Cerebras's IPO goes vertical, and the death of OpenClaw? | E2287

This is one of the highest stakes games I've ever seen.

A 750-megawatt deal with OpenAI in January. It's not just insanity, it's insanity with a twist.

The question here is, what is that deal with OpenAI worth? And then, is that deal, uh, guaranteed or not? If the revenue doesn't show up, you're out of business and you go bankrupt, can't pay your bills. Is Cerebras going to, or Oracle going to sue OpenAI? The acquisition by OpenAI took the wind out of the sails.

This Week in Startups is brought to you by NetSuite. The business landscape is very chaotic right now. That's why you need NetSuite by Oracle. Get the free business guide Demystifying AI at netsuite.com/twist. Deal. Founders scale faster on Deal. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deal.com/twist. To learn more. And LinkedIn Jobs, hire right the first time. Post your first job and get $100 off towards your job post at linkedin.com/twist.

All right, everybody, welcome back to TWIST. I'm Jason Calacanis. He is Alex Wilhelm. He's @alex on X.com. I am @jason. We have a packed show, so we're going to get right to work. Let's go, Alex. What's first up on this docket? I mean, before we start anything, I think I got to just get my plug.

Ah, yes.

And I'm going to just hit the record button. Oh, I get the nice haptic, I get the light goes off. And this is the one you put on the back of your phone. What is that? What's the difference between this one and the pin? Well, this one, if it's on the back of your phone, you can record phone calls. Now of course you have to make sure you're making people aware of this, or maybe if you're in one-party state, all that other good stuff. But if you're jumping on a Zoom call or a conference call it actually can record the call and not on speakerphone. It just does it. So an amazing product. It's MagSafe. It goes right in the back and easy breezy. If you want to charge it, just this little charging cable, bang, MagSafe goes onto it and you're charging. Boom. Easy. You put it on the back of your phone. You never forget it. Of course, you like to wear it on your wrist. So you have the wrist attachment for the Plaud Pin. So there's the Plaud Note. You have the Plaud Pin.

I think everyone here needs a Plaud Note Pin S, which is what I have. If you want to join us in keeping track of everything that happens in your working life so you don't forget a thing. Keep your boss happy by going to plaud.ai/twist, P-L-A-U-D.ai/twist. Use the code TWIST, save 10%. Look like a genius.

Absolutely.

Work like Jason and get, be productive like Elon. It's all there, friends. Just get Plaud. And as Jason says, we applaud our friends at Plaud.

Let's get started with the show. What's up first, Alex?

All right. First up today, we have Ori Goshen. He's the co-founder and co-CEO over at AI21. It's an enterprise AI company, Jason. They make both language models, LLMs, maybe even some SLMs in the mix. And they also provide a platform called Maestro that helps orchestrate agentic activity for the enterprise. This is not your general company that's building an AI model to do everything. Instead, they want to make AI actually work in the enterprise. Ori, welcome to the show.

Tell us a little bit about what you've built and how it works.

Yeah, so we've, we've built a platform called Maestro, which is basically based on what we call a meta model. A meta model is a model that learns the behavior and patterns of other models, could be frontier models and open weight models. And when I say behavior, I mean what is the cost, latency, and accuracy characteristics. So it tries essentially to predict what would be a successful call at what cost or at what latency. And that way, it can basically activate different models and tools in the most cost-effective way.

Got it. So that's what the orchestrator does. Just like in Perplexity, it has that routing to— like Perplexity has the ability to route to whichever model it thinks is going to do best or based upon cost, right? And so enterprises are trying to save money here. And they're also trying to get the best answer. So how does it actually pick that? Like, what is the logic there?

It's a dedicated model that we have, a model model that actually learns these probabilities of whether calling model A or model B would be more successful and at what it tries to predict also at what cost, and then it can do that smart routing. But also it can use different, more sophisticated techniques. Like it can call a model multiple times and try to fetch the best answer or try to call different models at once and try to fetch the best answer. And so it's actually more than just routing. It's executing different inference strategies at scale. So—

Ori, on this point, we know about model routers. We're all familiar with them to some degree. Why does this matter in the enterprise in particular? And I think about enterprise companies, they have lots of money. So to me, they're less cost sensitive, but this seems to be a cost forward technique. So tell me why that matters.

Right, yeah. So cost is definitely a consideration. And these days, token bills are going through the roof. People are deploying these other agent decoding or just other agent workflows. Some companies we see, they're token maxing like crazy. They're just incentivizing their employees to use more and more tokens. But now it gets into the phase that it's not just about the quality of the tokens, but actually whether the ROI is there, whether it's actually being produced efficiently. And I think a lot of companies are now looking into it.

So this brings us to your own models that you guys have built, most recently the Jamba line. Now, we've had a couple of people on the show talking about SLMs versus LLMs, and people seem to have different definitions of how many parameters kind of is the threshold. Tell us about the JambaLite, and then also, do you consider them to be SLMs or just very small LLMs?

Yeah, so Jamba is actually a model family. It has small versions and it has a large version. The largest version is 400 billion parameters, probably not an SLM, but the small one is a 13 billion parameter model with a mixture of experts. So that probably counts into the category of an SLM. And I think the interesting part, and this is what I predict we'll see in the industry in the next couple of years, is innovation also on the architecture side. I mean, all these LLMs are based on transformer networks. You know, that invention was started back in 2017. We haven't seen many innovations on that front. I think the interesting The interesting part of Jamba, which is a combination of two architectures, one is attention, the transformer-based architecture, and the second is Mamba, which is a new architecture. It's highly efficient. It's really good at processing long sequences. So that's why when we built Jamba, we built it with long context processing in mind. That was the idea behind this. And we're seeing actually now You're seeing in many of the new open weight models from China and actually also in the US, a lot of folks are innovating on the architecture side, and I'm very bullish about that.

So does this mean that Jason and I spent all this time learning about how transformers worked by reading the Attention is All You Need paper, and suddenly that's going to make us behind the curve? So now we need to go learn new architecture for models. Dang it.

Jamba is an open source model, correct? Yes, it is. Now that is your model, or that's an open source project that you are hosting or supporting or you're forking?

No, it's an open-weight model that we've actually built. We pre-trained it, we post-trained it, we shared it with the community and we have some commercial deployments for private deployment companies that need to deploy it on their premises and they're limited by resources like memory and compute. So Jamba is a highly efficient model that they can use.

Anybody can take Jamba and then fork it, uh, or use it without going through your company. It's just an open source, open weight model out there. The same way people can take KIMI 2.6, or they can take DC4 and they can run it on their own. But your company, like wordpress.org/com, will help enterprises deploy more efficiently if they need services and, uh, hosting and that kind of stuff.

Yeah, that's exactly right. And in today's world, with all the variety of models, our focus is actually more on the orchestration layer, the Maestro system that I described earlier.

And the Maestro is called what?

The Maestro, that's the orchestration system.

Like, does that have a name or is that another open source project? And that's a small language model itself?

No, so the Maestro is an orchestration system. It can use any model, Open Web, its Frontier model, et cetera. It just has an internal model, that's the meta model, that learns these probabilities and know how to route and how to activate these.

And that's your proprietary one, or is that also an open source one?

That's our proprietary system.

Got it. So when we think about where value accrues and the business you're building, hey, the language model, Jamba, is open to anybody, but the orchestration model, you keep that as the proprietary tool that is how you provide value to enterprises, correct?

Exactly right.

Adding a new member to your team is a crucial decision, and you don't want to rush into a hiring situation that you will regret. But if you're a busy founder, you also don't want to spend a ton of time in the trenches looking for that perfect candidate. Instead, you need a partner and you need a trusted partner. And LinkedIn Hiring Pro is that partner. How do I know this? Because I use LinkedIn Hiring Pro. You want a real-world testimony? All right, here you go. We recently hired a new customer success manager. They handle all the advertising accounts here on This Week in Startups, and they ensure that we're keeping all of our wonderful partners happy. LinkedIn helped us connect with an amazing candidate right here in Austin, and he had exactly the combination of experience we were looking for. It's magic, it's alchemy, it's everything for me to find a great team member. So hire right the first time. Get started by posting your job for free at linkedin.com/twist. Terms and conditions apply. That's linkedin.com/twist. Got it. What do they pay for it? How do you price it?

So it's still early days for the product. We just launched it last year, late last year. So we're experimenting with different business models, but basically it's a combination of a fixed license fee and a consumption-based model.

Got it. And so you're going up against Mishra or Claude or any of these companies in the enterprise. And we saw today OpenAI, Alex, has bought a company. Uh, I don't know what the name of it is. You'll totally want it in, but tomorrow, tomorrow, but without a W.

Yeah, it may be tomorrow now that I think about tomorrow.

And so this is the future here, Ori, is this sort of, um, teaching companies how to deploy this internally.

Yeah, I, I think in many ways, um, this is the future because there is no one model to rule them all. It's actually very interesting to see the coverage, like how many of the tasks be dealt with different models. And it's, you as an AI builder, the first thing you do is you go and you pick a model, right? And that's, and then you try to optimize it with various techniques and so on and so forth. And that's what we see. It's a very painful process in the enterprise. So the idea here is basically say, hey, you know what? This process can be automated. You just need an intelligent system that learns these priors of performance of different models and then know how to activate them automatically. And then a new model comes in. You mentioned Anthropic or Mistral release a new model. It's compatible with the system. The system keeps learning how to use new models and how to basically activate them in the most cost-effective way.

Who are your customers?

We work with large enterprises. We have, for example, one of the largest retail in Europe, a company called Fnac. They're using this for many of their kind of mission-critical workflows. We work with several US-based, some of the largest actually US tech companies. We work with Israeli companies that are kind of home base, so we support this industry.

Ori, can you tell me just how you guys went from building Wordtune, a consumer prosumer tool to help me improve my writing, and we've ended up here building orchestration layers for enterprise agentic AI? I presume there's a thread between them, but from the outside, it's a little bit hard to parse how you guys have made this impressive progression.

Oh yeah, thank you. It really gets back to the evolution of the company. And we kind of going back, we released Fortune back in 2020. That was pretty early in the super cycle, right? Like pre-ChatGPT. And so this was back then, like Grammarly was the, you know, the writing assistant application. And so we launched, we back then we actually built our own foundation model. And it was, yes, it was. It was pretty big at the time, which was 178 billion parameter model called Jurassic-1. And so we took that technology and we said, what is the best application? Back in the days, the best application was reading and writing assist. So we put that out there and in a couple of years, we had tens of millions of users and pretty significant revenue. Yeah, this is the paper. And so that application grew significantly, but then ChatGPT came out and I think it really, it changed the modality. People suddenly started to work with chat versus these writing, these plugins that check your text and rewrite your text, et cetera.

Yeah, it became a commodity really quick. So you had a demo that you wanted to show. Go ahead and show us your demo.

Just to give you a sense for how our system works, our system actually simulates what would be the cost accuracy or latency accuracy trade-offs. So you can put it, all the models, let's say you're an enterprise company, you put all your models, all of the open weight models, frontier models that you have as a resource, And then using that portfolio of models, our system will give you this chart. And what you see on the chart on the left side, on the x-axis is the average cost. And on the y-axis is the success rate. And this is a deep research task, for example. And what you have here as the gray boxes is actually baselines like GPT-5 or GPT-5 Mini or Nano, different sizes. And what you see here on the red is actually a result of combining different, like mixing. Some of the results are going to the large model, some of them to the small model. So what you're seeing here is actually a new Pareto frontier.

Right, 'cause the top of the curve here is showing that you can combine models to essentially get to perfect completion at a lower cost than some individual models that cost more.

Exactly, and same goes for latency. That's exactly right. So you can use basically, you have a lot of configuration. The action space is wide, like the number of models, the way you activate these models with prompts is very, very high. This system actually lets you as a builder to explore that space automatically and surface points, like operational points that would be impossible for you to discover.

How much money does this save an enterprise company with a standard token budget? Is it like a 10% savings or is this something much more dramatic?

It depends on the task. It definitely can get to 50%. We have some results on different benchmarks that show significant savings. Maybe one, just one more chart that I can show that it gives you an intuition. Here you see, this is kind of a Venn diagram showing you how many of the queries could be solved by each model. In this case, like a Minimax, that's a small model, a GPT-5, we've certain retrieval technique and another GPT-5 with a different retrieval technique. And you see that there are actually 7% of the problems could be solved only with, for example, in this case, a GPT-5 model uniquely can be solved. So if you have the combination of models, if you use them as a portfolio, you actually get better accuracy because of better coverage of the types of problems that the models can solve.

Does that work outside the enterprise use case or is that only possible when we're thinking about kind of the enterprise workflows that we automate, which I presume are relatively rote?

Yeah, it's typically most effective in enterprise context because in that area you have very specific evals, you have way to measure the success and the cost and latency, et cetera. And so basically the performance. So once you've a way to measure, a setting where there's a very good way to measure its success, then the system really shines.

Well done. Continued success and we'll drop you off and keep running through the docket.

Thanks, Ori. Appreciate it, man.

Have a good one.

Next up, we are going to go to the oceans, Jason. We are going to bring up Alex from Magrathea Metals. He is the founder and CEO. This is one of the coolest companies, I think, in the entire world today. We don't talk about magnesium much on the show. We're not really a metals podcast, but if you can pull important metals from the sea and not ruin the environment while you do it well, Sounds like a hell of a startup idea. Alex, welcome to the show.

All right. So, Alex, do you have a demo to show us or charts or anything, or you can just explain to us what this is and why it's important?

Magnesium is a pretty insanely critical material in every aluminum alloy. We cannot make anything out of aluminum without mag. That's cars, planes, helicopters, construction materials, etc. We use it to make titanium for aerospace. We use it in steel production, hafnium, zirconium, beryllium, boron, like producing all these different metals for defense, aerospace, national security applications. Nations. And, uh, the US has zero production. China controls 95% of the world's supply. And we've built pretty much like, I think, the biggest pilot in two generations for making magnesium metal in the US. Um, I will show you a quick little video.

And is this your Oakland facility?

This is in Oakland, yeah, just across the bay from San Francisco.

Founder scale faster on Deal. That's the deal. You can grow your company without borders, and you can set up payroll for any country in minutes. Hire anyone anywhere like a modern startup or large company does. And DL is gonna get all the visas handled fast so you can get back to building. There is a great talent war that's going on right now, and you need people with superpowers for your startup to be competitive, to beat your competitors, to get your products to market. But anytime you try to grow your team with overseas hires, oh my Lord, you've got to reinvent the wheel and you gotta navigate a tangled web of international laws, regulations, and you can't get these things wrong, folks. You wanna onboard new staffers in other countries? You wanna get 'em set up on your network? Nice and secure, IT access, all that good stuff. You want to manage their benefits. Trust me, this is all a nightmare unless you partner with Deel. They are the people stack for startups. They're going to take care of all the onboarding, payroll, HR, IT, benefits, everything you need quickly in one place, done perfectly. So visit deal.com/twist. That's deal.com/twist.

And then just to make sure to sportscast as we go along for everyone on the audio version.

For sure. I assume that means explain what I'm showing.

Yes, please.

Here I've got a video of our molten salt electrolyzer pilot in Oakland producing the first new primary pure magnesium metal in a generation in the U.S. That's our tech, our R&D technician, Brian, yanking the metal out of molten salt. So this is like an actual, like, you know, reasonable scale pilot, not bench scale, not, you know, not small tech R&D on a table. It's a pretty sizable facility and we're now in the process of scaling it up to commercial scale.

So I think the most important thing here, Jason, to understand is that getting magnesium out of seawater is actually not a new idea. But what Magrathea has done is somehow put in some secret sauce halfway through the process when it comes to— and Alex, back me up here— drying magnesium salts that makes it environmentally less disastrous, which means we can kind of onshore this capacity to the States.

Yeah, the largest magnesium smelter ever built was built in Freeport, Texas, in World War II to, to fight Nazis. And, and it was, it was built using seawater as a feedstock from the Gulf of America. So magnesium was precipitated from seawater to make a concentrate that was then converted into a chloride salt concentrate. That's the one you're referring to, which we then, which they used to then dry and electrolyze into metal. There were a lot of issues with that process, and a couple other times this type of process was built in Utah, uh, in Israel, and, and in Europe, in Canada, um, and every single time it was built differently. So we identified what were the historic drivers of CapEx and environmental impacts that would have made it impossible to build this type of process in the West today. Uh, you know, probably possible in China, right? But, but not possible here. And we solved all those issues, and drying the salt was actually really the hard part because Magnesium salt really likes to hold on to water. So if you need to make a dry salt to electrolyze it into metal, it's very challenging.

And this is a venture-backed company. And then how do you make money?

This is a venture-backed company. We've raised 3 VC rounds, an angel round, seed round, just closed our Series A a month ago. Yeah, I'm super proud of that. Thank you. We'll make money by selling metal through a joint venture. So we formed a JV with a publicly traded company called Tetra. They're an industrial minerals company that serves the oil and gas industry. And working with them, we're building a smelter in Arkansas.

So how do you get— where do you get the seawater from?

So, so we can use seawater or brines. So any kind of salty water with magnesium in it, we can take the mag out and turn it into metal. And in Arkansas, there's a really high-grade brine 10,000 feet underground that we use.

Ah, so you're taking it from the water table, like through a well. Is that right? For this first situation, it's a production well that goes 10,000 feet deep.

It goes like super far below like freshwater aquifers that we would use for water. And, and this same brine is already used today to make bromine, calcium fluoride, magnesium hydroxide, a whole bunch of different minerals.

What do you do with the output of it? If you— because that's always like the concern here in the US, right? Is like the environmentalists are very concerned about what happens with the water afterwards or the byproduct. So is there an issue there?

There's no issue. So the brine comes up, the minerals get extracted, and then the brine goes back underground with virtually nothing added. And this is done every single day today in Arkansas by companies like Albemarle and Lanxess making bromine. And it's been done for calcium fluoride and other minerals too before. So, so we'll be using pretty much the same process that is really environmentally benign and really well understood.

So where is all this magnesium currently coming from?

Almost all of the world's mag production is in China using a super dirty process called the pigeon process. It's basically a Rube Goldberg device of coal. It's really nasty. It works for like very specific reasons. They've tied in magnesium production with, with basically coal processing to use waste heat that is essentially free. And the CCP subsidizes all of it. So that's what the West has been basically trying to compete with for the last 30 years.

And how did you discover this process? What's your background?

Uh, so I'm a chemical engineer by training. I studied chemie and philosophy in undergrad at McGill in Montreal, and then went for a PhD in chemical engineering at Northwestern. I dropped out with a master's and got into mining about 10 years ago.

And how did you— so you got into mining, how did you discover this, uh, or where did this innovation come from? Is this like some foundational study that needed to be applied? Is it a patent out of a school? Where, where did this specific innovation come from? And then how do you protect that IP, or do you protect it? It's just more about the you don't bother protecting it, just about the execution layer.

Our—

so our intellectual property is the product of a 2-year technology review of the last 110 years of different smelting technologies for making mag. So we put together some really smart engineers who could understand all the physical chemistry from first principles, and we identified whitespace that no one had ever really kind of taken advantage of before. That's resulted in a big IP portfolio with 11 patent applications filed and a huge amount of trade secrets. So, so yeah, we've, we've, we've, we've reduced the entropy of a very large amount of information to be able to make a process that actually makes metal.

Jason's question about like venture-backable, I think, is reasonable because we're seeing a lot more companies in the startup world begin to work with and help industrialization here in the States, which has not always been the highest gross margin product. It's not always been the most growth-oriented. So help me understand the economics of this. If you guys get this process up and running at scale as you hope, what's the margin you can get on the mag you create? And then also, how much demand is there domestically for American mag versus Chinese magnesium?

The price of mag right now in the US is something like $7,000 a ton. We're going to make it in Arkansas for around $3,000 a ton. So it's a pretty big spread. The US market's around 100,000 tons per year. The defense industrial base need is actually a tenth of that. It's 10,000 tons per year. And that's also exactly the same size we're building our first commercial smelter in Arkansas. So the goal is to take the risk of foreign, foreign dependence off the table for the US with that first commercial plant in Arkansas.

Okay, I got— yeah, Larry from Oxcart, one of your backers, was on our venture roundtable last week, and he sang your praises. So I'm glad we got to go a little bit deeper on this. What's the website and are you hiring for any particular role you want to shout out into the void while you're with us?

Yeah, the website is magratheametals.com. There's not too much on it. It's a little 1.0 right now, but we'll be hiring more folks in the, in the coming months after, after closing this capital and kind of expanding our joint venture.

Good. Well, I'm really glad you're working to bring industry back to the States. I believe, as Jason says, much continued success and come back when you have that smelter up and running. We want videos. And we want to see the magnesium rolling off the production line.

All right.

Well done.

Well done. Wow. We got hard science today. Really, like, these minerals are super important and looks like American independence from China and the CCP and what they're doing is a high order bit here in the United States. Yeah, it continues to be.

You know, this is one of those companies, a bit like Knox Metals, that is not cosplaying industry. They're not like putting a thin patina of CNC milling machines on top of their business. This seems really like in the weeds, nitty-gritty. That's why I like it. I like the companies that are really doing the hard work. All right, do you want to see some final sidebar demos?

Yeah, that would be great. We of course have been doing this project. It is a simple project we wanted to have, or a seemingly simple one that is hard to execute, but people are making progress on it. We wanted to have people create a real-time agent that could listen to a podcast and then, based on persona, do something like real-time fact-checking or real-time jokes or roasts or real-time, uh, deep research to come up with the next question. In other words, could you in real time watch a transcript go by and then feed it into an LLM? So when we were just hearing about magnesium, uh, they could preemptively give us data. Now, this is what happens on all live shows. Uh, if you're Howard Stern doing a live radio show, there's Jackie the Joke Man, uh, or after him Artie Lange, were feeding Stern jokes. How did they do it? They have a little monitor here and they'd be sending jokes in. So what we said was, hey, we'll do this. Then I stupidly said, I'd pay $5,000 if somebody could figure this out. And then of course 15 different people started building it. A little bit of bad feelings from maybe 2 of the first people who did it. Because they were like, I thought I had this in the bag. No, if I'm going to give a $5K bounty, I want to see some people make something production-worthy, uh, that we can use. And so we decided we would do like 2 more weeks of this and then pick a final winner. And then I reserve the right, since it's my bounty, to cut it in 3 different ways, 4 different ways, give it all to one, whatever it is. Uh, so sorry we didn't make the rules clear from the beginning, but it was just an off-the-cuff comment, and we had to work backwards into how we were going to run this as a contest of sorts. And I think we have to give you a 1099 or W-2 or whatever, because yeah, you're winning a prize. Yeah. Uh, so if you're not using AI tools to make your team more effective and increase your productivity, trust me, you are being left behind. Sitting on the sidelines is no longer an option. No more waiting. With NetSuite by Oracle, you can start putting AI to work for your company today. NetSuite is the the number one AI cloud ERP, and it's trusted by over 43,000 businesses. It's not just an AI add-on or a chatbot that sits in your browser. No, this is a unified source of truth that brings together all the data you need to run your business. From software to IT services to healthcare, equipment manufacturing, financial services, and many other great American industries. NetSuite NetSuite delivers a customized solution for your business. If your revenues are at least in the 7 figures, get your free business guide, Demystifying AI, at netsuite.com/twist. The guide is free to you at netsuite.com/twist.

So we have 2, uh, for you, Jason. One is Glass Sidebar by Oliver Choi. He's a member of our Noti gang. We love this guy. And then we also have a major update to Sidecast by Pat Hughes. And I'm going to show you Glass Sidebar right now. Quick aside, we spend a lot of time on the show talking about using AI, right, Jason? You're like, you have to be AI first. Everyone at Twists has to be AI first.

Trying to, yeah.

Yeah, well, I took the GitHub link for this and I was just feeling lazy and I threw it into Codex and I was like, can you just build this for me? I'm in a hurry. And it did. So this is my Codex-built version of this app straight from the GitHub code. And there's one technicality with it that's a little bit struggly, which is that if you wanted to listen to your tab, you have to run Zoom inside Chrome. And Jason, Zoom today did not want to actually make my camera function inside of the tab. So instead I'm going to talk and you will see this machine is doing a transcript of me as I talk. And then as we go through the fact that—

Oh, so this is a live demo.

Oh yeah, this is live right now. I'm doing this for you.

Okay. So it's in a browser and it's capturing your voice in a browser window. And because you're running Zoom in a browser, which most people don't know you can do, but you can run in a browser window, um, it's kind of janky. Um, it is now doing a fact checker and a comedy writer, and that's what we're seeing here. Yeah, yeah.

So I was talking about Zoom, and as you can see, the, uh, the comedy writer said if Zoom had a mood today, it'd be camera shy, riffing off the fact that I said that Zoom wasn't working out so well.

And so what's the difference between live cards and what's happening on the right-hand side?

I believe on the right-hand side, you're getting kind of all of them. And then these are the most recent. So that way, if you're watching the video right here, Jason, for example, you can put in a stream, then it has these. You can also collapse the interface. There's a lot of like little things you can do up here that make it pretty neat. Oh, and by the way, you can also run quite a lot of different AIs through this. We're using my personal OpenAI API for this, but it's very cheap.

Amazing. So you're able to like actually take these and run them. But it does work. It does take an audience.

It does transcribe, and it does do all the important things that we asked for. So I was very impressed with this, Jason, and I can build it myself. Yeah. Oh, wow.

Amazing. Okay, well done. So this is one of our— how many applications have we gotten?

Oh, Jacob, what's the— hit me with the final count that we have. I know we had a total of 6 or 5 that were demoable, and we had some others that were, um, inscrutable.

10 or so, he's saying. He's saying 10 or so. And what's the URL for people to see all this? We always make an easy-to-follow URL. Do we have one?

Yep, we do. It's going to be in the show notes, but I'm just going to pull it up here. So this is the GitHub page. And if you would like to go to it, you can go to, um, well, that's—

or do we have it in the docket notes? Like, remember I had asked Jacob to make a page where we could have all the contestants so they know we reviewed them. So Jacob, is there a page for that? And can we put that into the docket? thisweekinstartups.com/docket will take you to the docket. You look at it today. Okay, looks like we don't have it. Okay, so we'll make that and put it in the show notes. This week in startups.com/docket, and in the show notes you'll find something for this project. What's the official name of the project? Did we give it an official name?

Glass Sidebar, and we have links to the GitHub, the website, uh, a video overview, and the—

No, no, that's for this project, but for the overall, um, yes, Bounty Sidebar is what we're calling it, I guess.

There is a link. It is in the section right in the Sidebar Bounty section. So if people want to go to the docket, they can find it, and it is linked here. I'll just make it in big bold letters.

What's the plan here, Alex? When do we actually pick a winner?

Well, it depends on how you want to go about it, Jason. Would you like to be the final say on this, or would you like the production team to have a moot, put our heads together and pick our favorite?

Well, I don't know. We had already set dates. We had set dates and we had explained that to the audience. So I don't like to change the dates because then people feel like we're rug pulling or changing the rules of the game. So what did we already tell the audience we would do? Memory serves, the last demo is going to be on Friday and we were going to hand out the winner either This Friday we were going to do, so today we did an update of two of them and then Friday where we're going to give the winner or next Friday.

I believe it's this Friday is going to be the, the winner. Um, I'm, I'm hedging ever so slightly because we might have a bank show that day. So there's some, there's some scheduling questions that would be better taken offline.

All right. Okay. So, uh, we're making progress. Go to the winner actually is going to be a week from today on Monday. This week in startups.com/podcast. Bounties or bounty, bounty. Let's just go with /bounty. So we have two bounties going on. One is for the sidebar and the real-time sidebar and one is going to be for annotated. So we're having some fun with these. And next Monday we will pick a winner. We want something that's production ready that we could actually use. And so I want to be able to share a stream at maybe or at least be able to see a working version of this. So that we can actually apply it. And I think the one I wanna have is I wanna take out the jokes. I just wanna have a fact checker. So I think what we'll do is we'll reduce this down to one thing. Let's, for this next sprint where we pick the winner, let's remove everything but fact checking. Because you know what? It's, we, we gave like all these different personalities and I feel like it's so hard for somebody to just refine this down to a working product that will make it even easier for the people who are building these awesome, uh, products. Let's just make it fact-checking, real-time fact-checking of podcasts. Makes it super easier, easier on the builders, and we'll judge you based just on that real-time fact-checking of a podcast.

Moving right along, you're looking, you're looking healthy today, Jason. Maybe even slim, perhaps even svelte.

Well, um, yeah, doing okay with the weight. I'm adding a little bit of muscle. I've been really focusing in on my sleep. I've been focusing in on my nutrition. And that's part of this holistic journey I've been on. Ro.co/twist. If you've been thinking about taking a GLP and you want to, uh, follow your boy J-Cal down this adventure of getting in shape. I peaked at 213. I'm 5'8.5", 5'9" on a good day. That was just way too big. I was fat Cal. Uh, and just over 3 or 4 years of using these GLPs, I just slowly shed the weight. And now a svelte one, you know, low 170s. I probably have 5 more pounds to go for my ideal weight. Not putting a ton of pressure on myself. The GLPs helped me break through, and you should take a look at it if you're— let's just call it what it is— if you're obese, if you're struggling with your weight. I did. I gained 2 pounds a year for 20 years.

Adds up.

It's not— it adds up. 1 pound every 6 months is like 3,500 calories. 3,500 calories in 25 weeks is about 100 and change extra calories. In other words, if you eat 2 Oreos a week more than you burn, you're going to be 40 pounds overweight. That's how slim the margin for success is here. If you have a Frappuccino instead of a black coffee, and there we are, all of a sudden you're massively overweight, unhealthy. Now I feel great, uh, you know, dare I say, I look better. And, uh, yeah.

Go to ro.co/twist for your free insurance check. That's ro.co/twist. Jason, did you know that 1 in 8 Americans is currently on a GLP-1? 12%. That blew my mind.

It makes sense to me because the United States has a broken food system. And, you know, we have things in the food system we probably shouldn't have in terms of chemicals. We have fast food where, you know, we had a food pyramid. That was completely ridiculous, portion sizes that are absurd. I think everything is just stacked against you. You know, you walk through an airport, you walk through a mall, you're walking to work, whatever you're doing, you're presented with choices that are bad choices, and it's really hard. And then people are under stress. The food system is designed to make you fat. It's designed to make you consume more calories. And whether it's fast food, whether it's bread, And what I think the GLPs have done for people is it's really hard to reverse the cycle, but once you reverse the cycle and you learn how to eat smaller portions and the food noise goes away, well, you can then taper off of them. You can take a microdose. Everybody's, uh, and then there are some people who are just taking GLPs who think, uh, there are other benefits to it, and you can look up the other benefits. But yeah, the food system being stacked against you is the challenge here.

Yeah.

All right, we have some news. Yeah.

Oh yeah, we absolutely do. So Cerebus, uh, everyone's favorite enormous chip company because they make enormous—

is it Cerebus? How do you pronounce it?

Well, now I'm really doubting myself. Cerebus makes way more sense than what I was saying. Cerebus is the famous historical, uh, dog, right?

I, I don't know if that's the origin of it, but, uh, sure, we'll go with that.

Spelled a little bit differently, yes, but the hound of Hades. Anyways, Cerebus, we'll go with that, has raised its IPO price range, Jason, from $115 to $125 a share, which was going to put it at a valuation above its last private, all the way to $150 to $160 per share. An enormous repricing of this IPO before it goes out, which frankly, very bullish. You love to see it. We could use a big winner of an IPO right now. And if folks are curious, the company's simple and fully diluted valuations at its $150 to $160 price could be as much as $34.4 billion or $48.8 $8 billion.

So what's the revenue on that? Does the revenue support this kind of, uh, valuation? What would be the price-to-sales ratio? Usually you've got like some of these on the back of your hand ready to go. What, what is it looking like in that regard?

So Ceribus reported $171.4 million in revenue in Q4 of last year, the most recent number we have. On a run rate basis, that's $686 million. Now, how does that translate to a multiple? Well, I'm glad you asked. Okay, at $34.4 billion, which is its simple valuation at $160 per share, it's 50x run rate at $48.5 billion.

50 times top line revenue, not 50 times their earnings, price earnings ratio.

No.

And the average price earnings ratio in the market is typically a low of 15, a high of 25. It's probably right now the markets are trading at all-time high for P/Es. These companies are probably— a lot of these new companies are money-losing, but these are getting disconnected with reality, huh?

What matters here is that the valuation, uh, multiple goes up if you think about the company on a fully diluted basis. Then it goes up to as high as 70%. But a little history—

70%, you mean 71 times, not percent?

That is exactly what I meant, Justin.

Okay.

Yeah, just so we're watching ourselves right off the plank there. No, no, we want to make sure because it is confounding to people. I, I was talking to some people and I was doing this exact math, and sometimes people are like, oh, it's a— oh, it's growing 70%. You're like, no, no, no, no. Or it's 70 PE, we just have to make sure it's 70 times. So every dollar of revenue is $70 in, uh, the market capitalization, which you would think means this thing has to be tripling revenue year over year, quadrupling revenue year over year. We don't have that data yet. I don't believe.

What do we— I can pull that for you one second. But the thing that matters here is if you go back to the first time it tried to list, it had interesting revenue, interesting margins, but essentially one customer, G42, right? The company has then took its IPO back, waited a couple of years, did a lot of work, expanded its partners, expanded its customer base. It's still pretty heavily weighted right now towards the Middle East. but the company has done a lot of big deals so far this year. So, uh, 750-megawatt deal with OpenAI in January. Uh, signed a deal with Mistral in February. Collab with AWS put its chips inside of its data centers in March, working with Cognition. So the way that I think this is being priced is not that today this company is worth this amount of money, but instead the amount of major companies that are saying, okay, we don't just need GPUs for training, we need fast inference chips Now we're dying. That is bringing this company to the forefront of the market. So if you think that Agentic AI is going to keep driving compute demand, especially on the CPU side and GPU side, then I can see putting money into this company. Would I bet the farm on it at that price? No, it's expensive. But I do understand a little bit of the logic behind it. It's not just insanity. It's insanity with a twist.

Well, I mean, the question here is, what is that deal with OpenAI worth? And then is that deal, uh, guaranteed or not? So the headline number, I believe, is like $10 billion, uh, over some number of years, like 3 or 4 years. Um, so the deal structure will be the one that matters. And is it actually guaranteed or not? What we saw from the Nvidia deal where NVIDIA had the option to invest in OpenAI, that was an option. So a lot of these deals are going to turn out to be optional. Now, does that mean they're going to happen or not? It's trending towards— we have so much compute usage and so much token demand that you would think these would be fulfilled. But, uh, these are big numbers. It requires the OpenAIs, the Anthropics of the world to actually have the money to pay for those. The cynics in the markets are saying, well, and you actually had Brad Gerstner say this directly to Sam Altman, how does a company with $20 billion—

oh, that was awkward. Yeah.

You know, able to do hundreds of billions of dollars in buildout. And he said, well, we project our revenues will keep growing. Sure. And that has come to pass. This is— and then you had Anthropic. And Dario say, well, you have to be very careful here because the amount of compute you're building, if the revenue doesn't show up, you're out of business and you go bankrupt. Can't pay your bills. So this is the, one of the highest stakes games I've ever seen. And we don't know in all these contracts what's, uh, written in stone. And then what happens if it's written in stone and they don't have the money to pay? Is Cerebras going to, or Oracle going to sue an OpenAI? Because they can't pay their bills? Or do they say, hey, we'll just push out the contract? And do they have the option to push it out? Instead of buying these things in 3 years, can they say it's 9 years?

Do they pull a SpaceX AI? And if there is excess capacity from your oracles, from your Stargates, from your whomevers, do you just rent it to whomever needs it? Because in the case of Anthropic, there may be demand somewhere else that is quite high. But I think that Cerebus is a wager that training compute as a percentage of global AI compute is going to decrease as we drive more inference products, as we see more workflows brought into the AI world. And so from that perspective, I think it makes a lot of sense. I just, I just remember when we paid like 70x for SaaS companies and that was terrible. Well, I fret, but I know that I'm boring, but like it does concern me a little bit, Jason. Maybe we're being a little bit enthusiastic.

We are going to find out in the next 2 years if we're ahead of our skis and if the buildout can keep up with demand, or if the demand is going to level off. And open source models, you know, making more efficient models, running models locally, you know, there are so many different ways for this to play out. That no one knows exactly what's going to happen except that this transformation is already happening and that it's moving very quickly. So who knows if people have ordered too much compute or not enough? It's starting to feel like not enough. Uh, and if you look at what's happening in the Mag Seven, the ones who have cloud computing, uh, which is Google, Amazon and Microsoft, all their services are growing massively. The cloud computing, the Neo Clouds are all growing massively. They can't keep up with demand.

So yeah, CoreWeave got dinged for only growing like 25% from Q1 to Q2 of this year. People were annoyed that their growth rate wasn't higher. I'm like, they're adding like $500 million from a $2 billion last quarter result, but it wasn't enough. And I think in that case, Jason, it's more concerns about debt. Uh, last note on ServiceNow. Benchmark is one of the leading 5% or greater shareholders. So another winner for our friends over at Benchmark.

Yeah. Uh, well, that's the other thing about these IPOs that are happening, these crazy valuations, is that we're going to see, uh, an answer to the question, is venture capital a viable investment category still, uh, with all these late-stage investments or not? And it seems like between SpaceX, Anthropic, OpenAI, Cerebris and others, we're going to see some great, um, returns flow through to LPs and keep venture moving.

Yeah, speaking of that, uh, an apology. Did you know that Fervo Energy is going public? They're a venture-backed geothermal company.

I did not know that.

Yeah, yeah, I, I got halfway through the notes to prepare a little section for us, so maybe we'll get to it on Wednesday. But, uh, venture-backed company raised more than a billion dollars, going public, and I was just— missed that one. Anyways, next up, OpenAI has formalized its AI PE push. Talked about this on the show before. Basically, OpenAI and Anthropic are partnering with investment firms and private equity giants to take their AI and shove it into portculls. We have some information about what this actually is now. They've named it the OpenAI Deployment Company, which I believe Jason wins an award for the lamest name of all time because blah.

So this is another company from OpenAI that is designed to help private equity firms— TPG, uh, Warburg Pincus, whatever it is— go out and deploy AI in enterprises. And this is majority owned by OpenAI, but they've raised money from TPG, Warburg Pincus, Bain, and a bunch of other people, and Bain, Capgemini, McKinsey. They're working with them to use this company to push OpenAI's products into the market. This makes no sense to me. Um, why would it be a separate company? Why are they— why are these investors investing in a separate company as opposed to the main company? This seems absolutely convoluted.

Oh, convoluted. Yeah, rat's nest. But welcome to OpenAI. It's, it's a SAM company. Come on, Jason, there's going to be some deals, some side deals, and some arrows going all over all over the place. The way that I understand is pretty simple. You got OpenAI over here. They make models, they serve inference, and they don't really want to be a services company, which makes a lot of sense. So you need to put forward deployed engineers into these enterprises to figure out what you can automate and where AI fits. You need to put the experts in, but you don't really want to take all that spend and dump it onto OpenAI the year it wants to go public. So what do you do? You go out to people with money, you take some of it, you promise them a 17.5% annual return. Yeah, that, that broke a while ago. And then you get a bunch of private equity companies that want to improve the efficiency of their portfolio. You take the bucket of money, plug in some FTEs, take the demand from the PE side, send the engineers in, inference calls go back to OpenAI. Huzzah! More usage. And then everyone kind of gets some value out of it except for the employees who get fired.

So, you know, the problem with these things always is who are they serving? And this reminds me of in the dot-com era, people would spin out from Barnes Noble's barnesandnoble.com. So you'd have the Barnes Noble legacy company trading as one, and then you'd have barnesandnoble.com. You'd have Toys R Us, toysrus.com. And what that did was it just confused the market. Where is the value accruing to? This to me is extremely confusing because is this company eventually going to go public? Is this company going to give dividends? Is this company, uh, is the value in this company going to accrue to the model, the frontier model owned by OpenAI, or is it going to accrue to the balance sheet of this company and their shareholders? It just feels super complicated and like financial, um, engineering that is not necessary. If you actually believe there should be a company like this. It shouldn't have the OpenAI name on it. It could just be a separate company. Now they're acquiring a company to put into this company. Yes. And who's going to focus on this? It's the same thing with OpenAI's Johnny Ive project. They gave Johnny Ive some percentage ownership in OpenAI. Okay, so now they've got a hardware product. Now they've got this PE spinoff. Is the hardware product going to spin off? Very confusing. What is the board of directors at this company? Who do they answer to? What if they want to use Claude, right? Yeah. What if Claude's a better solution? What if the best advice they could give to their customers is to use KIMI, you know, or Deepseek?

Sure.

Are they not gonna give that advice to them because they're loyal to OpenAI? The best advice might be OpenAI's not a good piece of software for this application, or you don't wanna feed your data into there. Super convoluted. I think these things are all gonna wind up getting unwound. And then put back into the mothership, uh, is my best guess.

So Anthropic is also putting together a very similar effort. That's a $4 billion push from OpenAI, $1.5 billion from Anthropic. But you would also say in that case, same thing, bit of a boondoggle, gonna get unwound?

I just don't understand why these are being done as spinouts as opposed to a division inside the company. If you look at Oracle, or if you were looking at IBM, they have services divisions and They're not spinouts to the best of my knowledge. IBM, you know, or Oracle, pick the company, they don't create a separate entity to do this. So it sounds like they don't want to have the costs of these implementations hitting the balance sheet, as you're saying, to make it look less profitable, which then is not a good reason to do it this way. They should just have a business unit and be able to explain to people, hey, a certain percentage of our revenue comes from services, or they should enable a channel. So Microsoft has channel partners to implement the hosting of Office 365 or whatever other software products and the implementation of it. Those people generate 5, 6, 7, 8 times the revenue as Microsoft does from selling the licenses in a long tail of lower margin, 10, 20% margin services businesses. So that would be typically how you would do it is have a channel strategy. So they don't want to create a channel. They want to create another company. Maybe this is to speed it up or something.

But the only reason why I can see it making sense from, from the perspective you're arguing, and I think you make some very solid points, is if the private equity companies can go faster at automating their portcos, which are in many cases a lot of SaaS companies that aren't worth much because they burn cash or aren't growing quickly, you might be able to rescue your portfolio. So there might be like this, this demand from private equity to please come fix our software companies. And if that's the case, then I can see a lot of like more pull than push. And from that perspective, having a separate CEO of this OpenAI development company could make some sense because you don't want everyone being too spread out.

Yeah.

But last thing I'll say about this is I just pulled it up. IBM's price sales trailing basis. Jason, give me a guess.

Got it. Okay. So I would guess they trade at 2.5 times.

So close. 3.2.

3.2. Okay, so if they make a dollar, they trade at $3 of enterprise value. So that's the reason to take this out of there. And so just, you know, my best advice to folks is, um, financial manipulation or engineering is generally, um, not a great idea. It just feels like a suboptimal this feels suboptimal, is all I'll say. Yeah.

Okay. Yeah. There's been a lot of skepticism. You're not absolutely alone in that. People are skeptical, but they have enough money. I mean, OpenAI raised $122 billion earlier this year. Anthropic is raising from everybody. So if they end up losing a little bit of cash on this, it's not the end of the world for them. Maybe it'll work. All right, next up, I want to show you something. We're going to talk about agents. Here really quick. Jason, this is the viral tweet that was going all around over the weekend. People were talking about it. And what we have here on the screen is a Google Trends chart for OpenClaw. And as you can see, Jason peaked sometime in the middle of March and has been in pretty rapid decline ever since. I've been keeping tabs on this trend as well, tracking both the decline in token usage by OpenClaw on different services and tracking the rise of Hermes agent in several services. So I'm kind of curious if you think that OpenClaw was exciting and directionally correct, but too early to sustain its hype. And if so, what does that mean for OpenClaw in the next, like, couple quarters?

Yeah, you know, I think the acquisition by OpenAI took the wind out of the sails where there was this enthusiasm with Peter and this was going to be like the revolution and then the Empire bought it. So it'd be like, Okay, yeah, now Luke Skywalker works for the Emperor, you know, like, okay. Um, it's kind of hard, I think, for folks to look at all the competition coming out, whether it's Perplexity Computer, Claude Co-Work getting better, or just people writing their own. And I can tell you internally for us, what happened was our top 2 people who are working on, uh, OpenClaw instances, I had to kind of redeploy them because they wanted to make software. They wanted to basically vibe code solutions for a number of internal projects we have. And so I was like, okay, go do that. And then we'll go circle back around to agents when we get to it. And then other people in the organization got a little excited about Perplexity Computer and Cowork, and they're working on that, which has a better interface. So yes, the wind, I think, has come out of the sails a bit for OpenClaw. That doesn't mean it's over. I just think it's. People are promiscuous when it comes to tools. And when 5 tools drop in 90 days, I think everybody has to go check them out. And then there'll be a bake-off as to which one is the most supported and which one do you trust the most.

Yeah, I would— to me, well, one, yes to all that, but also I think that like OpenClaw was the signal of where we're going, not the final piece of technology that's going to get us there. Yeah. The first airplane was like, ah, we're going to do flight now. It did not invent the F-16, right? The Wright brothers were not there. And so when I think about what we're seeing from Cowork, from personal agents coming from both Google and Meta, from, you know, set up deployments of Claw from Chinese AI companies, it's clear that everyone wants to crack the personal computer use agent. No one's there yet, but at least OpenClaw showed us that there are these regular moments of magic amidst the API fails and other crashes you have. That really show us where we're going. So I, I think OpenClaw won, even if it's not currently seeing the same explosive growth that it saw, because it showed that this is what—

it may still have growth. I mean, what we looked at there was just Google Trends and people searching for it. Doesn't mean that the people using it, um, aren't using it 10 times more. It just means people aren't searching for it. So searching and Google Trends is a way to know that some new trend has come out and people are looking for it. If you were to type Airbnb or Uber into Google Trends, you know, or Lyft, of course those things would go down over time because people know those products, love those products. If you typed in Amazon Prime, I don't know how far back this goes. Google Trends is, is how interesting, how it's new to people. So look, here's Lyft in, um, yellow. Nobody's searching for Lyft. There's Uber during some peak time period.

Yeah, this is the pandemic right here, really brought down. But it actually— it's Jason, look at that.

I mean, it's still actually— people are still putting it into trends, so it's not a great example, but it's not surging.

In other words, no, no, definitely not.

Like, these things are still popular. People still type in Airbnb into Google. People still type Uber. They don't type Lyft anymore. The point is, it's a proxy for people who, who are the laggards. So anybody searching for Airbnb and Uber today are laggards. Or Amazon Prime might be like a really good version of this. Like, if you don't know about Amazon Prime by now, like, yeah, so it's just not a perfect thing to look for, uh, because it's, it's also like, um, the top 100 apps in the App Store— when something's a new app, they give it extra pushes in the App Store, or a new podcast gets an extra push because they want it to not be just a self-reinforcement top 100.

I've covered up the branding on where this data is from to avoid, you know, making anyone unhappy. But here's a chart that I track literally every day. It's 30 days of Open Claw usage from an unnamed API model router.

Okay.

And it has been kind of consistently coming down. And I've been tracking this because it feels so counter-narrative to me as agents get better. I mean, we talked about agents last year and they didn't work. Work. This year they do, and we're seeing this trend. So maybe, maybe, or maybe people are just going to co-work in other managed properties that are easier to use. But I think we're seeing a lot of people try to get that Open Cloud magic, and I think they're stealing away some of its, um, some of its hype and momentum.

Yeah, I mean, you just have more options. It's like somebody opens like the only sushi joint in town, and then there are 6 of them, so everybody goes and checks out the other ones. We'll see who wins at the end. Uh, but I do think I think the competitors in the market are really good. Groq has just come out with their downloadable desktop tool.

With a K. Yes. I was like, no, no, no. NVIDIA bought that.

Yeah. No, no. Groq is an XAI's product. They now have their cloud cowork competitor. So everybody's going to have one of these. Google will have one. Microsoft's going to have one at some point, a desktop app where you can start doing cron jobs. And, you know, listen, they work really well. and I think that they're much easier for somebody to use them. They have a much better interface. The open— it's very similar to, uh, using like Linux for your desktop. I don't know what those wrappers were like. There was at one point Lindows where they were kind of making a play on words, Linux, Windows desktop, and those things were just too hard to set up, you know? So the ease of use of OpenClaw is the issue. It has to be easier to just go to a website, download it, install it and have it work. Somebody needs to build that, an OpenClaw client that you just download and it works.

Well, last note, cause I, we haven't gotten to bring this up in the last couple shows, but there was a tweet from, maybe it was Peter, someone else from the OpenAI Codex team, and they were talking about like, yeah, we're gonna bake OpenClaw-ish things into Codex, which is interesting because Claude co-work, Claude code distinct, right? But if you take the co-working stuff and put it into Codex on the OpenAI side, you have kind of a super app. And so we're seeing either bundling or unbundling at the same time. I, I don't know which is the better way to serve the market, but it's cool to see different approaches to it. We're not going to get the same thing warmed up from every company. Uh, and then Jason, is this the Groq feature that you were talking about a second ago, these, um, connectors?

Yeah, so I think the connectors are in the app, so you're going to be able to connect in whatever app you're using. You know, this has existed in the other work products for a while, so they're playing catch-up now. But, uh, Notion, Outlook, you know, Google Workspace, Gmail, all that stuff, you know, you need to be able to connect so you can say what's on my calendar today or make me my to-do list, etc. All right, so, and then there was an interesting story. I don't know if you saw this, but TikTok is testing a $4 a month product, $3 a month product, and I thought that was interesting, an ad-free product is something that I think gets them out of some trouble with privacy issues. And so I like this as a concept. YouTube also has an ad-free product. I think 3, 4, 5% of people are, you know, if you're addicted to these products, getting rid of ads has got to be fantastic. Uh, if you consider this a new form of entertainment, so interesting to see TikTok experimenting with ad-free. And a subscription model.

As I, I normally, Jason, it takes me about 2, here it is finally.

Okay.

Cut. Jason here. I have a copy of the story of what's going on here. £4 a month for no ads in the UK is one price point for this. That's incredibly cheap. I would absolutely pay for that. Like right now, I think any service that wants to show me advertising, I hate. So for me, I would definitely pay that, but I do feel a little bit like I'm being charged to have them stop hitting me. You know, it feels a little bit unfair.

I mean, it's— I have to say, I could never use YouTube with ads. It's impossible to use YouTube with ads for me. So I've had the YouTube Premium— they used to call it Red, I don't know what they call it now— but I pay for it. I could never use YouTube with the amount of ad load they're putting in there.

What are they doing? Are they trying to make it so unpleasant to use that they— we all become premium subscribers? Because that's a much smaller cohort of the world.

One of the goals to get people to move to a premium service is to make the non-premium service painful. So this is what the airlines have pursued for a long time. Let's keep making it more painful to be in coach so that people go for whatever Coach Plus is called, or business class, or business class light. They're trying to get people who are business travelers So they make it so you have to pay for the first bag, right? You have to pay for drinks, you got to pay for internet, just you have to pay for everything, and it's painful.

Nickel and dimed.

What's that?

Nickel and dimed.

You're getting nickel and dimed to a level of absurdity. And so I think what's happening here is if people were to complain about privacy on TikTok, TikTok can now say to the EU and the UK, which have very granular filters. Hey, consumers have choice. So if consumers want a free service, they can be tracked and have targeted ads. And if they don't want to be tracked, we will know nothing about them and they can just pay us. So it kind of gets you out of jail from the regulators. That's, I think, the backstory here.

Didn't Meta get in trouble for charging in the EU to avoid ads or something like that?

I, I had advised it to Meta, but they were talking about a paid Facebook version version. So ad-free Facebook.

Oh yeah, here we go.

EU says Instagram's paid ad-free option breaches rules. July 2024. So, okay. But I think what happened here was the EU said, you're breaking our rules with your advertising techniques. So Meta said, okay, we'll just give you no ads at all for £11 a month. And no one liked that either. So, but I, but I, I'm fine with this.

Oh wow, you know what it was? They were forced to do it. Um, so I think the EU Commission, uh, forced Meta to do it as part of their fine. So they wanted to have a less personalized version. This is what happens when you get into this like crazy regulation stuff, is you start doing unnatural acts to appease them. And nobody really wants to pay for Facebook except for like lunatics like maybe me and you would be like, oh, my time is worth that. I do think every product's going to have ad-free, privacy-free versions just to appease, um, the regulators, the regulators, and the top, call it, 5% of users. I pay for it on my NBA League Pass. There's like an extra $20 a year, I think, to take out the ads. And then you get to watch the in-stadium stuff, you know, like the in-stadium, uh, when they have somebody come out and do a balancing act with the teacups on their head, or they're throwing the shirts with the air cannons. Like, yeah, I get to watch that stuff on my, uh, my NBA League Pass.

All right, Jason, good news is that the Philadelphia Sixers not only lost— oh, they got swept kept by your Knickerbockers. How the hell did you manage to crush them 4-0?

Uh, crazy. Um, we have— our team is surging. I got to go to the third game. I sat courtside, uh, with a friend of mine, Ben Stiller. So shout out to, uh, Ben Stiller, the actor, uh, producer, director. We had a great time. And, and that was a great run for me. I got to see my dad for a week. My mom stayed at their place in Brooklyn. and just slept in the basement of my childhood, uh, home, the, the, the brownstone, and went to the Atlanta Massacre courtside, 2 MSG games, and then a final, uh, courtside. So I did 4 games in a row, 4 wins in a row, and now I'm back home with my family getting work done.

But maybe, uh, okay, Pistons-Cavaliers, right? The winner of that series is your next opponent. And I think everyone out there, you know, who also appreciates the, um, the odds side of sports might be curious to know, uh, who do you think is going to win Pistons-Cavs? And then when you face them in the Eastern Conference Finals, what's going to be the series record?

Yeah, well, we are— these are the top 3 teams in the league, uh, or in the East. Most people consider those— the Knicks, Cleveland, and Pistons— the 3 best teams. Boston was also up there, but these are considered the 3 best teams. Boston was a work in progress, so we don't care who we play. I think we'd rather— I would personally rather play Detroit because Detroit won the regular season versus us, and I don't want to have anybody complaining that we didn't play the, the best team. But they're, they're, uh, they're up 2 games to 1. Obviously we want to see them go to a 7-game series so we can rest and get prepared. Um, and then I think we'll win that series in 6 games. So Knicks in 6 for the Eastern Conference Finals. I think we play the Spurs I think we'll end up playing the Spurs in the finals. I think we win that in 6 games as well.

Haven't the Spurs not won a finals in like 30 years or something?

Um, yeah, whatever. Tim Duncan era. Uh, you had something you wanted to add to the list. I had challenged you with, hey, what are you doing off duty? Everybody wants to know what you're doing off duty, Alex. What do you got?

Yeah. I'm going to bring up a book that I'm just finishing that I think is so interesting that everyone who appreciates startups, technology, AI, and thinking about the future should absolutely get into. It's called There Is No Anti-Memetics Division. This is just the Amazon page. And as you can tell, I bought the audiobook. So what, what is this about? It's actually not a novel originally, per se. Instead, it turns out there's this collaborative speculative fiction website wiki. I know, called the SCP Foundation. And this author under a pseudonym, Quantum, wrote a number of essays, part of this project that were incredibly popular and were put into a book that's become a smash hit bestseller. Everyone loves it. What's it about? Well, it's about a secret organization that kind of keeps the world safe. Not very interesting.

Okay.

But what's cool is there's a group inside of it that deals with anomalies or weird stuff in the world that consume memory and information, hence anti-memetics, because other things spread quickly because they are memetic. Antimemetic things, consume information. And this makes them hard to see, hard to track, hard to know what's going on. So the book deals a lot with memory, collaboration, technology, and a kind of near-future science fiction element to it. I can't quite do it justice. It's one of those things that I just— people should just read it. Okay, that's good.

So there is a wiki that one author is making, or a group of people are making, about this world.

Yep.

Okay, so here it is. A group of individuals is working on this SCP Foundation, which is a clandestine worldwide group that is securing and protecting humanity in the near future or far in the future?

Near future. So the technology that you see is very much right now. In fact, some of the story goes back to like the Got it.

So they then made this into a book based on the wiki, and some anonymous author or authors are the authors of this bestseller?

One guy. Okay, one guy, and he has his real name out there. I'm not saying it just because he, I think, wants to be unknown by his pseudonym QNTM, or Quantum.

Wow, this is weird. Um, I wonder if this has been bought to become like the next Nolan film, if it's dealing with memory and time and near future. But, uh, wow, what a— you're thinking interesting.

Stellar.

Well, I was just thinking more like, um, um, Memento and his other memory. He's always, uh— Tenet was about memory, Memento was about memory and time. He's always into those kind of concepts. But I wonder with this, the IP ownership of it, if everybody's getting involved in a community, does that mean the authors own this or anybody can own it? There's like an author's page on the wiki And there's a lot of people in here. It looks like, you know, I don't know, a couple of hundred people have contributed to this. Very strange. Uh, this is like super interesting.

When I heard the title of the book, I thought, oh, I don't want to talk about memes. I'm so sick of memes. Memes, memes, memes.

Wrong.

I was entirely wrong. I, it's one of those books that like 20 minutes in, I was like, oh, okay.

There you go. Check it out. It's linked there. And you had another one that you wanted to share?

Yes. Really quickly. Uh, if you are a person who likes to do, um, anything that involves long time horizons. Stroller walks, running, playing complicated video games with the sound turned off, I cannot recommend highly enough the Fall of Civilizations podcast. I'm going to pull up the website for everyone right now. For example, what they'll do is they'll do 3 hours on the rise and fall of the Persian Empire, replete with quotes and interesting stories. And it's one guy who writes and narrates the whole thing. You can find it on YouTube as well. It is absolute mind candy if you want to better understand the history of the world. For example, they have an entire series on the Aztecs and how they rose and how they fell. And it turns out that when you look at the historical record of civilizations, you find, interestingly enough, that it's where technology fails as the world changes that kills them. Usually drought.

Drought playing a big one. Yeah. And so this has been a popular theme for people, I think, with the American Empire being criticized. And hey, are we in the final days of our empire? What's going on? China's rising, and Russia fell, and the UK fell. Where we stand. Ray Dalio talking about this as well. So yeah, this sounds fascinating. And, uh, have they talked about the American Empire when they, when they talk about stuff? Do they try to— oh, The Fourth Turning is the other book. Yeah, that's, uh, something.

Okay, should I read, should I read that? Because my dad read that when I was a kid.

I don't know. I mean, I, I feel like people are obsessed with the end of America, and I don't buy it's happening. I do think we have problems, obviously, but I don't, I don't think it's going to be the end, uh, of the empire.

Yeah, I do think we're going to have a bad 10 years because we have to figure out how to spend less money at the federal level. Yeah, exactly, exactly. But once we get that sorted out, pretty, pretty straightforward.

You know, I, I've now come to the conclusion, uh, at my age, I think I'm moving into, you know, the, the last third of my life that I really appreciate the, the, um, uh, the United States of America. I love the idea. I am against federalism and I like the states, uh, and I like states having individual freedom. And I like the idea that we can move from one to the other. I love this idea of like giving a really hard intractable problem—certification of universities and education, universal healthcare, I like the idea of splintering that and letting 50 different states really go at it and then eliminating the federal version of those things. So if something becomes intractable, yeah, why not give the states the ability to try it? Now, this one's going to be quite controversial. Um, Roe v. Wade was something I did not want to see changed. I kind of like the status quo of it, but many other people didn't. And now I happen to live in a state where women have less protections and less options. Okay, do I feel good about that? Absolutely not. But what I figured out is, uh, people now are working really hard to change it in the state. It's become unpopular in the state, and people also are routing around this state to go get their healthcare that they need in other places. in other words, getting an abortion or morning-after pills or whatever it happens to be. So it's almost like when it goes to the state, you have 50 different states processing it in real time and you can get to consensus. And I think you and I just talked about this with poker regulation. Yep. Uh, and cannabis regulation and reclassifying drugs, psychedelics. Yeah. Also being done on a state-by-state level. It's almost like a way of us state by state working through the issue in our local communities. And then maybe at the federal level, at some point cannabis gets reclassified, which I think just happened.

So generally speaking, yes to all that. I think, though, that the way we have kind of uniparty states now makes me a little bit worried about, about, about this because it feels like it's not that Texas is going to be able to do a lot of experiments per se because it's kind of run by one dude. We know who in Texas is the governor and so forth. But it's kind of like one person, one party, which to me feels less democratic in the way that I think we're shooting for in this. So absolutely, but I'm a little worried about implementation of it. That said, in Florida, they're doing their own accreditation, I think for universities and— Are they? Is that true? I am, they're tickling the back of my head. I think they have some sort of alternate system. But yeah, we're doing some of this. Can I push on this in one particular way? Sure. Curious. Good. Okay. So anti-discrimination laws. Let's say you're— let's say you're a state. Let's say perhaps in the South and you want to roll back federal rules that prohibit school segregation and it passes. To me personally, Alex, I would say get bent. But, you know, states' rights do have a say here. So hit me with where you draw the line.

Yeah. So that one feels like basic human rights. So that's where you get into like interpreting the Constitution. And I'm no constitutional scholar, obviously. Um, but yeah, so that would be one where the federal government got involved. Um, is it nonfunctional right now? No. So I think when things are nonfunctional, the abortion debate was nonfunctional. Cannabis, uh, poker, online gambling, education, healthcare, certification of colleges and housing. You know, these things feel like they're non-functional at times. And if you look at housing as an example, Texas is killing it, Florida is killing it, Nevada is killing it. Yeah, New York can't build houses, and they don't seem to understand that you just need to add more units and supply and demand. They're like, no, we're going to freeze the rents. And so, okay, great, mandami, but that's not going to work because if you freeze the rents, nobody's ever going to build another unit. Because capitalism.

Yeah, okay, we can cut this out of the final show, but I have to yell about something really quick when it comes to housing, NIMBYs, and rent. Okay, so in Providence, we were— some people were trying to pass rent control. And yeah, I've lived in San Francisco in a rent-controlled building. I have lived the experiment and the dream. And let me tell you, if you want building maintenance, don't live in a rent-controlled building because they don't care. Because you did. Anyway, so this all came out And my local city councilman, John Goncalves, whom I know and I really like, and I voted for him and I have a sign of him in my front yard. He was the deciding vote to not get enough votes to override the mayor on rent control in Providence. And I wouldn't bring this up because it's just a small city issue, but there's now stickers going up around my city that is saying, like, John Goncalves raised my rent and I'm just ripping them off everything that I can Yeah, it's—

he saved the city. We— you, you also have the option to move to the state that you like best. And I think with personal freedom, you know, that's the, that's the incredible story. There was, um, Ezra Klein— I just saw a clip going viral. Ezra Klein had the guy who wrote the book Abundance with them, uh, or the book that talks about abundance, which they kind of cribbed the term from Peter Diamantes in his Abundance book a couple years ago. And they were just referencing, man, how incredible the Dallas and the Austin story was. You know, you basically have rents going down for 3 years in a row, and Dallas imported like, uh, essentially 100,000, 200,000 people, and housing prices stayed the same. So you can make this work. All right, everybody. Another amazing episode of This Week in Startups. We'll see you on Wednesday. Bye-bye.