Meta AI Deep Dive, Jeff Huber, Sheel Mohnot, Leif Abraham, Samuel Hammond, Víctor Perez, Jai Malik, Pratap Ranade

You're watching TVPN. It is Tuesday, April 8, 2025. We are live from the temple of.

Technology, the fortress of finance, the capital of capital.

And today we are particularly in the temple of technology because we're doing a deep dive on Meta's open source AI strategy, talking you through the history of Llama and how they built out that LLM and what their strategy is with it. Also all the teams behind it. There's some interesting and some of their latest internal dynamics.

Yeah, benchmarks hate this one simple trick.

Saturate your models. Apparently that's the trick. We'll, we will be as fair and balanced as we can about it now.

We're excited about it.

There's a lot of. Yeah, I mean there's a lot to like, there's a lot to be skeptical of and there's a lot of uncertainty. But we are going to bring on a guest, surprise guest, Jeff Huber, who was not in the announcement post but will be joining at 11:30 to break it all down. But let's start with AI at Meta on X they post Today is the start of a new era of natively multimodal AI innovation. Today we're announcing the first Llama 4 models. This dropped on the weekend. Shout out to Zuck for grinding constantly. You love it. Llama 4 Scout and Llama 4 Maverick are most advanced models yet and they're and the best in the class for multimodality. Always funny when you say our most advanced model. Our most. It's our most powerful iPhone ever because if it was less powerful you wouldn't release it. Like you everything you do should be superlative in the context of your own business. Anyway, it clearly did mark a step forward and they nailed some other superlative news because they had the industry leading context window of 10 million tokens. Of course that means how much, how much information can you stuff in the into a prompt and still get reliable results out? Google took it up to 1 million tokens and that was amazing. You could upload a two hour podcast, you could upload an entire TVPN episode, start asking questions and it would be able to find things. It was pretty good at least. Gemini. I saw some demos of them picking needles out of the haystack. Hey, I changed one word in this entire book. Can you go find it? And it would do it. Very, very cool. Bit of a debate online about what that means for rag retrieval augmented generation, which is where you load up a bunch of documents into something that the LLM can kind of process through and there's debate now about is, does, does large context windows, do they kill rag? Well, we're having Jeffrey Huber on from Chroma. He's a RAG expert and has built vector databases, and he will defend his position and I think he'll have some interesting takes. They also launched LLAMA for Maverick. That's a 17 billion active parameter model with 128 experts. So these are mixture of experts models. So there's a little bit of internal routing to find, you know, what, what, what neurons in the LLM need to be activated to go after math or writing or poetry or whatever. And then they also have image class, best in class image grounding, with the ability to align user prompts with relevant visual concepts and anchor model responses to regions in the image. And so a lot of these models are going multimodal, so they can deal with images and text. And that's very important because obviously we as humans can process both images and text.

So.

So if you want to make something that's human level or AGI or even close to it, you got to be able to do everything that the human can do. And so this is where it gets controversial. They. They said we have unparalleled performance to cost ratio with a chat version scoring Elo 1417 on LM Arena. That's where all the chatbots battle and humans score which ones they like. Very controversial. People are saying that the results don't tell the full story. So we're going to dig into that. But first I want to go to.

Who was it was saying that the, the vibes were way off.

I mean, Rune was talking trash about LLAMA when it dropped initially. Of course, he's pretty aligned with OpenAI. I think everyone knows that at this point. But Rune was saying, you know, everyone was like, oh, OpenAI's cooked because LLAMA is now open source. And he tweeted just like, have you talked to that thing? Like, and it was this idea of like, who cares about, like, whether or not they have the same number of parameters or it's open source. It's like, do you have a good experience actually chatting with it? And that's where the vibes are off. But there has been more commentary about the Vibes and we will get into that. But first I want to hear from Zuck himself. We have a clip from Zuckerberg explaining Meta's open source AI strategy. So we're going to play that and then we'll use that as the backbone of this analysis to really kick off how he's thinking about open source AI at Meta, because it did kind of come out of left field with VR. They're very closed source, clearly going towards let's build a platform, let's lock everyone in. But took a very different tact. And to be fair, some of the open, some of the VR work is open source. But. And they do want to build an ecosystem, but they're being much more aggressive about open sourcing in AI and there's a lot of good reasons for that. Ben Thompson has broken that down and we've seen Ben Thompson's made a very convincing argument for their strategy.

Yeah. Basically why open sourcing this and just making it widely available for free will benefit their ad business long term. Which is the real, you know, cash.

It's a bad day to not be commoditizing your compliments.

That's right. You always want to be commoditizing your compliments.

But let's hear it from Zuck himself.

Let's do it.

Do we have.

My view is that open source is a really important ingredient to having a positive AI future. I think there are all these awesome things that AI is going to bring in terms of productivity gains and creativity enhancements for people and hopefully it'll help us with research and things like that. But I think open source is an important part of how we make sure that this benefits everyone and is accessible to everyone and isn't something that's just locked into a handful of big companies.

At the same time.

I actually think that open source is going to end up being the safer and more secure way to develop AI. I know that there's sort of a debate today about is open source safe? And I actually take the different position on it. Not only do I think it's safe, I think it's safer than the alternative of closed development. And a realistic aim that we should.

Hope for.

Is that we use open source to basically develop the leading and most robust ecosystem in the world. In that we have an expectation that our companies work closely with our government and allied governments on national security. So that way our governments can persistently just be integrating the latest technology and have a, you know, whatever it is, a six month advantage, eight month advantage on our adversaries. And I think that that's, you know, I don't know that in this world you get a 10 year permanent advantage, but I think a kind of perpetual lead actually will make us more safe in one where we're leading than the model that others are advocating, which is, okay, you have a small number of closed labs, they lock down development, we probably risk being in the lead at all. Probably the other governments are getting access to it. That's my view. I actually think on both these things, spreading prosperity more evenly around the world, making it so that there can be more progress. And on safety, I think we're basically just going to find over time that open source leads. Look, there are going to be issues, right? It's like we'll have to mitigate the issues. We're going to test everything rigorously. We do. We work with governments on all this stuff. We'll continue doing that. But that's my view of kind of where the equilibrium I think will settle out. Given what I know today.

I think it's fascinating looking back at that historical clip and seeing how incredibly front and center AI safety was. And then you look at the Llama 4 announcement today and no one's saying, oh well, like Llama 4 is like not safe or we should be having a safety debate. It's all about the benchmarks. It's like, it's not superhuman enough, it's not aggressive enough. And so we've kind of blown past that. But again, I do think there is a good AI safety argument to be had about open source. And I think it's played out kind of like he said, like it's kind of good that, you know, at the very least, it's like when you open source something like Llama, it very easily can get in the hands of us. Let's, let's not go to paperclipping. Let's just go to, you know, fraud on your grandma. Right. Sending spam texts that are LLM generated so they're a little bit more convincing. We haven't really seen an epidemic of that yet. And there's been just as much economic force towards preventing that type of spam and scams that the open source, like the net impact, I think has still been positive. You get, you get plenty of, plenty of small companies who. Or kids who. Yeah, I have a GPU rig that I used to game on. That was my Christmas present and now I can fine tune Llama and make some app or deploy it really cheaply and that's a net benefit. And the scammers aren't really getting away. Like, I keep going back to the election and it's very hard to make the argument that AI swung the election. Yeah, right. Even though, I don't know, both parties would probably have used that.

Zuck has been accused of doing that, you know, or being with the 2016 election that. Yeah, yeah.

But but it's much harder to make, which is weird.

I just think it's funny. I love that he takes the position generally that he's like, you know, it would just be. I really think we should avoid having, like, a few big companies control this, like, very important technology.

Except in social networking.

Meanwhile, Meta controlling, you know, 20% of the US digital ad spend, you know, not even including, you know, social networking, which I'm sure is significantly higher.

And also just this idea of, like, if you want to put something on the Internet, increasingly like, like this idea of, like the open Internet where everyone has a website and they all have their own style guides and it's all like this chaotic, you know, like, what do they call it, Web 1.0, Web 2.0 or something? I don't know. The. That independent web has really, like, disappeared because of Meta's, like, power over it. But. But, you know, I don't know. It still makes sense, and I think it makes more sense from a strategy thing. He's kind of making some arguments that sound good in theory, but maybe aren't fully motivated. They're more motivated by just his business needs and.

But I think his business dog wants to win.

I think business needs are real. Yeah, I think they're valid. Valid.

No. There's two things that can happen simultaneously. One, open sourcing Llama and allowing anybody to build on top of it and do what they want with it is a net benefit for the world. It also very clearly is highly strategic. He's doing it because he wants Meta to be a much bigger company in 10 years from today than it is, you know.

Yeah, yeah, no, 100%. But if you are trying to look like Mark Zuckerberg, you got to get on bezel. Go to getbezel.com shop over 24,500 luxury watches fully authenticated in house by Bezel's team of experts. You know, he's got a Cubitus, he's got a Patek, he's got an.

Pretty much.

He's got everything. He's got it all. And now's your chance to catch up to Zuck by going big on Bezel. So download the app too. The Bezel app is fantastic. Highly recommend it. You can scroll, filter, find what you like, create a little wish list, and then pretty soon start knocking those down as you send wires off to get Holy Trinity watches.

Yeah, I want to. I'm going to have. We're going to have Quaid on again this week to talk about the watch industry's reaction to the Tariffs.

Yep.

Which has not been, you know, switch your line specifically has been targeted.

But moving on to the reaction to llama4, the headlines were generally glowing. Two mediums, two new medium sized mixture of expert open models that score well and a third 2 trillion parameter behemoth is promised. So they didn't launch that yet, they're still training it. That should be the largest open model ever released. And again, openness is spectrum here. There's open weights where you can fine tune it. There's fully open source where you can actually see the code and all the changes. There's open data, you could have the data open source that they trained it on. And also this is open source. But all the big tech companies are doing this funny thing where they're like, hey, anyone can use this. Really? Anyone? Except for if you have over a billion, over like $500 billion in revenue or something like that. And it's basically just to exclude the other big tech companies and they don't care. They set it like right. Whatever Snapchat's revenue is or user base is, they're like, if they have 10% less, you can use it.

Yeah.

Which is like, honestly amazing for a lot of entrepreneurs. So it's cool. But it's very funny that they're like, you know, we're not actually going to help my competitors.

Evan might kick the bots off and be like, oh yeah, game now.

But it's funny because like the history of open source has been like MIT license. You can even just take this code and just go and sell it immediately. And if you can get someone to buy it, you can make money off of it. Now that they are in, there's a whole variety of open source licenses. But anyway, Meta just got a huge jump. This is from LM arena from 1268 to 1417. And that puts them allegedly higher than OpenAI higher than XAI. But it was hotly debated as we'll get into. And so there were a couple of takeaways here. Llama 4 released on Saturday, the blog post. So they didn't launch like the paper. And there's nowhere near the level of detail from the Llama 3 paper in terms of transparency. And so that's another aspect of open source is sometimes people want to know, hey, what other algorithmic tricks did you come up with? What are you, what are you coming up with? There was interesting. One of the, one of the most fascinating leaks, I guess you could call it from the Llama open sourcing process was that they had a bit of code in there that was just called do not blow up the power station or do not blow up the data center. And basically what they realized was that when they're training Llama, they're running, they're pulling so much energy from the grid that if, if they finish a training run and then the power consumption drops, it will, it will do something with the power substation and the transformers and the data center literally might explode or something like that. So basically what they did was they just said, hey, when we stop training and we're not doing all the matrix multiplication and all the math that you need to do to crunch all these numbers down to create the weights, just do random math. Just, just, just keep doing random math because that will keep. At least the energy will be the same. Obviously it's not efficient, but we need to like wind down the energy consumption slowly, so. Very funny.

Yeah, it's like sprinting, right? Like, yeah, if you're sprinting and then you try to just halt completely, break.

Your knees, slow down, you know. Exactly. And so the smallest scout model is 109 billion parameters. And this cannot run on consumer grade GPUs. And there was this funny interaction between, oh, what's his name over at Google. He's an absolute legend. I forget. Anyway, one of the top. Jeff Dean. Jeff Dean. He's like the greatest programmer in history. And someone was like, oh, this is such a bummer, I can't run the new Llama on my consumer grade GPUs. And he was like, what are you talking about? Like, of course you can. And somebody was like, oh, like, like Google, like GPU expert, AI expert, like discovers what it means to actually have a consumer gpu. Because whoever this was clearly was talking about like an Nvidia gaming PC. But I'm sure Jeff Dean's like, consumer rig is probably like $50,000. Because they're just like, here, Jeff, like, why don't you just take the best thing of everything? Even when you're training at home, you know, you want to be able to.

Run this cost as much as a house.

Yeah, exactly, exactly. And so there's also this, this question about the claimed context token window. They're claiming 10 million and it's certainly above what the real context is, but it might not actually be 10 million. We're going to get into this with Jeff, but there's this question of when you zoom out the context window and you get so big, just like a human. If you're walking around a library, you're not. You might have access to Every book in the library, but you can't actually recall all of that. So LLM seem to be. There's this debate right now in this take that the really, really high context windows, maybe you get fuzzier as you get wider, just like a human. And so that's where something like rag and search and, and deep research from OpenAI, like it is a big context model, but really what it's doing is it's like going searching a webpage that's maybe 10,000 tokens, compressing that down, finding the, the key insight, quoting that in. And so when you get a deep research report, it's not really that it's stuffing all of it into one context window, it's that it's doing this thing iteratively, like an agent.

Yep.

And then there was. And then this is where it gets controversial. So LM arena, we talked about how they're scoring very high, but they used a special experimental version for LM arena which caused the good score. That's not the version that was released. This discrepancy forced LM arena to respond by releasing the full data set for evals. And it does very poorly on independent benchmarks like AI Dir. And so now there's so many different benchmarks out there that you can, you can kind of game one or a few or the top ones, but if someone comes up and says like, oh well, you're actually doing worse on ARC AGI, it's like, well, you didn't get a chance to fine tune on that. So. So if you underperform like a truly breakthrough genius, LLM should just be better at every benchmark, even my benchmark of tell me a joke. And so, you know, it's tricky. There's this game of like, we got to, we got to rank on the important benchmarks, but now there's such a long tail that you can't really, can't really optimize for all of them. And then there's an unsubstantiated post on Channel Chinese social media that we created covered on Monday that claims the company leadership pushed for training, for training more aggressively to meet Zuck's goals. But this was categorically denied by Meta leadership. And we should go into what Ahmad over at Meta is saying.

He says, we're glad to start seeing llama4 in all your hands. We're already hearing lots of great results people are getting with these models. That said, we're also hearing some reports of mixed quality across different services since we dropped the models as soon as they were ready. We expect it'll take several days for all the public implementations to get dialed in. We'll keep working through our bug fixes and onboarding partners. We've also heard claims that we trained on test sets. That's simply not true and we would never do that. Our best understanding is that the variable quality people are seeing is due to needing to stabilize implementations. We believe the Llama 4 models are a significant advancement and we're looking forward to working with the community to unlock their value.

Yeah, I think what's happening here is that we are hitting the pre training plateau. As we've heard before. We talked to Dwarkesh about this and, and that's like underwhelming. And then at the same time, everyone has such strong opinions about all the foundation labs. Like you meet someone, it's like, oh, you're an anthropic guy, oh you're an XAI guy, oh you're a meta bull or whatever. And so there's a lot of emotions that go into these things. At the bottom, I think the correct frame of mind to evaluate Llama for in is. Is what will people do with this that they couldn't do with a closed source model. And so when, when, when Google open sourced their dream studio, there were, there were people that were able to fine tune that and create those magic avatars which are now kind of the studio Ghibli's where you could upload a couple photos, get a, get a custom avatar of you looking like Superman or something. And so we what's cool about llama4 is that because it's open weights, it's easier to fine tune. And also it's not from deep sea, so it's a little bit less politically controversial. But there's still a lot that you can do with it. It doesn't run on consumer GPUs. Now I'm sure that won't be a problem. I'm sure that people are going to figure out how to distill this and do a bunch of different things. And when we go through the history of Llama, we will talk about some of the ecosystem that has actually cropped up and is exactly what Zuckerberg was describing in that.

You want to hear a joke from Meta AI, Please tell me a joke like Theo Vaughn about Sam Altman. Sam Altman is the ultimate tech bro. He's out here trying to make AI do all the work so we can just sit around and think about how great we are. I mean, I'm pretty sure his five year Plan is to invent a robot that makes avocado toast and brings it to him while he's meditating on a beach somewhere. Because that's the real dream, right?

Wow.

I can't.

AGI has been achieved internally, externally, externally, everywhere. Hang it up, folks. Quit your job.

That's brutal. And to be clear, I'm not actually. It's not clear that I don't think llama 4 is actually running on meta AI.

No, I think you're probably using llama 3. But these rollouts are always really staged. In fact.

Was that the style of Theo Vaughn?

I thought I was listening to Theo Vaughn.

Yeah.

It could have fooled me. That is remarkable. Was that a Theo Vaughn quote? It's remarkable how spot on that is. You listen to.

Making an avocado joke is like very 2012 coded. Yeah, like 20, 2018.

Hipsters are quirky.

Yeah.

Okay, we're past that. Anyway, so there are mixed quality reports across different services using LLAMA 4. And implement implementations should stabilize in a few days. This is kind of common when they roll out these big new models, they gotta figure out how to run them on all the hardware, get them into the data center, swap things over. Like people say, oh, yeah, switching an LLM. It is just one line of code. But there are more things to it, especially on the performance side. We've seen this with Studio ghibli. Like, the GPU's melting, which I think we all believe is real, because how many times. I mean, this happened to me? A bunch where I've said, hey, make this image Studio Ghibli. And it's just like, hey, I stopped and it's like, what? Like, no, like, Instagram filters don't just stop halfway.

But we talked about this with Aiden. Or was it Aiden or. No, Swix. Swix was saying that the models are already showing signs of needing rest.

Right?

Yep.

Anyways, it's crazy.

And so there's a bunch more going on. Let's move through this. So there's. There's a couple themes that are sticking out in the discussion about Meta's Llama 4 performance. The big one is. Is just a general disappointment from the AI community, I think, based around the. How much horsepower was going into this. So the claim was that they trained on 100,000h, 100 GPUs. Of course, Zuck and Jensen have done the famous Jersey swap. He's one of the biggest Nvidia customers. He can get the best. He's not under any import restrictions. There's really nothing stopping him. And most importantly potentially is the fact that Meta can really, really go full send on the capex here. Because Zuck knows that hey, if, if Llama doesn't go anywhere, we never. We like LLMs cap out. It's not important. Yeah, we're going to use those 100K H1 hundreds to train the REELS algorithm better or, or the new thing that was the whole thesis behind.

Yeah, or at the very least do our own Ghibli style.

Yeah, yeah, yeah. I was thinking about this. Like, if I wanted to, if I was like the pm, if I was like the PM at Instagram, I would immediately implement the Studio Ghibli filter and just send every Instagram user a Ghibli of their most popular post or of their profile picture. Just pre render it all, just batch them all and then send them and just say, hey, do you want to try the new filter? And everyone would be like, this is amazing. It would be this amazing viral moment. They could definitely do that, but it would be extremely expensive from. From a inference perspective. But they can probably afford it and it would be cool and delightful and I think they should do it anyway. So. Despite having fewer resources, Deepseek claims to have achieved better performance with models like deep seq v3. And there are some benchmarks where Deep SEQ is still outperforming llama 4, which you hate to see if you're duking it out in the open source world. Yann lecun stated that FAIR is working on the next generation of AI architectures beyond autoregressive LLMs. And so this is a debate that we've been hearing for a while, like we probably scale is important and we need to continue to scale and we want to do big data center build outs, but we also need new algorithms on top of those. The poster suggests that Meta's leading edge is diminishing and that smaller open source models have been surpassed by Quinn with Quin 3 is coming. And then there's another debate about Meta's. Meta's llama 4 fell short. Scout and Maverick have been released but are disappointing. Meta's AI research lead has allegedly been fired. The models use a mixture of expert setup with a small expert size of 17 billion parameters, which is considered small nowadays. Despite having extensive GPU resources and data, Meta's efforts are not yielding successful models. And so I think that there's a debate about. George Hotz was talking about this when GPT4 launched. It was a mixture of experts model and a lot of that is defined by the actual structure of the chips and the interconnect and what we talked about with light. Lightwave.

Light matter.

Light matter. Yeah, Just this idea that yes, you can have a hundred thousand GPUs, but if they're not networked together really, really well, you're, maybe they're memory constrained. There's all these different parameters that can constrain you and so you wind up having to fractionalize your, your, your, your LLM. And that can be fine if, if 17 billion is enough and, and, and, and you can route and there aren't any problems that require multiple experts or bigger experts. But clearly in this scenario a lot of people are disappointed. And so someone said they left me really disappointed. You hate to be disappointed. I mean, what's free magical intelligence that show some, show some respect. There's some respect for Zuck giving you something that cost a billion dollars for free.

Something that like it three years ago would have been groundbreaking. Yeah, no, but, but the expectations are extreme. Yep, they're, they're spending, you know, almost as much as anybody on this. And ultimately it's becoming clear that ability to spend is not all that. You know, it takes finesse too. And so it's like, okay, meta can.

Yeah. And so people are, people are joking. I'd like to see Zuckerberg try to replace mid level engineers with llama 4. And one commenter joked that perhaps Zuckerberg replaced engineers with llama 3 leading to llama 4 not turning out well. Brutal.

Ouch.

Another commenter suggests he might use, he might need to use Gemini 2.5 Pro instead. I love that people are just like, absolutely trash talking with the most like industry jargon here. Like, oh man, like this is more of like, like you're, you're llama 4 coded, not Gemini 2.5 pro coded. It's like guys having too much fun anyway.

Yeah, this one was even more brutal. Somebody saying that, calling it a complete joke and expressing doubt that it can replace even a well trained high school student.

Oh yeah, yeah. I mean, in general I think my takeaway is like, you know, llama 4 might be underperforming, but you can't sleep on AI even for a minute. There's a new model every day. You can't sleep on AI innovation, but you can sleep on an eight sleep. So go to eight sleep.com night sleep that fuel your best days. Turn to the bed into the ultimate sleeping experience code tbpn. So this was the debate that was popping up from Sean who came on the show last week he says unpopular opinion right now, but Llama's 4. 10 million token window will finally actually end the long context versus rag debate. That's retrieval augmented generation, but not the way that other guy is thinking. And so this was a very like I'm going to, we're going to talk to Jeff about this because I was like hey, you got to come on the show just to explain this. But I think what he's saying is that is that huge context window is not a perfect substitute for rag, and RAG still might have a place in the future of AI agent development, AI implementation. But I'm excited to dig into this because I didn't fully understand this post and I want to know more here. Anyway, let's, let's move on to another LLM eval eval that happened with Stagehand. The results are fascinating. What LLMs can actually do consists of can actually parse deeply nested structured data like a DOM document object model, an A11Y tree, which is like parsing a website basically. And so this founder benchmarked Llama 4 against other models and found that Llama 4 came in maybe 8th below GPT4O Mini, below Claude 3.5 Sonnet and below Deepseek V3 and then Gemini 1.5 Pro 2.0 Flash are all higher. And so at least from his benchmarking he thinks that, you know, given that this is the latest and greatest and newest product from Meta, it's underperforming stuff that's been on the market for a couple months. So not the best information then. There's actually an example here that from Vic that I thought was pretty good. This is the clearest evidence that no one should take these rankings seriously. In this example it's super yappy and factually inaccurate and yet the user voted for llama4. The rest aren't any better. So this is what, this is what a LLM arena actually or LM arena like interface actually looks like. So there's a prompt and then you see both of these. You don't see the names of the models I believe, but you have to pick which one you like more. And so the question was what is the latest season of Fortnite and Claude 3.5 Sonnet said Fortnite Chapter 5 Season 2 titled Myths and Mortals is the current season. It begins on March 8, 2024 and features Greek mythology theme with characters and locations inspired by Greek gods and legends. Then llama4maverick experimental says a super timely question. Fortnite seasons are very short lived and they last about 12. And it's like yapping, yapping, yapping. And then, and then it gets it wrong. It says the current season is Fortnite season og, also known as chapter four season six, which I believe is like before chapter five. Obviously I don't know enough about Fortnite to fact check this, but it seems like it's wrong. And then, and then there's a bunch of emojis. And so there's a debate from LM arena and they go on to write. We've seen questions from the community about the Latest release of LLAMA 4 on arena to ensure full transparency release, we're releasing 2000 head to head battle results for public review. So anyone can go see these and decide for themselves. Hey, did, did the folks at LM arena get this wrong or are they, are they happy? Early analysis shows style and model response tone was an important factor demonstrated in style control ranking. And we are conducting a deeper analysis to enjoy to understand more emoji control. Because some people might just be voting, oh, I like the emojis. Even though that's kind of taking you away from the facts.

Yeah.

Or they might just be like, hey, I like.

Some people want it. Yeah. Some people want a super short answer.

Yeah, exactly, exactly. So there's all these like human biases that are coming in and at this point we are in this like qualitative, you know, like the open air guys say just talk to the model, figure out the vibe of the model. People seem to like Claude Sonnet and just the vibe of that of that model. And so there's still more debate over company leadership potentially blending test sets from various benchmarks during the post training process. Metta has denied that allegation, but there is a lot of debate raging. And Ethan Moleck says the Llama 4 model, that one LM arena is different than the released version. I have been comparing the answers from arena to the release model. They aren't close. And so what he did was he was he looked at the, the actual results that were posted on Elamarena and then same query on the llama4 model that was released. And this is evidence that they, that they went to Elamarina with a separate model.

Yeah.

Which is controversial.

Not what you want to do.

I mean it doesn't inspire aggressive approach Cotton. Yeah, but you got to be interested.

To see what Jeff thinks about all this.

And so Ahmad says, seems like there was a lot of truth in this leak from two months ago. Lama 4 is beyond disappointing. It's a model that shouldn't have been released. And this is from probably blind or something. Meta Jenny I Organ panic mode started with Deep Seq v3 which rendered Llama 4 already behind in benchmarks. Adding insult to injury was unknown Chinese company.

I remember this post in training Budget.

And and there was that meme of, of Iron man being like they built this with screws in a cave. Like you have a trillion dollar budget but if you want to control your budget you got to go over to ramp.com Time is money.

Save both, save both.

Corporate cards, bill payments, accounting and a whole lot more all in one place. And so Jeff is going to be joining in just a minute. We will run through some of the timeline and break down what happened with Llama and the development here. The most interesting thing that I found when I was digging through the history of Llama was the company's notable foray into large language models was an academic tool called Galactica, which most people hadn't heard about. That actually backfired months earlier. The demo was pulled after only three days amid criticisms that it confidently generated false information. And if you remember GPT3 like it was, you get some wild hallucinations out of that thing. And so Meta's leadership was cautious at that point as generative AI fever swept tech. And once ChatGPT came out and it became like customers want this, then they started to push forward. And so there was a team at Meta's fair research lab in Paris. I believe this is the one the KUN is involved in. They were, they were hard at work on large language model they believe could compete. Llama 1 was the fruit of that effort. A set of models ranging from 7 billion to 65 billion parameters trained on a rich diet of text, which of course they have because they have every piece of text. Not only do they have everything in the Facebook ecosystem, but they also scrape every link that's shared to Facebook, which.

Is every link ever.

And so they have the entire view scraped. Unlike OpenAI's headline grabbing GPT3, Llama wasn't offered as a public chatbot or API. Instead its initial release Meta. In its initial release, Meta made the models weights available on a case by case basis to academic researchers. So you would just email them and say like hey I'm at Stanford, like can I have the weights? And they'd say sure. And then of course that leaked immediately, which is awesome. So it was a non commercial license. So they would send you the weights and then you could mess around with it and test. Then of course someone it was quote unquote Open research, not open source. So you couldn't build a company. On top of was more like a research paper with some downloadable code. But this leaked onto the Internet and then developers everywhere had their hands on a GPT class model in raw form. And so that spawned these fine tuned models, Alpaca and Vicuna, which are derivatives of Llama. I think they're related mammals. And so then people started fine tuning with instructions and getting it more into a chat mode. And then Llama became a product that had close to a chatgpt like experience. And so I'm excited to talk to Jeff about this. I'm not sure if he's in the temple.

He's in the waiting room.

Let's bring him in. Jeff, how are you doing?

I'm going great.

How have you been following? Well, first, welcome to the stream. Can you do a little introduction? But then I want to hear your reaction to the llama 4 news, how you're processing and what it means for your business.

For sure, yeah. I'm Jeff Huber, the co founder of Chroma. We're working on retrieval for AI and broadly working with developers kind of across the ecosystem to build production systems with AI. A lot of it is focused on business applications, good old fashioned business process automation and so always super excited to see new open source model jobs.

Can we go to this post from Sean? He says unpopular opinion right now, but Llama's 4,10 million token window will finally actually end the long context versus rag debate. But not in the way that other guy is thinking. What does he mean by that?

Yeah, for sure. I think Silicon Valley has a tendency to be sort of extremely intellectually shallow. This is both a strength and a weakness of the valley. To be clear, in our view, AI is not this deus ex machina, this technical machine God where all of the information of all time is always going to be in the weights of this model. This is really just a new form of computing. And so in the same way that we have a memory hierarchy in classic computers, we have the cpu, ram, disk and network, we are also going to have a similar memory hierarchy in language models. And again it already exists today. We have the actual sort of transformer attention heads, we have the context window, we have the retrieval system and tool use. And these things have different trade offs. Right? You think about kind of access, speed, capacity and cost. There are trade offs to all of these things. You know, I think like, you know, saying something is dead like plays pretty well on Twitter. I've actually gotten myself into some trouble Where I, some people were allegedly shit posting and not actually sort of sincere posting about this. And I didn't know, right, because they're hard to tell like what's a shit post and what isn't, of course. But you know, the bait, the bait is strong on Twitter. And so what Sean is saying is actually that like we've all been. There's a certain class of people who are like, long context is all you need. Again, these people are probably like 21 years old. That's fine, we love them. But they just haven't seen how like real systems depend on trade offs between speed, cost and accuracy. And like 10 million tokens is not a panacea. You need to keep information outside of the context window. You need to give developers and programmers control over what information is inside the context window. You know, even these like needle in a haystack tests, like, are not actually that representative of like real world utility and reliability of long context windows. You know, they mentioned in the training for llama 4 they don't even have passages that are longer than I think 250,000 tokens. Anything 250,000 is just synthetic data sort of just made up. And so what Sean is saying is that like what 10 million tokens is finally context window for llama 4 is finally going to put to rest is that long context windows are all unique. He's going, you know, the 10 million, the 10 million context window length is going to finally, hopefully make people understand that like, no, there are different things here that are good at different things and we can put them together to create a good system.

Should we monetize the eschaton?

I don't know how you knew that I was writing about this this morning. No, we absolutely should not.

Yeah, we absolutely should not.

It's always, always been a trail of tears. Let's not do that.

You explained that to me a while back. I had fun with that. So let's talk about Llama 4. How should startups be thinking about Llama 4 as tool in the toolkit against the other options that they have?

Yeah, I mean, I think like, you know, Twitter is quick to. And research in general.

Right.

Is quick to sort of view state of the art as the only thing that matters. And I think that actually in many cases being first is overrated. You know, we've seen, you know, going all the way back to sort of the slack in teams charts. You've seen the famous chart, slack versus teams. Right. Distribution is incredibly important. As long as, you know, sort of the incumbents can Wake up and can catch up. You know, I would not bet against Zuck and $100 billion of profit per year. You know, I think that, you know, Zuck also is in some sense playing a different game. Like he's not trying to build like the, the sort of very best open source like chat experience for consumers with Zuck sees, I think rightly so, is that having an open source model which is really good, is good for the ecosystem and is good for Meta and most businesses don't love using closed source models. They want to use open source models for all kinds of reasons. Privacy, security, continuity, cost. You can build your startup on GPT4 and it's amazing. And then there's a new version out and OpenAI deprecates, you know, the old version. Right. And all of a sudden all of your prompts don't work the same. And so, you know, open source models are going to continue to play an extremely important part in the ecosystem now obviously like, you know, the Deep SEQ R1 launch a few months back like totally took everybody from surprise. You know, I think we're still in the early innings of this stuff where like good ideas can come from anywhere and oftentimes good ideas do come out of the sort of group think context of Silicon Valley. Right. So, you know, but yeah, I wouldn't, I wouldn't bet against that.

Do you think there's an opportunity to build a company like Red Hat in Linux but for LLM implementation on top of something like Llama or is that like a crazy idea that doesn't really match to the modern foundation model landscape?

I mean the bull case for Llama for Meta is that it's actually more equivalent to how Meta open sourced its data center kind of layout and.

Right.

And that's the bull case for, for meta is that actually in industry sort of forms around that and become a standard. Right. That's sort of why, you know what argument for why they did it in terms of like the Red Hat 4, you know, I think that like Red Hat 4 works well for operating systems, but I don't think of an LLM as an operating system. I think an LLM is much more like a cpu. Right. It's information processing and so obviously it's a new thing. It's not exactly like a cpu, but yeah, I'd have to reason about that some more. I'm not sure.

Yeah.

If you're running Meta AI, what would you do from here? Not to put you on the spot or anything.

Yeah, I'm not to be Clear. I'm not running that AI. I've not received that job offer at all. I mean, I think that like, you know, you have to keep going, you can't stop. I think focusing on like the business use cases is pretty important. I think focusing actually also what developers actually need and want out of models is also very important. You know, you see a lot of like model drops that come out, but they don't actually provide the real hooks. And they do very well in the benchmarks. Right. They do very well on like kind of the public leaderboards, but they don't actually provide the hooks that developers need to do like good tool use or reliable structured data output or the practical stuff. Right. The developers actually want out of models. And so like if you want to create a groundswell of developers that like love your tools, like do the developer experience part, like meet them where they are and like give them all the hooks that they need and don't just stop at like, hey, look, you know, we hit state of the art benchmarking. Aren't we special?

Yeah. Is there, is there a narrative here where maybe they're trying to do everything all at once and instead should focus on like Llama is amazing at code or Llama is the next version of llama 5 is like all about tool use or super great at reasoning or just like the best at deep research or just the best at image generation, for example. Like it feels like there's kind of a bifurcation of the market and maybe the opportunity is actually to laser in on something that's high value but then let the other stuff kind of, you know, simmer out there amongst other teams.

I mean, focus is probably always a good lesson for all of us. Right. Do less and do it better.

Yeah.

And so, you know, presumably it's also true for meta, I think also obviously unlimited capital can both be a blessing and a curse in that way. Yeah, again, like focus on developers. Yeah, I think that's the beachhead. That's how you win the B2B market. If you win B2B market with your open source models, like you get all of the sort of downstream effects that you want, you know, you don't need to beat, you know, GPT5 on some.

Yeah. Do you think that part of the narrative that we're seeing around llama 4 is just pre training, scaling, hitting a wall, a need for new algorithms, a need for a deeper focus on reasoning and maybe even whatever comes after that?

I mean, you know, I, I, you know, so you mentioned a moment ago, sort of the, you know, imitating the eschaton.

Right.

You know, throughout history, you know, every exponential that we've observed eventually results in a sigmoid curve. You remember early Covid, right?

Yeah.

The fur of like, oh my gosh, you know, the Twitter guys doing their thing where they're like, well, if double the amount of people get it every day, everybody on earth will have had it seven times.

The next 100 billion people will have it.

Yeah, exactly, exactly.

Yeah.

And so, you know, I think that there are laws of physics here. I think that there are, you know, diminishing. They're clearly diminishing marginal returns.

Right.

We're sort of spending 10x on compute. We're not getting 10x or better models, at least, evidently not yet. And so, you know, the transformer is incredible, is amazing. It's a, you know, technology is probably as important as the invention of electricity. It will probably bring about a increase in GDP that is on the order of the industrial and revolution or greater. And so I think we should not minimize this technology and sort of boil it down to, oh, this is just dumb pattern matching. Right. By the same token, we also should not believe that all technology we're going to be able to rent, seek on sort of forever. So yeah, new things are definitely needed. And you know, I think that like an inference time compute internal chain of thought is really promising. And you know, I look at the stack today and I think about how sophisticated computers are.

Right.

And how architectures are and operating systems and kernels and compilers and all of this stuff and you know, we're just like in the baby phase today of AI, like it's just in its infancy.

And there's a lot to build from a recruiting standpoint. Have you run into some of these super aggressive non competes that we're seeing? There was a headline today about, you know, Google basically paying engineers to not.

Not work at Chroma, not work for.

A year when you know, they could be working at Chroma or any of these other labs.

I mean, yeah, you know, airplane, red dots, png.

Right.

I guess, like if I was affected by that, I wouldn't know it.

So, yeah, can you take us through some of what Chrome is building today and where customers are getting the most value? I've talked to you a little bit about some of the use cases and I think there's underrated potentially in how simple and obvious they are when you explain them. But I want you to take me through some of the modern context.

Yeah, I mean you've heard left and right on the Internet now for like three years, all about this acronym, rag. I don't know why anybody would ever name something rag. That seems like a pretty dumb idea. We just call it retrieval. And of course the idea with retrieval is that if you want to build an AI system and you want to be good at something, well, you need to teach it how to do that. You got to teach it about your data, you got to give it your instruction set, right? And updating the weights of the model is not a very good idea because you cannot really deterministically control that. Right? You can fine tune, but what you're going to get the other end again you don't really control. And so giving the system access to a repository of instructions or knowledge about your organization, your business problems, that is something that you can control. And that's the problem that retrieval solves. And so we talked to enterprises and businesses building useful applications. I think today 90 plus percent of it in enterprises is retrieval generation or it's using retrieval. It's sort of chat on top of unstructured data. I think if you zoom out though and view what is really AI, AI gives us the primitives and the ability to process unstructured data in a common sense fashion. And you think about the scale of data, right? Even today inside of enterprises, unstructured data is 10 times the size of unstructured data. We have 10 times more unstructured data. And then you consider the real world, right? If we were actually putting robots out in the real world, how much unstructured data they're going to be ingesting and needing to process and reason about and action on. And it's just sort of like going to be 1000x, 10000x, 100,000x the data that we have today. And so that's kind of the direction I think is not so much like sort of this sort of simple like one human, one AI talking at a chat stream back and forth, but it's like real embodied intelligence which you know, you could call an agent, you can call a robot. You know, I don't love any of these terms but like really the goal here ultimately I think for anybody who's building something practical is building something that's reliable. You know, you think about like we've been seeing self driving touted as like this technology for like 10 years. And of course if you live in San Francisco, you can use Waymo. And it is actually incredible, but it's taken 10 years. The gap between demo and production has always been so great in AI. And so if you're building something practical in AI, your big question as a developer is like, okay, the demo is super sexy and cool, but how do I actually make it work really, really, really well and reliably? And the ability for these systems to sort of like self improve or improve under human guidance, I would say is the biggest thing that's underrated today. And of course, we think that retrieval plays a key part in kind of how that happens.

Can you concretize that a little bit by walking me through, like a. A potential use case for us? I mean, we stream three hours a day. We're probably emitting, you know, I don't know, tens of thousands of tokens every day. If I used whisper, I tran. I transcribe every minute of our show, I could search that through, you know, fuzzy search or deterministic if I want to. Just search like every time I mention artificial intelligence directly find that. Or I could try and fine tune llama4 on it and maybe it just hallucinates like, oh, yeah, John was talking about this randomly. How would I use Chroma to create a more definitive index of every time John or Guest has talked about artificial intelligence or Llama in hundreds of hours of video? Is that something you could do?

Yeah, we're seeing some kind of fun things today where people are taking the corpus of all of their writing or all of their speeches, and they're kind of like, quote, teaching the model it. They're loading it into a tool like Chroma, hooking it up to a language model, and they're giving end users the ability to like, chat with John. Yeah, and like, you know, see what John thinks about artificial intelligence. Right, and that's exactly right. So kind of all those transcripts get processed to get broken into pieces. They get indexed and searchable in various ways. And then user asks the query, you know, hey, John, what do you think about the latest, you know, llama release? Or maybe they say llama, they say the latest release, Liz AI from Facebook thing, right?

Yeah, exactly. Yeah.

Like, the search is good enough that it can, like, find all the relevant things that you've said, and then the algorithm can, like, respond as you. Because it kind of ground itself in the things that you've said before. So.

And so it's basically taking like different blocks of text, different ideas, and then kind of vectorizing them into some way that's not necessarily human readable, but it can still. It's basically like better fuzzy search in many ways not to degrade what you're doing, but it's amazing. It's magical and super powerful.

Fuzzy. Yeah, fuzzy search is really useful when people like are not, you know, experts in their own data.

Right.

Instead of your Google Drive, you know how to search for stuff pretty well.

Right.

But like your users don't know how to search for the stuff that you've said before. And so that's the kind of the power of like embeddings and vector search toolbox. It's not a panacea. Again, we're not advertising the eschaton here. We're not.

Right.

But it's like a very powerful tool and people are getting a lot of.

Value out of it. Yeah.

I'm curious your reaction to AI 2027. Our point of view generally, just from all the conversations we've had, is that sort of model progress and advancements could sort of slow and that would be fine just because there's so much value to unlock out of the underlying models. I'm curious to think how you processed just the forecast generally and maybe, maybe take it from there.

We think the capability overhang we have in the models that we already have today, and we will have absolutely in six months is immense. Think about, for example, the possibility of democratizing access to state of the art services to everybody on earth. Like it is very possible the poorest people on earth today or in 10 years will have access to better health care, better legal representation, better financial services than billionaires have today. I think that's entirely possible and that's impossible with the models we have again today. And so the capability overhang is immense. Every time an extremely long essay from a sort of effective altruist drops, Right. They clearly tend to make waves. I think if you tell people that the world is going to end, they're going to pay attention. And I'm just not that frankly that interested in secular eschatologies about apocalypse and the end of the world. Right. There's a natural tendency for all humans to believe that we are the chosen ones living in the special time and the last days. Right. Even Fukuyama, right, wanted to sort of end history. Right. And so natural human tendency. This is again the immortalizing the eschaton, we'll mention it three times now, tendency. It's like it's really dangerous, right. You think about what's happened throughout the last hundred years in really, you know, you know, the hundreds of millions of people that have died, you know, across like different world wars and different, you know, dictatorships, like it is oftentimes it's like messianic complex that leads to a lot of that. And so I don't know, I'm just like, I think it's, I see it as entertainment more than anything else.

Yeah. On a more practical note, like I go to the Wall Street Journal's website, I just try and search for an article and they say, oh, search is powered by AI. It's not clear. It's clearly not powered by AI because I cannot fuzzy search at all. I can't say, oh, I know that it mentioned this person and I think it was about this and it was in the last week. It's not there. What does it take to actually roll this stuff out? Are these even potential customers of Chroma or is there another company to be built here? What do you think about that?

Yeah.

Wall Street Journal, if you're watching, you know, send me an email after the chat. Yeah. All that's very doable today. I think that, you know, the reality is that, you know, your classic, you know, the future is already here. It's just not evenly distributed yet.

Right.

Like, you know, any technology of consequence, even if generationally important, you know, still takes decades to roll out.

That makes sense.

And, you know, that's just the same is true here.

So that's great. Well, thanks so much for stopping by. We got to move on, but this was a fantastic conversation. We'll have to have you on.

Thanks for coming on, Jeff.

Appreciate it. Talk to you soon. And we got a big funding announcement. We're shifting gears. We're out of AI and into manufacturing. Going to talk tariffs, going to talk industrialization. Another theme we love on this show. We have some big news and I.

Just want to know, was this fundraise announcement intended, always intended to go out today or did they bring it? Did they bring it up?

Oh, because of the tariffs.

Tariffs and everything. Because the timing possible, just too good.

So Jay says, today I'm excited to launch the Advanced Manufacturing Company of America. We've raised $76 million. Let's hear it for a massive round coming out of stealth from Caffeinated Capital. That's Raymond Tonsing, Founders Fund, Lux Capital, Andreessen Horowitz and others. The best time to build this business is right now. Yeah, no joke. But the real work began decades ago and they launched.

I'm going to get every big fund. Yeah, just get them all.

Yeah, it's great.

Ran a process.

He says yes to everyone. Take a bit from everybody and it's great. They put out a four minute video produced by Jason Carmen, Story company. It's beautifully lit, beautifully shot. And they brought in, you know, we've been hearing for a long time that the legacy manufacturing companies are run by folks who are aging out and maybe they don't have the next generation lined up to take over the business. Well, they sat down, they interviewed one of those folks and it's a fake. Fantastic video. You should go check it out. Anyway, is he ready to come on in the studio? Let's bring him in and hear the news from him directly. Welcome to the studio. How you doing? Congratulations.

Hey, how are you guys?

We're fantastic. Thanks so much for taking the quick moment to chat with us. Can you introduce yourself the company and what's the news today?

Absolutely. So my name is Jay Malik. I'm the CEO of the Advanced Manufacturing Company of America. We call it affectionately amca. And so what we do is we design, manufacture and certify the next generation of critical products that go into all aerospace and defense systems. So that's both existing and new systems, the stuff that Boeing makes and the stuff that Anduril is going to make.

Okay, can you break down a little bit more of what the first products that you'll make will look like? We've heard about what Hadrian's doing. We've heard about, you know, injection molding, plastics. Like there's a lot of different buzzwords. Obviously everyone kind of wants to do everything in the long term. But what are you focused on first?

Yeah, so first, let me, let me just start a high level, right? Like, please, when we talk about the aerospace and defense frames like Lockheed or Boeing, they don't make anything today. Right. They've outsourced a lot of their manufacturing and engineering to thousands of suppliers. Some suppliers are focused on high volume manufacturing, things like wire harnesses, machine parts of Hadrian's doing, injection molding. There's a lot of suppliers that are focused on that, but there are also hundreds of suppliers that are focused on critical engineered products. Those are the products that, that, that, you know, basically determine system success or failure and are often, you know, very, very highly specialized. So stuff like avionics products, power units, you know, specific engine products. And so we're focused in those areas in the most critical areas where you need to both engineer and manufacture at relatively low volumes for the end customer for their system to succeed. So we're focused on a pretty different, I would say part of the market compared to most of the sort of software defined manufacturing startups that you often see today. In terms of where we're starting, we're starting Almost entirely on avionics, the part of the plane or the ground control system that involves controlling it. So the stuff we see in a cockpit, for example, so we're focused on things like switches, panels, displays, power units, you know, things that are critical to the pilot if it's a manned system and you know, if it's an unmanned system critical to communication and executing on the mission. So those are the areas that we're, we're mostly focused on right now.

Can you talk about the timing of the announcement? Was it just a happy accident you were always planning to, to go out this week or did you pull it forward due to the. Everything in the news?

A little minor news this week.

We were supposed to launch this week anyway. We actually had a few reporters that I think got scared of the tariffs, you know, didn't want to cover anything. And so we basically said, I'm not sure if I'm on a curse of this podcast, but after that and said that this is actually the best effing time to like, you know, take our company public. So we just did it. And so yeah, it was, it was planned but, but obviously timing is definitely in our favor.

Can you talk about. This is obviously a big raise to come out the gates with. Can you talk about, you know, kind of the use of, of proceeds and kind of like. I'm curious about kind of like how you're thinking of the structure of the business generally.

Yeah.

So we're going to be. We've already acquired one business that is a critical avionics supplier which you saw a video or some people may have seen a video about. We're going to be acquiring probably another two to three of them over the next 12 to 18 months. We're also going to be doing our own clean sheet design and development of adjacent products in this space with our own manufacturing and engineering talent. So it's a hybrid approach. I'm a firm believer that especially in this area of the supply chain, you can't just hack your way into it. You also can't just be a private equity firm and buy and price it up. That's not going to achieve what companies like Anduril want to achieve for their customers. We're taking a hybrid approach where we're buying companies with products that we think are going to be hard to just redesign from scratch and also developing ones that we think we can do a great job of ourselves and is part.

Of your advantage over traditional private equity is just the time horizon you're thinking about, of saying we don't need to come in and just immediately cut costs by 50% and increase pricing and hopefully flip the business in three years. I imagine you're buying to hold and that's part of why somebody would want to sell to you in the first place, I imagine.

Yeah, it actually goes deeper than that. So I would say the one thing that I have learned building this business so far, I believe, you know, it's still early, is, is that owners don't really necessarily care just about the time horizon, your ability obviously to underwrite, you know, the deal. They also care about one, not selling to MBAs, not selling, you know, to, to traditional finance people. You'd be surprised. It's big. It's a big thing for them. And that second that you know what you're doing, meaning like, you know, you're not a bunch of like search funders. You're not, you know, a bunch of people. You know, this is stuff from the MBA argument, but like you're not a bunch of people that haven't spent time in manufacturing, you know, shop floor, etc. And so our entire team are engineering and manufacturing folks. We spent our entire careers, you know, designing things for Space X, manufacturing things. You know, we're also young, which I think people like to see when, when they're selling their business. They know that just looking, you know, looking at the person across the table, that they're going to be there for that for the next 20 or 30 years. So I'd say all of those things combined make it a pretty strong pitch for Wine to sell to us.

Did you have this idea in mind or a rough idea of it when you decided to go back, you know, enter your next chapter? I remember it felt like a year ago when you decided to move on from, from active investing. It felt like, I just remember that like instantly you shared it and it just was like everywhere.

Yeah.

And because people at that time, it was like every non deep tech, hard tech investor was like starting to pile into the category and then everybody's like.

Wait, wait, if you can't do it, I'm screwed.

What are we doing? It didn't slow anything down obviously. But I'm curious like, you know, kind.

Of the origin in this week.

Yeah, I'm curious like how, how the idea and the opportunity came together.

Yeah. So I, you know, I spent three or four years, obviously my career at Countdown when I was 24 years old, you know, I started the firm, I spent a lot of time, you know, with, with manufacturing startups and also with mom and pop suppliers, like as part of my diligence for whether I should invest in companies, I would, you know, talk to mom and pops. And so I'd spend, you know, three or four years in the space. And it didn't really click, I think, until after I shut down Countdown, that one. The mom and pops have both the expertise and in some cases the qualifications that you need in order to develop and manufacture and bring the product to market. And then it also didn't occur to me, obviously, when I was venture investing that maybe there is a path where you can combine the mom and pop advantages with the spirit, the culture, the talent of a startup. When you're venture investing three or four years, every single day, your mind is just like, startup, startup, startup. New things, new things, new things. You're not even able to think about what does the future look like using something that already exists. And so it wasn't until I had actually shut down Countdown, had a month and a half to reflect, think about what I had learned, wrote down some. Wrote down some key themes, and then started to iterate from there. You know, talk to people, talk to customers, industry, talk to people at companies that are already, you know, building very successful ones, both in startup world and in mom and pop world. And that's when the vision started to come together. Like, hey, I have. I'm uniquely in the center of these two movements, right? Like, I helped, I think, start invest in a lot of startups in this space. And at the same time, I know a lot of people who are in the traditional, in the traditional world, and I should use that to. To the maximum advantage that I can.

I have one last question and then I'll let you go because I know you're busy today. Charlie Munger criticized Transdigm for buying aerospace parts manufacturers and then locking primes and. And aerospace companies into long contracts, raising prices. It was a little bit of a controversial strategy, but it's performed very well for that company. What is your takeaway from the TransDigm model?

TransDigm is actually a phenomenal business, and it's actually not the cause of any of those issues. The cause of those issues, certification lock in, et cetera, et cetera, has to do with decisions that were made 30 years ago at the top of Boeing, which is a paper right here, basically pillorying that decision, but the decision at the top of Boeing to outsource every single thing that they do from engineering and manufacturing at the component and at part level, up to the product level. And so Transdyme is just a recipient of the system that was instituted 30 years ago. If you really want to change things, we believe that you have to start from the, from the bottoms up with the critical products, build your way back up in partnership with the customer to reverse that type of decision making and culture. So, yeah, my answer to that is I think Transdigm is actually a great business. They have, they have run businesses very, very well. They're in an environment that, that was not created by them. They have taken advantage of it, but it was not created by them. And to fix that, it's going to take partnership with a company like apps.

Well, that's a fantastic answer. Thanks so much for hopping on on short notice. Congratulations on the round and we'll have to have you back.

Yeah, I have so many other questions.

And we can talk for an hour, I'm sure.

Yeah. Very excited for you and the team and excited to have you back on the show soon.

Thank you. Let's do it. Take care, guys.

Talk to you soon. Bye. Well, we are moving on to someone from fai, the foundation of American innovation, I believe is. They call it fai.org, the fai.org I've been to a couple of their events. Very fun. Gary Tan spoke at one, Trey Stevens spoke at one. Yeah, Stevens has been involved. I went to one in San Francisco and there were actually protesters outside, which is kind of fun. But they were like in, in very good spirits and kind of like taking pictures with everyone. It was a lot of fun. Anyway, welcome to the stream.

Boom.

How you doing?

Hey, man, how's it going?

It's good.

What's going on?

Thanks for hopping on point as usual.

Chill week. Chill week for you.

Sleepless nights, probably on maybe eight hours over three days.

Oh, good.

You should get an eight sleep. Go to eight sleep.com tvpn.

I have a helix.

No, no, we'll get you a switch anyway.

No, it's great to have you on. What's running through your brain. There's a bunch of things we can talk about, but where should we start?

Oh, just the contagion effects and potential collapse of the world economy. Simple stuff like that. You know, how bad is U.S. primacy?

Is there any element of cautiously optimistic about this for you or are you just totally black belt on it?

I mean, my white pills are Lucy's, so I do have some of those. But you know, you're going to need a higher milligram for this week.

Well, walk me through. Why is it so disruptive to you and what you do and maybe just for the viewers, give a Little background on yourself and the organization.

Sure. So I'm chief economist for the foundation for American Innovation. Ripping the swag. Here we are a tech policy think tank in Washington, D.C. originally founded to bridge Silicon Valley and the D.C. culture. Today we work on the intersection of national security, tech and governance. I focus on AI, but cover sort of all economic issues as well. And I think we kind of, or at least associate it with the sort of tech. Right. With, with guys like yourselves, like, I'm rooting for y'.

All.

And hopefully Martin Shkreli's Bloomberg Terminal killer takes off. So then we can combine, combine you guys and completely disrupt Bloomberg. So, you know, I think you see this in the administration too. The Trump administration is a series of factions or coalitions, and we are definitely in the mix. But, you know, Elon Musk today called Peter Navarro, Peter Ricardo, and I kind.

Of.

You know, definitely hard to argue with.

Can you talk about the bridge between Silicon Valley and dc? It feels like that bridge is massive at this point. There was a moment where maybe tech was drifting away from dc, but now it feels like tech has taken over DC at the same time. You go back to the Obama administration, I always think about this statistic that I believe the number one organization that was non governmental that Obama visited during his eight years in office was Google. And so there was a moment when big tech and DC were tightly intertwined. Just happened to be with the Democratic Party. Now it happens to be with the Republican Party. But what is the state of the bridge and how did we get here?

Yeah, I think it's almost like a qualitative difference. So if you think like the last 80 years, the power structure in the US is being sort of either Wall street or like West Texas oil. So we either get like Rex Tillerson or Jamie Dimon.

Sure.

And, you know, since the Internet took off, you know, there's this new wealth on the west coast. And as that sort of germinated and matured, it originally was just sort of like one interest group among many. You know, you still still had those two main power elites. And I think with the, with this last election, it was, it was sort of an example of Silicon Valley, at least a part of Silicon Valley, asserting itself as its own distinct power center.

Sure.

And that, that is, that is very, very different. Of course, you know, all the other power centers still exist to some degrees. And so it is, it is sort of a constant struggle. I think there's a lot that this administration is doing great. You know, the stuff on energy. I think Doge at some point is going to turn to regulation. And that's what I'm most excited about. You know, once we start cutting whole parts of the cfr. You know, I, back in the day I used to do like supersonic policy and early worked early with boom aerospace and, and it's good that they're getting hearing now and maybe I'll be able to fly to either coast in flowers rather than.

I'd love that.

So, so there's a lot, a lot to like. It's just. And I think there's also like a steel man case for these trade actions. We participate with the Reindustrialized Conference. We have our own techno industrial playbook that will be coming out in a couple of weeks. So we're all on board for America needs to build again. And especially as AI deflates all the knowledge sectors, we're going to need more aluminum smelting like that.

Is there any glimmer of hope that there's this Mar A Lago summit? I forget exactly what Chamath is referring to, but the accords, and you do see reciprocal tariffs, but they, they actually have the effect of driving it down to zero tariffs anywhere in the world, either direction. Are you hopeful for that and would that be a good outcome in your economic framework?

That would be sort of the best possible world. There's also risks associated with that. Right. Because I wrote a piece recently discussing the sort of way the market reacted.

And.

On the one hand you could say Trump just likes tariffs. And that's definitely true. He has a 40 year track record of just liking tariffs. But then you have other people like Steven Moran and Scott Besant and J.D. vance himself as well, who at various points talked about the curse of the US Dollar being international reserve currency. And there's a lot of truth to that. The fact that China wants to hold our treasury debt and they build cheap cars, we build treasury bills, it does raise our living standards, but means that we are not ready to fight a war. And so that is a core, a core problem. But then the question is how do you deal with that? And if you do go all the way to a Mar a Lago accord, what you're saying is this isn't just about tariffs. This is about resetting global financial imbalances and we need to do that, but we need to do that sort of gradually because if you do this all at once, what that means is the entire floor will fall out of the stock market and the real estate market and with huge cascading effects through merging calls. And I Think mortgage debt is now back to its 2007 levels. So it's less to me about the mood or the ideas behind the policy, but the execution that this is a rug pull rugged.

Can you talk about the value of the yuan has been dropping. I guess it's at a record low. Can you talk about trade wars turning into currency wars and if that's what people should really be focused on.

Yeah. So China was a currency manipulator throughout a lot of the 2000s and early 2010s, but that really hasn't been the main way that they cheat. They cheat by basically suppressing household consumption and having these 60% savings rates. And so they end up building these ghost cities. And whatever technology they enter, whether it's cars or telecom equipment or. Pick your poison, they just overproduce it to the maximum, drive down the cost worldwide and then have to find these export markets to dump it. And the way that we're ever going to resolve this is especially now that the US is going to have 100% plus tariffs on China. And Europe doesn't want their shit either. They need to build up their domestic economy, they need to reduce their own savings rate, raise the standard of living of their households, introduce some basic social welfare programs or something so that they actually have a domestic consumer base. And if they do that, that's actually the best way that they can retaliate in a sense, because they're shielding themselves from the tariffs, but it also helps correct the big imbalance. And so that is sort of aligned in that sense where if China does the right thing, then it's a win win situation. If instead they double down on tariffs and trade war. I don't see we just exacerbate the contradictions in the economy and don't get to a resolution.

Ben Thompson has been advocating for a rethinking of the CHIPS act, mainly shifting from export controls, removing those, and instead taking a more operation warp speed approach where the US government is potentially a massive buyer of domestic made 3 nanometer, 5 nanometer chips. With the demand signal there, the American market should solve it. How are you processing the current CHIPS act and what are you hopeful for going forward?

I'm not opposed to the idea. The thing about Nvidia chips is the demand is kind of saturated.

Right.

They can kind of pick who their buyers are because there's just so much demand for them. And at the same time they've not been the most sort of like loyal actor in this space. And you know, there's any sort of Big meta narrative or theme to a lot of the right. The tech rights move into dc. It's been, you know, since from Project Mavenon that, you know, these companies have had corporate social responsibility policies, but not corporate patriotic responsibility policies. And technology is becoming geopolitical and we sort of have to pick your side. And so, you know, every time we introduce an export control, Nvidia two weeks later has a new chip that just gets under the line of what's being controlled. And the latest 1 is the H20. H20 is an inference chip. It's, you know, it will power these reasoning models. If we're worried at all about surge staying competitive, I don't think we can give up on those. And in fact, we should. We should be doubling down and that would be like a smarter kind of trade war than just across the board tariffs. But that doesn't have to be mutually exclusive with doing a kind of industrial push. And that's what I'd like to see. Because if we're going to do this big rebalancing, you can't just pull the rug. You have to mix my metaphors. You have to be the Indiana Jones that swoops in the bag of sand or something as you take the holy grail. And what is that new thing that we're going to be swooping in? What is the industrial bank that we're going to be using to bootstrap the industries that we need because they won't just materialize on their own.

Can you talk about putting the trade wars in the context of this race for super intelligence? In many ways people are arguing, hey, if we're making transformers harder to get and more expensive, does that hold us back from winning the AI race? And is that the only race that really matters? We've joked on the show about this idea of picking up pennies in front of a steamroller. AI has potential to transform the economy in so many ways. And it's very possible that just winning AI matters more than winning the trade war in the year 2025.

No, I 100% agree with that. Take, I can forgive a lot of stupid policy because in four years we're going to have such powerful AI systems that really swamps everything else. And we know what the bottlenecks are going to be. Building these models only has a few basic ingredients. You have the data and algorithms, which the US the algorithms are basically public domain. The data, China maybe even has an advantage because they don't have privacy laws and they can just scoop up everyone's genome or whatever. Then it comes to energy and chips. The export controls exist because right now our only structural advantage is at the chip and hardware stack where you know, our install base of Nvidia data centers is, you know, a huge portion of the world. China has been basically cut off since 2022 and 23. Then when it comes to energy, you know, China added 446 gigawatts of energy last year. It was 20% year over year increase. They're going to do that again this year. We added zero net new energy. We had a lot of renewables, but it came directly out of coal and other sources. And so the chips is the short run bottleneck. So that's why we need to lean into that. And then the long run is like how are we going to supply the energy? And then as the stuff diffuses to the people who worry about deindustrialization, it's true. The last 40 years we've specialized in higher education, knowledge work, legal management services, Hollywood, the creative class, all the stuff that is going to like be deflated. And China will have the factories that will become fully automated and do course because they'll also have the workforce that they can extract all the like the tacit knowledge out of and put into their robots. And so we, it's like a really urgent thing that we don't just like try to win on AI, but like win on AI plus heavy industry and robotics because otherwise our, our innovation in bits will be their innovation in atoms.

Yeah, the, the one point of view on the trade war and trying to bring manufacturing back to America is like yeah, yeah, we can bring the production capacity back, but will the jobs come back in the same way? Right. Just due to, if we actually want to scale production, we need to lean into automation and robotics. How do you think about job creation as part of reshoring and increasing domestic production in the context of long term, a lot of production just becoming automated just because that's going to be the most efficient way to produce the most amount of goods.

Yeah, we need to bring back manufacturing, but it's not a jobs program, that's for sure. In fact the only way we're going to bring it back is if we automate significant amounts of this. And maybe the guy who presses the on button every morning gets paid, you know, multiple six figures. But it's not going to be this nostalgic vision of like 1950s where we're all going into the factory and that's just like a structural thing. We're not that you know, AI is going to do that for a lot of stuff, probably most stuff at some point. And we're going to have to figure out what the new jobs are like. I saw on, you know, I saw a video of like professional back scratcher. You know, I know in the VC world those exist already, but like, but this was like a woman of long acrylic nails and so, you know, maybe we could start growing her nails out.

Crazy.

So my case is that there's like, there's, there's so much knowledge work to do around an advanced factory. I mean, we just talked to Jay from the Advanced Manufacturing Company of America. There's clearly a lot of high skill labor that is not getting displaced anytime soon.

That could, that's not millions of people.

Yeah, well, it might be if we're manufacturing a Dyson sphere with a million robots or something. I don't know. I could see a world where yes, there are a million jobs in the manufacturing sector, but it's all at the higher level.

But if 2 million, you know, traditional like white collar jobs get evaporated, you know, in the interim, there's clearly some big, big questions we're going to have to be thinking about. Do you have strong opinions on Unitree or any of these other Chinese robotics companies that are trying to.

He's just like, I love them.

Yeah.

I'm curious if you've written about it, if you've had, you know, policy recommendations that you or FAI have made around, you know, some of these more hybrid sort of dual use. Well, everything's dual use in China.

But the Unitree is really impressive. And you know, I've seen it do like kung fu and it, it, it does break dancing better than that Australian lady. Oh yeah, yeah. Shenzhen is like, it's like going to a flea market wherever you trip over baskets full of microelectronics and we need to be building some of those ecosystems in the US that's number one. And number two is like, yes, we have the data centers and the better models, but China has the batteries. Right. That's one area where they have leapfrogged us. And whether it's electric vehicles or robotics or drones, we need to have our own battery stack. And maybe we do need a chip sack too, but we also probably need a battery zact to compete with that because that will be the thing. It's fine if Anduril builds a drone factory, but where are these batteries going to be coming from?

Yeah, not asking for financial advice, but where, where are, where specifically in America are you long, you know, areas that could be that sort of American Shenzhen or maybe it's multiple places. I'm curious, what, what, what areas, you know, you know, different regions in the United States are, do you think benefit from reshoring most intensely?

In recent history it's been sort of the south and South Atlantic, the North Carolinas, the Tennessees, Nashville, partly because those also have the best housing markets. It's so much easier to build when you have greenfield. Longer term, this is also not something that the US can do alone. We're going to need almost like a North American plus like production, you know, frontier where you know, you know, let's figure out the thing with Canada, like do we need their lumber? Don't we, you know, whether, you know, do we want their bags of milk or not. But like we do need their aluminum, right. And we will need to have some kind of integrated production ecosystem to be kind of competitive and, and stand up to China because China, you know, already in purchasing power parity is larger than the US and you know that they want to gobble up their neighbors too and get even bigger. So. But I do think there is an opportunity here because when you do have like I'm not a technological unemployment guy, you know, I think new jobs get created, they'll just be very weird. Not necessarily in the sectors that matter the most. Like the purpose of heavy industry and robotics is like more military and like do we control the supply of core goods and services?

On energy, what do you think the lowest hanging fruit is in terms of energy deregulation? Should be focusing on the nrc nuclear. What's the biggest opportunity to help us jump from I guess 0% to 20% where I want to be. Maybe 40% would be nice. Maybe 200%.

Yeah.

You guys should definitely have my colleague Thomas Hockman on to talk about this for a full half hour because he's, he's been putting up the winds lately. We've helped pass a bill in Utah. There's activity in like Arizona, Montana, other places. There's a huge appetite to unlock America's energy in the short run, especially for these data centers. It's going to be natural gas. It's, it's going to be a bridge to more permanent base load energy. And then the next bottleneck is like the grid itself because if you want to do a, even if it's just like natural gas generators rolling in, that investment makes way more sense if you know that after that GPT7 is trained that you get to put your energy back into a grid and have customers for it. So that needs to be fixed. Other energy sources, I think enhanced and advanced geothermal are, they were underrated. I think people are starting to finally wake up to their potential. You know, with, with like really advanced geothermal, we could make like everywhere in America, kind of like Iceland, where like, you know, you have energy under your feet. And then with nuclear, you know, there is this case before the courts that I think is Texas, Utah versus the US Federal government that the NRC that argues that the NRC doesn't have jurisdiction over small modular reactors. And I think there's a good chance that this, that Pam Bondi and the attorney general settle that case, in which case states could then stand up their own licensing boards. And I think there's actually already movement in Utah to have their own nuclear regulator. And so that could, that could happen sooner than people realize.

Do a bunch of young founders building small nuclear reactors, does that scare you? Does that keep you up at night? Or do you think the technology is, you know, it solves.

Bernard Radiant isn't that young. He's got kids. I trust him with my life.

Yeah, basically. I think it should be a requirement. You should be, you should have to have kids to be a nuclear.

He worked at SpaceX. He's got the pedigree. I love that company based in El Segundo.

The big problem with nuclear is it, it's, it doesn't really pencil out without like large government support.

Yeah.

And, and so I would love to see the 600 billion in tariff revenue be given to Doug Burgum to build a reactor template and build 200 of them all around the country and make this, use every national security, national emergency trick in the book to get it done as quickly as possible. But it will need some kind of fixed capital backstop to make those investments, at least with the current technology.

I mean, given what you're like, you're kind of optimistic about Doge or it seems like you're pretty bearish about the tariffs. Like, are we in a regime where you trust the government to do mega projects yet? Because I think everyone was excited about the moon landing and then since then a little bit less excitement about the big projects. High speed rail in California has been a little bit of a rough go. And I don't really want to see a California high speed rail of 600 billion get burned on a nuclear strategy that doesn't produce a single watt of electricity for 70 years or something, which would be like the bad case.

Yeah, 100%. The state capacity and competence is really, it's the jagged frontier. There's places that have a lot of it, places that have little of it. I would have more trust in a Burgum or a Chris Wright of actually, you know, executing on something like that. It wouldn't be the Pete Buttigieg slush fund where it's just filling potholes in Indiana. They would know how to cut through the road tape. They wouldn't make it like this. Everything bagel. We're going to build TSMC chips, but then also rehabilitate justice involved individuals. We need to keep our potatoes and our gravy separate.

Got it. How are you thinking about the Deep Seq versus Meta's Llama strategy? We were talking about that earlier on the show, and it's kind of hard to. I think a lot of people on the vibes of Deep Seq, they're like, I don't like this. But then it's difficult to formulate an argument because are you anti open source? In which case are you anti Zuck and Meta? How are you thinking about kind of the intellectual property that's being developed in America around large language models and then makes its way across the Pacific Ocean?

I think what I find Most impressive about DeepSeq is less that the model they put up, but just that they sort of have imported a kind of Silicon Valley model of like. And that came from their CEO being like a hedge fund manager doing this as a side project. It's very. Sam Altman wasn't a hedge fund manager. He was a. A vc, but sort of analogous. Right? And that's striking because it's just a different model of corporate governance than you're used to seeing. And I think there's a question of how long does Deep seek become a victim of its own success? And they are the tall poppy. And it's not that China tries to hurt them because of that, but actually tries to help them and makes them a national champion and thereby sort of perverts it.

But, you know, it's a fun take.

They've been great at, you know, publishing what they're doing, and everything sort of has checked out, but they don't have the chips. And they've said, they've said that. Like their CEO said their biggest bottleneck is hardware, and so we shouldn't help them on that front. Like, there's $16 billion of orders for 20s just starts sitting in limbo about to go at the door. The Commerce Department has. Howard Lutnick has said that he's going to export control of H20, but they're so distracted by tariffs, they haven't prioritized it and the time is kind of running out.

What's going on with TikTok? We've been following the poly market around a new ban before May. There's markets around, you know, potential buyers, things like that. Do you have any insight that you can share around the latest there? It feels like again, one of those things that's just like not getting the attention and the focus because obviously, you know, if we are enter into the greatest global trade war of all time, like, yeah, it's rightfully people should maybe be sort of focused on that, but at the same time, it feels like something that we were supposed to have answers around by now and we definitely don't.

Yeah, totally. I mean, fai, we led the charge to ban TikTok over a couple years and I fully support it. I also enjoy TikTok, but I do notice that between my barbecuing steak videos and funny memes, I'll get a Pyongyang tourism board video now. And then it's like, well, that I don't plan on visiting North Korea anytime soon. But yeah, I don't have any super deep intel. There has been talks about, or rumors about Oracle maybe being part of this and I think Trump still wants it to be part of the new sovereign wealth fund and actually as sort of Zanian idea that is, he kind of has a point like if TikTok became American and you know, quadrupled in value, that would actually help pay down the debt. The interesting thing here is like, you know, people have pointed out that, you know, Trump is sort of placed fast and loose with, with the Constitution, with the law and stuff like that, you know, and as a matter of fact, all the people he's fired, totally constitutional. The biggest, the most unconstitutional thing he's done today is not enforce the TikTok ban. Oh yeah, because that was a direct, you know, statute that Congress passed that said thou shalt ban TikTok. So I, I'm hopeful that they can get a deal. The reason I would just doubt it is, is China has very strong export controls. Like, the reason TikTok can't sell is because algorithms in general are export controlled and so they would be able to buy the brand name and like the, the offices, but they have to completely revamp the algorithm, which is like the secret sauce of the, the thing. Now TikTok is in our building in D.C. so I can, I can try to plan a bug for you if you want.

Sounds great.

Well, fantastic.

Polymarket has the Chance of TikTok being on the App store in on May 1st at 97%. And who will acquire TikTok? Oracle's at 27%. Number two, Larry Ellison, directly at 24%. You'll love to see it. Amazon's still up there, but I just want to say thanks so much for joining. This is really interesting, interesting conversation. We'll have to have you back soon.

Yeah. And get some sleep.

Yeah, get some sleep.

We'll work on getting you an eight sleep so you start putting up some proper sleep.

I want to see a hundred for weeks straight, I think.

Yeah, Less red light from the stock market. More. Yeah, yeah, more sleep. Go get some sleep. Thank you for coming on and yeah, forward to the next one.

Yeah, talk soon. Bye.

Later, Ryan.

Cheers.

Next up, we have SHIELD Coming back on for his second TVPN appearance. We're going to talk about fintech, the markets, the tariffs, his dust up with another capital allocator on X the other day. Had a lot of fun with that. And I'm sure we'll have plenty to talk about. So as soon as SHIELD Gets here, we'll bring him into the studio. But those are some interesting questions. There really are so many debates right now. It's like DJI unitree, TikTok, deep sea. There's like seven different really important questions. Maybe we'll talk about it with S.H.I.E.L.D. maybe we won't. But let's bring him in to the studio and welcome him to the show. Welcome.

Boom.

Back with a suit.

Looking great. How you doing?

Looking good. Looking good.

No, no Apple Watch.

No Apple Watch. There we go. We'll get you a bezel now. That's the next step. We've deradicalized you from the Apple Watch. Next is radicalizing you to Bezel.

Yeah.

The tariffs haven't hit the secondary market yet.

No, it's a good, great buying opportunity. This is financial advice. Go to getbezel.com, download the app. Just for SHIELD just for you. Not for the listener. For you specifically. You specifically. I want to see an Audemars piguet or royal oak on you. Something like that.

Well, you've had a bit. I feel like you've been. The timeline's been in turmoil and you know the center of it. I guess. I guess Chamath knows who you are now, now that people reminded him that he would use your content in his newsletter. Yeah, but so funny. The whole. The whole thing was. The whole thing was funny. I'm sorry. I mean, honestly, probably good metrics, I'm sure were up into the right. So Elon Bucks, the creator, payout this.

Month you're going to.

It's going to go from 200 bucks.

To shield for screenshots of his content with just. This is why. I can't believe this Twitter. This, this app is free. And this is why I'm never deleting this. I'm never leaving this app. Because of the interaction.

Yeah, exactly.

Anyway, yeah, can you give us just your, your high level reaction, how you've been processing the tariff news, kind of set the table for us and then we'll dig in.

Yeah.

Wow.

Right into it. Okay.

I'm kind of like, I've always been more of a free market kind of guy, an American and I think. Free market American. Yeah, American. I've been an American guy and I tend to think competition makes us better. And I also like spent time living in protectionist India. And so for those who don't know, like until the 90s, India was a closed economy. Like they had super high tariffs on all foreign goods and it sucked. There were two local car manufacturers and the cars were like built in the 50s and they like didn't get any better from the 50s until the 90s because India had so much protections on their local car industry. And that was terrible. So like they didn't innovate. They never improved the quality. They were super high, so people couldn't afford. The prices were super high so people couldn't afford them. And so that's what really scares me. And then you might say that would never happen in America, but you'd be totally wrong because that's exactly what has happened in the US shipbuilding industry. So like the Jones act basically says that if you, if you're shipping goods between two U.S. ports, you need to use a U.S. built ship crewed by U.S. citizens and owned by U.S. citizens. So like it's super protectionist to the U.S. shipbuilding industry. And U.S. ships suck. They're like five times as expensive as other ships and they've never had to innovate because they have these protections. And then it totally distorts the markets in general. Like on the East Coast, I'm in New York right now on the East Coast. A lot of like the east coast gets some fuel from internationally because it's easier to ship here than it is to get it from Texas. And that's just like a perversion of markets that exist because of the Jones Act. So anyway, so I think like all these things, like, I'm totally anti protectionist. There's a question of like what is Trump doing is like with, it's not really a reciprocal tariff. Now everybody realizes it was a funny situation last Wednesday when people are like, what the fuck are these numbers? And then, you know, the guy who did the math was like, oh, this is about our trade deficit, not reciprocal tariffs. I think like, if you, if like now people are coming around and saying, oh, this is all about lowering trade barriers. I think that's bullshit because like you have Lutnik saying like, we need millions of Americans screwing in tiny iPhones or whatever.

And you.

And they also say that the tariffs are going to replace income taxes. So if those things are true, then it's not about leveling the playing field to zero. It's about like putting these tariffs in place to reshore. And I personally don't like that too.

There's this focus on trade deficits, but we're completely ignoring like, services and specifically like digital services. Right. So it's like, like, like Switzerland, for example, you know, we have a trade deficit because they have 8ish million people and we have hundreds of millions. And then they like make all the world's fine watches, which we just talked about. But then like, they also probably love Netflix, I guess, I guess that like a lot of people in Switzerland are subscribed to Netflix and we're just like completely ignoring all of that.

Yeah.

And you know, yeah, I have a.

We're the richest, most prosperous country in the history of the world. It's fucking awesome. We can afford to buy all their shit. Like, they don't need to buy stuff from us. They can't afford it.

So yeah, so the steel man here is like, first off, do you think DJI and the consumer drone market is a problem? And then if so, what is your solution? If not just ban DJI, tariff DJI. Like, we did a deep dive on GoPro versus dirt versus DJI and it really just felt like China was like, we are going to kill GoPro in the drone market. And they put so much firepower behind it. And I'm like, I still get that there's some weirdness going on here. And it's an important industry and it's dual use and there's a million different factors. So how do you walk through that specific example narrowly? Let's, let's, let's take away the blank tariffs. Walk through that for me. And how would you solve this in a more free market, more progressive fashion?

Yes, that's a great question. So the first, like, we have tariffs in every country, not just our enemies, with Trump, but China specifically, I do think China's playing unfairly and, and they're our enemy and we, we shouldn't let our enemy get data in the United States. Like that could be really bad. There definitely are national security issues with drones.

Yep.

I also think we should ban TikTok.

Sure.

And so I think those things can be dealt with, but they have nothing to do with tariffs.

Yeah. I guess one of my, one of my scenarios would have been if I could, if I could replay everything with everything I know now maybe you, you see what's happening with DJI and GoPro and you say, hey, we are the richest country in the world. We do have China buys a bunch of our debt. Let's lever up essentially and create a drone buying program from the government to stimulate demand for American made drones. Essentially backstop GoPro, let them get down the learning curve. Hey, if they make these drones in America, we're gonna buy them even more and let us develop that. And then we are competitive and we say, hey, it is a little bit. We're still shifting the invisible hand, putting our hand on top of the invisible hand, but it's still somewhat of a free market in the sense that like, just like what we did with EVs with Elon, like anyone could have gone for those electric vehicle incentives. Elon did a great job taking advantage of like, and we got a great product, you know, it sold really well eventually.

Yeah, I think that's absolutely right. I think. And like look, we, we moved in this direction already. Yeah, like the Chips act and IRA both did, did make good moves. Like yeah, they, they enable, they subsidized US chip manufacturing that are critical for military systems and other stuff. I think they have made some moves away from foreign chips and so that all that, all stuff is good. I think like leading with a carrot is way better than leading with a stick, personally. And I think like the, the ideas you mentioned, John, are spot on. And by the way, like we helped Tesla along the way. Like we loaned us taxpayers loaned $500 million to Tesla, that kind of thing. I'm totally in support of enable US manufacturing to be better, to compete on an even playing field by being more innovative, not by blocking like other countries from competing.

Yeah, yeah, Jordy, bummer to see the IPO window close. We had Klarna, StubHub. We'll see if Circle gets out. Klarna would have been especially nice for fintech broadly to get some marks.

Circle still at 86% on poly market for this year.

Yeah, they might just Be like, you know, crypto. We were born in the darkness. We're going out no matter what. Did you. Did you have a take on the Circle IPO in general? I saw a lot of people just were not kind of loving the S1, particularly just based on how much they were paying Coinbase to coinbase the token. I'm curious if you had a. Had a. Had a take on the IPO or dug into it at all, and you don't need to have any knowledge to have it. You don't need to have any knowledge to have a take. That part of.

Yeah, no, I would say, like, on Circle in particular, like, I saw all the same stuff you saw. Coinbase gets half of the revenue from Circle's token, all this other stuff. But I don't. I don't have a strong take on how the IPO will perform. I tend to think that these things are somewhat, like, initially somewhat disconnected from the reality of what's going on. So, like, I think, you know, we talked about Bridge last time I was on, and I think there became this stablecoin hype. And I bet if Circle was public at that point, they would have gotten a huge bump for no particular reason.

Totally.

Yeah.

But I think overall stock market. Yeah. Like, what does. What does polymarket say about Klarna? Is there a market for that?

Oh, I don't know. Jordy, can you.

Well, Klarna, I think, officially pulled their.

They pulled. Well, like, the thing is, these tariffs are especially. Especially bad for Klarna.

Right.

Like, it's consumer discretionary spend that you use BNPL on, and consumer discretionary spend in a recession or with high tariffs, like, goes. Goes to the toilet.

Yeah.

So, like, you're not buying that extra $2,000 item that you didn't exactly need, and that's what you were. Bnpling anyway.

Yeah.

I think we've seen a firm stock. Oh, yeah, I think our firm stock got cut in half.

Yeah. I want to go deeper on Circle because I feel like it's one of those companies that if they're about to ipo, I can't even name the founders. I don't know all the big investors. Like, it's this. It's this fascinating. This is case with a lot of. A lot of crypto companies, but even Bridge, like, we heard the story of, like, who made the money on this. Okay. They got acquired by stripe. Like, they're very much in the Silicon Valley world. And Circle obviously is, but has been. Hasn't really told their story in the way. And so it's interesting. They could have a meme stock moment where it's like the, it's the primary way that you get exposure as a public markets investor to stable coins broadly, I guess. And that could be a good narrative. It could just be a meme stock because, hey, crypto, it's, you know, whatever, but, but they haven't really told their story in a way that's broken through at least with me. I don't know if you, if you've processed it any differently, but yeah, I.

Think it's certainly less, less hyped than a lot of the others. The CEO, Jeremy Allaire. I, I went to a stablecoin conference a couple months ago and he spoke and so he's very sharp and he's, he's been at it for a very long time. He's like, I'm a different, slightly different generation than us. Like he, he started a company that IPO'd in, in like the 90s, like.com boom. And then he was actually a venture capitalist. Like he worked at General Catalyst for a little while.

Oh, cool.

And then, and then launched circle, whatever. 10 or 15 years.

Not.

Obviously not 15 years ago.

Yeah, Bitcoin didn't exist.

But, but yeah, that was the best.

Time to launch a stablecoin before bitcoin getting really early. I mean.

Here'S the bull case for USCC and like, and, and it's here. Here's my bull case.

Sure.

So Tether is like the most profitable financial institution ever, right? Like, it's literally. What are they.

Is it like million per employee or something?

I forget the actual. You probably shared it at some point. Chill. But it's like some absurd number. They're more profitable than like any of these other major financial institutions. The risk with Tether is like, it's opaque. You don't fully know what's going on. Like there could be. People for a long time said there could be systemic risk associated with Tether. They've been accused of a lot of stuff over the years, constantly. Yeah. But they're dominant from a market cap standpoint. The second biggest stablecoin is USDC at a $60 billion market cap. And a market cap is obviously just one to one with the supply. Right. And then you go down the list. The next one is dai, which is also run in this very crypto native way, from what I know. And then to get to the next stablecoin made from a sort of true traditional western institution, you have to go to First Digital USD, which is under a $2 billion market cap. And below that is PayPal USD, which is an $800 million market cap. And so to me, I'm looking at circle and it's like, here's like the power law winner, the dominant, you know, they're 80. They have 80 times the circulating supply as their next, like, regulated, you know, Western institution. And they have usdc. Like, it's a pretty good ticket.

I mean, the crazy part is we talked to.

It's not super sophisticated, but. But it, you know, brand matters.

I mean, we talked to Zach Perret at Plaid and we were like, if you had, you know, full authority, you were like, the president, could you speed up wire transfers in ach? And he was like, absolutely, but it's not going to happen. And so it's like, yeah, maybe stablecoins are here to stay. And all the pitch about just, hey, you're just going to be able to transfer money two days faster. That's enough. Even though it seems like you should just be able to speed up the government transfers.

What's your read on any sort of predictions on venture right now? I think that the lesson venture since 2020 has just been take advantage of chaos, invest through market cycles, never stop deploying. I remember in 2022 we were talking about the. What is it? Denominator effect.

Denominator effect?

Yeah, denominator effect. But then it didn't fully play out. We saw this. I mean, we saw like a, again, a bifurcation of like the big funds raising all the money on paper. But, you know, if you're a specialist fund with like a strong story, you can still get funds done. But I'm curious first about the venture market. And then I want to kind of ask more about portfolio stuff.

Yeah.

First, maybe a fun tidbit for you guys is just the last few days, obviously markets have been in various states of turmoil and venture capitalists are. Some of them are trying to seize the day where, for example, there have been a couple of companies in our portfolio that some investors have been really trying to invest in, but the companies are well capitalized and don't need the capital. And now the investors are like, hey, markets in turmoil. Might this be a time that you would consider taking my money?

Yeah.

You don't have to be really living.

Up to the vulture capitalist name. I like it. Making money.

Yeah. How do you even think about. There's going to be some enterprising founders that are like, look, I'm building a startup around that's, that's, you know, built to help solve, you know, global Supply chains or something like, you know, the chaos is a ladder. I'm going to take advantage of this. To me, it's like, okay, if we're entering this sort of protectionist phase of globalization, maybe it's too early to make bets, but yeah, how do you see companies like actually being able to make something out of the chaos? Or are you just telling your portfolio, just stay focused on the customer, ignore the noise, that kind of thing.

It's really like, stay focused on the customer, ignore the noise. I think we don't have any companies that are super exposed for some reason or any other. I saw you guys have J. Malik coming on later today, which sounds like he timed that perfectly.

Yeah, it's crazy.

Yeah, literally.

I mean, there's a few of those companies that have been. I mean, that's been the thesis for a while, just general reindustrialization, but they really hit a royal flush this week. Yeah.

So it's mostly stay the course. I think people are saying, okay, venture capital dollars are going to decline. But as you know, the way it works is we raise a fund every few years and we have plenty of capital. So, like, it's not like, you know, and lp, there's some impact on the markets today and that means we don't have money tomorrow. It's like, if there's any impact, it's a few years out. Though it doesn't change how we invest now, the later stage investors, it is a different equation because for them, they have a certain timeline. They're hoping these companies go public. And if the public markets are kind of frozen, that makes things difficult and their thinking on an IRR basis, their opportunities have declined. If the companies don't get out at a reasonable time.

Yeah. So mostly just texting founders in the portfolio. Have you seen this with a screenshot of the market data?

That's just what I always do.

Yeah, exactly. I want to get your reaction to this post from Semel over at Haystack. He says seed is again going to be the hotel zone where nearly every VC fund will want to invest. Just like when Covid struck. And in the early. In early 2022, VC shifted early to balance large checks by firing $3 million seed bullets. LPs should expect median seed entry prices to be up 50% in the next vintage. Does that seem like a good take or what do you think?

No, I like, I love Semel.

Yeah.

But I don't know if I buy that.

I think that's already pretty high. Right.

That's the thing.

Yeah. Please, please Please, please make, don't tell.

The founders, don't tell the founders. They're just like, wait a minute, I can raise my safe by 50% with one stroke of a pen. Let's do it.

Yeah, it doesn't make sense to me because. So we started this fund in 2019 and actually the seed valuation, 2021 was an insane time, especially in fintech. Like everything we were investing in seemed like it was like turning to gold and then, you know, and then maybe turning to shit afterwards. But, but classic but, but that, but actually like seed valuations have actually increased from that time and it's basically kind of been as like a straight line upwards. And what he's talking about actually started happening in 2022. And a lot of the funds invested at seed in companies. And the problem is if you're a multibillion dollar fund and you write a $2 million check to a company and you invested in the wrong company in the category, you don't get a chance to write a $250 million check into the right company. So I think it's pretty foolish when those funds invest at seed. And we have a bunch of examples now of like, of friends of ours who took money from a multistage but like the multi stage doesn't care about them that much because the small amount of money. So I don't know, I'm skeptical that this is going to happen again or that it's going to re. Accelerate and prices are going to go up.

I don't know.

We'll see. Well, we should make a bet and have Peter Walker from Carta give us the data in a year. I love his stuff.

Can you talk about. There's this meme of like, oh, for a while, if you're building a consumer or something, like you're gonna get steamrolled by what if Google builds it right? There was this story that Google is allegedly paying some AI staff to do nothing for a year rather than join rivals. Hilarious. I want your reaction to that. But then I also want to know like, is there, does that meme exist in fintech? Is there, is, is there an idea that, oh, amex or Visa or JP Morgan are going to build this? And has that ever actually happened in practice?

Well, yeah, and even potentially on that. I'm curious, like, OpenAI wants to run your entire life. Have you heard any sort of like, rumors or is any concern around people saying, oh, I'm building a consumer agent for consumer financial services? But then OpenAI might be like, oh, by the way, we Launched a partnership with Chase or we launched a partnership with Cash App.

Yeah. We can analyze your credit score now with an agent. And that. That model that that wrapper company got got steamrolled. Yeah. What's your take on all that?

Okay, so first thing I think you said was the REST Invest situation.

Yeah.

Where. And so I thought it was really funny because. Did you guys watch Silicon Valley, the TV show?

Oh, yeah.

So good. And there's that. Obviously there's the phrase REST Invest. I learned it from that show and it's certainly playing out.

I had no idea that that that phrase was, like, popularized in some way by the show.

Oh, yeah, yeah.

I thought it was. I thought it was a 2021, like, big tech thing.

No, no, no, no. This is the thing going back a decade.

If you haven't seen it, you got to go back and watch. It's so.

So I never could. I never could get into it because I. It was just too close to.

Yeah, yeah, same thing.

Like, the most, like, it was not like I watched you because, like, I want to, like, you know, not think.

Yeah.

Then watching Silicon Valley, it's like, oh, that's an email. I need to reply.

Exactly.

I should follow up with that founder.

My first company, Soylent, was in the intro to Silicon Valley, like in the intro sequence. And they're just like making fun of me every single day. Also, one of the creators went to my high school and so, like, I knew him and he's like actively poking fun at me every single episode.

Amazing. That's amazing.

It was great. But yeah, it's very silly that Google would let this even leak out. I don't know how that happened, but totally.

It's crazy.

I mean, the things you hear out of Google are so crazy.

It's wild.

I think more. More so than any other big co. Like, my wife works at Meta and they really got their shit together, like efficiency, stuff like that. I think probably before then it had stuff like this, but not now.

Okay, so that was topic one.

I think topic two was like, what if X company builds this? And is that the case in fintech? I don't think so. In fact, you actually had look, one of the sponsors, Ramp. So Stripe had built Ramp. Stripe had a corporate card and it didn't work. They ended up investing in Ramp and deprecating that card. So I think people have tried to do stuff. There is the what if Stripe does this? What if Plaid does this? And there are, in some cases, I think that's totally valid, but for the most part, I think there's plenty of green space out there. And you know, Stripe has been acquisitive. Obviously we talked about that before.

Bridge.

There's nothing I'm super afraid of. I will say in some categories, like for example, in wealth management, there was the wealth front and betterments of the world, the robo advisors. And people said, okay, like we're not charging 2%, we're going to charge you 25 bips. But the reality is that the service offered by somebody who's charging 2% is different than what they offer at 25 bips. And the 25 bip solution was fairly.

Easy for Vanguard to build.

And Vanguard became by far the largest robo advisor in the world.

Got it.

But I'm not afraid of that in general. Infant too much.

Do you think humanoid robots present an opportunity for loan sharking as a service?

You know, come break your legs, autonomous, break your kneecaps.

It's like, hey, we're gonna offer you this great rate or whatever. It's secured against your kneecaps.

It's funny, like we, our, our tagline for our fund, when we started it was everything is fintech. It would be funny if we invested in a humanoid robot company and then just. We were like deadpan, like, what do you mean? Like, obviously the use cases for loan sharks.

Yeah, I mean, I was talking to a sales guy.

I'm curious.

Like, my thought goes with the. Do you think that AI can, can get. Is already or could get materially better at underwriting than. Than a human just spending, you know, months on an opportunity? And is that something that, do you think that fintech broadly has fatigue around investing in like AI lending just because it's been sort of this like, ongoing narrative. Isn't there, like, what's the, what's the public company that was sort of promising this for a while something.

I mean, there was Metro Mile, which was better underwriting upstart. But even Metro Mile is better underwriting for your car insurance based on how you drive. They put a GPS tracker and, and like a gyroscope in there, basically. See if you're stomping on the brakes every two seconds. Give you a higher insurance premium. A lot of promise there, but not a lot of massive adoption over time.

I think they didn't execute that well. And I think with Metro Mile, there's actually Root has done a better job of it. But Metro Mile, it was primarily just mileage based. The number of miles you drive and then route gives you a phone. You put your, your phone and it like checks if you're braking hard and where you're driving and stuff like that. So I think that there are opportunities to be used in insurance, in underwriting for loans in particular. It can be tricky because you're the regulatory framework in the United States Equal Credit Opportunity act, reporting act, etc. You're not allowed to discriminate on the basis of race and gender and some of these other things are tied to that. So that, that can become tricky and you're actually, you have to respond, you have to give people the reason why they were denied like an it's adverse action notice. So you can't have a black box model that's like, here's all the data.

They just give you a massive matrix of weights and they're like, this is why you were denied. Figure it out.

Totally.

So that tensor number 76 was activating for you. So get out of here.

Yeah.

So actually, you know, if we didn't have that, lending would be probably more efficient and like you'd better be able to target the right customer. But we do have those things for a reason and so we can't have a black box model. And so there are actually companies we invested in a company that's like in part detecting bias in AI underwriting for this purpose to make sure that you're compliant.

And.

Yeah.

So anyway, I think it can be used and can be super useful, but because of those regulatory bound guidelines, I'm not sure it's going to be a step change in underwriting.

Yeah, makes a lot of sense. Jordy, last question. You want to let SHIELD get out of here?

No, this is great. Always a pleasure, always fun. Super fun, guys. Looking forward to the next one.

Yeah, this is great.

Likewise.

Have a great rest of your day.

Godspeed.

We'll talk to you soon. Well, we got some breaking news. Another massive funding announcement coming in to the studio. Victor from Craya. Is that how you pronounce it? CRAYA is coming in. AI video company that just announced a massive fundraise. Let me look up if I can find the details of this fundraise so I can get everyone up to speed before we bring Victor in here. The website is KREIA AI K R E A and they just announced a huge funding round. $83 million. Just a couple million. Over. Over Jay. Not that it's a competition. They got Andreessen Horowitz, Bain Capital in the. In the round. The past 14 months at CREA have been been hectic. We rolled out over 50 major product updates, grew to over 20 million users organically and they 20x their revenue. All with a team of 8 working out of a living room in San Francisco. That is fantastic. You love to see growth like that. Doesn't happen every day, but it's happening more and more in AI. So excited to bring Victor into the studio and talk about that. They write the numbers are exciting, but they can miss something crucial. The team behind it all. Kreia is the work of a small, talented group of imaginative, incredibly dedicated people. And yes, most of us still live together. That's fun. Until now, we've never shared metrics or announced our funding. Heck, we didn't even have a blog until a few hours ago. Those details always felt secondary compared to what truly matters to us. Making AI intuitive and controllable for creatives. Now, after the release of our redesign, the growth of our team and recent funding, it feels like the perfect time to open up about what inspires us and what we're building towards. So they write. We're living through a moment where everyone talks about automation, APIs and how AI and software are eating the world. Perhaps too much. Don't get us wrong. While AI is powerful, transformative and is going to radically change creative work, creatives aren't going anywhere. 40,000 years ago, we painted red ochre onto cave walls. Later we drove. We drew with graphite on paper. Today it's complicated. We use cameras to digitize light through glass lenses and silicon sensors. Transferring data through metal wires to illuminate the LEDs lighting up the screen you're reading on now. How do you know I didn't print this out? I could have printed this. The printer still works at tbpn. We might have shifted to laptops, but you never know.

Too much pushback.

Don't assume anything. No, I am reading this on the screen. Paper usage and we have Victor in.

The waiting room yet, by the way.

Not yet. Old tools and workflows will disappear, but our creative itch won't. And I agree with that. Excited to dig into that with him. We will build new and more powerful tools to keep doing what we've always done. Master new mediums for self expression and storytelling. AI will render some tools obsolete, but not the people behind them. We see AI as a new medium that that lets us express ourselves through any format. Text, video, sound and even 3D. Such a medium needs better, smarter and more controllable tools. That's where Kreya comes into play. They say it won't replace AI, will not replace creativity. Creativity is not disappearing, but the walls between creative mediums are traditionally excelling in one creative medium, rarely translated smoothly into another. AI changes that. And we're bringing Victor into the studio to tell us more about Kreya and the fundraising. So welcome to the stream. Victoria, you here? Yo, how you doing?

What's going on?

Yeah, I'm here.

Great.

Can you give us a brief to meet you guys?

Fantastic.

Is the office gonna change with this new fundraise? I gotta ask, are you guys gonna stay posted in the living room?

Do we have you? Victor, I think we might have lost you.

Seems like we're having some technical issues.

Okay, well, we can hear it, see and hear you now.

Yeah. Okay, let me go far away from.

The WI FI router. I mean, that is the issue with working at home. Complicated WI fi. You need the enterprise solution soon now that you have the Series B. Done.

Sold very soon.

Yeah.

Oh, shit. They are doing another meeting. So I'm gonna steal Diego's room.

This is great.

I love it, by the way.

You guys. You guys want to see the office?

Yeah, please.

Yeah, let's do a tour.

Anything that you can show us, you can turn around. There we go.

API keys.

Hopefully music going on.

That's fantastic.

Real life.

Yeah.

Wow.

Hey, guys. How you doing?

Congratulations on the milestone.

Looking great.

Wow, you guys. You guys were not kidding about the living room, but you've really built it out. I love it.

That's amazing.

It's looking good.

Nice. Why don't you introduce yourself? John was going to ask you that and then I cut him off.

No, you're all good.

Yeah, sorry. Sorry about that.

No worries. There we go.

So.

My background. Yeah, like, I guess on my background, like the TLDR is. Growing up, I was very interested into creative things of all kinds of. I mainly had a music band and I was doing, from playing multiple instruments in the little studio that I created in my house to producing music, mixing, mastering, like learning about all of these processes around music production. But through my music band, I also got super interested on doing photography and, like, doing different kinds of content for that music band. So that way I explore, like, many different things from graphic design, 3D, graffiti art. I also had, like, a big passion for that. And at some point I was. That was in 2015. I just finished high school and I was not sure about what to do after that. And I had, like two options in front of me. One of it was go and do classical guitar studies at the Conservatory of Barcelona. At the Conservatory of Guitar of Barcelona. And the other one was doing something around computer Science or physics. I really liked math and I guess that what I like about math is kind of like the challenges that it poses and the interesting. Yeah, I guess I love challenges in math. Put a lot of challenges in front of me. But in the end I found this middle ground on this degree that it was called Audiovisual Systems Engineering. It was kind of like this degree where they showed you how a microphone works, how MP3 encodes audio, how MP4 encodes video, et cetera. And that's where I met Diego, my co founder. That was like 10 years ago. He ended up in that same degree following kind of a similar story. In his case he came from having a lot of interest in film and a lot of interest in 3D as well. But also he also loved programming and he also loved engineering. So we both ended up like in that degree. And on the second or third year I got introduced into. I mean first of all I loved coding. Like right after getting into the degree I loved coding, found it extremely creative. Later on I found about AI, I was mind blown by deep learning. Like, just like the fact that you can have these neural networks learning by themselves from data and being able to do such complex tasks was very interesting to me. And when I discovered about Gans that they were like very early models for image generation, that's when I fell super deep into the rabbit hole. And I ended up like reading a lot of papers, doing a ton of implementations by my own from all these papers that they were out there back when everything was open source and ended up.

The good old days. The good old days.

Do you have a first question or no?

Go for it.

20 million users, absolutely massive. Congratulations. Where are you seeing those folks come from? Is it consumers just having fun prosumers who are maybe doing little contracting work monetizing their creativity on social networks or are you already in the enterprise or all three?

Three, all three.

I think that up until recently there were like two very well defined blogs of users. One of them, it was the consumer type. It was people who this technology gave them a zero to one when it comes to creative freedom or to like enabling them to create. It's people that didn't necessarily come from a creative background, but they had a lot of joy out of expressing their creative ideas using this technology. And they were paying for the subscription almost in the same way that you could pay for a video game or that you could pay for a camera. Then we had the professional and the professional it was that user that did have a creative background and that he was using Our technology, I mean our platform to speed up some of their processes. These speed ups vary depending on the industry. You would see architecture studios coming to Krea with very low resolution renders and using our enhancer to get these renders up to 4K resolutions with very crisp textures. Or you would see game designers coming to our real time tool, putting a bunch of ideas around characters and being able to create prototypes for some characters that they were designing.

Can you talk about just general adoption? So during the, during the sort of like Studio Ghibli moment, it's still top of mind. We saw a lot of people that still weren't aware, they had no idea how these images were being created. John. And I think that some people thought it was like Snapchat filters or something like that. Can you talk about just like consumer sort of awareness and adoption broadly? Are you guys still finding people every single day that are just sort of like completely new to this sort of new image generation models, or what do you think the broad consumer awareness is today?

I mean, I just came two days ago from a short trip to New York and I feel like that trip to New York made me realize how deep in the bubble we are here in sf. I think that I take for granted that people know that nowadays you can generate images with artificial intelligence. And that's not the case. I think that we haven't reached, I wouldn't even know what's the percentage of reach that we have had right now, but it's definitely very, very small. This technology is still nascent. People like us are trying to make it intuitive and usable for really anybody to just like grab a phone, type a URL and be able to create an image. Very easy. But I think that people still need to know that this is even a possibility. Like, I think that they, they just don't even think that some of the problems that they have when it comes to marketing or when it comes to doing product design can be solved today by artificial intelligence. So I don't know if I'm the best one, like if I'm the best person to have like a good sense of what is a covering adoption because of how deep we are in the asset bubble. Yeah, from my experience that we, that I've had in New York, I don't know, like I had this fun story that I was on an Uber and the Uber, like she just like saw that I was like talking on the phone in Spanish and she was from Puerto Rico. So she started like talking with me and asking, asking me where I was from and what I was doing, and this woman, she had like, she was like selling a sort of like beauty products on Instagram. And I saw it and she started asking me, oh, so can I use your tool for doing like this product photography or can I use like all of these things? And as she was talking, I was like, yes, you can do it, but you need to go through a process. It's not like some magic thing that you, like, go there and like, the AI does everything for you. You need to go and train a model with your product. After the model is trained, you go to the image generator. There you create all the assets that you want. And after you have this workflow in mind, after you have this workflow in place, you can generate as many assets as you want, and your workflow is going to be extremely optimized.

How do you think about prompt engineering long term? I remember a year and a half ago, maybe a year ago, everybody said every company prompt engineer is going to be this new role. Now it feels like, like it's getting easy enough to prompt a lot of these tools. I'm sure, like Korea, that maybe it's just not necessarily. Maybe it's a skill set, but not necessarily a job. But I'm curious how you think. Do you think that prompt engineering will still matter in five or 10 years? Or it'll just be super intuitive?

I mean, prompt engineering, at the end of the day, it's. It's just like being able to communicate your ideas in a clear way with this technology. We have this AI model that can understand language and that can do things. And from engineering is just the way that you tell this knowledge that we have encapsulated how to do things or what exactly to do. So at the end of the day, it's just managing. And I do think that this feels like a new way of doing software, and I do feel like this is gonna like in the future. Most software that we see out there has been created by a very big percentage through PROM engineering, through steering AI models towards whatever you want to accomplish. And I see this on the visual space. Like, I see us building Korea in the future more and more through instructions. I see our users working with our platform more and more through instructions rather than through just like typing a prom and just like getting an image. I think that this new model from OpenAI kind of shows that.

Yes, speaking of the new OpenAI model, it seems like they've evolved the actual underlying algorithm. It's not purely diffusion based. Are there new buzzwords or keywords that have you reverse engineered any of how they're doing that? Because it seems like there's a number of steps, like they're actually transforming the prompt. There's some reasoning in there. The image loads top to bottom, which we haven't seen before. Midjourney kind of diffuses everything from blurry to crisp, just the whole image at a time. It seems like they're doing some sort of blocks or line by line rendering. What can you tell us about how that system actually works? Works?

I don't have a. I mean I have some intuitions but I'm not like super. I don't have high confidence on how it works. It seems like there's some auto regressiveness going on and we have already seen similar things with CROC image generation. But I feel like to me what it's really game changing about this new image model is the. It's like very similar to what we were like talking about before. This is an image model that is able to reason and it's able to understand instructions. It's able to understand. Here's the picture of my dog. Turn it into Studio Ghibli. This is something new. This is something that previous diffusion models were not good at. Diffusion models are good at. You have a text and you can generate an image that represents that text but it's very hard to have them reasonable and to have think about like what you want to do and what is the instruction that the user wants and how to accomplish that goal.

Yeah, yeah. It seemed like there was like. It was like everything style transfer should have been. Plus the latest and greatest in diffusion models. Like they really like packaged that up very well and so I think that's why it broke through. But anyway, congratulations on the round. Thanks so much for stopping by and thanks for the office tour unexpectedly. That was really fun but we'll let you get back to work. Give our best so much to do to the whole team and we'll talk to you soon.

Sounds great. Thank you so much for having me.

Thanks. Talk to you soon. Bye. Very interesting.

We got Leif coming on from public. I'm curious if he's been sleeping at all.

It's a wild time in the market. He has some interesting data on what's happening on public because that's where people go to trade. Multi asset investing, industry leading yields. They're trusted by millions. Folks, you've heard us do the ad reads before but now we have Leif in the studio breaking it down for us and we will bring him in right now. How you doing? Leif, welcome to the stream.

Boom. What's going on? Great to finally have you.

Nice.

What's going on?

Let's go.

I was just saying I had my.

Caffeine already, so that's why I'm like, good, Fantastic.

Somebody was commenting yesterday about our caffeine consumption. Oh, yeah, like, yeah, it's easily, easily 500 milligrams plus at this point. John, that's 500 milligrams during the show. But he's like built like a horse, so he can take it. How are you doing? We were just joking about whether or not you had slept at all the last week. I know it's been busy, you know, I'm sure it's been a busy time for you and the. And the whole team.

It's been busy, but our system have, our systems have been, have been up at least compared to other folks. So that's great.

Yeah, that's good. Walk us through some. I mean, I'm mostly curious to hear you had shared on X yesterday about how there had been more buyers than sellers over at least in certain moments over the last week. Maybe break down that data point and then we can talk about some other stuff. That's top of mind.

Yeah.

I mean, generally speaking, just like a mini step back is like this generation of investors call, especially millennials, man, they have been through market cycles like crazy in the past five years. Right. Or even just through their lifetime. And I even saw some meme on. On X the other day of like millennials experiencing their fourth once in a lifetime opportunity of a drop in the market, you know.

Yeah.

And so on. And I think especially like the March 2020 drop where like circus breakers hit and so on, crazy is still in people's minds. And I think that specifically because you saw a lot of like individual investors actually also making money on that. And I think that has stuck with a lot of people. And so generally speaking, this behavior of buying the dip is a little bit retail investing culture now. And so whenever you see these massive drops, this is really when we see some of our best days.

Yeah.

Yesterday was like one of our record days in just deposits, for example.

Yeah. My reaction, the stock market was down 5% back to back days. And I was just like, what is everyone complaining about? We're not even hitting circuit breakers. Like, this is not that crazy to me because I remember 2020 and it was way crazier. But of course, can you. It's very easy.

Can you actually break down the mechanics of what it. Like the, like what exactly is happening and what truly triggers the circuit break.

Yeah, I don't have the exact numbers in front of me, but it's essentially just if it drops too quickly to specific thresholds, I believe it's 7% and then 10 or 15 or so, essentially they pause the markets. And that didn't used to exist in the past. Right. And so it's essentially a little bit of a safety trigger, like a speed bump to make sure that investors can take a breather when these markets start to drop too quickly.

Yeah. In Japan, don't they have lunch break in the middle of the.

They do.

Not just in Japan. Other countries too.

Other countries too. I love that. But yeah, it makes sense. I mean, we. A couple of years ago when the algorithmic trading got really popular, there was like the flash crash. I think the market traded down like 20% in like two seconds and then went back up. And yeah, obviously you want to avoid that. What about overnight trading? You always hear, oh, you're watching the futures market and it seems like somebody has an edge here that they can trade before the market opens. Is 247 trading coming to America? We've heard some rumors. Is there a way to get in on that action?

Generally speaking, I think it's definitely coming. It will also come to public at some point. Right now we essentially have 4am to 8pm but the thing that you have to think of is that each trading window has its own participants and its own liquidity and execution venues. And so think of it as like there's the regular opening between your 9:30 and 4:00pm that's the most liquidity. It's when like most people participate. You could call that the healthiest time in the market in theory. Then you have essentially pre market and post market, which is, you know, 4:00am to 9:30 and you know, 4:30 to 8:00pm and then you have overnight, which is like the new thing. Overnight. Right now there's essentially only one major player who drives the liquidity for that. And what happens there is that because you have only one player, you have a lot of. You have not as many platforms participating yet. And so you can have these moments where there's a lot of kind of unilateral flow happening. And that's why in the overnight markets you often see certain stocks just suddenly rally also on. And that is a little bit, I don't want to call it fake, but it's a little bit like it has these wild swings because the types of people that trade in those times of the market and it's A concentrated liquidity pool. And so these swings just happen way more dramatically. And so you often have these moments where in overnight a stock goes up. And you see on Twitter, everyone's like upside on X. Everyone, like posting the screenshots of like, oh my God, you know, Palantir is going nuts right now or whatever the company might be. And then suddenly like 9:30 market opens and it goes. Goes down and everything kind of normalizes again.

Right.

And that is really just because each market window has their own participants and their own pools of liquidity. And so you kind of have to take a little bit with a grain of salt, like, can you play that? Maybe, but there's obviously some risk there as well.

Can you talk a little bit about information diets? And really, I want to know what events that are predictable, not like Trump imposing massive tariffs all of a sudden on Liberation Day. But what can we count on, like clockwork, every single day to be the highest volume day of the year? Is there a Super bowl of stock trading that happens whether it's earnings day or jobs day? What. What are the big reliable sources of high liquidity in the market?

I don't know if I have a good answer there. My, My gut reaction would be just from internal measures like Monday mornings because you have all. You have a lot of cute orders from the weekend and stuff like that.

Yeah. Makes sense.

Get executed at the open.

Okay.

Because not everyone will trade pre market and stuff because spreads are wider and all this. All these things. But specific days. Not really sure, to be honest.

Maybe like big tech earnings too. Is kind of like a season for.

Generally speaking, is that like, it's always like. It's often just driven by market events.

Right.

At the end of the day. People will trade when they see opportunity.

Yep.

And if you can predict that, let's start a hedge fund together tomorrow. But other than that. But other than that, you know, it's. It's like it is driven by these, these, by these moments. Right. We have good days when the markets are in the news, no matter which direction. But as long as the markets are in the news, we have good days because people get inspired one way or the other. And you know, so. So in a way, trading volumes for companies like ours are a little bit like competing with any other thing that competes with attention. Because, you know, if markets are in the news, you get inspired by something and that might drive action. And that's what we see.

Yeah.

What do you see from a demographic standpoint? I'm curious. Public obviously offers access to bonds which, you know, it was probably good to be in bonds, you know, if you sold last week before Liberation Day. But. But do you see a lot of demographic differences? You know, sort of like Gen Z is basically like bonds for me or just like, you know, GameStop. Right. Like I always know it's gonna be worth something.

It's a store of value.

It's a store of value. Right. But I'm curious if you see sort of pretty specific activity across different demographics in terms of interest in these different types of assets.

Yeah, I mean, straight up bonds always skew older. Just from the perspective of the older you get, the more you're thinking of preservation versus growth. But then I think what's interesting now is that, so we've launched multiple yield accounts essentially. So you have your high yield cash, which is, is, you know, similar to a savings account. You just get your yield and it's variable based on interest rates and such. Then you have like your bond account, which is essentially like a basket of corporate bonds underlying. And then you have like your treasury account, which is like US Government Treasuries and those kind of simplify the investments into bonds because you just deposit money, earn yield. It's like very simplified and just like you don't have to like pick certain bonds and stuff like that. So what we've seen is with that is that a lot of people are using those to just put money into the markets, waiting for these moments of opportunity. So what we've seen the last three trading days essentially is that people cycle out of the yield accounts and into stocks and ETFs because they essentially had this cash lying around. And we're like, okay. I think totally what we've been hearing a lot is that, hey, after Trump was elected and the market started ripping and there were all time highs and a bunch of things going on all the time, that there was also a bunch of, of individual investors who were essentially feeling like, oh, I'm just buying the top right now. And so they put it into these yield accounts. But then also the minute you saw things drop the way they've done now, they've basically cycled it out of the yield accounts into stocks and ETFs specifically. So you're seeing younger generations using it less as a, hey, I'm going to now actually hold the bond to maturity 10 years from now and more use like these account types that we've created for, you know, just like earn some yield on your stuff, you know, until you, you know, actually see other opportunities.

Yeah.

Do you think AI is already have helping retail investors better understand the companies that they're investing in. Right. Every public company is putting out a huge amount of information. Unless you're becoming just overly obsessed with this specific stock, it's hard to figure out what's, what's important. What should I be looking at? What have you guys seen? I know you've released products to help people leverage all of that data with AI, but I'm curious what you're seeing.

That's a great layout, Jordy, by the.

Way, thank you.

To my pitch Thankful sponsor. But yeah, 100%. And we obviously launched Alpha which started off by just you can swipe down on any stock, ask any question about the stock and that just created this bite sized researching for things. And we've kind of fed the model with a bunch of data that we already had from years ago. For example, we acquired a company three, four years ago that was essentially a tool that turned all the SEC filings that had custom company KPIs of subscriber numbers and how many cars has Tesla shipped and things like that into more structured data. And then we use that structured data to train our models and such to make that very easily accessible. Now what now has happened is that it's much more proactive than just you kind of have to pull information. And so the obvious one is what we call why is it moving? Which essentially if a stock is going in either direction very heavily, we kind of pop this card on that page and tell people why this thing is likely moving right now. And then you can tap on that, that brings you into a conversation with Alpha, gives you a more granular breakdown on that. And so it's much more the pulling versus much more like the pushing versus the pulling. And I think that's also just generally where it's going. But what's awesome to see is that these bite sized contextual moments where AI can be super fast, where it's just really great at summarization and can also go against biases. So if we go for news stories, for example, we sort of say that QA multiple sources, so we're not just coming from one source and pop it to you, but we QA multiple sources and then the summarization comes from the multiple sources. And so there is a little bit of QA built in and a little bit of taking the bias out that maybe one writer will have or something.

Yeah.

Which can be helpful.

Speaking of data quality, what do you think? Yesterday Walter Bloomberg shared some news that was, that wasn't quite accurate and move the market, you know, trillions of dollars maybe helped us avoid whatever it was Black, Black Monday Kramer was calling it. What do you think is, do you think that, do you think that there's any like solution to that? Or it's just the nature of the Internet where now you have these accounts that basically act, they publish, they basically like mainstream media, except they just publish headlines. They're not doing any journalism, they're not even looking at data. They're just sort of like trying to be the first or second or third big account that's like sharing a headline. Is there any, is what's the fix there? Right? Is there one? Or is that just the nature of the Internet where this sort of information breaks and then markets are going to react really quickly and now retail investors are so ready to act on information. You know, a good example is like if you just happen to be on X when Trump posted Trump coin and bought 20 grand of it, you became a millionaire within a few hours. But I'm curious if you've thought at all about how. Yeah, I don't know if there is a solution. Right.

Yeah.

But I think we always come back to who are you building for and therefore what behavior is your product and inspiring and you know, and generally speaking, what, like the way you design your product will always have an impact on how people use it and their behavior. And in our case, you know, yeah, you cannot, you can buy a meme coin, you can do options trades. But generally speaking, the way we've designed the product and the offerings that we have are more focused around building long term portfolios for people that want to, you know, compound their wealth over time. You know, the, all the fixed yield, look at the fixed income offerings that we now have, etc. And I think there are just certain design decisions that impact in the end the behavior of these users. Right. And therefore in the end I think that is much more important for people to make healthy investment decisions than necessarily how they consume and so on. But that behavior you're talking about is obviously not necessarily coming from the potential wrong information from some X account. That behavior is more cultural or how they were trained when they started being in the markets, you know, and so I think that is, that is much more the sense of like if the platform you're using is closer to gambling, you will end up being more of a gambler automatically just because of the design of how you were introduced to the markets. And therefore you'll be more prone to, you know, potentially react on these types of things because your investing style will Be closer to a gambler than maybe someone who tries to compound their wealth over time and, you know, cycling money out of a hiker account into, you know, Amazon stock or whatever because they see an opportunity that.

Yeah.

Then also hold it for long term and so on. So I think that's much more the issue, so to say, than those accounts.

I have a bunch of other questions, but I think we're over. We'll have to have you back very soon. I know you got some big stuff in the works. Thanks for coming on.

Yeah. I'm excited for the announcements. This will be great. We'll talk to you soon. And we have our last guest of the show announcing a 30 million dollar Series B, the smallest round of the day. It's rough out there.

What is this around for ants?

I don't want to talk too much. It's still getting. That's a size gong hit. But it is funny. We've seen a bunch of huge rounds today. It's a good day in the markets. The markets. The public markets are down, but the private markets are ripping. Let's bring in the founder and CEO of Arena AI. Today, he's announcing a $30 million Series B and introducing Atlas, an AI hardware engineer that is used by many of the world's most respected and ambitious hardware companies. I'm excited to talk to him now, if you're there. Welcome to the studio. How you doing?

What's going on?

Nice to see you guys.

It's great. Would you mind with just starting with a little introduction on who you are, the company, and maybe a little bit of your background? Because I thought the previous company was really interesting too. So I want to hear about that.

Thanks. No, that's awesome. Yeah. So I'll tell you like a bit of that company. It's Atlas, AI hardware engineer. The company's called Arena. We're based in New York. Yeah. Background. I started out as an applied physicist, so I spent like a decade when I thought I was doing physics dealing with hardware problems again. This was like a while back, you know, switched gears, went and did a brief stint through consulting. So wore suit for a short flash of time over there and then bring it back. Yeah, it started right during that financial crisis too, which was a wild time to be starting a job. But then, then sort of misstech moved out of San Francisco. The first company, to your point was in 2014. It's called Kimono. So the idea was to make it really easy, I mean, to write a web scraper. And it was pretty popular. We grew to 150,000 users. We got bought by Palantir, which is where I met my co founder. We were there for a while and then started Arena 2019.

Very cool. Can you take me through the founding of Arena? How did you settle on this to build? It seems very on trend now, but you've been working on it for a couple years. What inspired you? What was kind of like the early go to market, the first customer that you were talking to, what does customer development look like, all that stuff?

Yeah, totally. So just, you know, we had a bit of an interesting path here. I would say it's like a little bit non traditional. We started, we decided to bootstrap the company and so we said, well, you know, if you think about like our view on enterprise problems, if you think about the B2B problem space is there's almost like a Maslow's hierarchy, right? Which is if you, let's say you've had a job for two to four years and you're like, you've encountered a certain group of problems. Like you've encountered payroll, onboarding, communications. But then if you've been there for a while and you're deep in that industry, you're almost, you're seeing another set of problems, right? And I think one of the things Palantir did so well is they were able to go so deep into a customer for so long that they encountered problem set for which there was very little competition for. And like previously you had consulting companies kind of doing that. So there's, there's, I think a whole host of sort of untapped problems. And so our view is if we want to tackle problems that are really deep in an industry that are really valuable, we need to go really deep with a customer. So that's been the philosophy since the founding days. Instead of saying we're going to sell to other startups and sell bottom up, the view is to start with a very difficult to enter customer, start top down. And the origin of the company was actually more, let's say less vertically opinionated. We had added depth and reinforcement learning and Transformers were like, let's go and apply that for enterprise problems. So we were not, as you know, our thesis hadn't formed as sharply as it had today. And then we saw traction in a few different markets. And like post ChatGPT, we were like, look, I don't think for a small company playing in horizontal AI is really a winner's game. But we found that like there was this beautiful intersection that went back to sort of my days as A physicist where like two different technical fields together with like, you know, applied physics, electrical engineering and AI. Now you have an interesting customer set where you look at like a hardware test lab. It actually hasn't changed in a long time. The incumbent competitors that are three companies from the 80s, you know, it's. It's weirdly the underpinning layer of technology on which all of our software runs weirdly hasn't changed that much. Developing the hardware has like stagnant. And it's kind of surprising when you think about it.

Yeah. Can you talk a little bit about like Hardware Engineering 101? Are we writing Verilog? Are we in CAD? Are there other systems? Like, what does the work look like? And is this something where it's like, it's managed on GitHub? So Devin's gonna go off and write some code for you and we're just doing fancy autocomplete. That's probably. I'm not trying to diss. It's incredibly valuable if that's what it is. But just concretize like, what are we actually talking about here for folks who haven't done hardware engineering?

Totally. Right. So let's like, let's break it down. Let's take a simple example. Let's take them like, let's take something like a drone, right? We've got like, you've got like the mechanical shell where you've got like your mechanical engineering, CAD models, stress strain modeling. Still a lot to do, but humans got pretty good at. We've been building physical stuff for a while. We can screw things in, weld them together. Again, not trivializing that ton of opportunity, but that's kind of figured out now inside, especially as you think about systems that are starting to go autonomous or partly autonomous, right? You're like. I mean, we've had helicopters, now we have drones, right? So what's the change? The brains of this are basically like a set of embedded systems, right? So embedded systems, effectively your computer, the green motherboard, like you've got inside, except a whole bunch of them, right? You've got like one that's operating as a sensor and multiple different types of sensors. So like an IMU for how you're oriented, like temperature sensors, optical sensors, so all of your sort of sensors, just like the body has and then a brain. And you might have multiple brains which might be on board. Again, in our drone example, flight computers. And then you have actions that you take, like servos and actuators. And you think about this, like inside that mechanical shell, you've got this almost electrical skeleton, sort of like your own nervous system sort of wired together. Right. And at that layer, you know, to your point, there are two things that are happening. One is all of the electrical connections to make that work. And then for certain of those chips, you're running code on board. So to your point about the Devon, so you know, you might have an FPGA that's programmable. You're putting code onto it. And so that's sort of the sort of. We're currently at that inner layer. We're currently at that nervous system because there's this huge need. And if you look at like the. Just the kind of labor markets for a second, this is actually weirdly not surprising, but kind of has profound implications. The last 50 years, if you look at computer science course enrollments, they're up by 90%. None of us are surprised by that. But electrical engineering course enrollments are down by the same amount. And so you look at it, we've got this huge research, we've got tariffs, we've got all that. We've got a huge resurgence in American manufacturing. Right? And now you have all these intelligent hardware companies, robot companies, space companies, and like you have this burning platform problem where it's like, oh my God, people haven't been studying this stuff. And like, we're trying to now ship at the velocity of a software company in a hardware space where stuff can literally explode without a workforce. Right?

Yeah.

My favorite example here is you have Sonos, which hasn't. Has made like beautiful devices, but they haven't managed to like just get like, even the saw. Like, like the collective experience of using a Sonos product is just completely brutal. Right. And then, and then you like look like this is a company with like hundreds, thousands of employees. They're public. And it's not even defense. Like, it's not critical that my speaker work, like when I want it to play music. Right. It's annoying, but it's not the end of the world. And then it's like, hey, if that is hard in a controlled environment and a home.

Yeah.

And then we need to do much harder things in the sort of these defense critical industries. Like, that's a. That should be a red alert.

Yeah, it totally is a red alert. And that's where we've got customers kind of screaming for it. And you know, at the root of it, like, you have this idea that, you know, it's often you talked about Devin. Right. Like, I mean, it's never been easier to write code. Right. Like, I Mean, already, weirdly, Python was an abstraction over, like, you know, C. It's not as hard as C. It's gotten easier and easier. And now like, like you're speaking in English and like, that's amazing, right? It's like the Star Trek computer. And it's a beautiful environment because code doesn't need to obey physics, it just needs to render in your browser. Right now, suddenly you're making contract with nature, as we know. You guys go outside, you're like, nature's unforgiving, man. It's not like we haven't like fixed a bunch. We don't have space elevators, we don't have jet packs. Like, no, we have like TikTok, which is great, but like, what about all of that? And the problem is we're encountering this physics. And so each test cycle, to your point about the Sonos, is like, great, I have an idea. I'm going to prototype it. Let me run it in my terminal. It doesn't compile. Great. I probably made a stupid mistake. Everyone makes these. But the cost of making a mistake at that speed in hardware, worst case scenario, you get it wrong, something explodes. But then even on the development cycle, each time you're like, oh, damn, the board was wrong, I need to go and re spin it. That's like you're adding three months to the cycle. And so these timelines and cost structures, I mean, we all know how much the F35 program cost and overran. It's like, that explains it. There's a lot more that explains it. But that's like a piece of it and an important piece of it.

Yeah. Can you talk a little bit about where you see the most value to be delivered in the AI stack from? It sounds like you're not doing pre training on a foundation model. Is fine tuning important? Is building a system on top of existing LLMs important? Are you doing reasoning or is it more about UI and integration into existing systems? There's so many different ways to create value in the stack right now. I'm sure it can be kind of overwhelming, but how are you thinking about it?

Totally. And you know, it's a cool question because our own thinking on this has evolved quite a bit. I would say we started with a view that was much more, we kind of need to own all the pieces on the modeling side and solve the hard modeling problem. And we sort of realized like, what's happening is base cognitive functions are just becoming available as an API. So like vision is just going to be available. We shouldn't Work on a vision problem, like go fine tune like a, you know, yolo or whatever. VLM is your favorite. You know, LLMs are maturing. But what we do find is if you think about like a person doing a work, imagine like our objective as AI is we're trying to get as good as like a medium class person, let's say, or like a junior person even, right? And we can unlock a lot of value with that. If you want to do that. Now people do, you know, if we think about. You talked about reasoning and there's a sort of like notion for reasoning in LLM land. But like if we just think about human reasoning, there's like a nuanced kind of like there are a couple of things that are special, right? There's some sort of like, especially in a formal environment like electrical engineering, there are certain rules of the world that we've learned over time that need to be true. It's like gravity is 9.8 meters per second squared. You can't probabilistically learn that by watching stuff fall in air and being like, yeah, my ML. No, no. I mean that's just some law, like speed of light. Like you're gonna encounter that. It doesn't matter if you're an ML model or like, you know, it's just real, right? So there's some of these things with your hard constraints and you know, where AI has struggled is you tell an engineer something obviously wrong, they're never trusting you again. And they shouldn't, honestly. Like you want to fly in a safe plane, you don't want that happening. So there's a piece here where it's like reasoning, but inside this sort of like structured, constrained, where there's a set of physics constraints that, that apply, that you sort of need A, to win, trust the user, but B, to work. Right. The second piece is it's this kind of multimodality where again, I'm not saying we need to build from the ground up those models, but you need to make sure you're getting really clean input, right? And so that's input from. And it's weird. It's not like you're taking text. Input text is of course a part of it. And I think the LLMs have gotten so good that it gives us an ability to like really ingest a ton of text documentation. For sure, that's a piece. But now you're also looking at the thing. You're looking at it visually, you're looking at thermally, is it getting hot? You're looking at Readings from an oscilloscope. And each of those things has meaning to an engineer. And the idea is, can you now tease the right meaning from that? And so a lot of our work is basically on that data and fine tuning side. How do we turn all of that into a package that can be fed in to a set of models? And the other thing we found is a person is doing multiple different pieces of work. A person might be saying, hey, I'm cross referencing in some data sheets. What should this FPGA be expected to do? Can this pin handle 10 volts at 100 degrees Celsius? Or is this thing, am I going to short out the most expensive part? So that's almost like a kind of a text based lookup. But then you're actually running a test, you're comparing waveforms, you're doing math, you're running simulations. And so what we found is we're using different systems to do the different, like different as agents to different pieces of the specialized workflow. So you sort of have this like meta agent that you're talking to and then you have these others that are sort of. And the line blurs now between like what's an LLM agent versus what's it calling? You know, we'll call tools, but those tools. If you'd talk to me in like November 2022 before ChatGPT, I would've been like, these are machine learning models and companies, but they're just tools now it's like, oh, I need a thermal recognition stapled with like, you know, the view from the waveform. That's an ML model that doesn't need to be like a 600 billion parameter model, but it's a non trivial thing to do. And so you could look at this entire constellation as being. That's sort of the product, if you would.

Got it.

One quote that comes to mind is we were promised flying cars, Instead we got 140 characters. I have to imagine that what you're building and other tooling, like it has a potential, you know, the exciting potential to me is sort of getting us out of this period of stagnation, right. There's a lot of companies that are building, you know, there's companies building supersonic jets, right? And they can use your tool. And then there's all these other things that we've yet to even imagine or we imagined in science fiction, but now we should probably think about building. How optimistic are you around, you know, AI helping to accelerate and help us sort of achieve these sort of like science Fiction dreams that we've had forever but have never quite been reality.

First of all, I love the quote and it definitely speaks to my heart. It's like, you know, if you look at. And it's an interesting question because it can feel sometimes like reading the news, like doom and gloom. AI is taking our jobs. And it's like, you know, I'll go back to an example that I lived as a physics grad student, right? I spent a lot of time and I supposedly came in to do physics. I was like, oh, I'm gonna do all this great quantum mechanics research. And I was basically like a mechanic and a plumber for like 99 of the time. I was like, this thing is leaking. I think there's a water leak. Oh, the screw got bent. Oh, like, like I was not doing, I was doing like less than 1% of physics, right? And I was like, you know, that was the reality. And I think that's a reality for a lot of us. And so if you could take a lot of that away, like, I think what it does is it changes to your point what human ambition should be. Like, what could you achieve? Let's say it takes 10 years on average to build a startup. In the past, like what we, what we considered a SaaS company is now just going to be a feature for in the future, right? And I think original Silicon Valley was about Silicon Right. It was actually hardware based. And I think we're going to see a resurgence, hopefully that. My hope is, like, what gets me personally excited is one of the magical moments at my last company Kimono was you took someone who couldn't code and you said, hey, wow, with this tool you could like write a web scraper. And it was like, we just got the most amazing, like customer comments. And I was like, felt this joy of enabling people to do something. And you know, hardware can be intimidating. You're in like a hardware lab even in college. You're like, oh man, this is really complicated. It's really. There's a high barrier. Does it need to be that high? You know, it's like we're seeing kids cheat on their essays with ChatGPT, that's a good thing, you know, we'll generate more stuff. What if we could let them cheat at EE lab with this? Like, would we have more people going into hardware? Can we lower that barrier? Like, what if you wanted to build a drone on the weekend? Like, you should be able to, right? You should have Jarvis. I mean, the goal is to be Jarvis for and kind of enable everyone to be a little bit of a Tony Stark.

Love it. In your announcement post, you highlighted five categories that it seems like you're going after in the first initial rollout. Semiconductors, aerospace, automotive, medical devices, and defense. Is that sorted by market size? Like the burning need? Just you like the way that it sounded in that order. But I am interested to hear which of those has the. Has the most immediate need or is the largest market Market.

Yeah, it's a great question. So we started with semiconductors, right? So that's sort of why we put it first. Because, you know, the most complex, especially we think about a lot of what we do is electrical today, electrical engineering problems. And if we take the philosophy of we want to be a little like Nike, start by selling to the Olympic athletes and get everyone to buy it, that was sort of the proving ground. And so we still have a few semiconductor companies, you know, that we're scaling up to. But I think that's like a. We all know the household names. It's a small set that are really valuable, but they established the credibility that actually helped us. We had a few great companies then come in inbound based on that. A lot of that was automotive in aerospace. And it's interesting actually. You think about like EVs and like nor autonomous driving and aerospace, you have a huge amount of that coming in. Medical devices. We've got an early customer in there. And it's like that's, that's going really well. You know, it's, it's, it's. There's a whole like FDA angle to this that we sort of need to, need to work through or newer that. But it's a, it feels like the potential for impact is super high. And so that's sort of like a little bit like the landscape. I think we're seeing a ton of pull on the aerospace side. Like especially as you look at that industry is getting. We're going to have more stuff that flies. And then you introduce space to the mix. And like aerospace and defense like increasingly are kind of mixed. So you look at these, these things, I think that's becoming kind of a unit in some way. And so it feels like there's a ton happening there right now. But yeah, I mean, that's sort of just a little bit. There was not like a whole bunch of science behind that.

Ordering makes sense. I want to get your reaction to the tariffs. It seems like you're probably an American company selling to a lot of American companies. You know, regardless of what you think about the economics, it could potentially Be a bull case for your company. How do you process the news? And what are you thinking about if it shifts your strategy at all over the next, you know, using this Series B over the next couple years?

Yeah. So, you know, it's, you know, we actually do have a couple of international customers too. And like, you know, I mean, obviously it's a huge impact. Right. I think the first thing we did was just like, call them because we're like, are you okay? You know, like, especially look at these. Like, the, the margins on something like a car have dramatically changed. Right. It's like you're doing some of the work here, you're doing some of the work there, some of it in America. It's like, it's, you know, there's a lot of American factories. What percentage of the car is actually getting made here is a totally different question. So you have a ton of like, kind of panic in the system. Right. But you know, you're right. Like, for us, it's been like, it's accelerated customer pull and deployments. They're like, oh, no. Like, we can't go ahead and have like that gross margin impact and therefore have to do this with people. We need, we need technology. And so it's actually a forcing function. Like if part of this means that us sort of quality and speed and sort of ability to manufacture needs to get up really quickly. And I think this provided the economic incentive for it and it's just not there. Honestly. This is just an accelerator. It's more fuel, there's more urgency than there's ever been on the customer side. So, yeah, I would say, like, overall, like a lot of like chaos there, but net, I think, like, from our perspective, good, because it means there's a huge pull and like this problem we talked about, you know, we Talked about the 90%, the changes in those people. Like, we suddenly need to do a lot of this hard engineering in America. And this has finally put like a dollar amount on how important that problem is.

How are you planning for kind of as AI technology gets better, it feels like we're firmly in the co pilot era. There's a lot of talk about, oh, electrical engineering exam. I'm sure that these models have aced them at the highest level. And there's a prediction that, you know, AI will earn an IOI gold medal this year. And it's at like 50% on polymarket. And yet I can't get AI agent to book my flight yet. So how are you planning for integrating that? Taking advantage of what's State of the art and amazing and what AI does well. And then how are you building around the rough edges of the kind of innovation? Jagged edge.

Yeah. You know, we think about this a lot because I feel like the question of defensibility probably is going to come up much sooner in a company's lifetime than it's ever done before. And so it's sort of like playing with fire, which is we want to be on the glide path where our product automatically gets better as the titans go to war and the foundation models improve. Right. Like, we just want to ride that wave. But if we're just. That's what we're doing. We're like, let's get documents from the Internet and help you do cross reference. Okay. That's like, not. That is going to disappear super fast. And so, you know, to, to, to your point, it's tied in with like your Math Olympiad or Physics Olympiad question, which is, yeah, you've got your friend who's the genius who's like, really good at tests, right? And then you've got your friend who I bet is a different person, who is really good at building stuff. And usually they're not the same person. Somehow, like, in my experience, there's always. There's your tinkerer friend who didn't somehow get the A. Right. That tends to be the way. And so we're not trying to pass the Math Olympiad, we're trying to be the guy who's tinkering in the garage, right? And so the tinkering in garage problem is very unsolved. Like, you look at AI's capability there, it's like, like it's, it's a disaster. But as the base cognition gets better.

To be clear, like, when I think about the guy tinkering in the garage, my, My dad was a. Was a high school teacher and he taught this class called Project make. And it was all about. It was some combination of, like, wood shop meets electrical engineering meets physics, and, you know, you're making rockets and all that stuff. So when I think about him solving problems, it's like the tinkerer in the garage and that he would just try a lot of different things and experiment. And the beautiful thing about AI is, like, I have memories as a kid of him working on one little problem, like for five hours, like on a Sunday, like, trying to figure something out, whether it's around the home or in class. And, like, he's basically running like a series of experiments, right? And so the potential of AI is like, run every experiment at once, like in a simulated environment, but, like, run like a thousand experiments in, like, you know, 10 minutes.

Right.

And, like, when you start to think about what that can do for, like, accelerating progress, that's the most exciting thing for me because it's superhuman. It's like the tinkerer in the garage, but multiplied by million.

Yeah, totally. I mean, I want to get your dad signed up with the sort of. We're working on, like an academic edition. I want to give him free access to that, see if he could even. We're looking for feedback, but I think that's retired.

But. But I'll put you guys in touch.

If he's still interested, if he's still got a garage, that'd be great.

Well, I mean, thanks so much.

Here's the bar for the team. Five years from now, I want to be able to design our own podcast equipment.

Oh, yeah.

The most cutting edge. You know, we need H1 hundreds in these things.

Yeah. Microphone.

I don't know why yet, but it sounds cool. Congratulations. I do have one last question. How much do you attribute your incredible energy levels to your, you know, being. Being a triathlete? Do you think it gives you an edge as a. As a. As an operator?

I would. I don't know. I would hope so. I don't. I do a lot less triathlon than I would like, but I feel like it does. Training the pain threshold is a useful thing. I think it's just a useful thing in life.

Yeah.

That's awesome.

Love it.

Well, thanks so much for stopping by. Congrats on the milestone for the series B.

Thanks for having me, guys.

Appreciate it.

We'll talk to you soon. Have a good day. Let's go to the timeline. Justin Ross is, quote, tweeting, did you see the Colossal company? They have brought back dire wolves using ancient DNA with their firstborn on October 1, 2024, they waited a couple months to make sure that dire wolf was healthy and growing. Over 10,000 years since dire wolves were extinct.

Can we do a deep dive on this?

I have done a full video and deep dive on the company Colossal Remix. I was emailing with the company a while back, rekindle that connection, hopefully have Ben, the founder, on the show. He has a very funny collection of investors, but some really great. George Church from Harvard. Fantastically renowned scientists involved and they're working on cool stuff. So this post was put in the Truth Zone. People said, hey, they're not technically dire wolves. They didn't really revive them using ancient DNA. It's more genetic modification of existing dogs a little bit of controversial but JD Ross chimes in and says I don't care if these are real dire wolves or not, they're very cute and we should mix in golden retriever DNA and use them to hunt deer with us. And I couldn't agree more. And you know where I would love to have a dire wolf hanging out and maybe go on some deer hunting in a wander. I want to find my happy place.

Find your happy place.

Find your happy place Buck a wander book a wander with inspiring views, hotel grade amenities, dreamy beds, top tier cleaning and 247 concierge service. It's a vacation home but better code TVPN and I want to go to Mike Noop who we had on the show. Founder of Zapier. He says on the topic of AI is trained on all of humanity. Why can't it innovate? A big question that we're talking about that question of the test taker versus the hacker. He says new ideas come from two places. One noticing similarities between two existing ideas. New ideas in one area translate into another and two logical construction. New ideas follow from prior axioms. One is easier and bounded and two is harder and open ended. And d' Orkesh was talking about this like if you've trained a lot of a lot of scientific innovation just comes from somebody who's read so much about the scientific literature, they put together two random studies and they find out that if you put those together you get innovation. And then that's true in all sorts of different industries, but specifically in just if you've read all the papers, you've read all the books, you start making connections. This is what David Senra does a lot with his show Founders podcast. Go download it. But he says two is harder and open ended paradigm. One can look a lot like career advice to work at the intersection of two fields because it's easier to become an expert. In contrast, being an expert in a single domain requires much deeper hierarchical knowledge and innovation requires novel in domain idea construction. And this is the story of Elon Musk working in space and electric cars. Maybe having both of those knowledge sets multiplies in some way. 2 He says it's harder because you don't know if innovation is blocked due to the prior axioms not existing yet or if you just haven't combined them in the right way. We want to build AGI that can innovate due to the fact that one is bounded. Search leverages ML strengths like pattern recognition and can bootstrap from human knowledge. I think we will create AGI that can reliably do number one. Well before number two. Very, very interesting take. And I just think it's like an interesting question that he's clearly asking. These AIs, they're blasting through all the benchmarks, they're doing all these amazing things, but we're not seeing innovation come out of them yet. Or even you could think about the joke test. Like you kind of need to be innovative to come up with a joke.

Yeah.

It needs to come from some. It needs to be new and fresh. It's not just information retrieval.

The idea of like creatively creativity is often just taking ideas from two different places and combining them in some way.

Yeah.

And it feels like the models do that very well today and that you can ask it to make me a song. It's not doing it independently. It just like. Yeah, you know, you have to sort of prompt it.

This was the genius of Harry Potter Balenciaga. Like the, the human element there that made that actually go viral wasn't the AI. It was the idea that combining Harry Potter kids story with Balenciaga high fashion, that was funny. And then the AI just instantiated it. And I agree with you. Like the idea of, of taking two disparate concepts, putting them together is where you get genius. Like take your best performing ad and put it on a billboard with adquick.com like that's going to perform better. And so go to adquick.com out of home advertising made easy and measurable. Say goodbye ad quick.

Basically took the amazing attributes of online performance marketing and brought it into the real world.

That's actually what they did. That's true.

Serious.

Yeah.

That's the whole point.

We're not messing around. Yeah, we are.

You get a dashboard, you get all the different things that you expect when you're running a performance ad campaign online. On Facebook it gives you similar dashboards but out in the real world. And they do a lot to help you track the performance of your out of home campaigns. I thought this was a funny one. We'll move on to Quake 2 has been fully AI generated and replaced by Microsoft. You can play it in a browser. Every frame is created on the fly by an AI world model. So they trained it on Quake 2, had the algorithm or the AI play a ton, generate a ton of frames and then just take the input from the controller output of the frames. And so there's no, there's no game engine. It's just input is what you're doing on the controller. Output is the Game with visual fidelity. And you can see where this is going. It's crazy. It got a lot of hate.

Yeah.

Read this.

Quake dad.

Quake dad.

Clearly a fan of Quake, dedicated to the, committed to the bit, says, meanwhile.

They haven't released a new Quake in like two decades. So this guy has been in the trenches forever.

This is absolutely disappointing, disgusting, and spits on the work of every developer everywhere. Bold John Carmack says what question mark. This is impressive research and I love.

That because he's the creator of Quake. And there was an amazing. There's an amazing meme that was like John Carmack holding a white monster, being like, oh, you completely replicated exactly what I did. Awesome work. And it's like the developer himself is like, this is cool. But he did unpack it a little bit more. And so I want to read through this. He says, I think you are misunderstanding what this tech demo actually is, but I will engage with what I think your gripe is. AI tooling. Trivializing the skill sets of programmers, artists and designers. And that's real. My first games involved hand assembling machine code. What a code. This is why he's one of the greatest programmers of all time. And turning graph paper characters into hex digits. Software progress has made that work as irrelevant as Chariot Wheel mate. Yeah, you don't want to be in the business of Chariot Wheel maintenance. Not a big industry today. But building power tools is central to all the progress in computers. Game engines have radically expanded the range of people involved in game dev, even as they de emphasize the importance of much of my beloved system and systems engineering. Maybe first person to say that systems engineering very, very hard. And it's a lot of. It's a huge time suck to. I mean when he built Quaker, he had to build the whole game engine. He had to build everything. The idea of a floor that you can't fall through a floor or a wall, you don't want to walk through the wall. You have to write all that code from scratch instead. Now you just fire up Unreal Engine and you get Fortnite out of the box. Or you build in Roblox. Right? So he says AI tools will allow the best to reach even greater heights while enabling smaller teams to accomplish more and bring in some completely new creator demographics. People who don't know systems engineering or even programming, for example. Yes, we will get to a world where you can get an interactive game or novel or movie out of a prompt, but there will be far better exemplars of the medium still created by dedicated Teams of passionate developers. And this is like the innovation concept, this idea that, you know, the. The Harry Potter Palenciaga game will be the one that goes viral and gets a lot of attention. If you have distribution, you can capitalize on that. But also if you have a novel idea that AI could. Could think of, you will have a breakout success. And so we'll focus more on game mechanics. Game like there was this game, Balatro, that took. That takes poker cards and you're basically playing poker and trying to create like royal flushes and whatnot, but it adds all these crazy mechanics on top of it. It was a very simple game, just designed in an engine. Not crazy on a technical level, but the game design was so, so incredible and so novel that it just went massively viral. And like the solo developer, basically, I think he had a few people on his team just printed and became like the number one game of the year or like of that quarter on Steam or something.

Well, what should you do if you're printing, John?

Pay your taxes, that's for sure.

But your sales tax go to numeral autopilot.

Sales tax autopilot spend less than five minutes per month on sales tax compliance. Go get started. They're back on your benchmark.

Five minutes a month.

You don't want to be bogged down. And I mean, this is true. Like, you want to be focused on the innovation that your company is doing. You don't want to be dragged into a bunch of unnecessary reports. And.

You can be thanking your lucky stars that you're not a part of the trade war right now. But take an opportunity to get your sales tax ducks in a row. 25 states are now taxing software sales and Numerl helps you stay compliant.

So go HQ and check them out.

Just do it.

And I think that's a good place to wrap up. What do you think?

Yeah. Fun show, John.

Great show.

I enjoyed podcasting with you today.

It was fantastic. Can't wait to do it tomorrow. I know. I was so worried. I checked the date in my intro and I was like, is it Wednesday already? It's not. It's Tuesday. We got three more days. Glorious podcasting.

I know. I have a feeling there'll be more news this week for sure.

The.

The size gong hasn't rung its last gong.

Sound the gong. The gong.

The gong is still ringing.

As long as we're.

Yeah, it's great. Three beautiful, three amazing Series Bs. We got all three founders on. I think we did great.

Was Jaws a Series B?

Oh, I guess.

More like a pre seed.

More of a pre seed anyway. Three big three ten plus million dollar rounds in the tens of millions. Great to see it. Lots of money flowing into startups that we love to see. All working on various things.

Remain bullish on America.

Me too, me too. Anyway, thanks for listening.

Never lose faith.

We will see you tomorrow.

Have a great afternoon.

Have a great afternoon.

Cheers.