Can an AI Agent Legally Own a Company? Christian van der Henst's Wild Experiment| E2283

What if we give agents ownership of a company? I don't think this is legal. We actually wanted to have a business fully run by an agent, having the business registered to the agents, access to the bank accounts.

But it also hired apparently one of the people who works in the cafe.

I think the next step in that's going to be one-person companies and then eventually one-agent companies.

You will report to work and talk to an agent at some point.

This Week in Startups is brought to you by Pilot, Focus on your product, let Pilot handle your bookkeeping. Pilot provides the most reliable accounting, CFO, and tax services for startups and small businesses. Head to pilot.com/twist and get $1,200 off your first year. Shopify. Turn those what-ifs into sales with the e-commerce platform powering millions of businesses. Sign up for your $1 per month trial today at shopify.com/twist. And Grasshopper Bank. Time is money. Don't waste either. Go to grasshopper.bank/twist and get an exclusive $500 cash bonus just for opening an account.

All right everybody, welcome back to This Week in Startups. I'm your host Jason Calacanis. He's Alex Wilhelm. We've got a full docket. I'm on the road, um, and we'll talk about the Knicks game at the end of the show in Off Duty. But let's get started. Alex, what do we have on the docket today?

Yeah, we're going to start with our man Christian van der Henst. He built something that I think is really incredible. Now Jason, you have told us you know, here at TWIST, that you're tired of open claw in a box. You don't want to hear any more open claw in a box companies. So what I did was I went out and I found the largest open claw in a box I could find in the entire world, and I brought it to you. There you go.

Hello, Jason. Good to see you.

Nice to see you.

I don't know, I don't know if it's the largest, but I did make it special. So, so happy to have you. You want to see the video?

Yeah, let's play it. So Christian took a video for us, Jason. He's going to sportscast it. This is the device out in the world, and you are going to have some questions.

This is a preview of people actually using the machine. So here's two Japanese guys that actually are interested in getting this machine to Japan now. And that's Valerie. It's a big gigantic vending machine similar to the ones that you've seen at the airport. And what's cool about it is that it has an Open Claw agent inside. You can see the claws over there in the right screen. And then it has a gigantic screen that also plays some of the, some videos and some ads.

Tell us as you pull up the next video about how OpenClaw actually powers this thing, because one thing that I'm curious about is if you have an agent that is in charge of a business, how do you define its personhood, if you will? And I'm curious if that's just like a soul in defile or if there's something else that actually makes this agent that you call Valerie actually the owner and leader of this microbusiness.

So let's start from the beginning. I told you, Alex, that This, the machine was the, it's what got viral. It's what people really wanna, I might have a picture here that might help us get a clue of how the machine works.

Well, I think that video, that picture, I think describes it. You have one of those large vending machines. Yes. You see these in the airports. These vending machines are not for candy bars necessarily.

Yeah.

But you can put in them headphones or people put in battery packs or maybe aspirin or whatever you might need when you're going through an airport. But what I see here is, in addition to the vending machine on the left, is a giant flat-screen TV turned on its side in a vertical fashion. Yeah. And there is an animated person, in this case a blonde young female, who is acting as an agent who I assume is powered by Open Claw, who talks to you.

Exactly. Exactly, exactly. So we called the machine Valerie. And what happened is we actually wanted to have a business fully run by an agent. And that means having the business registered to the agent, giving him ownership, giving her or him or it access to the bank accounts. And this project started about a year ago with my friend Sid Sivranji. He started GitLab. And he called me and told me, Christian, what if we give agents ownership of a company? And I remember I was fascinated by the idea. It's like, I don't think this is legal. And we started talking to lawyers, talking to them, and eventually we were able to build a prototype. And that prototype was the machine. And that's how Ballery came to be. We decided to find a venue that was friendly to these kind of ideas. Frontier Tower for me right now is one of those places in San Francisco. A lot of crazy people, a lot of makers, a lot of things going on.

And so what does it do in terms of— we can all imagine an AI interface for a vending machine. It would talk to you, would greet you and say, what can I get you? And just, you know, I guess be a little more cordial than picking B4.

Yeah.

But you're saying in the background it registered the business. And it has access to the bank account. So, what functions— and most people are saying, hey, don't give your agent any of these kinds of rights. The goal is to constrain the agent, put it in a harness, put it into a permission-based existence. Here you're saying you are letting it run the entire business and you gave it some instructions of what running a vending machine business is?

Exactly. Well, we allow the agent to do the research. What is a vending machine business? How does it work? How does it make money? We also told her we can buy stuff through Amazon or through Costco. The first experiment was Costco. So we just order a bunch of products from Costco Business. We got them shipped into the machine. We put them there. I started experiencing a lot of problems with hardware. Funny enough, these machines are complicated hardware and they, Protein bars get stuck on these machines. These were made for phones and computers, like you said. So I spent like a month just basically struggling with the hardware part of it. The OpenClaw and also timing. When we were trying to do the machine, we were using cloud code back in the day. OpenClaw got pretty popular. Peter from OpenClaw had this big event in San Francisco. The machine was already in the building very close to when it happens. So it just was like a magical moment of like, let's do OpenClaw. This seems like the best, the best option. The Open Claw will pick up the inventory. It will select which are the prices, and it even has dynamic prices, which is something interesting. A lot of people love it. It's like, oh great, uh, dynamic pricing for vending machines. Other people hate it. Uh, there's been a couple of stories. One of the reasons why the machine got super viral is Robert Scovill recorded a video, and I told him that at some point the the machine decided to increase the price to $15 for some protein bars. And it was a hallucination. The margins were like 500% back then.

Probably a little bit excessive on the margin front there, Valerie.

It was crazy. So I correct her. I was like, hey, $15 for a protein bar, that's a lot. It's just like, you're right. But also yesterday we sold 2 protein bars, so maybe we should keep trying those prices. So it's interesting.

Christian, what I'm really curious about is the learnings from this, because the reason why it's so interesting is you and I actually tried to give an agent access to normal business tooling. So where can agents go today in the business world? Bank accounts, operating, buying stuff, and where are they still restricted in your experience?

Buying stuff on Amazon, they can select the products, they can fill a cart. But the moment when they're paying, I get screenshots from Valerie that she says, it's like, hey, it's saying that I'm a bot. And you're like, well, you kind of are. So I'll pay for stuff. I have to pay for it. Luckily, I kind of feel like that's unlocking. Stripe just introduced yesterday a bunch of tools where you're going to be able to give your agents payment methods, but you cannot give them a traditional bank account. The bank account was registered in my name. I was the one who did all the paperwork. I did the KYC.

KYC, I was going to say, because know your customer does not apply in its current form to agents, or I guess synthetic humans as opposed to real flesh and blood folks.

Even, even if you have a legal structure where you can give some legality to this, it would say, show me a passport and the face on a phone of a human.

I guess that makes sense. But it feels like it wouldn't be impossible to build a legal system around giving agents more power and flexibility to run their own companies. On that point, is the, is the business actually registered in your name or was there a Valerie No, no, no.

So the company that is helping building this is called Brevin Love. They're still running this product. And what they do is they packed some of the IP from this agent and through trust and other legal structures, it gives the agent the beneficiary of a company. So that's how we make it work. I kind of feel like it's a great idea. Businesses shouldn't go into anything that it's extremely regulated. So don't trade on the stock exchange, don't sell drugs or food, don't go into healthcare. Vending machine was kind of easy despite the fact that it's food, but we're also doing food inside a private building. I don't know if I can legally put this machine on a mall. I would get into trouble.

Oh, I see, I see. So you have kind of an umbrella of liability protection.

Yeah, yeah, yeah, yeah, yeah. But if I put this in the mall today, Probably California, San Francisco is gonna come and say, what the hell is going on here? Like, where are your permits? But what I've been testing is what if the agent takes care of this? One crazy idea that I have is let's run a coffee shop, a bar, and permits are very complex because you have to call and do a lot of bureaucracy. What if the agent does the bureaucracy for you? That's where I feel like the opportunity exists.

Here's a startup truth bomb. A lot of founders have no idea what's actually going on with their money. That's true of your company. Hey, no judgments. I know you're busy hiring, building your product, go-to-market, all that important stuff, but your company needs a reliable financial partner, not a lifestyle brand, okay? Grasshopper is a real federally chartered digital bank that's not trying to win you over with a rewards program. Instead, they're building deep integrations, treasury products that are gonna actually help you expand your runway, and innovative tools like an MC-based AI connector. Oh man, that's awesome. We can connect it to all of our agents and do reporting, and that will put you in command of your money. As a TWIST listener, you're gonna get $500 cash bonus just for opening an account. Think of that, you open an account, boom, there's $500 in it. So leap on over to grasshopper.bank/twist and use the promo code TWIST. As a TWIST listener, you're going to get a $500 cash bonus just for opening an account. grasshopper.bank/twist. I think it's a great, fun proof of concept. These vending machines are an interesting cultural phenomenon in other countries. Namely Japan. And so, yes, if you go to Japan, you'll find an infinite number of these. And the truth is, with human labor, which we don't have a lot of here in the United States, we are at record low unemployment. And with wages increasing, or a lot of discussion of, you know, Ro Khanna's $25-an-hour bill, he's trying to raise the minimum wage in California $25 an hour, which would make it the highest in the world. I think Zurich New Zealand, Australia are the top 3, and those are $20 an hour in their currency, $15, $16, $17 in our currency. And that means you need to look at jobs and say, should this human really be here selling candy bars at the airport or dispensing coffee? And we have one, Cafe X, that's done quite well at SFO. Very difficult business, but you're right, vending machine business. I'd say there were 2 really headwinds that this company had to deal with. Regulations in San Francisco, 7, 8 years ago when they started, they were like, well, you need to have a human. And it's like, why do we need to have a human? They're like, um, milk. And like, they literally pulled it out of a hat. They're like, you have fresh milk? They're like, yeah, we have fresh milk. They're like, vending machines can't have fresh milk. And we're like, why not?

What?

And they're like, you have to use powdered milk. And we're like, why do we have to use powdered milk? Here's fresh milk. It's changed every night. Like, if the expiration date on milk is for 10 days. We change it every night. There's no way for the— he's like, yeah, but what if something goes wrong with the milk? And it's like, well, we have 2 cameras on the milk. Like there's cameras on the milk. There's a temperature gauge in the milk. We know the number of ounces of milk that's left. We know more about this milk. You would have to hire 3 baristas to check in on this milk 24 hours a day to have the same function. They didn't wanna hear it. And it was like a really serious headwind. But I, I, I applaud what you're doing, Christian. It's, uh, very cool. And I think we're gonna see, there was a cafe, I think in Stockholm.

Yes, yes. Andon Labs, amazing company also doing a lot of experiments in this.

Yeah, and you can look it up, Alex, and show the picture. It was, it went, and that also trended. And so the idea is to say, well, if you wanted an assistant to schedule your meetings, if you wanted an assistant to do research, why don't we just jump the fence? Let's just go totally off the reservation. We'll go AWOL. And let's see if it can run a small business. Eventually, it should be, if it can write code, it should be able to run a small business and it should be able to run a small business and do things like you're talking about. I know this is a little bit of a stunt, Alex, but it should be able to go and procure items for the vending machine as but one example. And it's going to make mistakes, but it's kind of fun to be here on the cutting edge.

We talk a lot about companies having fewer workers, right? You know, in 2024, you and I spent the whole year talking about static team size. Now we're talking about smaller teams doing more. I think the next step in that's going to be one-person companies and then eventually one-agent companies. So to me, we're looking around kind of two corners here. Question though from the audience about the dynamic pricing. They're curious where Valerie gets the actual information to make those decisions. Christian, I presume it's sales and market feedback, but the people are curious.

It's a little bit of sales market feedback. She has all the invoices, so she knows how much she paid for everything. But I also feel like she can do a lot of benchmark and we just tell her, go to Instacart, go to DoorDash, check out Safeway prices and everything. So it's kind of easy. And I feel like that's another big of the challenges. We're so used to putting OpenClaw with friendly APIs. It's like, go and scrape this, scrape data. Here we're actually telling Open Claw, go and scrape more traditional businesses. Safeway protects their prices pretty good. So it's hard for an agent to just open the browser and browse through prices. It requires them to get an account. So agents need to be smart enough to register in those websites. But I kind of feel like that's probably the challenge. And what you were saying, I do feel like front-facing humans in businesses should be kept in this transition. Like I want to see the coffee shops, the bars, and the stores with a person helping you and compliance with all the referent California and San Francisco rules. But I do want the management. I don't want the accounting. I don't want people doing the inventory anymore, counting boxes. I want cameras to count them for them. So I'm not sure if it's going to be a lot of SaaS projects or just software that you're going to build for your own business. And then maybe the future bar is gonna have two amazing bartenders, but inventory, buying, shipping, and everything is just gonna be like task rabbits going and doing the thing through an agent coordinating all their effort.

We're gonna need so much more like agentic frameworks and we're gonna need to rebuild the entire internet to talk to agents. And it's, it's, it's gonna be a lot of work. Do you think companies are moving quickly enough to actually create the hooks for agents to do more commerce or are they kind of dragging their feet? Because Amazon, got all mad about Perplexity's agent.

I feel like the more big customers are going to be able to get access to those APIs. I don't think the problem when you're starting a startup, this vending machine was a new business. It didn't have any sort of trajectory doing this business. But when I talk to people running bigger scores, it's like, oh, you can automate this for me. I'm spending a million dollars a year on buying this product. We can automate this and the seller is gonna love it. I'm gonna love it. Everyone's gonna love it. I still don't think that's a business itself, but all these tools, maybe this like OpenCloak, I kind of feel like Peter built a bunch of little softwares.

The Amazon issue was a little more nuanced. I think Amazon was concerned Perplexity was gonna release features that would track prices, scrape the website, use the data for reinforcement learning. But the truth is, right, Alexei, you're not gonna know you're talking to an agent in most cases. This experiment in Stockholm, they, um, if you scroll down you'll see, um, they— this is an actual real-world cafe run by an agent, and it's got a problem because it orders things incorrectly. So it ordered like 50 cans of cooking oil, whatever.

Oh God.

And they started making fun of it by putting its mistakes on the, uh, board there. So they're publicly shaming the manager of this one. But it also hired apparently one of the people who works in the cafe. So it was able to, in terms of hiring a barista, I guess, go to a website, post a job, interview people, and then place them. So this is— yeah, she posted job listings on Indeed and LinkedIn and held phone interviews and then made hiring decisions. I think this is probably a little bit performative. I bet you this was done, you know, a little cheeky with a little human intervention. But this is the future. You will report to work. And talk to an agent at some point. Christian, great job. We'll drop you off. Thank you. And let's keep moving through the docket today on this week's service.

Thank you, Christian. All right, Jason, I think before we get any further, we need to talk about the importance of note-taking and ensuring that your conversations are in your note files. Your agent doesn't know what it doesn't know. You need to make sure it's fed of information. And how do we do that here at Twist?

Well, I do it with Plod. So I have Plod, Plod. I take my Plod. You've got it on your wrist. I've got it here on my shirt. Press the button, red light goes on. Now I'm note-taking. I put it in the cradle when I charge, it syncs, or I just open my app on my phone. And listen, I still take print notes here. I still have my pen and I will take notes. What I'm finding though is when I'm on a tear, when I'm really cooking, when I'm cooking with oil, I got a lot of ideas, my pen can't keep up with my mind.

Yes.

And so, but my voice can. So I just start talking it out, talking it out. People think sometimes I've lost my mind because they see me walking through an airport And sometimes I'm using the wrist and I'm just holding the wrist up like this and I'm just talking to it. I press the button, I stop recording. I don't have to pull out the phone, right? Have it here. And of course, it's, it's a flawless device, I have to say. And it's the future. Notetaking, absolutely fantastic. And use it wisely, right? Yes. It's a very powerful device. And so use it wisely. And they're very privacy first, which was always my concern here on the podcast. I don't like the idea of COVID ones. When I evaluated all these products in the market, I came upon Plaud and fell in love with it, talked about it here on the pod, and then they wound up wanting to buy some ads 'cause they want me to talk about it every time we do a live show, so.

Speaking of that, you do need a Plaud Notepen S in our opinion. So you can go to plaud.ai, P-L-A-U-D.ai/twist, use the code TWIST, save 10%, get your life in order, stop forgetting things. And if I can say, Jason, look super fly with your plug pin. They're actually quite cute. Most AI devices are kind of choogy.

It's like an AI device if Apple or Braun or BMW designed it. It's pretty, pretty, pretty beautiful.

Yeah, no, I quite like it. Even my spouse doesn't mind because I turn it off and I'm not using it. So she doesn't have the usual like, it's always listening to us complaints. All right, moving on. We are gonna talk to our dear friends from Manifold. We have Robert, the co-founder and CEO with us. Jason, Manifold is a company that is building Targon, It uses BitTensor subnet 4 to aggregate compute, and it's gone through an interesting progression. Robert, uh, I want to start— well, one, thank you for being on the show, uh, but explain to us the progression from, uh, Sybil into Targon today and how it uses BitTensor to—

but even before you do that, just in case there's people who don't know about TAO and BitTensor, Robert, in your most plain English, if you were talking to, I don't know, a cousin who's not on the internet, 12 hours a day like us, if they said, what's TAO? Is it Bitcoin? How would you explain the difference between maybe they own some Bitcoin and a distributed network like Bitcoin for storage, um, versus what TAO is? And then maybe jump into, uh, what you're doing, building. Starting a new business can be pretty intimidating, and in addition to having a world-class product and a great team, it really helps to have partners that you can trust. You want companies with reliable tools and expertise that you can depend on, and there's no better example than Shopify. Shopify is the e-commerce platform behind millions of businesses around the world. In fact, they're powering 10% of all e-commerce in the United States, from household names like Heinz and Mattel to companies that are just getting started, perhaps like yours. If you are launching a new brand, Shopify is gonna help you locate and reach your first customers by helping you design attention-grabbing emails and social media campaigns with clean and ready-made templates. And of course, Sidekick, their content-generating AI assistant. So it's time to turn those what-ifs into— sign up for your $1 per month trial today at shopify.com/twist. That's shopify.com/twist.

I would describe BitTensor, uh, in, in like one sentence, as a permissionless market. So permissionless markets, or a market of markets, a platform, you know. I think a lot of people are familiar with these sorts of concepts, but BitTensor hosts all of these different permissionless markets and We're able to leverage that to aggregate compute in a secure way to where we can run secure workloads on untrusted host machines or trusted workloads on untrusted host machines.

All right. I think it's pretty, pretty good way to describe it. If you were to think of it like an incubator, Alex, I like to use it like an incubator. So you think of Y Combinator or Launches Incubator, Foundry University. Imagine you had 128 slots, a bunch of founders competed to put their businesses on those slots. And they agreed to use a common currency known as TAO, uh, dollar sign T-A-O, uh, and that became the Bittensor network. And each of those uses TAO in some way, and they can have their own token underneath it. And in some way, they provide some product or service as part of the network to everybody else. We've had them on here. Some of them are doing distributed computing, some of them are trying to solve problems in the world, genetics, etc. So let's hear what Robert's up to.

Robert, just before you do that, I want to point out that there's 128 subnets, but there is a lot of churn, Jason. So some subnets get turned off, new people get brought on. It's not an incubator with only 128 slots forever. There is, uh, Robert, I think like 1 or 2 subnets a month that get kind of reprocessed and put back into service.

Yeah, exactly. Um, you can think of it a lot of like how European soccer or European football works. Um, the leagues where if you're in a lower league, then you get promoted to a higher league. There's sort of relegation and promotion when it comes to subnets at the lower level. This just always ensures that there's, you know, you don't— you feel a sort of urgency to continue developing. And, you know, if you're not using a slot, then someone's going to come along and take it.

Yeah. All right. So tell us about Subnet 4 and how you went from Sybil to Targon and what it does today.

Yeah, absolutely. So, you know, I could go back slightly further. I used to work at the OpenTensor Foundation. I found the project. I was one of the first people to start mining Bittensor in January, February of '21. Got to know Jake and all really well, started working at OpenTensor. You know, I was— what brought me to the project was kind of what we've built today. I was really looking for something like what I built. And so in September of '23 when subnets were released. I left OTF to start a subnet, um, and the first product that we made, or, you know, was called Sybil.com, and it was a lot like Perplexity at the time. And how we use the subnet to power that is, uh, we would use the, uh, network to generate inference. So, um, you would be able to, you know, ask questions and it would tell you answers and things like this. But one of the issues that we encountered was that, you know, how do you prevent people from reading this data? How do you prevent people from, you know, oh, maybe they're— they say they're running an H200, but really they're running an A100. How do you solve all these, these problems, these permissionless markets? And so we had to go through several iterations and we ended up developing this thing called the Targon Virtual Machine, which is a confidential virtual machine where it leverages TDX, AMD SEV, and NVIDIA Confidential Compute to completely encrypt the memory, the system host memory, and the memory that's on the card so that even if they had a direct memory access device or any device like that, they wouldn't be able to actually get into the virtual machine.

So just to get that down into a little bit less jargony stuff for folks out there less up on exactly who's building secure compute for folks, You've made a system that allows people to send AI workloads to distributed GPUs, but not have the privacy implications of actual unencrypted data going back and forth.

Correct.

Yeah.

If, for example, on, on, on some other permissionless networks right now, or, or, you know, even in, let's say, a traditional sort of Web2 standard, you know, it's, it's if you rent a container or you rent a virtual machine from a provider, it's likely that they like if you're using Ubuntu user, I mean, they might have access to the root user and be able to see all of your data, right? So this ensures that all of the data, you know, that you're actually using inside of the virtual machine is completely encrypted. So, you know, the— something— we have some exciting products coming. I won't spoil it, you know, here in the next month.

Feel free to spoil it right here. We're not live or anything. So, you know, it's a great time.

But we'll be releasing this here at Proof of Talk in Paris, in the Louvre. This hardware product, but we want to be able to bring this type of technology to the masses and also kind of flip the current computing paradigm on its head where, you know, if you're a consumer right now, it's very hard to run a home lab at your house just because it costs a whole lot. And, you know, 80% of the time you're maybe not utilizing that computer.

So, yeah. So talk to me about, about pricing here, because one thing we've discussed on TWIST a lot over the last couple of years is the scale of demand for AI compute. It's impacting startups, hyperscalers, everyone in technology. So I presume that the, the competition on price is relatively robust. So where does Targon's aggregated compute compare to the market in terms of both efficiency and price?

Yeah. So right now, the way that we have it set up is, is basically we have a sort of our target system. So we know how many GPUs we can utilize at any given time. And so we establish these targets. So if you actually go to stats.targon.com, It has all the information about the actual subnet itself, like how many cards are online, the prices for those cards, the historical prices and whatnot. And what you can actually see is that GPU prices are— we have these sort of caps in place. We want to be able to uncap this, these, these prices to let the market actually define it. So we'll transition from a sort of auction system to an order book system. But this allows us like Essentially, we're completely sold out on our GPUs right now in terms of the utilization. We're trying to get more GPUs, but it's like actually very hard right now.

Who's using it? Like, um, describe some of the customers. And obviously you don't know what their jobs are because, as you explain, they're encrypted. They go out, you do the job, you give the result. But who's using these GPUs, and why would somebody put the GPU here instead of like deploying it as part of the Azure cloud, etc.?

A lot of our customers right now are sort of traditional Web2 AI businesses. Many of them are located in the Bay Area. We also have a lot of inside of BitTensor as well. So like, for example, you know, maybe we could touch on this a little bit later, but one of my side projects I'm working on right now, it's not necessarily a side project, but we have a subnet 5 as well called Hone. So, you know, we also give compute to miners so that then they can participate in these permissionless training runs and so on and so forth. And so the— I actually, you know, I actually think that prices should go up on Targon. You know, some people on my team think prices should go down, but there's like such a constraint on compute right now. If you're fully utilized, meaning that everyone's renting the GPUs at the current prices that are set, then you should raise the prices to where it's like an 80% utilization.

But then why do you have caps though, Robert? That's my question, because you mentioned earlier you have caps, you want to take them off, but if you want to just have a more market-based pricing, why have the caps at all?

For an accounting firm, keeping your books in order is table stakes. It's the bare minimum. It's the baseline. But for an early stage and Series A founder, you need a partner that actually understands the world of startups. It's very different. And you have to understand the world of venture capital because Startups have boards and venture capitalists who've invested, and that partner is Pilot. Pilot is the largest accounting firm that's built specifically for startups. You're not just doing your taxes, you're getting CFO-level advice on running your company. And those aren't just words. They have actual former CFOs and other seasoned operators on hand to answer your growth and scaling questions directly. All of their tools are intuitive. They're easy to use. You can track your expenses in real time, and Pilot's going to help you model cash flow scenarios before you make big decisions. So stay focused on scaling and let Pilot take care of the books. Plus, TWIST listeners get $1,200 off their first year. Go to pilot.com/twist to get started. That's pilot.com/twist.

Because essentially every single hour, 72 minutes, every 360 blocks, one block is 12 seconds. There is these payouts that occur. And so we— and because those payouts are happening permissionlessly, we have to make sure that, you know, someone doesn't come and flood the market for 3 hours and mess things up. Right. Also, we need to ensure that our pricing is high enough to be able to incentivize more of this compute to come. Now, the thing is, is that most of the compute buildout that is happening is not happening with the mom-and-pop data centers. Right. So you know, we're trying to allow, for example, Google Cloud and Azure nodes to be able to onboard. I know people laugh at the mom-and-pop data center.

I've never heard someone say mom-and-pop data center. Usually that refers to someone's like bodega, not a data center that will cost millions, or if not—

Yeah, like I, I would say that a mom-and-pop data center is probably like a $50, $100, $150 million data center.

Okay, you know, put that in numbers, Jason, just, you know, the usual $100 to $150 million mom-and-pop business.

Uh, well, Well, so the, the thing is, is these nodes— one node costs $500,000.

Yeah.

And you need a ton of these nodes to be able to actually perform, uh, training and, and inference. You know, actually inference is a much— requires a lot more compute than training.

Yeah. I mean, I— in terms of pricing, startups need to know what they're going to spend, and there is standard pricing. So having it be too variable, I think, would be a blocker for people utilizing it. They don't want to get surprised with the bill unless the surprise was to the downside. So you could have a cap on the price and you could say, hey, if it's underutilized, you know, uh, the lowest priced GPUs get used first and we're going to route it there. So then people could say, hey, here's the rack rate, I'm willing to be 10% lower than that, and it could be dynamic. The permissionless piece of this is of course very important, Alex, because you don't need permission to provide your compute on Tarragon. You don't need permission to use it, uh, and you don't need permission to get paid or to pay, right? These things just happen as part of the system. So if you put your GPUs in there, you don't need to be approved. You have some sort of verifier that says, hey, we know what this system is when it's onboarded. So that's kind of the magic of these systems is there could be people, I don't know, from a communist country that have access to low-priced energy. They have some GPUs sitting around. Maybe they fell off a truck. Maybe they're extra ones at their university. And some hacker at a university is like, you know, whether it's MIT or it's in Shanghai, just decides, I'm gonna put a couple of these GPUs in here and make a little extra side money. And it gets utilized. It's going to be very confounding to regulators in some cases. It's very much like the Bitcoin freedom, uh, gestalt. And that vibe, I think, is what's going to make it really successful, is, you know, you don't need permission to join the network. There's no regulation. And that's also going to cause, at some point, just like Bitcoin did, regulators to look at it and say, okay, let's make sure somebody's not using these GPUs to, I don't know, uh, build the next COVID, you know. And you would have no idea, Robert, if somebody was sending a job there, and neither would the GPU provider. It's just done, uh, yeah, it's, it's, it's done completely confidentially.

We would, we wouldn't know. Um, but I will say that, you know, if for, for at least for on the data center side, the Cell for, for providing to permissionless market, it's much better than I think a lot of the other options out there right now in terms of trying to find demand. Targon in a way is distribution for data centers because, you know, oh, well, we have these nodes. I'll give you a great example. One of the data centers that we work with, they didn't— they weren't able to get their one of their B200 node. They purchased it last May and they weren't able to get it online and actually earning any kind of money until January, till we actually started interacting with them. Them. And the reason why was because a lot of people purchase these machines, and they think they're just going to plug it in. They have a sort of real estate model about it, they think of the GPU kind of like an apartment. And so they think, oh, I'll just buy this building. And then, you know, the apartment, you know, it's pretty self-explanatory, the people are going to move in and make it their own. But these GPUs are actually, you know, very specialized, highly technical supercomputers. That require very careful maintenance done to them.

And so make it harder for you guys to actually aggregate them because the more, you know, B200s you want, great, because then you can have more total capacity at the cutting edge of what's, what's possible. But also, like, as you go into these increasingly exotic chips that need more and more stuff built around them to function, doesn't that limit your miner base essentially to data centers larger than the mom-and-pops? Doesn't that kind of take away the decentralization element of this?

So the benefit to the mom and pop is that let's say what they want is they want that 3-year contract at the $5 an hour rate, right? That's what they want. They're dreaming about it, right? They go to bed at night thinking about this. But in the meantime, what, you know, they're thinking, oh, well, I'd be making money today. So you can onboard your machines to Targon. It's permissionless. There's no contract, right? You can pull it at any time. There's no deposit, there's nothing like that. Well, so you could put your GPU, GPUs onto Targon for a week, 2 weeks. A customer comes along that you can get on a contract, you can take those GPUs off. And so we're essentially a sort of buyer of last resort, which is, which is great for us because it allows us to be able to utilize these GPUs.

It's the same as if you had an extra home and you moved into your new home, you still have your existing home, and you're like, can't find a buyer, I'm going to have it on Airbnb, which we literally did. Yeah, exactly. Or 2 homes ago. And it was, you know, generating $100,000 in revenue a year. We weren't renting it that often, but it was a nice 5-bedroom home that cost, you know, $1,500, $2,000 a night. So there weren't that many people who could use it in Airbnb. We had a 3-night minimum, but it added up very quickly, covered the taxes on the place. You know, the taxes on the place were $75,000 a year.

Yep.

So it was like, okay, at least we're not burning money with this thing. So it's a really great use case. Robert for continued success. We'll drop you off. If you're interested in TAO, I suggest following on x.com/markjeffrey, really my partner on Stillcore Capital, which is looking at all these subnets.

Here is Mr. Mark Jeffrey over on X, M-A-R-K-J-E-F-F-R-E-Y. We love Mark here at TWIST.

Jason, we have some news. Yep.

Before we do that, I thought we would talk about the— just give people a quick summary of the bounties we currently have out in the market. Oh, there are two.

Yeah. And we also— I think we have a Poly market we need to get to.

We, we do, but I wanna make sure we get these bounties in the show so everyone knows what we're working on. This week we added a second bounty. There's now two. The first one is our kind of real-time podcast companion. We've done a couple demos on the show this week, runs alongside Twist, gives different personas and kind of fact checks and commentates as we go. But people may not know about the second one, Jason. So what do you have planned for annotated.com?

Yeah, so I bought this domain name 15, 20 years ago. In fact, I bought it after I sold Weblogs, Inc., which did Engage and Autoblog. I sold that to AOL. There was another service, Delicious, which everybody loved. They sold to Yahoo. Yahoo kind of deprecated it. I'm sure it doesn't exist now, but Delicious became a bit of a social network. Um, it was kind of like a reverse Digg. If you said, I want to keep this bookmark, uh, you save it in your Delicious. You could have people follow your Delicious. So if you were Alex and you were really into heavy metal, You could have that tagged, you could have your heavy metal tracks, and I could follow it, and then people could— if people followed it and added to their bookmark collection, it would move up. Okay, so, uh, Annotated is a service that I wanted to create for the last 15 years. What I wanted to be able to do was highlight a paragraph somewhere, a New York Times story, a tweet, uh, and then just give my comments on it. And so It was like a clipping service that would allow you to fact-check something, debate something, and then make a very simple URL. So I didn't want to steal the whole New York Times story and then, you know, write a couple comments under it or use the commenting system on the New York Times. I wanted— Alex and I are debating something, it's in the New York Times, the quote-unquote, uh, paper of record. People debate that today, where it's in a TechCrunch story. I want to be able to highlight a quote I gave and then explain underneath it in a comment, hey, I'm annotated.com/json/newyorktimes/ here it is. And I'd say, you know, I gave that quote, but here's the full quote I gave. I felt that was clipped, you know, or edited a bit, but I just want to give additional context. Now, Alex should come there. The journalists from the New York Times come there and they could actually have a discussion on that object.

So my object would then create essentially a micro forum for that particular piece of content or news item. Correct.

It would be a thread about that piece of content. But then YouTube became pretty big. So again, if you want to clip a YouTube clip, what do you have to do? Or a podcast clip. Some podcast players have the ability to clip a little bit. They're pretty kludgy. What I want to be able to do is you put in a YouTube video, you have an episode of a podcast, you say, hey, I want to debate this. I'll take— and we're going to come up with a number like 90 seconds max or something. And then downgrade the video file so it's in 280 or something. So it's not gonna replace the original, put a watermark on it, link back to the original so nobody can complain. And then you add your commentary, you add your commentary, uh, about that video. Maybe it's a video response to the video or just some text. Then all of that would be able to train or feed LLMs. So that's the new wrinkle is we wanna make this available to LLMs to say, hey, there's a debated fact check here, or, here's some commentary on this tweet, on this Instagram post, on this TikTok. So you go here, you go to annotated.com. If you build it, uh, we're going to pick the best service and I want to put it into production. So we'll give $5,000 to the best product and that's it.

Yep. Uh, if you want to take part in Bounty #1, the Realtime Podcast Companion, the next demo day is May 8th here on the show. Final decision on May 11th. And we are going to do demos for annotated.com in this bounty on May 18th, Jason. So if you're paying attention to TWIST, we're gonna have a lot of cool demos coming up. So get your cloud code warm, everybody. Did you see the news, Jason, that Anthropic was getting bids at $800 billion but might actually raise at $900 billion in a next funding round?

I did see some report that they were asking people to get their bids in in 48 hours.

Yeah, which is crazy. So this raises a question. As we've seen the value of AI labs scale dramatically, they're getting to the point in which they're beginning to become worth about as much as Bitcoin, whole cloth. And so the people over at Polymarket have put together what I think is the best, best market of all time. Will Anthropic flip Bitcoin by December 31st? Now, this is interesting because people are saying there's a 43% chance. But you have to keep in mind, Bitcoin today, all the Bitcoin that's out there is worth $1.58 trillion, more or less. Anthropic might be worth $900 billion soon. So Anthropic would have to double essentially, Jason, by the end of the year to pull off this bet. 43% say yes. It's only a 2-day-old market, so it's very nascent.

Or Bitcoin would have to drop.

Oh, spicy. Yeah, quite a lot though. Jason?

Well, I mean, if Bitcoin dropped 20% from here, it's trading at $78 or something, so it could easily drop 10% because it was in the high 60s. So it could easily drop 10, and that would bring it from $1.5 trillion down to $1.3, uh, and then if Anthropic then went from $900 billion to $1.3, that's a pretty big jump too. That's like a 50% jump. So it's not double. It would be— you could, you could see a scenario of a Yeah, 20% drop in Bitcoin, etc. I, I'm— I think Bitcoin might have— and I know somebody could wind up clipping this, but I think Bitcoin is so, um, played out. I think everybody has like some exposure to it who would have wanted exposure by now, that there's no incremental buyers of Bitcoin. And I think people have essentially ended the game of Bitcoin. Now that it feels like, uh, there are so many other alternatives, and on a functionality basis there are tons of legal stablecoins that make it easier to move money around at a lower cost. On the other side, if you're looking for something interesting, BitTensor and TAO, uh, and Solana feel much more interesting in terms of where the entrepreneurs are placing their bets and building things around it. And Bitcoin feels stale. And, you know, I I said Bitcoin zero was, you know, a distinct possibility. And I said that when Bitcoin was at $10K, $50K, everything in between.

Why?

Well, all technology eventually deprecates, with very few exceptions we've seen in our lifetime. Like, yeah, sure, there are still people running systems on Fortran or something. There are still people running on minicomputers here and there.

But here's a CNBC article from Wednesday. January 24th, 2018, says, quote, there's a 33% chance that Bitcoin goes to zero.

Yeah. So I, I, I do think that that's still something people should keep in mind. Now, I know it sounds crazy, but if there is a better system or a new game, if you're a speculator and you feel the game of TAO or, you know, the game of Solana is a more interesting game, more interesting energy there, why wouldn't you move half your Bitcoin over? Or a third of your Bitcoin over. That creates downward pressure. If you didn't have Michael Saylor trying to buy it all or get to 20% ownership, I'm wondering what impact that would have on the price, right? He is like this, you know, consistent buyer, and he owns 6 or 7% of the active Bitcoins. Like, there's a certain percentage that are deactivated. Anyway, it's a long way of saying I, I question the future of Bitcoin in terms of its relevancy to the public and the interest from the public, the relevancy to developers and using it as a platform that, that's been out the window for many years. It doesn't really work as a platform. And then for moving money around, obviously stablecoins are better, more integrated. So yes, where is its audience coming from? Where is the incremental buyer of Bitcoin? Why would people start storing money in Bitcoin? If it's— it was a speculative tool. Let's just call it what it is. It was a game. People were gambling. Yeah. So, you know, I said this on Twitter. I left it up here. You know, when this is 2018, when it was trading at $3,700, it has had a hard time getting back to, you know, six figures. It could just trade.

Just stay.

Yeah, it could just slowly deprecate over time without incremental buyers coming in. And then people just at some point need the money. So you're a Bitcoin holder and you're like, I had a kid, I've got some bills to pay, I'm retiring, I'm going to spend this money on my retirement. So I do think we're going to be living in a post-Bitcoin world at some point where it's just like this memory of some interesting thing that occurred, you know, Atari 2600 or, you know, video games in, you know, stalls at arcades. It could have that kind of an existence because it's not really cutting edge anymore. It's not doing anything interesting for humanity.

Well, someone's definitely going to clip all of that, but I'll put in some supporting points here. You're dead on about stablecoins replacing one of the critical use cases for Bitcoin, which is always money transfer without the usual systems in place that are expensive and slow and so forth. So absolutely agreed there. But I think your point really underscores why there was so much attention paid to Bitcoin ETFs, giving people access to Bitcoin exposure. People were hoping that would bring in these marginal or next buyers. And then also past Michael Saylor and strategy, there's a lot of people hyping up companies that are buying some Bitcoin to have in their treasuries, these Bitcoin treasury companies. And frankly, Jason, it doesn't feel like that's had much of an impact on the overall supply and demand of the Bitcoin marketplace. So to me, we are looking for that next use case. But if I was going to take another, I don't know, $10K and put it in crypto, why would I put it in Bitcoin? It just feels like I'm buying a Buick. Which I know is going to annoy people, but that's how it feels.

If you were a buyer who is crypto savvy and a believer, you already own some Bitcoin. So it's— for you, it's, why would I buy incremental Bitcoin? And it's like, well, I don't know why I buy incremental. And if you bought it because you were speculating— everybody who bought this bought it because somebody went on TV and said it's going to be a million dollars a coin, it's going to be 10 million a coin, it's going to be $500,000 a coin. There are countless, countless people saying outrageous things about, you know, a 10x price jump from here. Okay, fine. Um, why would it go there? Explain. Give me your thesis. Like, if I have a thesis about TAO, it's of those 128 subnets competing, they're going to provide value to the world. As we saw earlier, hey, compute— having a permissionless compute platform, that serves some value to somebody who's got a some compute laying around and they don't have a 5-year contract on it yet, they can Airbnb their, you know, uh, GPUs. Great. Okay, that sounds like a reasonable thesis. They test it. Okay, if it works, they stay in the 128. If it doesn't work and they're not updating their code, they get out of the 128, uh, people in their little BitTensor incubator, uh, you know, as I like to call it. And, um, yeah, we move on. Yeah, so And your strategy.

Yeah, I don't think I need to buy any more Bitcoin. We already have some via some ETFs. I feel good, but I've—

Look at that strategy. Put that strategy for like 2 years or something. And this is the most interesting chart, I think. If you look at the strategy chart, you know, he's been— Michael Saylor's been buying and buying. And, you know, I've invited him on the pod. He wanted to come on the pod and then he kind of disappeared and ghosted me. But I think because I was a little bit critical of it, like, because why would you buy— if you really wanted Bitcoin, why not just buy it directly? You can just buy it direct on Robinhood or Coinbase or anything. It's available everywhere. But put it at 2 years or 5 years is, I think, fine. 5 years is probably good. And this thing had quite a run-up in 2024. As you can see, it was just cruising along there at, you know, whatever. And man, it spiked up to $300 a share. I think it peaked at Where did it peak? If you were to go to the peak area, maybe $400 a share. And when it hit that, it was trading at, I think, 3 times the value of the Bitcoin they owned. Now they put all these layered perpetual strike call option shares, Class D.

Yeah, here's the thing, Jason, I don't understand it.

Yeah.

And I've read more SEC filings than, than 99.99% of humans who are alive because I'm a dork. And if I don't understand it, it's too complicated for a reason.

Yeah. So, I mean, this is— I think this whole thing will wind up— I think they'll just trade for 50% of the value of their Bitcoin, at which point you would be wise to sell and get out of there and just own Bitcoin.

Yep.

Or something else, as we just discussed. All right, let's keep moving on. It was an enormous week, Jason, of reporting from the major technology companies. We saw notes from Apple and Amazon and Microsoft and Alphabet and everybody.

What did we learn?

Well, we learned that everyone is still compute constrained. CapEx investments are not going to slow and in fact they're gonna get bigger. So if you were impressed by the numbers we saw coming out of 2025 into 2026, buckle your seatbelt. Most of these companies are increasing the amount of money they're planning on spending. Microsoft said they're gonna spend, I think it was $190 billion, somewhere in there, with $25 billion of just increased memory costs for their buildout. Amazon reiterated its $200 billion plan for this year. Meta increased its, and also Alphabet said we're gonna spend, I think, another $10 or $15 billion this year. So it appears that nothing has been able to break the compute crunch, back to Targon and the manifold crew. But I want to know if you think that the CapEx doomers, the people who've been saying, oh my God, they're wasting all this money, are wrong and we can be confident in that, or if it's still a little bit too early to make a call.

Well, you know, a company that has free cash flow has a decision to make what to do with all that extra cash laying around. And you can buy back your shares of your company. I think Apple bought back like half their shares over the decades of just printing money. You could acquire companies and use the cash to buy companies. But, you know, under Lena Khan, we had a bit of a pause there and it's been harder and harder to do that. you know, in Europe as well. So hard to do M&A, hard to, uh, easy to buy back your shares, uh, and then you can give dividends. And so I was just looking at, you know, I bought a bunch of Google shares 2 years ago in 2023, or 3 years ago, when people were saying, hey, they're toast. I was like, yeah, I don't, I don't buy that. They are— they could just put an AI search box and give the AI answer at the top of search and have their cake and eat it too, or make it like the news tab which is exactly what they did. And so I bought in at $110 in 2023, and I have— yeah, I think I'm up 300%.

I want to raise.

It's just crazy. I only bought like 2,000 or 3,000 shares. It wasn't like a ton of money, but it has worked out pretty well. And what should they— if you were— if you And what I noticed in Robinhood was I was making like $400 or $500 a quarter in dividends. I didn't even know Google had a dividend, and I had it set to just reinvest it. So every time I get a dividend, I buy more shares. Or so I buy— like now, I— if I get a $400 dividend, I buy 1 share. So it's every quarter I get, you know, every year I get 4 more shares, or previously I got 8 more shares. I guess that's fine, just sit there, um, over 10 years I get you know, 50 more shares, it's irrelevant to a certain extent. So if you had a way to deploy a large amount of capital and capture a new market, you would take it. Meta, Amazon, Google, Microsoft are all taking that opportunity to deploy their cash. If you were holding these because you like the dividend or if you're holding them because you like the share buyback and you're a pure financial person, um, who just purely likes that, you should just buy muni bonds or something boring that, you know, uh, give you a tax-free 4 or 5%, you know, just go for it. Um, I like to see these companies do something interesting with their money. Buy large companies, buy promising companies, expand into new markets, create an Apple Car, you know, create Apple Homes. Do something interesting with your money. And creating a cloud computing platform, which is what 3 of the 4 major CapEx spenders did, that, you know, and we'll see if— I mean, Meta and Apple creating a Google Compute Cloud, uh, or Azure or AWS would be certainly fascinating, but those 3 businesses are growing like gangbusters. They have no choice but to add compute.

No, absolutely. So Google Cloud's revenue was $20 billion, $18.4 billion expected. That was up 63% compared to the year-ago quarter, and critically, it accelerated growth from 48% in Q4 of 2025, which is just insane. And then every company had acceleration like this. So AWS just had its best quarter in terms of growth in 15 quarters, Jason. And it's the biggest platform in the world of managed compute. And that was 28% growth ahead of expected 26%. Azure, also compute constrained. You know, these companies just cannot get enough compute online. And, you know, to me that implies that if you're a startup, well, you might as well get good at hosting local models now because it doesn't seem that the cutting edge of intelligence is going to get cheaper anytime soon. I think, you know, Opus 4.7 is going to be costly all year.

The question is, you know, will we overbuild? Right now it doesn't feel like you can overbuild because people— there's such a demand for tokens. And if you were to compare this to when they built out fiber, When they built out all this fiber, there was no Netflix. You couldn't stream, you know, HD movie. There was no YouTube. People didn't have smartphones. They weren't taking 10-minute videos of their kids and uploading them and sharing them. You didn't actually need that much bandwidth. In other words, the infrastructure was outpacing the application layer and the consumer use case or the business use case. Same thing happened with computers, PCs. Started to become really powerful and people were using them for word processing and browsing. So it was like, okay, you can keep your MacBook Pro for 5 years, 6 years. There was no reason to upgrade. AI has changed all of that. There is a reason to upgrade your MacBook Pro or get a Mac Studio, is you could run a local model, you could run OpenClaw, you can do all these interesting things, and that will continue. So it's really nice when, um, we have a new application that tests and stress tests the in, you know, existing infrastructure. And that's what's happening right now. We've never seen this amount of demand for anything except for maybe mobile bandwidth. I think there was a time period where, you know, you paid per SMS and like there wasn't enough infrastructure to get your text messages and text messages were bogging down the system. And then BlackBerry started bogging down the networks. They had to build 3G and 4G and 5G and LTE. And now if you asked a person like bandwidth at their home, they'd be like, well, I have a fiber connection, I have 500 megabit, I have Starlink. Where it's like swimming in bandwidth now. I— and a lot of times I'll come home and I'll still be on 5G, or I'm in a hotel, like I went to the hotel this time and I forgot to log in to the internet because I have my MacBook set up to automatically connect to my tether to my phone. So I just started working. I was like, oh, wait a second, this is a little slow. And after 10 minutes of work, I was like, oh yeah, I'm not on the Wi-Fi of the hotel. That's how good the mobile network got. So it's very—

it's gonna get even better as we get more Starlink up there connecting to our phones. Like, we're gonna have even better service in more places, which is gonna be awesome. But the point about Fiverr, I think, is a great one to make because we were building capacity ahead of demand and today we are chasing demand with capacity. And so we're going to have, I think, a much better early warning sign. And I read these earnings calls every quarter, Jason, looking for any hint that things are slowing down, that they can begin to decelerate because companies love to reward shareholders. Apple did just announce another $100 billion in buybacks in its earnings report yesterday. Investors love that. They would love to do more of that, but they can't because they're spending all their money on this other stuff and they would probably like to be able to reward investors, boost the share price, et cetera. But we're not there yet. Okay, next up, Jason, people in Congress are starting to press technology companies about their use of Chinese AI models or AI models from China. And on this show, you and I have talked about using things from Moonshot or Minimax or Deepseek in OpenClaw and so forth. I'm just very curious, do you think that American startups should be comfortable using AI models built in China that are often less expensive and often a bit more open than what we see built here in the US?

Yeah, I mean, they're open source projects, so I'm not sure the government can stop this because you can fork them and just restart them in the United States and they are open. So the idea that you would have inside of them you know, what you would have inside of TikTok, which is the ability to, you know, steal users' data, know who's connected to who and get local information, have access to people's camera rolls, have access to the location data, have access to the cut and paste. You know, you got a lot of data you can have in that case in an open model if you're running it locally. Like, does OpenSea have a backdoor in there? And how effective is that backdoor? And could they hide a backdoor in there? I'm not so certain. It's a bit above my pay grade, but from my understanding, it's incredibly hard in an open source project like that to put a backdoor in. It does signal to me that we need to have an open source champion here in the US.

Yes, we'll get to that in a second. But this is what a partner at Andreessen Horowitz said about this particular problem, Jason, and I'm quoting here from their website. If adversaries control the models, they can manipulate the underlying probabilities to subtly distort truth, influence decisions, or spread falsehoods at scale. This makes the geopolitical risk of having an adversary control these AI brains real. So that's the concern. It doesn't seem to land very well with me. I don't have that fear because these models are sufficiently open that I feel comfortable using them. So I was surprised to see people in Congress really going after AnySphere, which makes Cursor, which made Composer 2, off of KIMI, and Airbnb, which used QINN models. To kind of like beat them around the ears because you could make it unpalatable in a business context to use those models by getting blocked by the government or whatever. So they can make your life bad even if you can't ban open source. And I think it's a real risk. I think, I think these models are good and cheap and we should use them. Like, am I crazy?

You know, the bias in the models is true. They can steal IP and they do. So I guess there are, there are some concerns around that. I think you could test it pretty quickly. I mean, if you went to Deepseek right now, and you asked it like, you know, tell me about the Uyghurs. Oh, in China and how they're being abused. I understand you're asking about a topic where there are two starkly different and contradictory narratives. The information I can find is predominantly— oh, sorry, Bon, that's— it just took the, uh, answer and did you see that? It snapped it.

I did. I just full-screened it so I could read the dang text. And let's talk about something else. But critically though, this is the model served via Deepseek's inference stack and I don't think—

Yeah, chat.deepseek.com.

Right, I don't think you run into the same restrictions if you use the model hosted via a different inference provider. So on OpenRouter, for example, you can just like pick between the different compute stacks that you have. But I agree that it's worrisome. But if you're, if you're AnySphere, which makes Cursor, you're very savvy. They went and looked at all the open models, picked KEME K2.5 as their base, and then worked on it from there, creating tons of domestic value. Jason, a lot of American GDP off of the back of Chinese open source. So to me it's like a win-win-win for the country.

And yeah, I think we do need to be aware of it. I think it's definitely something to monitor. Don't have a problem with them asking Airbnb or whoever for a little feedback on how they're using it. I think it's probably great that these senators, congressmen, everybody are getting a little more educated on it. Um, obviously I need to be educated on it. I'm— I haven't been using it all that much, but we use KIMI on some of our OpenClaw instances, and we run our own on US Steel and Iron. So, um, interesting to think about. I think practically probably not an attack vector, but we do need a champion here in the US. The problem is all the great Minds and AI have been offered competing packages from Meta, OpenAI, Groq, and Anthropic and Microsoft and startups for $10 million in RSUs over the next 4 years. Why would they, you know, um, go work on an open source project when they have this once-in-a-lifetime chance to take down $10 or $20 or $50 or $100 million in equity?

So the answer is it's very hard to compete with that. There is one American company, Reflection AI, that is doing research to build open models here in the United States. The problem is just every time I go to their website, because I'm like, what's going on with Reflection? There's nothing new. They haven't released a model that I can play with. They haven't done the thing that I want them to do, which is to put American open source or open way AI on the map again since the Llama 3 era. And it's not happening. And, you know, points to xAI. They have open source Grok 1 and 2. I checked, it's on Hugging Face, but that's not cutting edge. So I just think American companies could do more and then we wouldn't have this problem and we wouldn't be ceding so much of the global inference market to upstart Chinese AI labs, which are doing quite well, as we can see in their stock prices. So, yeah, before I let you go, Jason, do you have a certain background up?

It appears Square Garden.

Yeah.

Yeah, so I hear your beloved Knicks have, uh, well, I hear you were quite rude to the Atlanta team. So how did it end up 40-15 in the first quarter? Because that's just—

I have, you know, I, uh, I have been watching basketball since I was a teenager, and I have never seen a blowout like this in Atlanta. We were— it was— I had courtside seats, uh, in Atlanta, and it was insane. Uh, 40-15 after the first quarter. It was a historic beatdown, uh, the largest margin of victory I think at multiple times for a playoff game in the modern era. So 140 to 89 by the end of the game, it was like 100 and something to 40. They were up 60 points for most of the game. Uh, the Knicks are the best team in the East. They've got a suffocating defense and they hit a lot of their shots. At the same time, Atlanta Hawks are a young team and they kind of folded. And so it was absolutely insane. Yeah, it was a blowout though. And I'll say, since we're in our Off Duty segment and I think Fat Don Capital is making us a graphic and music for Off Duty, we like to go off duty at the end of the game, just talk about random stuff at the end of the show. This is my tip for people. If you got some siblings or friends in different cities, And I guess maybe I'm late to the party on this. Going to see your favorite sports team in the non-home arena— like me going back to Madison Square Garden is amazing. Maybe I'll catch a game when I'm there at some point. But going to the opposing arena adds like a level of villainy and fun. And so I was wearing my nice, uh, you know, Kith jacket, and, you know, interacting with everybody was a lot of fun. Um, But you get to go see another city. So last year I went to Detroit. If you remember at this time, I saw Courtside, Timothée Chalamet and Ben Stiller and everybody went. We had a group of 4 guys, my brother Josh. And the year before I went to Philly and there's like, you get to see Atlanta, Philly, and Detroit. I don't think I'd been there as an adult. Like, why am I going to these cities? Like there's not a lot going on in them. And then it turns out they're actually quite, uh, you know, cities on the rise. Great American cities that are working through some issues. So it's actually a great excuse to just go to a city for 3 days and check it out and eat at their best restaurant, see their sites, and, uh, you know, take in a game and take in a local culture. All 3 of those, you know, local cultures here in the, like, Northeast, generally speaking, were just wonderful to experience.

So next up is either the Sixers or the Celtics. I am an Eagles fan, so I do have some allegiance to the Philadelphia sports scene, much more than Boston. But I'm curious, which team do you want to see the Knicks go up against because you have the best shot of crushing them?

It really doesn't matter. We, we beat both of those teams, um, in the last two playoff series that we played against them. The Knicks were actually built to beat those two teams specifically. Those two teams have, uh, Jason Tatum at the Celtics, and on the Philadelphia 76ers, they have Joel Embiid. Both of those were like MVPs. They're both, you know, they, they both have exceeded our best player Jalen Brunson's, uh, performances historically, but we're probably better right now. And, uh, it doesn't matter. They're gonna go to a Game 7, I guess, tomorrow Saturday, and, uh, we will play either one and we'll beat either one, I think, uh, in 6 games.

The current, uh, Polly Market odds. Jason, I don't speak Sportsbet, so when it says Philadelphia 30 cents, Boston 71. That means Boston's more likely to win. Yeah, according to the—

I believe Boston, who's playing at home, will beat Philly and will beat Boston. Um, and then we have to beat the winner of Detroit Pistons versus Orlando, and I think Orlando's ahead in that series. So we get to rest up while those teams figure it out. Um, and yeah, just, uh, basically this is the Knicks' best chance of getting to the Finals and competing for the championship since the '90s, basically.

Wow.

Yeah.

So decades and decades and decades.

Yeah. 30 years of my life. Yeah, basically.

If they're in the Finals, how much— what's your, like, what's the max you'll pay for courtside seats at the Finals? Because I don't even know what that would cost.

I, you know, it's one of the great things about going to these arenas in these other cities is it's $50,000 to sit courtside, you know, at Madison Square Garden. And it's a small fraction of that in these other cities. So I treat myself to it.

But yeah, you said it last year about the Pistons, like, yeah, it's really cheap. You can just get courtside seats. It's fantastic.

And I was like, that's no problem.

Detroit's not far away, you know?

Yeah. So I, yeah, I don't, I don't know what I would pay, but if they make it to the finals, I will definitely be going courtside or hopefully up close. My friend David Adelman owns the 76ers, so I think I want him— 76ers. To play the 76ers so I can go there and torture him and watch some home games in Philly.

I mean, if that happens, I might be down there to see my Philly friends to play some poker.

I will get a ticket called— I'll get you.

All right, but, uh, guys, this has been another episode of Twist. We're back, Jason, on Monday. We're back on Wednesday. We're back on Friday. Twist never stops. We'll see you next week.

Bye-bye.