in there built in that are not blockchain. And the MPP itself can utilize X402 under the hood if necessary. So I think that's important to understand the comparison of the two. And they are competing, right? Obviously, one is owned by Coinbase or was released by Coinbase. The other one's released by Paradigm and Stripe. So it's going to be interesting to see which one actually takes off. On a technical level, the biggest difference is that MPP ships immediately with a thing called Sessions. And Sessions is meant to answer this question of how fast we anticipate agentic payments taking place successively. So a single person, if we were to use an LLM session as an example, if we were to sit there with an LLM session and say, charge me for every token that I send back and forth to the server. That's obviously very, very rapid fire. You're quoting one agentic payment per second across the entire industry. We're anticipating, the world is anticipating, that this will probably scale to tens to hundreds to thousands per agent at some point. Right? Like that's going to, we're going to hit a wave where that happens. And if that is the case and you have these two protocols in the current paradigm with MPP, you would open up a session. It's like a state channel. Same thing. It's actually a technology that we use or initially ideated way back in the L2 like scaling roadmap. We say, let's open up this thing. It's like you open up a tab and you say, I'm going to start doing a bunch of stuff. You then start sending a bunch of messages and just keeps tab of what the ultimate price is going to be. And then you settle the price. So it's like really two on-chain payments, right? And initiating it and closing it and settling it kind of like lightning channels. And then with X402, you can't do that right now, today. Now, that said, X402 V2 has already come out and it has a few additional features. And they have a few more on their roadmap, stream being one of them. So they do plan on building something specifically for the streaming use case. But today, X402, frankly, is not viable for it, for that specific setup. And it's not even necessarily because of it not having a unique feature for stream. If you go and you look at, this is what I've been digging into the last few days. If you go and you look at like how this stuff is actually taking place, it's actually the underlying blockchains. They're slow as shit. But if you, I did, I've been running some tests. If you go to, I grabbed a ticker on live service today using X402. And the process for X402 is I go to the website. The website says, hey, you owe me a payment. I then issue a pay, or sorry, then I go to a facilitator. Facilitator then says, oh, you know, do the payment. They settle the payment. And then they say, okay, you're good for the payment. And then it's up, right? So it's like two parties, three parties total. And it goes back and forth a few times. And in that back and forth, because base has 200 millisecond block times with flash blocks and then two total seconds for their full blocks, it takes about an additional second for every single payment for this third party, this facilitator to settle the transaction to say that it's good. And if you're thinking of agentic payments, like, man, one second delay for every single transaction. It's not going to cut it. Especially if you have a longer chain of payments. Like, if there's a more complex hopping through things, and each of them takes this extra second, then you can reach a minute delay. And if that is contingent on something else, then this is not going to work. Maybe on a more personal level, are you at a stage where you're comfortable giving your agent or your agents access to money, to payment tools, to any sort of agentic payments at all? Yeah, I mean, like, it was required for me to even do those tests. So I actually, one of the fun things that I've had my agent doing is mining addresses for me. So, you know, I've already, you know, deployed a smart contract. I relaunched a project, all this shit. And before that, I said, hey, go, we need to find you a deployer address because people are going to see you deploying stuff. And so if you look at on MegaEath, if you see anyone deploy a contract and the last eight digits are bad code, then it's him. But yeah, I give him money. I let him do stuff. Now, it's not like a routine daily thing. I'm not saying order me, you know, a burger for lunch. I'm not treating like a personal assistant. I'm treating it like a research assistant. It's a little different. Yeah. And maybe one more question on that. Given how big this discrepancy is at the moment between the number of agentic payments and the number of human payments. Like, since we started this segment, 12 million something payments have been processed by humans and 350 agentic payments. Again, that is everything in the X42 world. So given how small and early this is and given how massive the investments of a Visa, of a Tempo Stripe going in there, the expectation has to be that this is going to catch up to the human side pretty, pretty quickly. Could you, would you wager any guess how far we are out there from the basically agents even getting to a, I don't know, to a more meaningful share? There's 10, 20, 30% of whatever humans are doing. Is this, is this months away? Is this years away? Is this a year away? Hmm. Are we saying, is this tracking, is this 20,000 per second right now? Is this on chain transactions? Is that how it's monitoring this? Uh, no, this is, this is looking at basically all the payments happening. Like everything that's going on. Okay. That's a lot. Yeah. Um, so this is tricky for me because I'm burned a little bit in the blockchain industry where I say like, man, adoption takes longer than you anticipate. Right. We're however many years in seven years in 10 years in, and we're just now getting to, uh, stable coins being adopted broadly. Right. Right. But the counter to that, which makes me think I should give it time. The counter to that is this agentic stuff is moving. So, so, so rapidly. I'm, I'm getting like anxious, anxious messages from my developer friends who I consider to be like wizards. And they're like, man, I'm stressed out. Like, I, I don't feel like I can keep up with the breaking news of this stuff and have a life. So, uh, if I, I would say 30% by the end of 2027. If I had to guess. Yeah. This is, this is, I have these two views as well, right? On the one hand side, burned from thinking about NFTs four years ago and how early we were to that tech and how NFTs and NFT ticketing and everything will integrate it in everything. But on the other hand, you know, the AI side and this agentic progress and just yesterday we talked about minimax 2.7 coming out just a month after minimax 2.5, just a month after minimax 2.1. And even that progression, how significant these upgrades are, how fast cloud code and entropic shipping, how codecs is out there, open AI is out there acquiring startup at a startup. We're going to get into one of these news a bit later in the show as well. So this, this progress that's happening on that end feels so much faster than anything else that I've seen before. And just thinking how significant these investments are, I would, I would agree with you that in 12 months, this chart will have lots of agents walking down on the right hand side of the screen on the agentic road and we'll see a massive, massive bump there. And go for it. It's a, it's a deviation. So if this, if you still have a stream of thought, finish that. I would, I would, I would go in a different direction. So go for it first and we'll, we'll take it from there. The Visa CLI thing, were you able to get in? It was a fun website. I put in on the wait list. I have not seen any acceptance. I was just looking through my email to see if I got anything. Did you get anything? Are you in? Not yet, but I played it in a strategic way because I, I clipped that and Visa's head of crypto labs saw it and started following us. So getting into a conversation there and seeing if we could get some sort of access to that because I'd be, I'd be curious to see how, how that set of thing works. And on the, on the, on the other note that you said, like your, your deaf friends getting stressed out about that. There was this clip from, I think it was from the All In podcast as well with NVIDIA CEO with Jensen Huang speaking about his $500,000 engineer. And basically the, one of the metrics he'd be looking at is if that engineer at the end of the year has not spent at least $250,000 worth of compute, $250,000 worth of tokens. He'd be pretty worried about that guy. And if they only spent like five care or so, he'd be like, some things, some things not right there. What are you doing? Yeah. And I feel like everyone experiences that. I feel like there's, there's, there's two camps talk to a deaf, a friend of mine who said he hasn't used charge GPT in two years. You said two years ago to say it's not good and never looked at it again. That's, that's not the metric to, or that's yeah. Just, just not matching the speed of the development, which we, which we're seeing right now. No. Yeah. And like, we're, we're leaning heavily into it too, at, at Megi. Like it's, it's, it's straight up a mandate at this point where it says, you know, if you don't have like a task that you have been doing, that you have automated week over week, then it'll show up. Like we, we are heavily leaning into it across the board. And even listening to a podcast yesterday from, um, from Tempo, it was George Rosen, uh, Ryan, they were mentioning that to increase the performance of Reth, Reth is their client that they developed. That's used by a couple of different chains and is the primary client for Tempo. They said at all times, they just have an agent in the background running benchmarks on their code. And basically effectively doing, if you guys saw Karpathi come out with the auto research, basically just have like auto research running constantly against the client and looking for optimizations, just constantly. And so many PRs saying, Hey, you know, we spent 12 hours on this. Here's a PR been 13 hours on this. Here's a PR. Absolutely. And, um, I think I saw someone from, um, Xiaomi coming out an engineer was working on that model. We had, which we had an open router, which was the, the Hunter model, which was revealed to be the new Mio model. And they also said to their team, okay, you're not using that enough. There's one request by tomorrow. Each of you who has not had at least a hundred conversations with the new model will be in serious, serious trouble. Um, maybe, maybe, maybe building on that because we mentioned Nvidia, um, and Jensen Huang before there's been some interesting price action actually on the, uh, uh, bit tensor site, bit tensor being up 46 ish percent on the month, but also I was up 18% on the month earlier on the day earlier today, crossed $300. Now it's back down at 200 to $280. And that had to do with, um, this clip coming out from Jensen's interview on the Lynn podcast. We're going to watch a bit of that on the show. There we go. Let me ask you a question about open source. So we have these closed source models. They're excellent. We have these open weight models. Many of the Chinese models are incredible. Absolutely incredible. Two days ago, you may not have seen this because you were busy on stage, but there was a training run that happened in this crypto project called BitTensor. Subnet three, they managed to train a 4 billion parameter Lama model, totally distributed with a bunch of people contributing excess compute. But they were able to do it statefully and manage a training run, which I thought was like a pretty crazy technical accomplishment. Yeah. Because it's like random people and each person gets a little share. Our, our modern version of folding at home. Exactly. Yeah. So what, what do you think about the end state of open source? Do you see this decentralization of architecture as well and decentralization of compute to support open weights and a totally open source approach to making sure AI is broadly available? I believe we fundamentally need models as a first class product, proprietary product, as well as models as open source. These two things are not A or B. It's A and B. There's no question about it. Bring any of the. A couple, a couple of takeaways from, from that interview for me. A, we spoke about how to use excess compute. We spoke about how to use our, even our DM credits, which we're maybe not using up on a daily basis. And BitTensor is obviously something we've, we've discussed in the show as well. It's honestly still an ecosystem that I need to spend way more time looking into and play, play around on the, on the subnets. It's something I haven't gotten around to do. So I'd also be curious to hear your, your thoughts on that. And then the, the other takeaway apart from this distribution of compute here is very, it's very, very interesting to hear Jensen not only speak to the need for both like proprietary models as well as open source models, but also how heavily he recently has been, has been leaning into, into open claw. And the necessity of OpenClaw. I think there was a, an interesting progression to observe here. So on March 6th, let's get that up on screen as well. On March 6th, he said, open claw may be the most important software release ever. March 16th, open claw is a new computer. Three days ago, in CNBC, he said, open claw is definitely the next chat GPT. And it's something he's been speaking about for, for quite a while that his vision for the future is exactly this AI companion and AI assistant to have. And this is basically also where I want to take the conversation a bit into if we look at this like companion assistant side. And if you look how open models, how proprietary models are fitting into that and kind of try to establish a bit of a framework for ourselves. Like what do we think needs to be an open source? What do we think needs to be more of a proprietary model? But also first getting your thoughts on the, on the Bitensor side of things first. So let's maybe slice it up like that. Yeah, I actually don't have strong opinions on Bitensor, but I see Funky's hands up. I know he's very intimate with the ecosystem. So I'm going to let him slide in here. Funky, go for it. GM guys. Yeah. I mean, I think the thing that I wanted to share is this, what we're seeing is incredible. And it's actually combines open claw as well, because I follow a lot of people in the Bitensor ecosystem. There are many who are running agents that are effectively doing the mining. So in Bitensor, you have a subnet, right? There's currently 128 of them. On that subnet, the way that the rewards work, which we had the first having for tau, now it only generates half a tau every block. 18% goes to the subnet. 41% goes to the validator. 41% goes to the miner. Now the miners are the ones that are doing the decentralized compute. And when you think about the world we currently inhabit, insofar as there's massive data center demand, we don't have enough compute to go around. We can't build these things fast enough. Look at the prices of RAM, everything else. I think this is why Bitensor will succeed in that world, because there's all this excess compute all over the globe. People just using it on their computers. And so the other thing that's really elegant about Bitensor is it's now created sort of this Darwinianism where it is feast or famine. Like if your subnet does not produce intelligence that is useful to the network, it dies off because the way that the tau flows to those subnets has everything to do with how successful you are. So that if you look at Templar, subnet 3, which is a legacy subnet that's been around for a while, if you can pull up on tau stats, you could probably see some of the growth in that particular subnet. It went bananas during that training run because all of the energy was being sucked into doing that. And what we're also starting to see is that many subnets are using each other. So think of them as like Lego bricks. One might do one particular thing, and then it's working with another subnet that does another aspect of what's required for AI. So I just think that this is really about to explode in the coming decade. I firmly believe, not financial advice, that like buying. I missed when Bitcoin was hundreds of dollars. I think Bitensor could be that now moving into the future. So this is definitely something to pay attention to. I think Jensen Wang is spot on. You're going to have a world in which both exist. It's A and B. But there's something to be said about open source and how it accelerates faster when you're using that crowdsourced intelligence. Yeah, I like that point a lot. I think that reminded me also of this one guy on Hugging Face. Basically, Hugging Face being this page where you see all sorts of open source models, customized, rebuild, quantized, content, meaning that they are repackaged for lack of a better word in such a way that they are easier to use on less potent hardware. And there is this guy who's been out there optimizing all kinds of open source NVIDIA models, other models to make them usable on Apple Silicon, on using MLX repackaging for that and doing something. Also, I think partially on rented, partially on distributed compute to achieve that. So it is exactly this interesting point where you have a ton of open source models. You have a ton of open source tooling that you also could develop or repurpose or make more accessible to others. But for the more complex things, you need so much compute that it is hard to do that by yourself. So obviously, one thing then becomes maybe like commercializing your compute, which you are not using. But then the other thing also might be just picking things that you find interesting where you want to give away free compute. And that, in my opinion, has to be local compute. It's not going to work in the sense that, oh, I'm not using all of my codecs. I'm not using all of my cloud code limits. I'm just going to give the excess away for other things because that's something that obviously neither OpenAI nor Entropic are built for, that everyone's maximizing their entire usage limit. So it is local compute. But nonetheless, super interesting how this is almost a necessity if the open source contributions want to keep up with what the prop models are doing, because so much compute is needed and it's going to get even more expensive. You have all of the things that we discussed earlier this week, oil prices going up, helium prices going up, LNG prices going up as well. All of that infrastructure, all of these materials that you also need to build semiconductors, to build AI data centers, is getting more expensive. And I just don't think that this is going to cause a slowdown of the evolution. It's going to maybe call more for that distribution of compute. Now, I also don't know what that means for Bitensor and not and what other systems will see, which will try to work on that and do exactly that. But curious to pause here, Brett, and go to you for your response to some of the things we said. Yeah, I was trying to think through all of it. What we're ultimately saying is that, yeah, these proprietary models, these frontier models require so much compute that it is not practical for any small to medium to just regular consumer developer to be able to compete with them because you just straight up don't have the compute to train the model. And what we're debating now or talking about is marketplaces, aggregation services for compute so that you can develop something in collaboration with each other to build something that can actually compete with these frontier models and actually create things that are open source, but also leveraging mass amounts of compute that no single or small or whatever amount of people can muster. And your conversation, that won't be a consumer thing. It'll be like a prosumer situation. Same thing as with Bitcoin. With Bitcoin, maybe it started out with consumer, like you have some people, but even then it was still really technical to actually contribute, to create a miner, to contribute to the network. And then now today it is definitely prosumer to industrial, right, to contribute and do that. And I would not be surprised to see the exact same path for this where it says, you know, now it's probably prosumer doing the BitTensor contributions eventually. And I would say there's probably maybe even a few people that are doing it as an organization or an entity that come together and say, our entire job is to contribute compute to things. And as more things become available to contribute to, it becomes a monetizable business. And that's just the thing to do. So just like you run a miner factory, those factories are already starting to convert to compute contributors, right? So yeah, it's an established path. What I'm curious, and I think it's kind of a question you just asked, is I am not fully aware of all the options that exist today, right? OpenTensor, BitTensor is one of them, as we can see. Are there more that I'm unaware of? Because it's likely to be a market of its own where there are competing entities trying to do this stuff because there's so much money into it right now, right? And there's demand. The demand is there. Yeah, I don't know. So do you know of any other competing compute marketplaces, which is what I'm going to classify this as? And I guess the question goes to Funky as well, if Funky is aware of any competitors. I don't actually. Something I want to look into, even for the selfish reason that I don't use all of my DM compute, I have the Mac Studio also not using the fullest of its extent, while at the same time I just want to get a couple more Mac Studios. So I need to figure that out as well. But you made such a fantastic point on the professionalization of this industry. If it is to become an industry, then I 100% agree with you that we're going to shift from this prosumer to, okay, industrialized mining, industrialized compute distribution and renting. Exactly the path that you described with Bitcoin. 100% agree with that. Funky, quickly checking in with you if you're aware of any other marketplaces to distribute compute. And then we're going to keep it moving into the next topic. There is one that's name eludes me at the moment. It was built on virtuals. I will find that name and I'll get it to you guys. Cool. Thank you so much. One more thing I wanted to get into before you're going to cover the headlines. It's basically building on this discussion already. If we take it from that open source versus proprietary angle, it is to do with OpenAI, who is acquiring the open source Python toolmaker Astral. Astral's most popular open source projects include UV, which is a Rust-based Python package manager, and a Python code formatter, Python type checker as well. So basically all of that Python scripting language that is very much used, not only in Vipe coding, but also like traditionally, is now becoming a part of OpenAI, a part of Codex's ecosystem. At the same time, OpenAI did a couple more acquisitions, but so did Anthropic as they have been fighting for dominance in this fast-coring market for high-powered coding assistants. Anthropic did acquire Bun, which is a JavaScript runtime, also more than 7 million monthly downloads. So it seems like that the things that once you start Vipe coding for the non-technical people, which you'll see all the time, you run Bun there, you have your install packages, you have your Python scripts that you're writing. And a lot of that initially having been open source tech is now being bought up, as well as the people are being recruited to the big AI labs. And it feels like you're watching OpenAI, Anthropic, even Alphabet Google to a certain extent, buying all the bits and pieces of the Vipe coder stack and trying to integrate these, more than the tools, because the tools they have been integrated, given that they're open source, but more than the tools, the humans and the intelligence, the minds, thank you, behind that into their own AI superlabs. And obviously Meta is playing the same game as well. So that is interesting to see that, to say the least, that fight for the human AI minds to get absorbed into the AI labs. Yeah, it's insane. I think about like you cited Google. Google did this notoriously across industries, right? They were buying up companies as they continued to grow and expand their business. And we're seeing the same stuff here. These guys just have so much money that they like know is, if there is a no, it's more ideological. It has to be basically just straight up ideological because like they're gonna come over the top with whatever you think you're worth plus two times two times a hundred, right? And just like that's what's gonna bring you in. So interesting. And I think with Python specifically, like the way that I had this described to me with from our CEO is, or at least the concept that I was made aware of is a glue language. And why it's a glue language for what we're doing is Python is actually not the most like efficient performant code, right? Like if it's trying to be as perfect as possible with strict typings in it, like I'll say it's gonna be doing Rust, it's gonna be doing fucking assembler, it's gonna be doing stuff that is unintelligent binary, right? Go down to the root. Like that's what these agents are capable of doing and would likely do. But they understand that the interface to a human, you lose so much readability with that that it is not what is done, right? So what happens is you utilize Python as a glue language and like they'll often showcase that to you so we can read it, we can interpret it, we can do whatever. And then if they can on the backend, they'll do, you know, they will do the Rust, they will do the assembler, they will do like the lower level coding because it is more efficient and creates a better end product for them or achieves whatever they want to achieve because maybe it's not possible at the Python level. That was, but yeah, so this plays into that, right? So like to acquire the people that are one of the most popular Python labs in the entire world. Yeah, makes sense. And they're gonna continue to do this, right? Yeah, 100%. Maybe a bit more of a personal question. How's your vibe coding stack in terms of tool usage, in terms of the frontier models looking like at the moment? Yeah, it's still the same as we did whenever we had the weekend workshop, I guess we'll call it. I still have all my agents containerized and I'm basically still just built on, I tried to decide if I wanted to do skills or if I wanted to do like actually in the initialization file. I still have them as skills, but it's containerized skills so they don't have to like search and think. They each have three or four skills max that they have in their container so that that's what they use and it's just based on best practices. Cool. Let's keep it moving. I would say we have 15-ish minutes left in the show. We still have three main headlines to tackle. So let's start with the first one, which is adding another layer to this AI fight for dominance. This is not a company there. This is more a country layer than anything else. And it has to do with three men, including a Supermicro co-founder, being charged in conspiracy to export NVIDIA chips to China. So Supermicro's co-founder and two other employees of the company, they have been charged in New York for allegedly violating U.S. export controls. That is according to Financial Times by smuggling $2.5 billion of NVIDIA's AI chips service to Chinese customers. As a reaction to that, we did see the Supermicro stock slide a fair bit. I think when Financial Times wrote that, it was 12% in after trading. And after our trading now, with the pre-market, the stock's down 26%. So there's a massive, massive reaction to that. Interestingly, there's another stock who is kind of benefiting from that, and that is Dell. Dell is up 5% on the day. But I think we had this conversation. That's the way I want to tie it in with that news, because on the news per se, I don't think there's too much to say other than reporting it. And that is this angle of Chinese models using American frontier models to train. We talked about this technique where you basically set up thousands, 10,000s of bots and ask models 100,000s of questions and then ask them to explain the reasoning behind them and try to distill the essence of how Opus, of how Codex is working and thinking. And use these learnings to ingest them into the own models that you're building. On the same side, basically all of the frontier labs or the frontier AIs on the American side also have ongoing lawsuits for having used copyrighted content as well. So that is maybe the software, the software dimension of this AI war or AI fight struggle for hegemony between US and China. And now you also see some of the hardware side of that, with these $2.5 billion worth of NVIDIA GPUs allegedly being smuggled as dummy servers under the pretense of being dummy servers into China to give them access to that hardware as well. Any, I see you chuckling along, Brett, any thoughts on that? I'm tying it to, so like this is mostly just around the thought of, yeah, these, these, it's not the thought, they proved it, right? These foreign entities coming in and asking questions to extract information on these frontier labs. I'm tying it into our conversation of how deeply subsidized those max accounts are and like, what is it, like 90%, 95% depending on who and what. So if you think about that and if I'm a malicious entity and I have, or I've managed to gain hundreds, thousands, tens of thousands of accounts and I am just peppering it, maxing those things out with questions to extract information. Not only are you extracting the proprietary stuff on the back end to like take that in and build your own models, you're also burning their compute and their gas at a 90% discount. So for, you know, the actual raw monetary demand that they're putting on these, these companies is a lot of money, we'll say. And then they're also trying to undercut them by taking their model and can create a competitor with it. It's a pretty, pretty dastardly. Yeah. Great, great point. Ultimately, ironically, I do think that this for us is like the, the retail consumer or customer of that. I think it is beneficial. Sure. Just increases the competition so, so strongly that we are able to get an insane amount of great compute for $20. or $200 a month, which otherwise certainly would not be the case. Yeah. And without that, bro, Claude and OpenAI could come to us and say, give me your firstborn child. And some people go, hmm, maybe, possibly, is that worth it? Because, because there's no price competition, right? Yeah. They could charge thousands of dollars for that and you wouldn't have people pay for that. If you've once played with the model and you've experienced the difference to others, you want to keep access to that. Yeah. Yeah. I've seen, I've seen some people on the timeline talk about canceling their, killing their claw, their open claw, shutting down their, their plans. And for me, man, even I'm on the, the max claw, I'm $200 there. You know, I have a few other subscriptions that are spinning up, but minimals. I think I've been like the pro Gemini and a few others. Worth it. The efficiency gain that I get and everything and how I'm able to contribute and build and do, worth it. No questions asked. Yeah. 100% agree. No discussion whatsoever. And I still think sweetest price point if you had to pick one in the $20 range is, has to be open AI, has to be access to GPT 5.4 because the amount of tokens that you get to $20 is insane. Whereas the $20, I mean, I'm on $200, all of them, but I have the $20 other accounts to test that. And $20 on, on Claude, you've learned through that to your five hour limits so, so fast. It's, it's insane. Let's, let's leave it here. Let's get into the, into the next one. Might look like an AI related headline at first is to do with DoorDash introducing Dasher tasks. So Dasher's can now get paid to do general tasks, not only the, a food delivery and DoorDash had a bit of a longer announcement and that ends by saying that we are also piloting a new standalone app where Dasher's can complete activities like filming everyday tasks or recording themselves speaking in another language. This data helps AI and robotic systems understand the physical world. Pay is shown upfront and determined based on effort and complexity of the activity already partnering with companies across industries including retail, insurance, hospitality, and technology plan to expand into more task types and countries over time. You also have some images to highlight how this is working. So you have your typical task for a shopper but then you also need to scan shelves in a specific store and then shop for items. So it's like basically giving you extra tasks on top of that and you go to the assigned aisles in the store and you scan all along the shelves using your camera and they're showing this from people not watching us on screen on a couple iPhone mockups to see how that works and that immediately reminded me of this rentahuman.di website that went viral where the idea was that there's a CLI, this command language interface tool that AI agents can use for things that they can't do because they are still lacking a physical body and when they would need the human in the loop to perform a physical activity they could pass it on to rent a human and get someone to do that and that did not really take off from reports that we've seen later on where nobody really was using the website despite it going viral and to me it looks like DoorDash who obviously has a gajillion of dashers working for them is integrating this at an industrialized level and clearly saying it themselves with AI and robotic use cases in mind so definitely wanted to report on that piece of news as well. I think this is great and now it makes me question you know what I wish they would have been doing this entire time while they had these people doing it yeah throw in a task for taking a picture and like putting it in a Google review so that like I can see what the building looks like what the different section looks like do they have like how are their shelves like I would have loved a shelf picture of like certain sections you know if they even have the thing that I'm curious about so whenever I go and do stuff like I wish that would have been introduced before but now yes like this does feel like the logical extension of that rent a human thing and now I can see if like because I thought I saw somewhere that they spun up an MCP for this is that was that am I hallucinating if I'm hallucinating what they should do is make sure that this stuff is have a skill have a MCP have a CLI so that agents can I haven't seen that yet I'm pretty sure that's coming sure and then you do that so that agents can spin up this stuff and actually order the things themselves and order things for agents that is like a like what you're talking about like the physical things the pictures the recording of themselves the you know whatever so I can do stuff like yo Allison can I my Allison's my girlfriend was like I got you this for your girlfriend but I talked to my agents like hey order Allison a thank you song from the mariachi band down in Mexico and go and get it spun up for this okay have them record a video and then you know send it over be great the use case user script is exactly one for the mariachi band that they've been highlighting right they've been saying businesses need to know what actually is on their shelves whether the layout at another location has changed since last week but getting this information at scale in real time is a challenge and it's a challenge day they spent over decades solving that so it's exactly what you've been describing yeah the businesses need that right like it's like I would I'm sure Walmart wants to make sure that all of their their stores or are compliant right and instead of setting a physical person to it every single time and having that person hired like bro outsource it have someone upload a picture every week every store do whatever and that's like the compliance check think about Mr. Beast and his feastables chocolate for example and he has maybe a specific layout how he wants to look like in Walmart but he has no way of checking how that looks like every week in every Walmart nationwide and if that is compliant to the standards so absolutely agree with that and I definitely think we're going to see a CLI or MCP for that talking about MCPs and going into the last headline for the day maybe this is what I saw yeah probably this is one for you to take away bread cloud code channels what do we need to know on that yeah oh yeah I forgot we were doing a lens so yeah this is cloud code introduced channels and they've released two initial ones one is discord and one is telegram for the ability for you to talk to your cloud code or cloud instance directly from those two social services and the obvious thing to take away here is oh my god this was like one of the big unlocks for open claw what does this mean for open claw like all of the points that we made previously still stand there are benefits to open call that that cloud code does not have but this is a pretty straightforward integration for both of these two things and it should be very pretty much one-to-one with your interface with with open call so if you're familiar with that you now have that ability through through cloud code and the thing that I want to remind people of and why the biggest edge for the flow the architecture of this is you're still this is basically this is just your interface again right this is a direct line to the llm which is meaningfully different than your open claw instance where you chat to your computer or to wherever that is installed and then it is forwarded to an llm and that process allows for you to get a little bit more efficiency out of it but it's still it's just a sign man it's going to engulf one-to-one everything that open call does at least like I would say up all pillars it will it will engulf all pillars of open call into cloud code officially yeah it's a matter of time and as you said like the limitations they they still stand for example one of the limitations that is to do with telegram spot api is that there is no message history no search if you send a photo to that thing you have or your agent has to download it immediately it can't recall previous attachments that you have sent and open claw by its architecture is working around a lot of these limitations so that you as the telegram user don't notice the limitations of the telegram api it's not something that hot code or entropic is not capable to catch up to but as you said to me also the main differentiation point at a time where dropping is doing that codex certainly will follow with that like codex started to copy other things or introduce other things like one of the recent codex app releases had the subagent features introduced which where you can spin up agent teams in the codex app which is something that cloud code has allowed you to do for I don't know weeks a month two months maybe something around that time frame and then if all of these things and talk to your telegram if all of these things can live remotely on your Mac mini or Mac studio whatever and have sessions running there you still would talk to all of the tools all the time so I still need that is where I want to position my my open claw thought and an orchestration layer because they want then to be able to talk to one single entity which ideally is my open claw which has a good enough memory and a good enough ACP agent context protocol tooling that it's able to pick out whatever tool it thinks whatever CLI whatever element things it needs to and as you said as a model being way better at prompting it and being way more efficient and more context preserving and talking to that model than I would be as the the human as the human in the loop who maybe also needs explanations what is going on what does that mean and that model to model talk certainly does not need that yeah and it spans all and as you as you're going through this I'm a big believer in like the collective like I'm not like a woo woo kind of guy but I have just noticed that it's impossible for single entities no matter the size to be able to keep up with the full weight of the world if the full weight of the world is able to like coordinate around things and that that is the follow on to like oh Claude is trying to kill open claw again OpenClaw is agnostic I I maybe I maybe go back on my stance a little bit that that that Claude will be able to catch up and integrate all of the things that open claw does just because man the collective has been made so much stronger with each of us being able to harness individual models and as we get to a point where you can start doing local LLMs and like you actually start having better proprietary stuff I I'm going I think I might go back on a little bit where say like I don't know that they can ever keep up with the collective I don't know I don't think they can so like it'll never be one-to-one but it will be they will pull the biggest things and they will be very very well curated products for their specific instance of a model or of a you know LLM but I don't think they'll ever keep up so what's what's the winning what's the winning ingredient there because then I'm thinking I agree with you and say cloud code or now's research with Hermes comes up with a killer feature that everyone in OpenClaw wants to have someone certainly either gonna port that build that and integrate you've seen that happen before you have to like the builder community that is the most massive one when it comes to OpenClaw being such a enormous GitHub 300,000 plus stars and then you have the the selling point so you have this type of distribution then the selling point of Claude and OpenAI Codex remains that they sell you the compute on top of that at a pretty pretty attractive price they need to get the subscriptions because otherwise people will not use that because you get a nice package where you get the compute and the tooling and at a time where the big three Gemini OpenAI and Enthropic are still fighting for users and have heavily heavily subsidized compute that then tells me that they have to keep this game going on for a while no matter how unprofitable that is because this is now the only standout advantage that remains for them yes they do build the models but you could also use them via API elsewhere and do not need to have an OpenAI Enthropic subscription which obviously would be good for their business as well but given that there are also platforms who want to have like hundreds of millions of subscribers that as well that then means that I think we'll keep to see subsidized compute maybe for a bit longer than I was initially expecting that to keep on going yeah I mean it goes with our other conversation too maybe there's a there's a point of convergence where this these shared compute markets gain enough traction to where it no longer is necessary to run through a a anthropic or a open AI because maybe there's a flow where I have I can piece together something work with my open cloud to piece something together and say get the compute from open tensor get the get the model from open router get the whatever and they're able to construct a workable cost efficient way for me to operate without having to go through one of the big big shops yeah that and you also said previously a lot of the things become so agnostic you don't need to rely on a certain infrastructure if if that building community is big enough you will you will get that feature you will be able to implement it yourself I've seen this yesterday last thing I want to share before we'll need to get into into the neat giveaway and wrap up the show is this post by Emozilla I'm a big fantasy nerd love my fantasy novels and Emozilla wrote that they fulfilled a long-standing dream of theirs building an AI system that can tell a compelling story did that with Hermes agent did that with um cup parties which you mentioned before other research as well basically outline the world have character design world building plot flows and then use other research as the refinement system until that reaches a state where it's good do the same for the covers for the actual typography in the book and everything and basically build all the research loop using Hermes that can then iterate by itself and autonomously write this this book which is pretty good it's called the second son of the house of bells given that this is a fully AI written thing it's free to read definitely take a look at that because I think that also shows you a you could use that and immediately tell your open call build that I want to do the same thing you don't need to use Hermes for that nothing against that product but it goes to your point of being agnostic and it then also goes to the point that other research has um is also completely agnostic like you see other research as self-improvement loops in minimax is 2.7 model you see it being used for prediction models you see being used for creative writing so you can build and piece together the stack that you want to work with and then I think what really remains is the distribution that you have whether that is from having a community like open claw has or having a ton of subscribers like the the big frontier labs have but everything else becomes composable and is is completely agnostic any any final thoughts on that have you have you tried using Hermes yet I have yeah thoughts it's good it's a good product I run into issues as well after some time as I do with open claw but that is to be expected it's probably the easier setup significantly easier setup and like getting into being productive that new when that new users then then open claw then again I thought like I put so much energy into open claw I want to stick with that then I thought is this a stupid thing to say because I'm getting like to switch systems but then I went back on the thought to be like it's just the amount of tooling that's being built for open claw and the speed how they keep delivering is way way faster so for the primary infrastructure I'm gonna stick with open claw but I'm also like monitoring the the Hermes ecosystem pretty actively and still have it up and running just to see is there anything in there that I could like take away from and implement for my own things yeah my sense is generally like yours where I the winners win always and the momentum that open clause built up is going to continue to put it out front for foreseeable future but I also don't want to like put on blinders and get complacent so I'm trying I haven't downloaded myself or used it but I've been monitoring the discussion of people who have been trying to compare the two yeah and what I what a great segue that is a winners win into the last thing that we'll do show the week and that is to get to our Yita board to see who is gonna win $500 walk away with that for the weekend we have 73k in volume the usual suspects being on top of our leaderboard at being a funky with his 50k in volume a phonics 12k Atari God with almost 11k in volume we did set up the wheel of names which we are gonna spin right now and see is winning $500 and just a reminder for everyone on YouTube play with amounts that you're comfortable playing with and this is phonics in between two fun keys taking away the $500 congrats on that and that also brings me to the end of today's show bread it's been a pleasure to do just an hour of AI nerding with you no difficult topics no geopolitical crisis just talking about the things that we've been really enjoying a lot over the last months or so so thank you for that thank you also to everyone in the audience thanks funky for your contributions as well we will upload the show as per usual to YouTube to Spotify Apple music we'll have clips coming out and we're also gonna be back as per usual on Monday 7 a.m. Eastern time it's gonna be the last week where you're off by one hour for non Americans we're really still starting one hour earlier for people who are not on Eastern time or not an American time and then we Europeans will also follow through with summertime and then everything is gonna be back to normal but in the meantime have a great weekend and see you back on Monday you