Grok 4 Launch Breakdown, OpenAI to Release Web Browser | Chris Paik, Will Bruey, Joel Becker, Dylan Parker, Eric Olson, Ghita Houir Alami, Elliot Hershberg, Karim Atiyeh

We are live from the TVPN ultradome, the temple of technology, the fortress of.

Finance, the capital of capital.

Today we are covering Grok 4 launched. We're going to break that down. The third browser war has begun. Every artificial intelligence company is getting in the game, launching new browsers.

A browser.

The new Volkswagen electric bus is a flop, according to the Wall Street Journal.

Ouch.

Apparently Linda Yaccarino was not fired for the Grok dust up with the crazy hallucinations that were going on. We have more details there.

And, well, and I don't. Why would anyone think that she was considering that?

Clearly that was the timeline that was for sure. In the timeline people were talking about that like, oh, like this happened and she stepped down within like six hours.

Oh, really? Yeah, My read on it was clearly Grok and Xai are not her domain.

Yeah, totally, obviously.

And, and Grok was saying some things about her that should never be said.

Yes.

And my read on it was, who knows when Elon commented on her post and said, thank you for your contributions. It's like the boilerplate text. And so I'm sure that their relationship is maybe not as good as it was day one, but I almost thought it was maybe the, you know, Mecca was the straw that broke the camel's back. And she basically said, look, like, you know, I can no longer bet my career on this platform.

Maybe. Yeah, I mean, we can debate it. There's more reporting in the Wall Street Journal about what actually happened. And that story, which we'll get into is kind of pointing to this idea that it had been in the works for a while. Yeah. And that was not the straw that broke the camel's back. That was like the, like the papers had been signed, everything had been signed before that. And then the dust up happened with Grok. But the bigger news is that they actually got Grok4 out and people are excited about it. So we'll talk about that. And then the other Grok, Groq, the CEO of which. Of that company we had on the show.

Was it yesterday? I'm losing track of time.

It was very recently. No, it was Tuesday.

Tuesday.

Apparently they're out raising at 6 billion and we have some more details on that company, so. So that's interesting. Anyway, let's tell you about ramp.com time is money save both easy use, corporate cards, bill payments and a whole lot more all in one place.

They have a new agent launch today. Kareem will be joining later in the show. We're excited for that, for him to break it Down.

So let's break down the Grok 4 launch. DD Das has a summary. Insane. That Elon Musk has pulled it off again. Absolutely crushing the AI wars with Grok 4. And we can go into some of.

The meta crushing the benchmark wars for sure.

And there's a question about like, are we post benchmark? Does this matter? What's the real question to be asking here? But there's a bunch of interesting takes. So just summarizing the core announcements. Post training RL spend was equal to pre training spend for this, for this release. That's the first time it's ever been like that. I think when you go back to the original RLHF stuff that ChatGPT was doing, that kind of unlocked like, oh, wow, this really, really works. Pretty sure the pre training spend was an order of magnitude or two orders of magnitude bigger. Now we are truly in this reinforcement learning regime. $3 per million input is tokens, $15 per million output tokens 256,000 token context window priced 2x beyond 128k. It's number one on humanity's last exam.

Which interestingly was effectively like postgraduate PhD level problems, but across a bunch of different domains. So everything from literature to physics.

Yeah, kind of like the hardest SAT possible. Interestingly, I believe that benchmark was created by scale AI. And so Alex Wang is now at meta trying to figure out how can we beat our own exam. And Elon's just like, I'm number one at your thing.

Interesting dynamic. Yeah. The real test would be Elon doing the same problem set himself and saying, look.

Well, yeah, I mean, I was talking to Tyler about this before the show. Like, you know, it's like humanities last exam. It's like really good at PhD level math, PhD level stuff. But like, how often are you running into those types of problems?

Yeah, I mean, I think that's the whole thing about there's this concept of like spiky intelligence.

Right.

Where it's like, okay, it's really good at this very obscure problem that I never deal with. But if I have a super long kind of like context window, like, or there's no kind of like long term. It just completely loses its footing and then it's like useless.

Yeah, we're kind of in like less of the benchmark regime and more of the agentic, like how long can the agent run? So it's like we are in the 15 minute AGI regime. Maybe this is 15 minutes of like even better AGI. But we want to go to 30 minute.

Well, and this, you know, takes me back to him talking about continual learning being the next problem that we really need to solve. Because it's great if you have a PhD level expert in your pocket that can solve any problem in any domain almost instantly. But if it can't learn and take feedback and improve on certain tasks, then it's basically like useless. Like if you had a PhD level, a PhD join your team to work on a specific problem. But it was hard restarting at the beginning of every single task with no prior knowledge, it would be almost impossible for that person to succeed. Humans still got it on that front.

But at the same time, if you are trying to just really establish yourself as at least an API for tokens that every business should check out against anthropic or the OpenAI APIs. Just saying, hey, you know, we're on frontier. We're Gemini. Yeah, we're on the frontier is a good way. And they certainly proved that with GPUA hard graduate math problems at 88%. The really interesting news.

Yeah, the interest, I mean it's worth calling out. So Grok got number one on humanities last exam at 44.4%. Number two is sitting at 26.9% and then going down this list of all these different sort of challenges, they are consistently well beyond the second place. So they are at the frontier now of all these different benchmarks.

Yeah. So Mike Knoop over at Arc AGI says zooming out on Arc progress. I'd say OpenAI's O series progression on V1 is a bigger deal than Grok's progression on V2. So far the O series marked a critical frontier AI transition moment from scaling pre training to scaling test time adaptation. And. And this was the O series progression. If you remember that OpenAI was spending it was like thousands of dollars of reasoning tokens generated in the test time inference to actually get a good score on the V1 of Arc AGI. And so it had to think a ton, but it was able to figure it out and at least it proved that throwing a ton of tokens and a ton of inference at a problem and letting it cook basically wound up producing progress there. So that was kind of like a new just a new Paradigm says whereas Grok4 mostly takes existing ideas and just executes them extremely well in my opinion. The notable thing is the speed at which XAI has reached the frontier. And that is really like it just can't be understated that this is crazy. You put a post from OHNN in the chat.

I'LL pull it up here. He says elon Musk is such a beast. I'm not even gonna pure. I'm not even a pure fanboy anymore. How does he. He's a lot of swearing in here. Own gotta keep the keep the timeline pg. But how does he come out of nowhere with a cold start late to the game and ship Grock 4 and do it alongside everything else he's up to? He's launching new political parties. He's literally magnitudes above every founder. It's humbling. So basically everyone agrees that it's almost like he was a co founder of OpenAI.

Yeah, I guess he was returning.

You would have to almost be a co founder over there to be able to do something like this.

Let me tell you about graphite code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. You can get started for free@graphite.dev if.

You want to ship like ramp get on graphite.

Yeah, Chamath was saying the same thing. Somebody in his reply says, seriously, how does this guy produce what he produces? Meta is buying talent at $200 million a year and Elon keeps his people at a fraction. It's mind blowing. Very deeply underappreciated edge for Elon, says Chamath. The retention of the best people happen when you can offer them a freewheeling culture of technical innovation. No politics and few constraints. And people in the comments are like no politics. What are you talking about?

Yeah, can get a little political over there. But.

But, but probably not within the engineering org at xai. Right, right. Like it's probably just okay, how do we build the biggest thing? Cool.

Well, you can imagine the politics of like who gets the best spot for their tent. In the office tent there's a hierarchy. A tent hierarchy, proximity to the house.

I want to be directly under the air conditioning unit. I want to be closer to my desk.

The windows can be nice too. So you can pull down your tent a little bit and get a little view. Morning light.

I wonder what the political structure is of the tent city.

The tent hierarchy.

Is it democracy? Do they vote for who runs the tent city?

I guess it's just the XAI tent city.

It's probably just Elon at the top, but does he have a tent?

Something about San Francisco Intense.

Yeah, very funny. But SWIX has been chiming in saying like we need community notes for LLM benchmark porn because in the in the Grok for launch they highlight this Aime competition math problem and and, and I mean it's. And so Matt Schumer is basically saying, AI aime is saturated. Let that sink in. Grok 4 got 100%. It made no mistakes on that benchmark, which is obviously very impressive. But there's this extra comment about the nature of aime and so it's a cautionary tale about math benchmarks and data contamination. Apparently, you know, like predictions was that the models weren't smart enough to actually solve these. But he says, I used OpenAI's deep research to see if similar problems to those in AIME exist on the Internet. And guess what? An identical problem to Q1 question 1 of AIME 2025 exists on Quora. I thought maybe it was just coincidence, so I used Deep Research again and problem three. And guess what a very similar question was on Math Stack Exchange skills. Still skeptical. I did problem five and a near identical problem appears on Math Stack Exchange. And so like at a certain point, if people, you know, put out a benchmark, then talk about it a lot online and then that gets baked in the training data. You're just memorizing the result. You're not necessarily learning everything. It's still cool. It's good, it's good to have everything memorized, but it really, it's not beating like the knowledge retrieval knowledge engine allegations and it's. And we're not really in full BHP intelligence.

When Scott Wu was on the show earlier this year, he was basically saying AI will win an IMO gold medal this year. He felt very confident in that. And I'd be interested to see how he thinks about. And I'm pretty sure new performance.

Yeah, I'm pretty sure the IMO gold medal questions are public once the IMO happens. So every year they're, they're developing new questions, but then they go out there and then they get memorized and the solutions become discussed and you know, there's all the context based on that. And so, yeah, it gets kind of baked in. So big question about how valuable are these? At the end of the day, it's really just about like adoption. And that's why, you know, we were looking at the poly market for the best. Which company has the best AI model at the end of July? And XAI has just surpassed Google, which was sitting around 80% chance for a while and then started dropping earlier this week. Last week started dropping and now XAI is sitting at 48%. Google sitting at 45%.

Well, yeah, actually updating, it's updating live. Google's back up at 49%.

Is Google planning to launch something new in July? Because it feels like this market particularly is more driven by Google's release schedule, because Google might have something in the lab, but they like to release things at specific times. It's a big company. They don't just drop it. Who knows?

Gemini team Logan over there might be fixated on this Polymarker.

I need this.

Yeah, yeah, yeah. Oh, during the wait, he was like, if you need something to kill the time. Google AI studio.

So, I mean, people were definitely memeing the production values on the Grok 4 launch because it was supposed to start at 8. I think it went live at 8:45 or something like that, maybe a little bit later at Pacific time. And eigenrobot was saying, yeah, this.

This market is based on lm, LM arena, specifically the text leaderboard. So currently they haven't fully updated it. So it's unclear right now. Gemini 2.5 Pro is still at the top, but I think the expectation is once they get Grok up there, it will be the top spot. So we'll keep following. Yeah, this market, there's over 2 million of volume already on it.

Yeah. It's so interesting that Anthropic is not on this poly market at all, because people talk about them as having, like, the best vibes, the best, like, big model smell, the best, like, you know, interaction. And Ellen Marina is, like, supposed to kind of like, test that with these A B tests, and yet, like, doesn't seem to be performing there. But it almost doesn't matter because they're just focused on, like, the business at this point as opposed to, like, the. The benchmarks.

So I don't know.

It's all changing.

We have a post here from Ben Hylek. He says, elon Musk on AI. So during the presentation, a lot of people were critiquing the presentation, saying that it didn't feel, like, super polished or whatever. I don't think that was the intent. And it was pretty fixated on the models themselves and what went into them and what they're good at. But Elon did have this one quote in here where he says, and at least if it turns out. So he's talking about what will, you know, what kind of impact AI will have on the world. And he goes, at least if it turns out to not be good, I'd at least like to be alive to see it happen.

It's like, if we get the Terminator ending, I want to be around for that.

Yeah, I want to experience it.

What does that say about those timelines? Because it's like, is he expecting not to be alive? I feel like most people that have been in the doom category have been like, the doom's coming soon. Not. Not the doom's coming in 200 years.

I didn't. I read into it more like, he will find it interesting if that is the outcome and it'll be entertaining. Less so, like, will I be alive when it happens? Kind of thing, but who knows? There was another funny quote at the end of the presentation where Elon kind of looked around at the very end and he's like, anyone else have anything to add? And one of the engineers, because it's a good model, sir.

Extremely online crew. Yeah, definitely, Definitely on brand. Well, Ben Hylak, as you know, he's been on the show, he's a designer, probably working in Figma all day. Think big, think bigger, build faster. Figma helps design and development teams build great products together. You can get started for free@figma.com and.

We have our first product coming out very soon with Figma make that Tyler has been cooking on.

I've been very. He showed me. He showed me it and I was like, oh, like someone built the thing that we were thinking about building, like, and he was like, no, like, I. I did generated this. This is in Figma. I was like, this is like an iframe on another website that, like, already exists because it looks like exactly what we want. But it looks so good. Like.

Like, it looks like he works on it.

He looks like it.

It looks like he worked on it for, like, a few weeks.

No, it looked like someone else did it. It looked like it was a professional product that, like, stole our idea, basically. I was like, oh, like someone else got to it. That. That was the vibe when I heard it. Yeah. Well, how is the. How has the experience been? I don't know if you want to leak exactly what you're working on, but.

Yeah, I don't want to talk about it too, you know, closely, but.

But how many problems it takes you to get where you showed me.

Yeah, I mean, maybe five.

That's so crazy, this thing.

The design is super. It's really great.

It's really good.

Yeah. The fact that it came out looking like, basically, like 90. Like 90%.

Yeah, yeah, yeah. And I imagine that there's probably like the last 10%, if we were really strict about, like, it's got to be on this exact style guy. Like, that might be something where, like, you know, Tyler winds up spending more time. Finalizing customizing stuff but in terms of like just getting a functional prototype out. Oh man, it was mind blowing. It was awesome. I'm very excited about the age of vibe coding. This is an interesting chart from Tracy Alloway. Been on the show. The cost to rent an Nvidia H100 GPU hit a new low this week with annualized revenue at 95% utilization falling from 23,000 at the start of May to less than 19,000 today. So that's not that big of a percentage drop. But it is, but I mean it is a 20% drop.

It's a consistent trend.

It's a consistent trend. I wonder how much of this driven is driven just by all of the frontier labs that are driving the most adoption are moving on from the H100 to the 200. I don't know what else would be driving this because if, if you can, if you can still get like, if you only take a 20% drop off of a full refresh of a new, of a new, of like a new hardware and it's the latest ingredients anymore.

Pricing drop, not a utilization.

Yeah, annualized revenue at 95% utilization. So this is revenue per unit utilization.

Is still very high. It's the price that these neoclouds are able to rent them for which is dropping.

Yeah, I mean the market's more competitive than ever. There's more neocloud spinning up and more people actually inferencing these things. And then I guess this is the question of like how, how stuck will certain workloads get? Like if you, if you have figured out a great use case for an LLM in your organization and it's something that's you know, not one shotting your entire stack or whatever, but it's just like, you know, we have data flowing through our systems and we are going to use, you know, LLMs are going to, you know, interact with every PDF that gets uploaded to our, to our website or whatever. And so we're inferencing a lot. Like you might not need to put that on the latest hardware or update the hardware forever. You might just like be like, yep, it's llama 3, it works. It's on H1 hundreds and it'll be on H1 hundreds forever. And that piece of our business will just stay there just like, you know, we have a postgres database that you know, works and we're not changing it every year. We're not changing everything. We're just like, we're just trying to cost, optimize that and just hopefully, the cost just comes down on that. But, like, we've solved this particular problem, then we'll go solve new problems with new technology. So I think that's probably what's going on here. But it gets to the point of, like, the biggest question with GROK is that, like, the model clearly is frontier. It works. The whole fine tuning on the actual X account is like a crazy final step of, like, system prompt. And people were joking about that. Like, oh, they got to fix that. It's like, that's not what they're demoing today. They're demoing, like, the underlying RAW model, which is clearly, like, just engineering focused, as you saw in the. In the. In the demo, the demo which was just like, you know, benchmarks.

Turns out the secret ingredient to crushing every benchmark is to have the bunch of data from schizophrenic posters on accident.

No, I actually think it's the design of the RLHF stuff and the design of the reinforcement learning pipeline. Tyler, you got anything?

Yeah, I mean, I think just, like, so far what I've seen on X, like, the overall response, like, vibe stuff, is that people are saying maybe it was a little too kind of overfit on the rl, like, VR, like, verifiable rewards. You kind of see this when even in the demo, I think it would sometimes respond in the answers with, like, in, like, latex formatting.

Oh, sure.

Which is like, okay. That means obviously they've trained a ton on, you know, math questions, stuff like that. Maybe people are saying maybe it was kind of, you know, benchmaxed. You see it like, you know, 100% on Amy is, like, kind of crazy.

It's like sauce. It's like, you don't. Yeah, yeah, yeah. This is the thing about democracy. Like, if you win, like, 80% of the popular vote, it's like, okay, let's say it was a blowout. If you 100% of the popular vote, like, probably not. Probably not a democracy. I don't know. I mean, in theory, these things should be able to do it. But I'm interested to know more. If we dig into ARC AGI, is there more stuff going on there? Are there any secrets? Because it does seem like kind of an outlier result. You can see it from this Aaron Levy post. Grok 4 looks very strong. Importantly, it is made. It has a mode where multiple agents do the same task in parallel, then compare their work to figure out the best answer. In the future, the amount of intelligence you will get will just be based on how Much compute you throw at it. I was joking with Tyler about this, that the individual models are mixture of experts models. So there's a whole bunch of parameters. Right. And then the individual parameters, like, light up the different neurons based on an internal to the model router. So there's kind of like the math section of the brain, the literature section of the brain. And so this was like one of the. This was one of the key breakthroughs in like GPT4. Right. Was a mixture of experts.

People think we're not super sure.

Yeah, we don't still. We don't fully know, but that's like an internal decision that happens within the model to be like, let's go. It's. This feels like a math question. Let's go down the math path in the model. But then Grok 4 is doing multiple. It's running the same model multiple times and then comparing the results. And so now you have grading it. Yeah. You have multiple agents running mixture of extra models. You have mixture of. Mixture of agents running mixture of experts models. And the next thing is going to be like, if you want the absolute best intelligence, you need a mixture of companies. And you need, like I said, one prompt and it goes to Grok and Claude and GPT and Gemini and a human.

Yeah. I wonder how open Router's thinking about this stuff. It is funny to think about the human version of that, where you give five engineers on your team build the same feature and then kind of compare notes afterwards. It's wildly inefficient. But with software, when you can do these things very quickly, there's incremental cost, but you can have more confidence in results.

I mean, it's basically like having a brainstorming meeting with the whole team and just throwing up a question and being like, hey, like, we have this hard problem that we need to solve. Here's my idea. What do you think? What does Tyler think? What does Ben think? You kind of like go around the table. Everyone kind of gives their input, their various expertise. They kind of think through the problem in different ways and then you compare answers and everyone kind of coalesces around one strategy. This is like how work happens in the real world with a meeting. It's kind of the same thing, but certainly expensive to do that. So it'll be interesting to see where companies like, how eager are companies to jump over to Grok? Because it seems like it's been a big lever for Microsoft to have Grok in the ecosystem as kind of a stalking horse for all the other models, because Satya wants Azure to be very model independent, serve them all. They have the. I think they have exclusivity for ChatGPT or GPT APIs or they have obviously a great deal there with OpenAI. And so if they can have Grok 4 as well, that's another tool in the tool chest to be like this top layer.

Satya is in such a good position, it's probably not discussed enough how much just by owning those end customer relationships and being able to vend in whatever model is hot at that moment and give people optionality and still get 20% of OpenAI's revenue, at least for now.

Yeah. He's also SoC2 compliant.

Of course.

If you want to get SoC2 compliant, head over to Vanta Automate Compliance, manage risk, prove trust continuously. Vanta's trust management platform takes the manual work out of your security and compliance process and replaces it with continuous automation, whether you're pursuing your first framework or managing a complex program. So yeah, eigenrobot was talking trash about the production values.

I don't know about that. They were just noticing.

I didn't think it was that bad.

Slides are worse than I'd create after getting roped into a presentation with one hour notice. You can tell the engineers made them themselves. I think this is just a reflection of the culture. Right. They're not. Yeah, very clearly. It's like screenshots dropped into a slide.

But it's light mode screenshots on dark mode slides. So let's do black slides and then you come with your white with your white screenshots that are kind of like misaligned and not really evenly distributed. They didn't do like the distribute evenly or whatever.

Distribute horizontally still gets the point across. And I think it's a reflection of their culture and, you know, it shows what they care about, what they don't care about. They're not trying to be the most polished, they're just trying to be the best.

Yeah, eigenrobat kind of did like a whole like live tweet here.

Yeah. So Elon Musk was predicting the model will discover new physics within two years. He said. Let that sink in. Long silence.

Is that one engineer. LAUGHS AWKWARDLY Is that sooner or later than his previous timeline? Because he was talking about AI discovering new physics soon. I don't remember if he was saying dating it two years or three years or one year before, because this could be that he's still excited about this. He still thinks it's possible, but he thinks it's going to take longer than he said previously. And that's kind of the more important update. I don't remember what he said originally.

See if Grok can find out.

But he was saying this at the Grok 3 launch that, like, that is the goal and if you can get there, you've kind of solved everything. And Sam Altman was talking about that too, that if you can create a super intelligence, that's probably the first thing that you'd want to do is like, hey, go discover all the new physics and really help us figure out how the world works so you can solve all this other stuff. I want to be clear. I love all you guys at XAI and only want the best for you, but I'm going to continue to live. Post Elon attempts to give a speech on alignment involving a very small child, a child much smarter than you. The monologue rambles with no conclusion in sight. A pause. Yeah, will this be bad or good for humanity? He says the, you know, at least if it turns out to not be good, I'd like to be alive to see it happen. Oh yeah, they had a polymarket integration that was kind of interesting.

Yeah, it's interesting. Basically giving the model access to real time poly market data so that it can help make predictions and sort of add context around the market itself.

Yeah, that's interesting, Elon asking the real questions. You say that's a weird photo, but what is a weird photo? I still don't understand why we're looking at weird photos of XAI employees. But they were charming. They're calling it super grok. Crazy features, 16 bit microprocessors. What is. I don't even understand what this is. Oh, they. Yeah, they built like a game in Grok. They had demo of a video game generated by Super Grok. It's a Doom clone. Every time the PC shoots an enemy, floating text appears reading Groktom. Eland is fabricating timelines for product launches on the spot. The engineering, the engineer sitting next to him is looking at the floor, face impassive, nodding. It's a good model, sir. For real, though. Congratulations on the launch, guys.

It's a good model, sir.

I thought this post from the actual X AI engineer Eric Zellickman was funny. It was like AI model version numbers over time. Did you see this?

No.

So it's this chart of the version numbers over time. And you can see that Grok is versioning fastest because it's like, at this point, what else are we measuring? At least they're iterating on the version number effectively as opposed. And I guess this is a shot at OpenAI because. Because they launched 4.5 and then went to 4.1 and they're kind of like there's this big question about like when will GPT5 come? The expectations are so high for GPT5 and so they've obviously with the GROK teams that like hey, at least every three months we release a new full number. So I wonder the 5 is a number that really no one has gone for. And I wonder if GROK will do it first. Like if you draw the line on this, they certainly should do it in like three months. They should have Grok 5 and there's no reason that they shouldn't. But maybe there's.

It's very possible that Colossus is the key to getting to five.

Oh, the new data center. Yeah. Well they'll need LINEAR to plan that out. LINEAR is a purpose built tool for planning and building products. Meet the system for modern software development, streamline issues, projects and product roadmaps.

They LINEAR app need LINEAR badly so.

Hopefully they've gotten signed up near said GROK on Humanities Last Exam Grok 4 I'm not sure I buy even in the general case that there's a given humanities last exam number which implies you discover useful new physics. How would one make a benchmark of the proper shape for this? You'd have to have a validation set of questions which are outside the scope of what we currently are able to do. You could choose things on the edge of our knowledge distribution and then try and exclude. Yeah, it is interesting. Like if you are able to memorize every hard math problem, does that allow you to memorize to discover new math? Like it's sort of a prerequisite because you have to.

I think where I've imagined these discoveries coming from are having a single intelligence that has PhD level intelligence across like a single mind that has PhD level intelligence across every human domain. Right. And being able to combine ideas from different domains. Like historically a lot of innovation is just taking something from one field, bringing it over here, making some combination of it. I think Elon talks about the potential of discovering new physics, but again didn't spend a lot of time breaking down how that would actually happen. But world is unpredictable.

Yeah, it's interesting. People are really pushing this idea of like, okay, like, like we are accelerating like the, the ag, the ARC AGI leaderboard is accelerating. But I keep seeing this in and feeling deceleration. Like I am not feeling acceleration right now. Are you Tyler?

Yeah, I don't Know, I think generally I'm kind of like, not that interested in a lot of these kinds of benchmarks. Like, I think ARC AGI is more interesting. But just like the humanities last exam, the kind of general math physics knowledge, it doesn't seem to be that, like, it doesn't seem to line up with like, you see GPT 4.5 kind of does very poorly on these things, but like writing, it does really great. So like, I think I'm. I'm more like, if I were going to long, short on like, different benchmarks, like the usefulness of them. I think stuff like hle, I'm kind of short, long. I'm like, have you guys seen the Minecraft benchmark where it builds the two different. Okay, you basically two models build like a Minecraft. There's like a prompt. It's like, build a house.

Yeah.

Then you can choose.

And then it's like their rank models for the mines.

But who's. Who's grading that? The human.

It's a human who picks between them. It's kind of like a elo, but just like general kind of creative tasks.

Sure.

I think stuff like that. Aidan Bench is good.

Yeah.

I think even in the Grok launch there was the vendor bench.

Which one's Aiden Bench?

Aiden bench is Aiden McLaughlin's benchmark. It's just like. It's kind of hard to describe how it works exactly, but it's just various, like, creative tasks. How like, kind of novel its thinking is. The, like, style of its text.

Sure. Wait, is it just like. It's just like whichever one he likes the most. At the end of the day, he's the only grader.

No, no, There is like an objective, like, function.

Okay.

You can like, run it. It's not just like, okay, the idea.

That like, open again.

It will be funny. You know, there come. There's a period of life where your SAT score, like, matters a lot.

Totally.

And it says something about you. And then a decade later it's, you know, what you can do, what you have done starts to matter a lot more. And so I do think we'll reach that point where it's like, yes, you can. One shot. Every hard exam question there is that you can throw at it. But, like, what can you do for me?

Yeah, yeah, totally. And I think that's. I think that's why, like, the bigger question is almost like, you know, chatgpt daus and like, and like actual revenue. Revenue. The final bunch of app installs and stuff. Yeah, I mean, the, the revenue thing is Interesting because you wind up in B2B cloud world which is valuable, but it's maybe less, it's more competitive because it's more commoditized.

Well yeah, if you don't have a lot of leverage in the enterprise. If Azure is able to offer infinite models that are infinite, frontier models, open source models that are maybe just behind the frontier but great at certain tasks, the leverage isn't quite there. There will need to be another pretty significant leap. Until then, you know, anthropic being really good at CodeGen, there's leverage there. We saw this yesterday with Llama switching over to anthropic models internally and then you know, just having a consumer app with a lot of users also very valuable.

Yeah. The other interesting thing about the, the foundation model layer commoditizing and it becoming like cloud and if you have a model you'll just be like vended in as an API to anything else like the token factory is that the, the hyperscaler clouds are extremely profitable. Like even though aws, GCP and Azure are all somewhat directly competitive and they're somewhat perfect substitutes for each other, they have not driven prices to zero. Such in the way airlines are like deeply unprofitable like AWS and Google Cloud are both profitable.

Yeah, or you look in other commodity sectors like oil and gas and I.

Don'T know if that's just because there's lock in. I'm not exactly sure, but there's something about where maybe the counterintuitive take is that yes, they do commoditize and there are a few major foundation models that are frontier and they all are roughly the same price but they all have decent lock in with their customers to the point where they're still able to extract some level of profit or they're just creating so much value that even if they're taking like a small marginal slice on top of the cost to run, that they're creating so much value that they still have 50% margins or something like that. Because like, I mean this was the story of aws, like no one knew how much money it was making. And then, and then they had to break out the financials in one of Amazon's earnings reports and it was like the AWS IPO as Ben Thompson put it. Anyway, before we get to the next story, let's tell you about numerous HQ sales tax on autopilot spend less than 5 minutes per month on sales tax compliance. So the big news is that the third browser war has begun. Google stock has dropped on the news that OpenAI is planning to launch a Google Chrome competitor within just weeks. And this is very interesting timing because.

It'S time to browse.

Yeah, time to browse. Certainly makes sense to become deeper, more deeply integrated into the user's life. Makes a ton of sense. There's a ton of benefits that come from having a web browser. What was interesting is we can go into what Google actually Launched or what OpenAI is talking about launching. But this news, this scoop leaked the same day that Arvind from Perplexity announced that they're finally releasing their next big product after launching Perplexity Comet, the browser that's designed to be your thought partner and assistant for every aspect of your digital life, work in personal. And so Perplexity launched this on June 9, and then OpenAI, the scoop, goes out via Reuters the same day. And so this feels like very much like. Let's not let Perplexity get a bunch of attention and drive a bunch of people to start daily driving Comet, the browser. Because even though we're not ready to launch our competitor, we want.

I mean, Arvind was on the show talking about Comet, but over a month ago, he said it was really important to the business. This was a big bet that they're making. And I'm sure both companies were racing to be the first to launch. But dia, the browser from the browser company, also launched out of. Or they're still in beta, but they launched like a month ago or something like that. So this is, you know, you're not going to be the first.

Oh, they launched a month ago with the DIA browser because I saw Riley Brown also posted the cursor for web browser and DIA browser, and I thought DIA browser launched that same day, but I guess it had launched earlier.

Yeah. So anybody that was an ARC user can download DIA today and chat with their tabs. But interestingly enough, Perplexity's browser and OpenAI's browser are both built on Chromium, the same open source project that underpins Google Chrome and Microsoft Edge. Yeah, so the cool thing here, that means that they're compatible with existing Chrome extensions.

Oh, interesting. Okay, that's cool. Yeah. I want to talk to more people who were active in tech during the earlier browser wars. The first browser war was Netscape Navigator versus Microsoft Internet Explorer. This is in the middle mid-90s, early 2000s. Netscape was super dominant. And everyone loved Netscape. It was originally the Mosaic browser. This is the Marc Andreessen project then. But Microsoft bundled Internet Explorer with Windows 95, and the distribution was so Powerful that Internet Explorer actually wound up winning and became really, really dominant. But then there was this lawsuit and went back and forth. But then basically by the early 2000s, Internet Explorer had over 90% market share. But then they got kind of lazy and stagnant, apparently. And I mean, I'm not exactly sure what happened, but there's a lot more competition. So Firefox, which was, I believe, like a spin out of Netscape or kind of like some of the same heritage there, began getting traction. And then Google chrome launched in 2008 and leapfrogged everyone. And Google Chrome was really focused on, like, speed. It was the fastest browser, and they, they did a whole bunch of work to optimize JavaScript so the pages would just load faster and run better on pretty much every computer that you had. And so, and then they had the open source project with Chromium, and so they were able to kind of standardize the entire industry. And so everyone's always been trying to draw analogies between, like, the browser wars and the LLM wars. And, like, what's the role of open source in that? Like, is open source a strategy to wind up maintaining your. Your dominance? How much does distribution matter? Like, Chrome was probably pretty easy to distribute because every single person was visiting Google just every day searching. And so you just put this bar, hey, want to switch to the faster browser? And people just do it. Because you can have basically like, you know, billions of ad impressions on your product every day. Will be interesting to see if ChatGPT can get people to download their own browser on desktop. I mean, I'm using ChatGPT on desktop in Chrome all the time.

Which ChatGPT model would you want to use as a default search engine?

That's the hard part because I always run into this problem where it defaults to O3 Pro, but that takes 10 minutes. And so then I have to go to 4.0. And then if I'm in an O3 Pro flow and I'm talking to O3 Pro and I let it cook for 10 minutes, it gave me a great answer. But then I want to just be like, okay, just like, clean this up a little bit or summarize this or do some bullet points, I want 4.0 to do that. So I have to switch over. So I don't know. I would imagine I'd go 4.0 as the default because I want speed. But even 4.0 could probably be faster before it truly replaces Google's.

Very fast. They've spent a very long time being fast.

Yeah, and I could imagine them doing a similar project to, I believe it was like the V8 JavaScript engine. They sent this team out to, I want to say like Iceland or something. They basically sent like a bunch of engineers to like an off site and they were like, just go optimize JavaScript for a month. Just go focus on this for a month or months and come back when it's done. You have no other responsibilities than just optimizing this compiler. They came back with the VA JavaScript engine. It created this whole node js. Boom. People were running JavaScript on the server. Then I could see Google doing something similar where they're like, ok, we have Gemini. It's good at looking stuff up. It's a good knowledge retrieval engine. Go figure out how to make it load all the tokens for the full response in 100 milliseconds. And that would be very, very cool. And I wonder if that's like a uniquely Google advantage. Tyler, you look something up?

Yeah, it was in, it was in Denmark.

Denmark, Okay. I was close. I was close. Yeah, I wasn't sure it was Finland or Iceland and Denmark.

Yeah. The interesting thing here, I'm realizing that tabs are definitely a light lock in to browser. It's not just the default, but if you have six to ten tabs that you just had open for a really long time and they're like from a bunch of different things and you can't exactly remember what they were if you had to list them all off. But you know, I personally end up using tabs as like somewhat of a to do list. And so if you're spinning up a new browser and you don't have your tabs, it's like, oh, well, do I want to just like get rid of my tab stack? I have a bunch of tabs that just have stayed there for years. And it's basically like a mini operating system.

Right.

With different apps that might be a Google Sheet or something else.

Yeah, no, I know what I mean.

So there's very real lock in. I could bring all those tabs over, but I have to then log in to a bunch of different services. And so it's really, really hard to actually win here.

I wonder if anyone's using, you know, in Google Chrome you can actually change those default search bar to, you know, when you type in the search bar and if you just type words, it just Google searches it. You can change that to search ChatGPT. Yeah, you can pass in a query parameter and it can just do that. But I haven't heard of anyone actually doing that. And I used to have, I used to be such a power user of Chrome. I used to have different code words basically. So if I, if I typed like I space and then a query, it would go to IMDb and search that specifically. So you could, you could have Chrome like route to any specific search. Any. You could press like Y space and it would search Yelp or, you know, anything else. But I don't know if people are, I don't know if people are doing that with Google. With ChatGPT. I think people mostly just like control command T and then hang out in ChatGPT.

Well, we'll have to ask Chris in 15 minutes about get an update on the browser wars because he was an early investor in.

I know one of those tabs that you have pinned right now.

What's that? Adeo, of course.

Customer Relationship Magic. Adeo is the AI native CRM that builds scales and grows your company to the next level. You can get started for free.

I've had Adeo open for thousands of hours in a row at this point.

Yeah. So Signal kind of breaks it down with the OpenAI launching the web browser. He says this is the oldest play in tech. Find product market. Fit with a single killer use case, then vertically integrate and horizontally expand until you control the interface layer itself. App platform. Once you own the interface, you own the defaults. Welcome to the next generation of browser wars. Yeah, what's interesting is there like Sam Altman at OpenAI and just the fact that OpenAI is a company, like there is kind of a mandate to like vertically and horizontally integrate, figure out code, figure out research, figure out devices. But every company wants to do everything. But then sometimes they run up against barriers. Like there was a time when Google was like, we want to win social networking and we want to beat Facebook and we're going to launch a direct Facebook competitor. And they did and it didn't go well. And then they shelved it and then they wound up producing trillions of dollars in market cap. Just doing the thing that they do. Great. And so the question is like the surface area of OpenAI, they have to explore, they have to experiment. It would be stupid not to see if they could get a browser and a device and chip and a nuclear reactor and everything. And sand, get the, get the sand, get everything. But, but there's no, there's no guarantee that they will win the entire vertical stack and they will be the one company.

Right. I think my question is, are these going to be like, is OpenAI's browser going to be an entirely new app other than their existing mobile app, is it? Or their desktop app?

I. Yeah, that is interesting because if.

They have to get people to redownload a separate app, then that's like an entirely. They have a good fly, they have.

A bunch of impressions. Wouldn't just evolve the apps that they.

Already have in production too.

Why are we doing that?

Is planning to release this as a new standalone app or it will be in the Perplexity mobile app?

Yeah. I mean, no, I think Comet's like its own thing because we were looking to download it and we need a code and you can't just get it if you're just on Perplexity. But I don't know. All I know is that you should go to Fin AI, the number one AI agent for customer service. Number one in performance benchmarks, number one in competitive bake offs, number one ranking on G2. So Arvind breaks down his philosophy of Comet, the browser that he's dropping from Perplexity. He says you can either keep waiting for connectors and MCP servers for bringing in context from third party apps, or you can just download and use Comment and let the agent take care of browsing your tabs and pulling relevant info. It's a much cleaner way to make agents work. So that is interesting. So I wonder how much like Puppeteering will be in this because ChatGPT and OpenAI have operator that operates a chromium front like a headless web browser basically. But you can actually see it working and it's clicking things. And so if they're like. There's also the value of like the training data. If you're getting people using all these websites, you have all this training data of like, okay, they clicked on the blue button, they clicked on the green button. They saw this, they entered the height. This is how they dealt with this form. This is how they dealt with that form. And so that feels like very, very valuable data if you can get it. So it's probably worth duking it out even if it doesn't. Even if it takes a long time for sure. I do wonder where, where else they will, where they will plug in. Like Cluli operates at like a higher level of abstraction with like the screen scraping. And I wonder if we'll hear rumbles about either Perplexity or OpenAI thinking about like moving up the stack to that level. I'm not exactly sure. Anyway, Dan Ivey says we believe Apple needs to acquire Perplexity for AI capabilities. Likely $30 billion range would be a no brainer deal given treadmill AI approach in Cupertino. Perplexity would be a game changer on the AI front and rival ChatGPT given the scale and scope of Apple's ecosystem. So people have been talking about this for a while. It feels like there were talks and then they kind of stalled out and.

Arvind is 300 and I think this would be a 375x revenue multiple.

Wow.

I mean the product sense is good. You use the product and like there's like Apple hasn't been able to deliver on the product side. They have the distribution but they haven't been able to get things.

We talked about this before though. The most expensive acquisition Apple has ever made was Beats by Dre for $3 billion which was a 3x revenue multiple.

Be a huge shift and huge ship.

I don't know that I think that Apple is embarrassed right now and feels a lot of pressure to deliver. I don't know if they're at the point where they would pay $30 billion just yet.

And even then it's like hard to.

Integrate or even 14 or whatever their last private valuation was.

Yeah. And the big question for me was like Perplexity is built on a lot of different clouds, a lot of different tools, a lot of different models. Is Apple cool with that stack? Yeah, because if all of a sudden.

Or do they want to just go direct to Anthropic or OpenAI which they are in conversation with and every once in a while these, these like scoops pop up around perplexity and Apple conversations and it's hard to read into that. Are these like, is this like rumor mill? Like what's driving that rumor mill?

Yeah. Well pull up the Mag 7 chart. I want to see where Apple and Google are sitting today. Apple at 3.2 trillion, Google at 2.1 trillion and Nvidia is holding strong at 4 trillion. Not bad. Yeah, I mean 1% of market cap. They're at 3.2 trillion. $30 billion acquisition to be, you know, to have a, have an AI product that clearly has a good roadmap. Is that crazy? I don't know. Well, if you're making bets on any of the Mag 7, do it on public.com investing. For those who take it seriously. They have multi asset investing, industry leading yields and they're trusted by millions folks. So in other Apple news, they're preparing to launch the new version of the Apple Vision Pro. They're just doing a slight iteration on the chip. They're moving to the M4 chip and they're launching a new strap which was something people were complaining about because the weight maybe it'll be better distributed. People were switching out for the pro strap like earlier.

It's so funny. So last week Elon Musk announced the America Party or I guess it came out on Monday. Stock dropped from $312 a share all the way down to $291 a share. That is when Dave Portnoy I think market bot he was saying Davey Day Trader is back. If that's not a top signal, I don't know it is. But he was market buying like 10 million of Tesla being like I just think it's going to go back up to where it was. It's just been climbing since then. It's back up to $308 a share. Almost, almost recovered. It's up 4.2% today. So on brand for Elon and basically gonna. It looks like it'll just recover the price prior to the America Party. And Dave Fortnite he literally was, he basically his thesis was like I think it's gonna go back up to where it was in about two weeks and I'm going to make 10% and I'm going to make a million dollars.

That was your thesis on Nvidia. You were like wait, it's like down because of Deepseek. Maybe it'll go back up. It was like the most basic analysis and it worked perfectly. It's just fascinating. Maybe that's broadly a top signal. Just the idea of simple analyses and not necessarily needing deep insight to call the market is good. I don't know. Who knows? Anyway, this Apple story is from Mark Gurman of course the master of scoops and Bloomberg.

He's on his fourth or fifth this.

He'S on his animal tear. I mean this one is a little bit minor. You know they're going to include a faster processor and components that can better run AI stuff. And so not that Apple has any crazy AI stuff that they really want to run in there. I don't think that's a major differentiator. I've been thinking about how is AI a key unlock for VR and like I don't think so at all. I think it's much more about the content and the use.

Entertainment.

Entertainment. I think it's a replacement for a TV for to start and they need to just make it dead simple to use.

I don't know. We got a demo of VR product a while back and it had some very cool native AI features, there's something there, but Apple's products doesn't feel like it's ready for just wearing while you're making dinner.

Yeah. So that version, the one that significantly reduces the weight of the headset, they're planning to launch that redesigned model for 2027, which feels so far away. I know it's only a year and a half, probably the end of 2027, but. So maybe we're talking two years. But that in. In the. In the AI race, where we're like, yeah, AGI tomorrow, AGI next week. What's going to happen next month?

I'm like, major news.

We can't ship a better. We can't shoot. We can't slim down the headset and take off the screen and create a lighter materials like this this month. Like, let's do it. But hardware is hard and, you know, the stuff takes time. So good luck to them. I'm excited for it. I'm very excited for the next quest.

Do you still have a Vision Pro?

I don't. I had it for a month. I took it back because I just wasn't using it that much. It was like, heavy and I couldn't find. And it had a bunch of things that, like, you had to do, like these crazy workarounds. Like, I wanted just like an HDMI cable that I could plug into it and then just be like, okay, my PS5 is in VR now. And I couldn't do that. It was like, you had to like, pull the. Pull the screen into the Mac and then screen share it in. There'd be latency. It was ridiculous.

The thing, the use case that I still see is people using it on planes. Yeah, but I just.

We gotta check in with Tyler. Give me your. How many times have you thrown on the VR headset in the last week? Did you play it last night? Break it down. Have you churned? Is it collecting dust?

No, it's. I've been playing a lot of Call.

Of Duty in the VR headset.

Yeah.

Okay.

It's a lot of fun.

There's like, no latency. I'm kind of surprised.

Really. No latency and you're doing.

The Cloud is really slow.

You're doing the Cloud?

Yeah.

Okay.

Online. It's multiplayer.

Multiplayer. So you play multiplayer and you play like the latest and greatest Call of Duty? Basically, yeah. Okay.

Black Ops, like six, I think.

Cool. So you have a controller and it's a big screen on the wall and you just chill there. But walk me through it is like 30 minutes a day.

Yeah.

Probably like 30, 45 minutes a day.

You're fired.

No, this is not a gotcha. This is. This is true research. So, yeah, I mean, I honestly think that the Quest Xbox, the Meta Xbox Quest or whatever. I forget the name, but, like, that. I think that's more. That's better to use than, like, a processor bump on the Vision Pro. Just, like, deeper integration so that you don't have. So you can just throw it on. What's the actual time to. You know, if you want to turn it on, throw it on. Plus, get playing, get into a lobby, actually get your first kill. Is that one minute?

No, it's like 30 seconds.

Maybe 30 seconds.

It's no, maybe like a minute.

Okay.

It's not, like, noticeably slow.

You're logged in. You don't need passwords or anything.

Yeah, it's not.

It's not like, a hassle. Okay, that's cool.

And I put the screen. It's funny. Like, I. I have a TV in my apartment, but I just put the screen right where the TV is. It's like the perfect spot on the couch.

It's a nice black square. Yeah, yeah, yeah.

I'm gonna have to get this back for you now. This sounds amazing. Now I don't have any time to do this, but I feel like the next. What's on the feature roadmap that you would want to see? Like, Apple is bumping the neural engine and trying to upgrade the chip. I'm not sure that that's the problem. Problem with the Vision Pro. What would you like to see out of the Quest four, I guess is the next one that's coming?

Yeah, I think the main thing. So I've tried the Vision Pro, and basically, I mean, the visuals are just, like, vastly superior. It looks so much better than even the. So it's the Quest Better Screen 3 Xbox Edition. That's what I have.

Yeah.

And the screen is just way better.

I think that's.

I would say that's the main thing.

So if they can just go find the supplier. If Meta can just go find the supplier for the Vision Pro screen and put it in the quest 4, you'd buy it yourself?

Depends on how much it is.

I'm kind of broke.

Okay.

I would definitely be inclined to not have to.

Not if you drop out and go.

Full speedrun all of Halo in order to potentially win one.

I would do that.

You would do that. You would do a very difficult challenge in order to potentially win one. Because you would. You would want it.

Yeah.

Okay.

No, I think that's that's. I would think that's definitely the main.

Thing is, are there any other, are there any other nice to haves that you think might, might shift people?

I don't know. I mean, the, it's, it's very light. It's way lighter than the, the Vision Pro.

Yeah, but still, I mean, I feel like light is very relative. Like, it's light to the point where you can do 30 minutes or an hour. I, you probably can't do like a full day or like, or like four.

Hours or any kind of like, workout stuff. I think I definitely would not do that.

Totally. Totally. But, but what I'm saying, you're not training. I knew, I knew guys, I knew guys at UCLA who, I didn't go there, but like, I. Friends that went there and they were so obsessed with Call of Duty that they would take a bunch of stimulants when the new Call of Duty came out and play it for 24 hours straight to get the max prestige. Because there was so many addicted to Call of Duty that they would just stay up all night chugging energy drinks and, and just to, just to beat that. And I just don't think you could do that in VR. I think after like two or three hours right now it's like too much and you have to take it off, get sweaty and tired. But so, so I feel like, I feel like screen first, then probably even a little bit lighter, a little bit more comfortable, and then just drop the price as low as possible. Because if the next one was 100 bucks, you'd probably buy it, right?

Yeah.

And I think stuff like, like maybe I'd want another screen, just this monitor. But that's just an issue with the, with like the visuals. It's a screen.

Yeah, it's, it's got to be, it's got to be competitively priced with the TV. And the TVs are so cheap now that you got to just be like, yeah, I'm just picking one up. Or the price of AirPods or the price of, you know, it's got to be down in low, low hundreds of dollars to really ramp that up. But I don't know, it'll be interesting. Anyway, our first guest is here. Let's tell you about A.D. quick, really quickly. Out of home advertising made easy and measurable. Say goodbye to the headaches of out of home advertising. Only Adquick combines technology out of home expertise and data to enable efficient, seamless ad buying across the globe. And we will welcome Chris pike to the show. Welcome back. Chris, fantastic to have you on the show. Thanks so much for taking the time.

Great to see you. Last time we got cut off. Great to be back in the temple. Last time we got cut off we were having a jump and I was like, I wish we had another hour. So at least we have another 30 minutes here. Yeah, let's get into it.

First off, what's top of mind for you? Have you been tracking anything in the news that's that, that's kind of updated your thinking? We were digging into Grok 4 and seeing is this an update to, you know, agent timelines? It seems pretty great in the benchmarks, but is there anything else in the last week that's been like, I can't get enough of the story? Just in your world.

It's a great question. I feel like every week is a total blur.

Yeah.

It seems like we're all waiting for not just these foundation models to come up, but like the next open source models, the big open source models to come out. I think that that's super interesting to me. The proprietary foundation models obviously are the frontier of research, but they're relatively inaccessible from a technology perspective because they're fundamentally rent seeking. You can't run them on your own hardware. It's significantly less accessible. And so I'm kind of waiting for the next generation of open Source models.

Yeah, one maybe underrated or under analyzed. Grok4 thing that happened last night was I don't know if either of you saw this, but they did this voice demo and they were really pushing the accents really far. I don't know if any of you saw this dialer. Did you see this?

Yeah.

And there's the whispering.

The whispering.

And so asmr.

Wait, did you think it was uncanny valley, Tyler?

I felt very uncomfortable.

Yeah. But at the same time I think it, I think it's a path where we're in the uncanny valley, just like we were with like six finger hands and stuff. And when they actually sort out the accents, the whispering, the intonation, the cadence, it's going to become a much more addictive companion potentially. So I want to, I want to bridge to your piece and talk about the different use cases that you see might people might kind of flow into with these like chat companions. Because you mapped out way more than just the normal take, in my opinion.

Yeah, well, so I guess it's worth asking ourselves where, where do we want other humans to exist and where will we accept like substitutes? I think the last time we were talking about it, you know, if.

If.

You can imagine a situation where humans are getting in your way of doing something, then you kind of. You hate them, right? Imagine you're in. In traffic, in gridlock traffic. That's the most misanthropic you could possibly be. You're like, if none of you existed, I could just get where I wanted to go without you being here. But then there are just total other times where we would refuse to accept anything other than humans as that thing. I know that it's very popular to talk about AI companionship. And I would never say that people who get a lot of value from AI companions, that value isn't real. But at the same time, I think that when it comes to the allocation of. Let's call it allocation of leisure hours, we really care about other people. Whether it's like, I think last time I was talking about, like, going to fine dining or reality tv, I think I mentioned that, like, we have the chess software that is way better than any human will ever be. And it's not entertaining to us. It's not entertaining to us because it's sort of like a. There's no drama, there's no emotion. Exactly.

Exactly.

And so.

Sorry, go ahead. I'm just thinking about, like, the shape of companionship because in your most recent piece, you call out, like, the imaginary friends that kids have. Calvin and Hobbes, Toy Story. These stories resonate because they poignantly depict how colorful, whimsical placeholders of our childhood slowly fade as society offers real alternatives. And I'm just thinking about, like, kids love imaginary friends, but they also love, like, multiple IPs, essentially. They like Batman and then they also like combination. And so I'm wondering, like, there's been this narrative for a few years in AI of like, don't build a GPT wrapper because you're gonna get. You're gonna get rolled. There's going to be immense concentration of value and there will actually be no middle class in this ecosystem. And I'm wondering, because there's this company, Tolan, that's kind of imaginary friend AI driven. And I'm wondering how we might actually. Is it possible that we're heading towards something where people are essentially developing new ip? And yes, there's still a power law in the companionship market, but these models and these products are, like, much more opinionated to the point where there actually isn't a one product to rule them all. And there's a variety of products that fit into different holes just for companionship, but also even just within the imaginary friend hole, there's 25 different options and yeah, there's one that's popular but then there's one that's 1/10 as popular, 1/100 pop. This is popular. But yeah, react to that.

Yeah, it's a really interesting. Let me first go back to this wrapper concept please. And I think it's really important to distinguish when wrapper strategies work and when they don't work. I would argue that the wrapper strategy works the best when the underlying infrastructure is purely commoditized. Like you can choose across many different options. Where the wrapper strategy doesn't work is if you're being, if you're, if you're basically building on top of a monopoly and that underlying landlord is basically just going to be increasingly rent seeking and squeeze you out of all margin. So sort of embedded in the wrapper strategy is the assumption that over time you're going to be able to distribute your product on top of increasingly commoditized infrastructure. So for example, like Snowflake, Snowflake launched actually just, just on Amazon and then over time distributed its product across Azure and Google and actually in doing so was able to expand the margin capture of its own product because it had a. Their vendors were competing to be their underlying customer base. Going back to this open source notion, this is actually why I am so interested in open source. The more that we have competitive fungible models, the more that the application layer on top can really flourish, the more that we'll see distinct unique applications and, and until we start to maybe S curve near the top of, of of the frontier, I'm sure you guys are familiar with the really popular essay the Bitter Lesson, which is basically the, you know, no amount of fine tuning or no amount of specific framing is actually going to out compete just the fundamental advances when it comes to more games with more compute. But if we start s curving, if we start, if these scaling laws break, which it seems like they are breaking on pre training and test time, compute and things like that, all of a sudden you have largely fungible similar capabilities at that commoditized layer and then we can really start to see the application layer flourish.

Yeah, my interpretation of the bitter lesson right now is that the impact of AI should be tracked less in benchmarks and less in individual tests of one model and more in the actual volume of inference tokens being generated by humanity. And it's fine that we don't have one central AI doing all of the work. If we just give everyone an AI copilot for every single Task, they all get better and we'll build more data centers to inference more and eventually that will compound and compound and compound until the overall impact of AI is remarkable and unmistakable in the same way the Internet has. But it won't be this like all of a sudden we unlock this one incredible algorithm.

Yeah, maybe.

So.

I'm going to try and map a comparison that's probably wrong for any number of reasons. Let's say we ported the bitter lesson to the rollout of PCs, right? I think there have been a lot of comparisons of like, AI feels like a new computer. So like, let's map the bitter lesson to PCs. Maybe the analog would be, say, don't work on building software that optimizes for like a computer speed of like 50 megahertz, because the computer that comes out that's like 100 megahertz or 200 megahertz or you know, 1 gigahertz is, is going to be, you know, is just going to blow out whatever software optimization you've achieved at that compute stop. And so I think about it a lot like that. Now I think this really begs the question of, okay, well, what happens when the vast majority of our use cases are satisfied by the compute threshold? I feel like everybody's like, who needs the next gen version of this? Because largely all of the applications that you use or you can run work. And so I, it's very clear that we're in this like rising part of the S curve. But when, when the S curve starts to taper off, that's really when it comes to the question of, okay, well, how do we think about this? The value delivery that sits on top of the underlying rent capture from, from these foundational models. You know, you could, you could think.

About it also.

Like video game consoles, the, the amount of, the amount of creativity that video game developers are limited by is actually how advanced the video game consoles are and also what the install base of the video game consoles are. And so I think one of the challenges right now is we have so few developers of AI applications.

You can.

Kind of count them on maybe a few sets of pants, right? Which is insane.

I feel like there's like thousands of startups that count in like the AI software developer world. I see market maps every single day. I'm Sorry, there's like 10 in every B2B category.

Okay, so maybe let's split the world between enterprise use cases and consumer use cases.

Oh, sure.

So enterprise. I would, yes, 1,000%. There's a lot of companies that are, it's It's a, it's a blue ocean sprint to vertical value delivery within, in different sectors. Because you're largely swapping. You are the spend, the addressable spend is, is opex, which is insane. Like, that's crazy. Your revenue opportunity is just like headcount spend.

Yeah.

And so that's for sure.

But also when it comes to tool spend, like, it's all the opex, like to your point, I mean, I guess not like real estate or rent or something, but basically everything else.

Yes.

Which is by far the biggest cost center for any company.

Yeah. Of course.

Anyone who's ever run payroll or anyone who's ever scaled a company is like, man, like, humans are expensive.

Yeah.

Humans are so expensive.

Yeah.

And I mean I, this is why I feel like everybody says that AI is the best candidate or the best argument that UBI is coming.

Sure, sure. Yeah.

Jordy.

I think the picks and shovels meme became too dominant and there were so many over the last decade or so. There were just so many amazing outcomes of people building infrastructure and like very visible outcomes. And it became cool to build infrastructure. Right. Like the Collison brothers made building infra cool. And you have like Parker Conrad is like a folk hero.

Sure.

Right. And like rippling. It's like, like. And you have the like ramp is a good example of this. Corporate spend management should not be cool. They've built a, you know, really cool culture around it. And there's this other, there's this like weird kind of pervasive meme of like, you know, you have two years to escape like the permanent underclass. And so I think people are like, well, I'm not gonna just build something weird and fun. I'm gonna build enterprise SaaS so I can make, you know, so I can escape the permanent underclass. And so I don't think there's been enough weird, fun attempts from people. Like Toland's, the one you brought up earlier is cool.

Like it's pretty rare.

Not the most rational thing to say, you know, if you just want to build a big business to be like, I'm going to build a little alien AI friend. But like, clearly there's demand for that. And we were talking with Scott Belsky yesterday around just wanting like new fun, weird consumer use cases. And I feel like, I think what you're getting at is like that whole area is like relatively under explored to date where like we've had a bunch of. We had two browser announcements yesterday and they're both built on chromium. And that's exciting and cool and we should talk about the potential for new browser wars. But I think the number of people that are saying yeah, you could call it a wrapper but I'm actually trying to create something entirely novel. Like the example we gave yesterday is like a dating app based on a digital twin that is just constantly dating other digital twins. And you know, I haven't seen. I'm sure somebody's working on that. I haven't seen it yet but there's like any pick any popular consumer app category and there's probably a way to entirely rethink it with this sort of LLM as a new computer at the.

Core of that consumer just seems so much more risky because it's like it's either a billion dollar outcome or zero. Whereas in enterprise it feels like well there's no way it's going to be a zero. It's going to be a $10 million outcome or $100 million outcome or a billion dollar outcome but it's not going to be. It's not going to be a.

So then you have 10 companies that each have 10, 50 million doing the.

Same thing where like they were trying to go for the hits driven business with the game didn't work and then they went into SaaS and it worked anyway.

I think that the additional challenge with consumer is inference is not free right now somebody has to pay for the inference bill. And so until you can run inference on device every developer is doing the mental math in their head of like how do I like if any one of your users can run you out of house and home if they abuse your service? How do you build on top of that?

Yeah, one of the OpenAI researchers I think in Zurich that's going to meta posted like I didn't realize that I had this thing running. It was like $150 a day just like every single day. Luckily I think he'll afford it.

You can afford to take that.

I did, I did. Before we fully leave like kind of consumer. What do you think? Do you think that AI companions and alums broadly present a real threat to traditional social media? Like the idea of a companion it A lot of people in our world are using these tools like very functionally like for doing research or getting answers or understanding topics. But a lot of people are using them as companions and that is somewhat of a social entertainment experience. And you see these charts ticking up of kind of user minutes.

So ChatGPT went from about five user minutes per day to over 30. There's no around 30 I think it's a 28, 29 over the last six months. There's no blip on any of the other social network. They're not declining yet. But personally I'm finding that if I'm doing research on a topic, I used to go to YouTube, I used to go to Instagram, and just people famously search TikTok for answers to things. And some of that is shifting over.

Yeah. And if you think that an analog sort of anecdotal sort of experience for me is when do I use social media the least is when I'm with my family, which is like companionship, it's this social time. Or when I'm with friends, like at dinner, hanging out, it's like rude to be, you know, using, you know, there's no point to go hang out with a friend and then use Instagram the whole time.

Totally.

I think I subscribe to this sort of cutting of our time as like, we're either allocating labor hours or we're allocating leisure hours. So we're either trying to be productive or we're trying to enjoy ourselves. And so I would say that all leisure allocation effectively competes against each other. I think maybe it was like the Netflix CEO that said that they weren't competing with hbo, they were competing with Fortnite. And that's largely true. We only have 24 hours in a day. We only have a finite amount of leisure hours. So if I'm allocating one hour to this leisure activity, if I'm watching an episode of Love island or whatever, that's time that I'm not going to be able to allocate to a different kind of leisure activity. So to that end, I would absolutely agree that AI companions, almost certainly firmly in the bucket of leisure, more consumption of leisure equals less consumption of other leisure activities. It's really zero sum. The only thing that is going to make it non zero sum is a fundamental advance on productivity that allows the leisure pie to be even larger. So maybe we definitely have more leisure hours as humanity now than we've ever had in the history of humanity.

Let's give it up for leisure hours.

Like proto human zero leisure hours. Now we just like amazing, amazing leisure hours. When it comes to labor hour allocation or research or utility, I would say that doesn't necessarily encroach against time on social media. And I think that all social media, whether it's YouTube or TikTok or Twitter, they care less about helping people get work done and they care much, much more about absorbing as much of attention as possible. Which is why the algorithms are so insidious of serving you exactly the kind of saccharine thing that you want to consume next.

I wonder if there will be an incentive. I mean there will obviously be incentive, but I wonder how it will play out in the LLM chatbot interface. Because right now OpenAI and basically anyone who has a dominant consumer AI app is probably seeing user minutes increase just naturally without putting in like growth hacks or retention loops or. But you could imagine world where to get from 30 minutes to 60 minutes the LLM has to not just give you the response, but surface. Hey, would you like to follow up and learn more about this? Click these buttons. That's already kind of happening.

Yeah. It starts surfacing you stories that it knows you're interested in by what you've.

You've searched for a bunch of stuff. Let's give you a new breakdown. Kind of pre populating a deep research report.

It is funny that the push notification hasn't really quite hit.

It hasn't.

That's right, chat apps. And it undoubtedly will.

I had this thesis that push would be very important versus like pull like you have to go to ChatGPT and ask it for something and someone is going to solve kind of like it's almost like an AI driven newsletter or something where it understands what you're interested in and then generates the report before you can even ask it. Because it knows that, you know, if Ferrari drops a new car, I'm going to want a table of all of the details because I like consuming information that way in addition to just hearing commentary and watching the Doug Demuro video about it. But, but OpenAI could pre populate that and just send that to me. But I don't know. How do you. How do you think that's going to evolve?

Question for you guys. Do you think we're going to pay for AI services forever?

I've been asking, I've been asking that.

I don't know. I mean I think, I think, I think that.

I think the better. The more important question is will the. Will the. Will the average American pay for an AI like an LLM? And then will they pay for multiple. Like the comp for. This is in streaming where Americans.

I was about to say Netflix Americans.

A lot of Americans pay for multiple streaming services, but they are incredibly ruthless about canceling them on average. Like a lot of. I'm sure a lot of people listening to this have like had some streaming service billing them monthly, you know, for years they haven't even watched. But but the average American is like, I'm not getting a lot of value out of HBO right now. I'm going to cancel even though they might sign up again in like four months next time there's a hit show that they're going to watch. And I think that, I think that there's not a. Right now there's this incredible demand for what's new and what's best. Right. Like GROK will drive a lot of signups today because it is. The Grok 4 Heavy is like a meaningful advancement. But I went and I haven't, you know, we've been busy this morning. I haven't had a chance to sign up and play around with it yet. And I was getting plenty of value from Grok3. Like I was able to just search.

Yeah, I was asking Grok3 about Grok4 as well. And it was actually doing a pretty good job, which is fun.

And I'm not yet. So I'm not, I'm not. I guess I probably get it through X Premium, but so I have some.

I have some data here. Netflix made 39 billion last year. They're on track for 44 billion this year. 1.8 billion of that was ad revenue last year. It's estimated to be around 4 billion this year. So their ad revenues doubling while their subscription revenue is growing by 5%. And so I think one my, my takeaway for ChatGPT would be, I would imagine that the ChatGPT paid subscriptions follow an S curve and we get to something where we see OpenAI making. I don't know if it'll be 10 billion or 40 billion, but they will soak up a ton of subscription demand for ad free Frontier models, the most advanced, the most expensive stuff. And then ads will eventually become the dominant revenue driver. But I feel like the subscription revenue will be a really hard tap to turn off just from, hey, it's people are paying and you know, it's a lot of money and we don't want.

It to go out.

Even in the enterprise it does. My bet is that we go, you know, continue down this trend towards paying for outcomes. Because a lot of people would just say, well, I don't want a subscription for this service because I only use it every now and then and when I get value from it, I'm happy to pay for it. I think a question to ask is, would you pay $20 a month for Instagram today to not have ads? And I would actually have to think about that for a while because like once a month I Get an ad for something that looks interesting and I discover a product that I wouldn't have otherwise discovered and I buy it and sometimes it's great. So do I want to just completely eliminate that and rely entirely on random organic or do I actually like that this ad platform is spending a bunch of time and energy trying to serve me the next product that I'm going to like, which is like actually a service? And like it's, it's, it's not a bad trade at all.

Totally. I think we, a lot of the way that we vote with our feet and what we vote with our wallets is that we, we're super happy to pay with our time and our attention if given the option. Yeah, like actually like the vast majority of people won't. Not only that, but people are more than happy to give up their attention and privacy for that matter, if it can save them money or instead of paying for something. One thought experiment I like to think about is just how much people will trade privacy for value. Imagine a checkout flow where you could get $5 off if you enter your Social Security number. Like how many people do you think would enter their Social Security number?

Everyone.

Or like.

An insane amount of people. And so people really don't value their privacy as much as I think maybe we say people value their privacy. And in aggregate, obviously that data is super monetizable. It's interesting. Search obviously is the big prize I think for AI and what I mean by search, it's intent. If you're the arbiter, you control this fire hose of intent. You can benefit by metering it out and having people bid for that intent. Obviously Google, maybe Google's like the best business, the best business model maybe ever invented. It's kind of insane. What's interesting is the most valuable searches maybe like not what people think. And certain kinds of searches are totally worthless.

So.

Knowledge based search, like fact based search, things like.

What is the market cap of this company?

Right, right. It's like what's the market cap of this, you know, who, who, who won this game or like, you know, who, who was president in like, you know, 1936.

Yeah, they're dead, they're dead ends. You get the fact that no value.

Yeah, zero value actually like there was, there was a. Google got subpoenaed and had actually share some, some documentation of like what their most profitable keyword. Super interesting. I suggest people to go check it out. The number one most profitable search for Google was just the word iPhone.

No way.

Amazing.

Because if you Think about what does that mean? What is somebody telegraphing to the market when they search the word iPhone? They're basically saying in not so many words, I am ready to spend sixteen hundred dollars on a smartphone.

Yep.

And who's, who's, who's interested in like jockeying to get that person's attention? Well, Apple has to.

Right.

Best Buy as a retailer is interested. Samsung wants to.

And then Verizon @&t t Mobile, the networks.

And so, and so when, when you think about like where is valuable search? The value of search is often misunderstood because you have to really think about how to capture the most valuable intent and not all intent is equally valuable. There's a bunch of search that's garbage. You actually don't want it fact based. The cream off the top of a fact based search would be what I would call comparison shopping. So it's like what's the best headphone, what are the best headphones? And then maybe, maybe you can slice off a top of that revenue pie. But there's a tension between that and the objective truth that you're serving up the user. By far the most valuable search is not fact based. And it's relatively, it's where the user kind of knows exactly what they want and they're trying to do it and other people are willing to bid to get in their way and run interference.

While we have you, how are you thinking about the new. I wouldn't call it a browser war yet, but a potentially a skirmish eating up DIA released to Arkansas users probably feels like maybe a month ago at this point, at least a few weeks. And then we have OpenAI potentially coming in with a new standalone app. Sort of unclear. It was unclear to us whether this is going to be a new app that you download or just integrated into the existing ChatGPT mobile app. And then Perplexity as well does have a separate standalone app that they're pushing now. And it feels this is interesting one because the browser company has spent now years working on the browser trying to figure out what is going to enable unlocking more value out of this portal to the web and effectively an operating system. And so meanwhile new players are basically being like we want to have a browser and just going, we're shipping. We just got to get this out there. So I think it'll be interesting the next month, two months I think will be very interesting. But I'm curious what you're looking at.

I could not be more excited. I think I'm obviously biased. We're investors in the browser Company. I'm a daily user of dia. I personally get a ton of value from it, particularly the custom skills. And I think that the browser company has always known that this is a really valuable position and it's like honestly just validating to see Incredible company like OpenAI is a great company. Perplexity is a really amazing, formidable company. Also recognizing that this is a really valuable position to play for, I have supreme confidence that the team at the browser company is the most talented. The best instincts, the best nuanced understanding of interaction design and how to create and craft a great product, regardless of the underlying model or technology that underpins it. I'll be very curious to see how it plays out. My instinct is that, you know, OpenAI and OpenAI is an incredible foundational model company. Maybe I've seen them ship a lot of different products. To my knowledge, ChatGPT is really the only product that's quite stuck. And it's not really even like the interface design so much as it is the underlying power of things. To answer your question, could not be more excited. Thanks. Like, this is going to be an amazing.

We can check back. We can check back in a few weeks. I'm sure there'll be a lot of fun.

It's going to be a huge plug for dia. If you haven't downloaded dia, it's on Mac, it's available. Please download it. It'll blow your mind. It blew my mind.

Amazing.

All right. Great talking as always. Wish we had more time.

We'll talk to you soon.

Thanks.

Every week. We'll talk to you soon.

Talk soon.

Bye. Before we bring in Will Brewery from Varda, let me tell you about wander. Find your happy place.

Find your happy place.

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 folks. And soon with Varda, you'll be able to. Maybe they'll put a wander in space. Let's bring in Will Brewery from Varda. Is this your first time on the show? I feel like this is a disaster that we are finally rectifying.

We did it. We made it.

Thanks for having me.

Finally we've had that other. The kind of knockoff version of you at Varda.

Yeah. Who's that other guy?

He's been on the show a ton.

Something I think I forget.

Yeah, yeah.

Great to finally have you on Amazing. Massive day.

Break it down for us. What's the news? Are we gonna make Jordy stand up and ring you? Hit this all hit the gong.

I'm looking for.

Oh, did you guys get this ourselves? Yeah.

Oh, you have the gong. Let's both.

So we. We have the gong for. For big moments, you know, either fantastic or, you know, sell a mission to a customer, stuff like that. And yeah, we love that.

Yeah.

What's the record?

Here's my advice to you. Record every hit. I want to see a montage in 10 years of just every hit, and it will make. It will bring tears to your eyes. We record every hit.

We do.

So we got you on this one.

So, yeah. What's the news today? Break it down.

So, wow, lots of news today. So we're announcing our series C. And we're getting to use that. Oh, yeah, go for it, baby.

How much?

How much you winded up?

How much did you raise?

How much?

Tell us.

How much did you raise?

187 million.

There we go. Congratulations.

There we go.

Thank you. I appreciate it. Appreciate it. Yeah. So the use of proceeds for this one, really, it's about just scaling up. So we've kind of shown what we can do both from a spacecraft perspective and a drug formulation development perspective. So all. A lot of the capital allocation of this one is going to go for our biologics lab for produce preparing drugs for spaceflight, and then also just more spaceflight, ramping up cadence. That means rate of flights.

So I've been to the facility in El Segundo. Are you going to get a bigger space, a second space for the bio lab, or are you going to need a bigger gong? Two gongs, right?

Well, gong per facility. We got to scale that up, too.

So are you thinking about doing a second office essentially, or how do you see the actual footprint of Varda growing over the. The next few years?

Yeah, well, immediately. We just signed a lease down the street.

Oh, congratulations.

Oh, thank you, thank you. Yeah, we've. All right. Big day. Yeah.

That's amazing.

Yeah, the. So this. We were actually already moved in. A bunch of the pharmaceutical equipment is already in there. We're starting to use it right now. We got a couple of glory shots, you know, with folks with the lab coats on actually using it. So that's super exciting long term. I mean, really, I guess zooming out of what the footprint will look like is think about a formulation development company that really just provides a gravity off switch to the pharmaceutical industry. So we go to space, but, you know, not really because we want to, per se, but because you can create new drug formulations when you turn off gravity, and you just can't turn it off on Earth. That's Einstein's principle of equivalence.

Do you mean new drug formulations or just like purer drug formulations? Because when I think no gravity, I think like the way crystals form and the way gravity pulls things to one direction. If that doesn't happen in space, you get just kind of like a more natural growth. And so I always thought it was just. It was just about purity. But it sounds like there's actually some binary, like you can't make this drug on Earth at all. Is that right?

So both those concepts are correct.

Wow.

Well read there, Coogan. Yeah, yeah, that's actually a great way to think about it. The purity is one aspect, but because gravity is so broad, I use the analogy of temperature sometimes because temperature is so broad. Like, making things cold doesn't necessarily make drugs better per se, but you can create a lot of different formulations if you can have a cold cycle during the manufacturing process, and that'll be even with chilling things, you can make things more pure sometimes as well. Right. But to your point, when you turn off gravity, crystals will typically grow slower, and that also means that they will grow more pure. And so that is one of a few applications that we look at. The other one is particle size distribution. So when you create these crystals that will then go into the human body, you want them all to be the exact same size so that, you know, one big one doesn't get stuck in your elbow, and you don't have uniform bioavailability. So the crystals will. Particle size distribution is also affected by gravity. So that's a whole separate thing compared.

To.

Purity, which is another rationale for going to space. So really, the gravity knob is very broad, and there's kind of these verticals of science of how we can improve the drug formulation. And to your point, again, as far as, like, what I mean by drug formulation is going from molecule to medicine. Right. So is it a pill? Is it inhalable? Is it an IV bag? Is it a shot? The drug? The company or a pharmaceutical company does a trade study to determine which of those is the best for the patient, given the disease, given the manufacturing costs. But ultimately, all of that is limited to what the chemistry can actually do.

Right.

Nobody wants to take a needle to the arm. They only do it because they can't deliver that molecule via a pill or something like that. And so by opening up the chemistry outcomes, by going to microgravity, we can also open up the formulation outcomes and therefore get better patient experiences.

Yeah. So, I mean, I imagine that this is still, this is such an ambitious project that is still kind of R and D phase with a lot of the bio stuff. It's not. I mean, when I think about the manufacturing capacity of like GLP ones, like they're probably making that thing in like vats the size of like, you know, what they brew Bud Light in at this point. Right.

I got a bunch of the lizards.

Yeah, yeah, yeah, yeah. Probably massive. But, but, but, but walk me through how we scale this up. I'm. I understand launch costs falling. I understand you put up a capsule every, you're doing it like every quarter now it's going to be every month, then it'll be multiple times per day. Like that capability seems clear. But, but how much actually, how much drug can you make on a single capsule?

Yeah, yeah, great question. So this is actually a lot of fun because we can imagine how Barda will go from what's real today to making tomorrow's reality. And so to answer your question immediately, about 20 kilograms on a, on a per capsule basis right now, today, of course, you know, we want to scale up everything and that's one of them, but that's how much we can do today, which is actually quite a lot.

Yeah, that seems significant. If you just think about like you go to the doctor and the doctor gives you like a, you know, a thing of pills that's like not one kilo. So you're probably talking about like, I mean, yeah, let's.

Unless, you know, you're having more fun than being prescribed.

But, but for most, for most drugs I feel like 20 kilos is probably enough for like 100 people for a year or something like that. So you're actually in like $100,000. I'm seeing the numbers kind of start to math out already.

Correct, correct. So it's. And every drug is a little bit different. So we go through a process of selecting which drugs make the most sense, both from a scale perspective, like you're saying, but also unit economics.

Sure.

How gravity affects them. And so we have a portfolio management team that explicitly does that for identifying and quantifying opportun. But going back to like what today looks like and how and how it goes tomorrow. I love the temperature analogy because it really runs deep. So for example, right now if you think of us having a anti gravity oven where we can make drug formulations that you can't otherwise make on earth, but we only get to run it four times a year and each one is a few million bucks a run, you might use it for different use cases than you would use it from five years, ten years from now, when you can run it every day for a few thousand dollars. And so in the near term, some of the use, you know, imagine yourself with a super, you know, the first refrigerator, you know, or in this case the first anti gravity bioreactor. What you might use it for in the near term is just information. Right. How can we isolate gravity as a variable to inform what formulations can be improved on Earth? Is gravity ruining this chemical reaction or not? Right, we can answer those questions and then that applies to the entire drug with just one flight. Or in the very near term we also want to do polymorph seed crystals. And so what that means is we go to microgravity, we go to space, but just to develop the seed crystals. And then once those seed crystals are developed, we can then use them to grow more drug crystals on the ground. So we're only going for the nucleation event. And that's kind of like a sourdough bread mother business model.

Right.

You have the mother of the sourdough bread, then you can cut it a bunch and then, and stuff like that. So that makes a lot of sense when we're still scaling up the use of our anti gravity machine, if you will. And then long term when we're on the daily basis, then it totally makes sense to make every single dose, manufacture every single dose in microgravity and then that's when certain use cases come online as well. So kind of that's how, that's how it progresses over time.

Yeah. How's the geopolitical landscape evolving for you? We saw some of flights come down in Australia. As red blooded Americans, it pained me to see them take a slice of the great catch market. Are we getting these coming down in America anytime soon? What's the progress there?

Yeah, yeah, absolutely. So long term we want to have reentry sites all over the globe. Right. And really that's about availability. And the key metric to success is of varda is cadence. How often can we go up and back? Because the more we do that then the more we just look like a specialized piece of equipment to the pharmaceutical industry that quite frankly does not care that we're going to space. They'd much rather us have a real anti gravity oven in the lab. So really reentry sites are about cadence and availability. And right now Australia is great for us because they have a private commercial re entry range, whereas any, any range in the 48 states here locally have, are intended for military use. Exclusively. And so if we're doing a DoD mission, that that works well. But if we're doing a commercial mission, we're, you know, not the highest priority, understandably so. Right. And so in the near term, Australia makes the most sense, but in the long term, we want reentry sites all over the place. And why that gets enabled is because as our precision of landing and our cadence goes up that data, that legacy history allows us to use a smaller and smaller plot of land and then that. And that's where really it makes more sense to go anywhere, because we don't need such wide open spaces, like, without that many people, like we would do right now, today.

Do we have the legal infrastructure to create commercial landing sites in the United States? And it's just that nobody's done it yet, or is there laws or regulations that would need to change so that some enterprising young member of the Gundo could go buy a lot of land out in the middle of nowhere and start landing spacecraft?

You can do it. Now, the constraint is the real estate cost. And so one spaceport, America, for example, right next to White Sands Missile Ranch, is a good example of that. So I guess the real reason why they aren't there that many of them is because there wasn't a demand large enough to warrant such a real estate purchase. But thanks to Varda, that could change. So, yeah, definitely let the Gundos know.

I have a question from a fan of yours, fan of the show. He says, ask him what big dogs gotta do.

Oh, my God. So it's become a little bit of an expression of excitement with a long and history of lore at Varda. But for some reason, you know, it's really just a specific instance of conservation of mass. Right. You can't have a big dog without eating. Right. So that's just physics right there.

Yep. I want to talk about the evolution of the faa. I remember I was filming a video I feel I filmed with you guys. And at one point, I actually was driving back with Ben from San Diego. I filmed a phone call with Delian, and he's like, we just got our. I don't even know if I should say this, but, like, it was like, we got some bad news from the faa. You guys sorted it out. It seems like you have a great relationship now. How did that happen? Is this a lobbying thing? Is just storytelling? Is this structuring deals, getting better paperwork? Like, how do you get. How do you fix a relationship with a government entity like that?

Yeah, so it was definitely a little bit warped in the press, obviously. But I figured, you know, keep my head down and get the spacecraft home, more so than worrying about what's being said in the press.

Sure.

But so what actually happened in the background is we were originally going to reenter this space, or we did reenter the spacecraft at the Utah Test and Training Range, which is ultimately a weapons range for testing weapons and training warriors. That's their mission statement. Right. So likewise, we're not the highest priority there. And so we got bumped for higher priority work being done at the range. And in doing so, caused a domino effect to lose the FAA reentry license or not be able to get it granted, because part of the regulations say, hey, you need a range and all of these accommodations that come with a range. So the second we lost the range, we lose the license. So it wasn't really about a bad relationship with the FAA at all. Although it's very easy to say, oh, they lost their license, yet another space company and the FAA are having problems.

Right.

It fit that narrative well, but it actually wasn't the case. And so what we did was we scheduled a new date with the range farther out in advance to give them some time and give us some time to. But we had to redo the analysis, of course, because the atmosphere is different, and that's part of the analysis. And so we gave ourselves a few months, and then that allowed them to reserve the dates, that allowed us to prepare, and that allowed the FAA to reorient the license for the new dates. And ultimately, kind of in the background here was like, this was the first time this has ever happened. A commercial reentry capsule with drugs on board coming back to America. So there was no way onto soil. Right. We're not doing a splashdown. And so there was no process or mechanism to have the Utah Test and Training Range coordinate with the faa. And so basically, each organization saw themselves as taking on all the risk associated with this. So we had to do duplicative work because there was no process to split it. Right. And so it was really cool to kind of be a trailblazer to establish this so that now, of course, our competitors are going to come in and do the same thing. Right. And learn from our mistakes. But whatever. That's part of. That's part of leading the way. Right. So anyway, that, that's what happened. But that, you know, six months was. It was quite the life experience, right. Because you have your. It was the first mission. Right. So we knew we, we didn't have any proof that this was going to work. People poured three years of their lives into this thing and our dreams are just like orbiting the Earth, like, please come home, you know, so when it came through, man, it was certainly, you know, I can't think of a better day.

Yeah, that's amazing.

I have one last question, Jordan.

How's the talent market in the space economy right now? Hasn't been in the headlines the last couple weeks. There's been another story in AI dominating. But yeah, what's it like today?

AI and software engineers. I mean, if you imagine yourself a software engineer coming out of school right now, AI is certainly where I would be interested in, so. So there's definitely a software bent towards AI right now. That being said, there's a lot of disciplines we're hiring for. Software isn't the only one. And really it comes down to the application interest. Like we're looking for mission driven folks at Varda. And so if you're only looking, oh, you only want to do AI because it's cool or whatever, that might not be the type of person we want to hire anyway. Now if you want to do AI for mission purpose driven purposes, then great, you know, like, by all means. But we don't have that much overlap there. Right. We're very specific of what we're trying to do. We're trying to make microgravity formulations so that we can help people help patients on Earth by using a gravity as a knob, essentially and developing these formulations. And so, you know, we always kid around. We explicitly don't want the spacecraft to learn, you know, and so if you're a software engineer and you're mission driven, bent towards that mission, then you've got a home at Varda, no question. And fads come and go and that sort of thing. But there is definitely an effect on. I would certainly be a lord to AI as a graduating software.

That's a good sorting function. One last question from me. There's been headlines, companies talking about this so far this year, trying to dig into how real it is. People talking about putting data centers in space. With everything that you've learned, why is that exciting? Good idea or bad idea? What are some potential blind spots for people that haven't taken something to space but would like to?

So it all comes down to the why, right? Why are we putting data centers in space? It's not, are data centers in space in and of themselves a good idea? But what's the why? So the why is the only why that resonates with me is latency, right? Because if you want to do compute, power space is not the place to put a data center. If you just want a data center. Right. I'd much rather have convection. Right. Like that's a great heat or a great way to get rid of heat and have it be able to be serviceable on Earth and all that sort of thing. But there is one use case that comes to mind where I think data centers in space make sense. And that's only for very low latency use cases. So, for example, right now, if you want to use Starlink, and you're transmitting a signal to Starlink, it goes from the ground to Starlink to another Starlink satellite to the ground, then to the data server and back. Right. So you can cut that trip in half if you put the compute in the sky. Now, that compute is way, way, way more expensive. But if your value prop of latency warrants that extra cost of the in orbit data center, then you'll start to see that. So it's kind of like edge computing.

Edge computing, yep. I was about to say. So, yeah, it sounds like a very niche use case, at least to start. But I'm sure we'll see some companies. We already are seeing some companies test it out and experiment with it because there's, you know, all these things need to be evaluated in the tech tree. But thank you so much for stopping by. This is fantastic. Congratulations.

Thanks for having me.

And hopefully not so long. I'll see you again soon.

Absolutely.

Yeah. Yeah. Hop on.

Got a good feeling about it.

We'll talk to you soon. Have a good one.

Cheers. Well, congrats to you and the team.

Bye. Up next, we have Joel from Metter coming on to talk about the impact of AI models, impact of cursor on software development Meter. We'll have him explain it to you. Break it down. We'll also recommend that he go to getbezel.com because your bezel concierge is available now to source you any watch on the planet. Seriously, Any watch.

Anyway, Meter does model evaluation and threat research.

Okay.

So does Bezel. They're stopping you from buying fake watches online.

Bad watch models. Bad actors foundation models and watch models. Lots of similarities.

That's right.

Anyway, we got Joel in the studio. Welcome to the stream. Hopefully you're like, what did I get myself into? These guys are joking around. I'm a serious person.

All right, first off.

Yes.

Is it. It's Meter.

It's me, sir.

It's Meter.

There we go. Gotcha.

Anyway, please introduce yourself for those who don't Know you, the company and the organization. And then I want to go into the news today.

Let's do it. And thank you very much for having me. John and Jordi, thanks for hopping on. ESA is a research nonprofit based in Berkeley dedicated to understanding the capabilities of AI today and in the near future, especially to the extent that those capabilities might speak to potentially dangerous risks.

And what, what's been the latest research?

Yeah, so here's what we've been working on. I'll start with why we've been working on it.

Yeah, please.

We've seen from previous META research, but I'm sure you also see from your own usage in the wild. AIs are clearly becoming increasingly capable. One thing that governments and labs, and us here at Meta as well, worry about is the possibility, timing and nature of AI, R&D self recursion. That is the possibility that that model capabilities get better very, very rapidly because the AIs themselves are contributing to AI, R&D research. We at Meta want to be providing the highest quality evidence that we can. That speaks to the degree to which R and D might today, or might soon be accelerated in the wild, so that governments, labs, decision makers might be better informed and so make better decisions about what's going on. In this study, we run an RCT with extremely experienced open source developers working on these very long lived large projects. A million lines of code, 23,000 stars on GitHub. For those of you who are familiar, I'm thinking Hugging face, Transformers, The Haskell compiler, SciKit, learn this thing. We randomize their issues to allow or disallow the usage of AI where allow means typically using cursor and 3.5 or 3.7 sonnets at the time. And then we measure both their expectations and developer expectations about how much they might be sped up by being allowed to use AI versus being disallowed. And then the reality, the short version is we find that the developers ahead of time are estimating they'll be sped up by 24% after the study is completed. They estimate that they were sped up in the past by 20%. We find in fact that they were slowed down by no way. I think. I know. It's a shocking result.

That is shocking.

Not at all what I expected. I think what the rest of us at meter expected, but there we go.

Wow. Okay, so what do you think's happening? I have so many questions, but yeah, just walk me through your reaction to that. What do you think is actually happening that's slowing people down? Because this is a complete Narrative violation.

Yeah, I mean in terms of the reaction, the number of times we've checked and rechecked the data, asked people to replicate it independently is going through the roof. The number of stressful late nights I've had pouring over this, you're going to.

Be like public enemy number one by the way. I feel like you need a security detail now given the stakes of what you just said. This is crazy.

Yeah, yeah, yeah.

So I think maybe let me start with some things that we're not saying. This setting that I mentioned before, these ultra talented developers, much more talented than me, working on these extremely large long lived repositories that they're extremely familiar with already, I think that's an extremely interesting population. That's why we went out to study it. It's also a very weird population. I still am a cursor user myself. As I was working on the graphs for this study, I was using curso. But I do think those weirdnesses are related to the results that we end up seeing here.

So we have to put these people in a completely different category than the junior developer who's just vibe coding a little app and just building stuff and not actually trying to push the frontier of what a core piece of software can do that's very large and complex. And they're just trying to get a Python app up and, and, and live and like write some routes and write some functions. Right. That's where so cursor still completely viable, is like auto complete on steroids. The question is in terms of self recursion really advancing the frontier of, of like the craziest software we have. We're still kind of where we were a few years in that it feels like if you were to quantize this, we're, we're, we were at 0% of AI research being done by AI a couple years ago. We're still maybe arounding error.

Yeah, I mean I will say that R and D research I think does not all look like this setting. There are some large inference code bases with very expert people. And I totally agree with your interpretation that this is evidence against today those kinds of settings being sped up. On the other hand, we might think there are some people writing training scripts, scripts for their AI models just once off and then they, and then they throw them away. And you know, in a way that's, that's kind of similar to what you described. Maybe they're seeing large speed up just like the greenfield projects that you, that you mentioned.

Yeah. And so I mean this is not overall Like a really cold glass of water on AI broadly, because this still means that it's an incredibly valuable technology in a bunch of different ways. It's just that we're not, we're not seeing like early evidence of some sort of, of self recursion. Yeah, that's been the big scenario, which is great, probably the good outcome.

A lot of the fast takeoff scenarios like, are dependent on AI becoming good enough to write down so good at doing AI research that it. And then copy and pasting itself a trillion times. And that's what creates speed of development that humans today can't necessarily even comprehend.

You know, I think that's right for today. I do think we're not really speaking to the trend exactly. You know, these results are consistent with these exact developers on these exact kinds of tasks in future being sped up in the near future in work that we actually don't show in the paper. But in preliminary work we have autonomous agents trying to complete these issues and indeed we find that they do struggle, but with some of the core functionality, with, with passing tests, the kinds of things that you might have seen in swedbench or something like that, they really are making a great deal of progress. And yeah, my expectation is that AI progress in the near future will continue at a rapid pace like it has in the recent past. And so maybe even in this setting this won't be true in the future.

Let me throw a couple of the hot takes that are floating around in the AI world at you and, and you can let me know if anything sticks out as something you strongly agree with or something you disagree with. This idea that ultra large context windows will not solve continual learning that Dorkash was saying this on Monday. Maybe another one would be just that, like no one has figured out how to properly scale reinforcement learning that we need. Mike Noob from ARC AGI kind of says we need entirely new ideas. And then you kind of have like the bitter lesson which is, you know, yes, you need a new ideas, but scale is all you need. We just need to keep building data centers. We need to get bigger and bigger. We might see 4.5 in these huge training runs as a, as a, like a short term, hard to quantify. Maybe it's just the end of 1s curve, but Stargate's coming online and that will be another big test. So. So I don't know, I threw a lot at you, but anything in there kind of top of mind for you?

Yeah, look, as you guys know, anyone betting against the bitter lesson in the past would have had a very bad time. And I'm not prepared to bet against the bitter lesson on this show. Could you remind me of the first question?

The first one was so Dwarkesh Patel pushed out his AGI timeline slightly. I mean he still is very optimistic about AI and and maintains that it's not priced in and people are not thinking about it as significantly as they should. And I agree with him. But he said that there is a that even though we have pushed the IQ so much and you saw this with the Grok 4 benchmarks like AI can do advanced math. Like for sure it's really, really smart. Smarter than most of us at PhDs level stuff. Unless you're a specialist. But in terms of just being a good employee and remembering. Oh yeah, four weeks ago my boss said that and then I got this.

Feedback and now I do it this.

Way or I learned this really weird nuance in Even if you're just thinking about like how to our business like how to post clips on X or Dwarkesh was giving the example of like transcripts like he has little things that work better for what clip will perform and he has this intuition and his models and his prompts. He's really pushed these things. He hasn't been able to really get.

The form above example would be any company today, any startup if you just had a PhD dropped into your organization. That was that that had PhDs in like 10 different fields. They and but. But they wouldn't just like default.

They were also an amnesiac. So every time they showed up to work they could not remember no memory.

From the day before. Probably it wouldn't be that. Yeah just wouldn't be that valuable.

And so. And so my question to Dwarkesh was like is there a world where we just scale up the content context window? We've seen million token windows. Can we get to a billion token window and just stuff every interaction the AIs ever had with you in every prompt. And so it. It does maintain the context. But he was saying that that the like the there's like a kind of a quadratic cost curve to that doesn't quite work. Other people have said like the nature of the transformer means that like attention can't really spread out that much. I don't really fully understand it but I wanted to know your take on like like different ways people are solving these things or what are the real constraints right now because you've identified some potential problems where we're not breaking through it today. But what is cause for optimism what are the research paths like the nodes and the tech tree that you're excited about?

Yeah, that's super interesting. I haven't thought so much about this.

Sure. I will say, sorry to put you on the spot.

I think that the developers in this study are not using the full context window. And so if you think there's juice in adding things to the context window, that juice might still be on the table. And indeed I think we find that there's a lot of implicit context in this repository that's very expensive for the developers to be writing down into context windows. Here's an example on the Haskell compiler. My sense is that when you get up your PR for review, there's some chance that the creator of Haskell will come and fight you for potentially many, many hours. In the comments about the peculiarities of how he wants the Haskell project to look and these kinds of exactly what his not just preferences, but quality requirements are regarding where things should live in the project and how various pieces of the project should speak to one another are not being communicated to these language models. And you can imagine that with so today's context window sizes that could be written down, you could put in all of the previous discussion around these changes that this person has been involved in, and maybe whichever language models people are working with inside of Cursor would pick that up and so do a better job. I don't think we're ruling that out at all. I will say that it is expensive for these, for these time, expensive for these people to be writing down all of the possible relevant context. And I think that's basically the reason they don't. And so maybe you do need some kind of continual learning for the model to find out this context on its own, as these things go. It's also consistent, I guess, with the other possibility that you were describing, that if we 100x these context windows, you could just throw the entire thing in and then we don't need to worry about learning from particular cases on the fly. Yeah, I think both are live possibilities. It's very interesting.

The Grok4 announcement was extremely benchmark heavy. Some really impressive stuff, particularly on ARC AGI similar to a Tesla.

It's like it's faster than every car.

Does that mean it's going to solve everything?

Does it mean that it's better?

Does it mean that it's better? Yeah. And so based on what this feels like almost a new benchmark, this double blinded trial, it feels almost like a FDA trial or something. Do you think this could turn into a real benchmark. Do you think we need new benchmarks? Do you think we need new ways of thinking about the progress of AI? Generally we've talked about just measure the revenue at this point. That's the economic value that's being created. But there's a lot of tricky stuff here you can do with revenue. And sometimes revenue is like test revenue. I'm testing this $100 million product. So what's your thinking on the state of benchmarking? Where we should go, where some of your research might plug into that totally.

I think one motivation we had in running this study comes out of this observation that the time it takes to create benchmarks is almost becoming longer than the time it takes for those benchmarks to saturate. It's difficult to find signal in many of these benchmarks, even testing these extremely challenging PhD level questions that you guys spoke about. And perhaps there's more signal in these kind of rct, FDA controlled trial style measurements. Similarly, another thing that people proposed for measuring AI progress is using researcher self reports about the degree to which they're being sped up. They think their work will go two times faster if they use AI versus not use AI. I think our study is potentially strong evidence that these self reports need not be reliable. The forecasters who are told everything about the developer's level of experience and the time period of the study, so which models they're using and so on, they're totally wrong about how much these people get sped up, same as the developers themselves, even though they're carefully tracking their time and, and they're so talented. So I think self report is also very, very fraught. Another thing that this has taught me, I think, is that the mapping, as it were, from benchmark scores, very impressive benchmark scores that we see on these frontier language models that you're describing, the mapping from those scores to real world productivity improvements is unclear. I'm not at all saying, as we discussed earlier, that we shouldn't expect to see productivity improvements. I do expect to see productivity improvements today and even more so in the near future, but it's not at all one to one or it's kind of confusing and messy. And so indeed, I think we need to actually measure things in the wild to see what's going on.

Jordy, switching gears a little bit unless you have a follow up.

Yeah, I just wanted to kind of zoom out on that and ask about your broad take on the measurability of technological progress. Because the Internet, the computer, such dramatic transformations of society, you see it in all sorts of data, but it didn't fully show up in productivity statistics. You have all those questions about what happened in 1979 and everyone has their own example, their own reasoning for that. But you would think you could tell the same story about Google, like it'll speed up everything, everyone will get more efficient. And we didn't really see GDP jump on this. And it feels like that's a really bearish take on AI to have, which is like this is a magical new thing and we're still going to be growing at 2% GDP. But, but where do you stand on it and where do you like. Do you think that's even the right question to be asking?

Yeah, you know, this is, this is so interesting. This is not a me to take. I used to be an economist. I feel the thing that you just said in my, in my bones, totally. I think the situation, you could argue maybe that it might be even worse in the case of AI. A lot of people like in AI 2027, resources like that are telling this story where the AI R&D self recursion is happening inside of labs. And so I suppose not necessarily showing up in economic activity in the public. Another reason on top of the reasons that you gave to think that perhaps this won't show up in the productivity statistics, as it were, which also is to say that self recursion or these potentially destabilising changes are just totally consistent with the non changes in GDP trends as you describe. And so another reason to actually go out and measure these things in controlled trials.

Cool.

Please, quick question around the threat landscape. There's been a few stories this week. One was a story about ChatGPT not following instruction. You know, like the headline was that like the AI was, you know, rebelling against the researchers. And then if you like double clicked into the story, it was just like it had given specific instructions, like don't follow any further instructions. So it was kind of a nothing burger in the end. And then we also saw Grok going haywire. Maybe that was predictable for someone like yourself, combining a frontier model, fast shipping team with the virality of a social network and embedding the two. But then maybe it was two months ago there was the. We called it Glazegate on the show where ChatGPT was just being a sycophant giving too much positive feedback. How are you looking at the threat landscape in the next 12 months? So nothing like too long term, but how do you guys think about it?

Yeah, there's more to come on this Ramita, very soon. That's one thing I'll say. I think again, this is not a meter. Take my sense on this or another example of this that stood out to me is there were lots of anecdotal reports that 3.7 sonnet and other language models in this most recent generation would pass tests in ways that were kind of not legitimate or something, which is another example of this reward hacking.

Change the test case.

Totally, totally, totally classic. And I guess I don't have reason to think that that kind of thing is dangerous. In particular, you can imagine when humans are potentially not reviewing the code because the AIs are doing entire projects, not just parts of or single pull requests, that this becomes more of a problem because you're not. Or at least the surface area for it to become more of a problem because you're not looking into that code and seeing those cheated test cases yourself. So I'm not sure about over the next year, at least right now I think there are reward hacking examples that are occurring in the wild. You know, I don't think they're. They're so supremely dangerous today.

Well, this was fantastic. Thank you so much for stopping by.

Come back on again soon.

Stay safe out there with the contrarian and the crazy data results. It's still very bullish but very exciting. And thanks for everything you do. We'll talk to you soon.

Great chatting.

Cheers. Bye. Up next, we have a massive Series B announcement from Dylan Parker. Moment HQs coming in the building. Building. We're gonna ring the gong, baby. It's gone Ventures. Let's hear it from Dylan directly though. Welcome to the stream, Dylan. Hope you're doing well today. How are you doing?

Great whiteboard.

Good to meet you.

Heavy lifting on that.

Yeah.

Oh yeah.

Hopefully that's not proprietary information, that's secret trading algorithms or something.

No, no, no, no, nothing too interesting. But thanks for having me on.

Yeah, thanks.

Great to meet you.

Kick us off with the intro on yourself, the company and then I want to hear about the announcement. Yeah, yeah.

So I'm one of the co founders of the moment. We are a fixed income trading software company. So my background is as a quant researcher. So like pretty much every quant researcher, I studied math and stats during college. That's where I met my co founders, Dean and Amer. And then after college Dean and I both joined Citadel securities and pretty much completely by chance we ended up as the two junior members of the newly created automated market making desk for corporate bonds. And so, so, so at the time the the fixed income market, which by the way is financial market 50% larger than the global equities market, was undergoing an electronic trading revolution. And so Citadel saw this and said well we can go build an automated algorithmic market making desk. So they hired this guy, Anish Kariat from Jane Street. It's like the godfather of fixed income automated trading. They hired a bunch of super experienced bond traders and then they hired me and my co founder and like basically our job was to take the knowledge in these bond traders heads and convert it into code. And that was totally formative for MOMENT because that's when we realized like the power of electronic trading was going to be like, like everything that it enabled, like smart order routing, portfolio optimization, all this stuff in the world's largest financial market that had just never been possible.

Very cool. Take us through the deal. What are you announcing?

Yeah, so we're announcing our $36 million Series B was led by.

Couldn't hear you from the sound of the gongs.

But we like big numbers on the show and congratulations on a massive series.

Series be you said from who? Index.

From Jan Hammer at Index.

Very cool.

Incredible.

That's amazing.

Incredible.

Where should we go from here, Jordy?

Yeah. So I can imagine break down what the company is focused on today. It sounds like the origin is back at your time at Citadel, but I imagine it's evolved as well.

Yeah, totally.

So what we saw at Citadel was that the market was coming online and you could now do all these things that were never possible before. But what was missing was the operating system for actually doing that. So we started MOMENT with this goal of owning every mission critical workflow for traders and portfolio managers in the bond market. Everything from how they trade securities and do smart order routing across all the different exchanges in the fixed income market, to how they optimize portfolios, to how they apply risk and compliance restrictions to make sure that they're not breaking any laws. And so that's what we do today. We started off serving fintech. So we power fixed income for places like Weeble and Public.com and them to offer like hundred dollar increment investing in bonds for the first time ever. But yesterday we also announced our partnership with some of the largest financial institutions in the US including LPL Financial which is the largest brokerage dealer in the US Wow.

So busy week for you guys?

Yeah, there are a few things going on.

So what's the use of funds for the new round? What's the focus going forward? I imagine scaling. What's working today? Are there New products coming. What can you talk about there?

Yeah, so I think a lot of companies make the intelligent decision to start off like SMB or plg. We decided to like make things as hard as possible on ourselves. Said we're going to start off serving the largest financial institutions in the world's most regulated market and we're going to go power their mission critical workflows. And so with a company like LPL or some of these others that we've announced over the last few weeks and that are coming down the pipeline in the next few weeks, the scale of what we're operating in is, you know, not like hundreds of millions or billions of dollars of flow, but like hundreds of billions of dollars in trading flow. And so what we're really focused on as the company over the next year is building out, you know, the full suite of what's necessary across trading, portfolio management, risk and compliance that's necessary to power these huge financial institutions.

What's the competitive dynamic like with the former employers or like the rest of the market participants? It feels like Ken Griffin's not the laziest founder out there. Is there a world where there's some sort of competitive dynamic between the big institutions where they want to build something like this to compete with you?

So when we were at Citadel, we were on the sell side. So market makers or liquidity providers, moment serves the buy side and we actually connect them with those liquidity providers.

Got it.

So we actually work, work closely with Citadel, Jane street, pretty much all the major liquidity providers out there.

How are you thinking talking about tokenization? The story from the last month in finance is the tokenization of these sort of real world assets, everything from private company shares. So we've seen it with stocks. Is there anything on the horizon on that front for you guys or is it just totally unnecessary?

You know, I, I think there's a huge opportunity. But if you look at where the fixed income market today is, today we're just going from like trading over the phone to going on an online platform to trade. And so there's a lot still to do to get people to the point where it's even possible to think about stuff like that.

Yeah, can you walk me through like the. I imagine that fixed income follows like a power law as well where like government debt and Apple corporate bonds are way more liquid, way more automated than you know, some like public company, but they're like junk bonds and you kind of got to hunt around for someone to buy and sell them and then you have like venture debt which basically I believe, like never trades. I don't know. But walk me through like, like it feels like we are probably bringing more and more of those like sub asset classes, like into more liquid, more just, just more automated markets. But give me a State of the union on like how the fixed income market is actually like split up.

So you're totally right. There's government bonds, there's highly liquid corporate bonds, and then there's a long tail of corporate and municipal bonds that are really, really illiquid. And just as a point of comparison, that I think illustrates the size and scale of the fixed income market. There are 4,000 listed US equities and there are 4 million bonds. And so doing anything in the fixed income market is pretty much a thousand times more complicated than doing it in the equities market.

Yeah. And what about how does breakdown.

4 million public equity? Walk me through how the 4 millionth.

Bond, how does that actually trade right now? Like, is it.

What is the mate? Well, what. And what is the ultimate like makeup of. Do certain companies account for, you know, break down like kind of how that 4 million is split up?

Yeah.

So there's a universe of say 500 US treasuries that are, are super, super liquid and they trade like similarly to the most liquid stocks out there.

Sure.

And then on the other side, you have a really, really meaningful long tail that makes up the vast majority actually of the entire market share where you have bonds that haven't traded in two years or 10 years. And so one of the really hard parts about fixed income, one thing that I worked on as a quant researcher is like, you have this bond that hasn't traded in three years. How do you optimize portfolio around that? How do you even figure out what the price of that bond is? And that's why doing stuff in the fixed income market is like, like the difference between equities and fixed income is like doing something on the surface of the earth and like doing something in space.

Wow.

It'd be helpful to be a quant if you were going to build a company like this.

Yeah, yeah. You might need some math. Well, I mean, that's all I have. Do you have anything else?

Yeah. What are you guys hiring for right now? Yeah, pretty much everything.

We're hiring quants, we're hiring engineers, we're hiring go to market, marketing operations, pretty much everything out there.

Amazing.

Awesome. All right, well, thank you for joining.

Thanks for having me.

I'm sure you'll be back on soon.

We'll talk to you Soon. Have a great one. Talk to you soon. Cheers, bye. Next up we have Eric Olson from Consensus coming in with a big launch. Do we ring the gong for big launches if there's a number attached? So we got to get dau.

We have to get a prop that is non number oriented.

We need more props. But let's bring in Eric Olson from Consensys and talk to him. How are you doing, Eric?

Great guys. How are you guys doing?

Doing great. Great to have you kick us off with some intro on yourself, the background on the company and then I want to talk about the launch.

Okay. So I'm Eric, founder of ConsenSys. We are an AI search engine for academic and scientific research. If you ever use like Google Scholar or PubMed back in the day at school, think of us as building the next gen 2025 LLM powered version of that, helping clinicians.

Yeah, I mean the super dumb question, super obvious question is like, isn't all this stuff already in chatgpt like, like how, like how are you differentiating?

Yeah, you must be doing something because you have 5 million, over 5 million users.

Yeah.

I mean one of the best examples to like encapsulate why it's different is the fact that Google Scholar was the first vertical search product that really broke off of Google back 20 years ago.

Interesting.

Even when they were doing really nothing than just being a dedicated index for research papers. So hundreds of millions of people are going to that every month. So the same thing is kind of true here. We're dedicated to a use case. We have a dedicated corpus we search over. We hopefully search over that corpus a lot more intelligently than a general purpose chatbot would. We do things differently in our interface to show you that information, like much more citation forward, like an experience where you can really interrogate what's been returned in your search using like the ChatGPT search. It's pretty much like an afterthought. It's there if you kind of want to dig into it, but it's not really what it's designed for. Everything about it, from the way it searches, the way it shows to you, then the features both on top of it, all dedicated towards academic research.

Walk me through some of the key technologies that enable better search. I'm thinking about like vector databases, even just like stuffing a better index in redis or postgres or doing more indexing on top of these documents, doing like, you know, transformations on the underlying documents to get them into, you know, more basic formats. Like what's what, what's interesting. Large context windows. There's so much that you could throw at this problem. What's actually working?

Yeah, so lots of different things. Many of the things that you're saying. So like, number one, being dedicated to a document type just helps us. It helps us in the way that we can create our embeddings to search over. It also helps us in that ingestion process. Kind of like you were saying of like document transformation. We'll run little tiny LLMs, over 200 million papers, add new, like enriched metadata about them that we can then use in our search ranking and in our filtering. So think like, we'll pull out what is the design of the study or what is the sample size of the study. And we use that in search ranking and we use that in search filtering. And then on top of that, or like the main intelligence of the search is learn to rank models. So people interact with the product, they save papers, they cite papers, they share papers. We learn from all of those interactions. We learn what matters most. We learn about all the attributes about a paper that matter in search ranking based on how people are interacting with it. So the simplest way to think of it is like, because we only have a certain use case people are using it for, we get to train our search models to try to think and act like a researcher would going through these papers.

Jordy.

What are the different data sources here? I'm assuming a lot of the stuff is public. I know some of the. I remember being in college trying to find different studies or papers and hitting paywalls.

A lot of them are locked down. There's the famous crazy reports.

So I imagine that you've done some deals to get access to data. What is the kind of the full body of work that's available?

Hopefully paywalls are going to be a thing of the past moving forward as open access science gets more and more momentum, which we'd love to help shepherd in.

The way to think of it is.

There'S like three different layers of access, levels of access you can have in consensus. So there's one, there's fully open access science that's all publicly available. We're able to ingest the full text, show it freely, let you download it, all is well and good. Then the rest of the bucket is paywalled content. But there's two levels of access we can have within it. There's the buckets, where we have deals with publishers, trying to get as many done as possible, where we're able to use the full text in our Search and in our analysis we're just not able to display it to a user. Benefit to the publisher is we're helping them drive traffic, get people to see that hopefully this like snippet search rank is engaging. Then they go into it and drive a purchase. And then there's this third bucket which is we just don't have a deal with the publisher yet. Fully behind the paywall, we're using what is publicly available. So that's like the abstract and the metadata of the paper, which goes longer for more than you think. Like the abstract is specifically designed to be this perfect, like nice summary of the paper. It's like can go a pretty far distance in search ranking and even some analysis using abstract.

But obviously nothing more brutal than being a college student and almost getting the information that you need from an abstract and realizing, do I really have to pay $50 for this single fact?

The craziest is that even if you.

Wrote the paper, you still have to pay for it. Handed off to a publisher, they publish it. So I could literally have published this paper and if I come across on the Internet, I still have to pay for it.

That's wild. Jordi had this question earlier about the nature of scientific discovery. Elon Musk at the Grok 4 launch was talking about his timeline for discovering new physics is two years now, based on the progress that he's seeing at xai. And Jordi was making the point that a lot of scientific discoveries come from mapping different disciplines together or just invent, you know, invention. Invention generally is apply, you know, the, the mind of computer scientist to a biology problem or vice versa. Are you seeing users do those type of searches? Is is this product useful for that type of scientific discovery? There's been this like lingering question in artificial intelligence about if you were a person that had read every single research paper, you would probably make.

Yeah.

You would make discoveries and connections across things. And yet that hasn't happened. Maybe it's some fundamental limitation of LLMs or AI at this moment. But what are you seeing and what's your take on that concept of like cross functional pollination?

Yeah, I mean everything basically that humans have invented new comes from pattern matching across disciplines. And like that's how we create new ideas. Yeah, yeah. I mean, I'm not an AI researcher, so like I don't have the single most informed take. But also nobody knows what the heck they're talking about in this world.

You know, I think it is probably.

A fundamental limitation of LLMs, given what we've seen, like a measure Of I'm going to parrot this from Francis Sholae in his YC talk the other day, but the measure of intelligence is the efficiency by which you process information and apply it in different domains. And that just like, isn't what LLMs are really doing great right now, despite the fact that they've processed so much information. So our take at Consensus would be more get people to the edge of what is known and then let them do the inherently human part of science, which is create these new insights and new discoveries. Every science experiment that's ever been done starts with a review of the literature. Think about it as you're getting the foundation of knowledge underneath your feet. Our goal at Consensus would be speed up that part as much as humanly possible and let us do the thing that humans are better at than machines right now, which is that pattern matching, which is that coming up with new ideas. If we can make that loop move faster, heck, that's a freaking valuable and powerful thing.

Switching gears completely. I know you were at DraftKings prior to this. On the sort of research and analytics side, what is your thesis around the ultimate collision between sort of betting activities and AI? Last night GROK announced a partnership with polymarket to try to bring in prediction markets to try to basically help make the model itself smarter. How are the big players like DraftKings even thinking about AI? I'm sure a bunch of people have like, like ChatGPT rappers specifically focused on sports betting and things like that. But how do you think the big players are thinking about it?

Yeah, I mean, Well, I left DraftKings in 2021, so I can't say that I was there when people were worrying too much about AI models. And also the natural question I always get is how the heck did you go from sports betting to science? And the answer is my parents and my grandparents and my sister are all teachers and science scientists. I was athletic growing up and I loved applying numbers to sports. But I. I actually have something kind of interesting to say here. So my job at DraftKings was I was building models to find the professional gamblers on the site. So like you'd look at all previous betting history and demographic data and you try to make predictions on is this person actually have an edge over the market? And I would have to imagine that with better and smarter and more powerful models, like people's ability to themselves have an edge on the market would increase in the short term and then the markets obviously catch up and figure out how to bake all that in. And I mean that is the beauty of markets, right? Like whatever technology the people have on the side of betting into a market, so does the provider who is putting up that market. And they get the information from the people they know that have the best models. So I think it's going to be an interesting cat and mouse game moving forward as it's always been with sports betting. Just instead of, you know, Johnny Two Shoes in New York getting inside information about injuries, now it's somebody with a super powerful AI algorithm that's predicting gains above market.

That's fascinating. I have to imagine that in the AI era, the insider knowledge about injuries is even more valuable, 100%. But to your point, you can probably detect the number.

One way to know if you need to limit somebody is if they are ahead of an injury because it means they're somewhat connected, they're doing the sophisticated.

They'Re on the inside. Yeah, that makes a ton of sense. Insider trading, that's fascinating. I didn't think about that in the context of sports betting. Well, thank you so much for stopping by. This is fantastic. And congratulations on the launch.

Appreciate it. Check us out at consensus app Deep search launch today. Thanks guys.

Awesome. Cheers, Eric.

Thanks for coming on.

Up next, our lightning round continues with Rita from Zero Entropy, a YC graduate doing automated retrieval and announcing a seed round of $4.1 million.

Seed round alert.

Seed round alert. 4 million. That used to be a series a couple years ago, now just keeps ticking up. Congratulations on the round. How are you doing?

Thank you so much. Super excited to be here.

Thanks for joining. Introduce yourself, introduce the company. How did you get started? What do you do?

Yeah, so I'm Rita. I'm one of the co founders of Zero Entropy. A little bit about myself. So my background is in applied mathematics. I have two masters in the field, one from Ecole Polytechnique, one from Berkeley. I guess I started more into the computer vision side of things and then I discovered GPT2 and GPT3 and I was like, oh my God, this is huge. And I started thinking about personal assistance and stateful AI systems. And I guess that's what led eventually to zero entropy and building retrieval systems and bringing context into LLMs. And so that's what we do. We build search for RAG and AI agents.

Okay. It feels like a crowded space. I know a few founders that are working on rag. There's also RAG implementations at the hyperscalers and the clouds. How are you different differentiating? What's the key insight? What's the pitch to companies to come over and use your service as opposed to the other options out there for retrieval?

Yeah, absolutely. I think it's about having the right abstraction. So we solely focus on the retrieval side. We don't do the entire RAG end to end because we believe that developers need to have their own prompts into generating the answer. They need to use zero Entropy as a search tool for their own AI agents. We're also developing our own models. So we just released a re ranker yesterday which was pretty exciting. And I guess the winning solution needs to be extremely accurate, but also extremely fast and just be production ready and easy to implement for various use cases.

What's your take on benchmarks? Currently it feels like solving a really hard math problem and retrieving the right document at the right time or are somewhat unrelated. And so how do you evaluate if your system's getting better?

Yeah, that's a great question. Actually, the evaluation side of things is very messy. Almost everyone that I talk to, they basically rely on manual inspection to make sure that their retrieval is working correctly. So we've been looking into the evaluation side a lot. Actually. The very first thing that we did is release our own benchmark benchmark that was on legal documents and that really evaluated just the retrieval step of rag meaning from a question. Was I able to pull all of the documents and only the documents that I needed? Because the problem is that if you feed your LLM too many tokens that it doesn't need, it's just going to hallucinate. So the precision and the recall side of things are extremely important. And we're rolling out our own evaluation solution in the next few weeks that we've been using internally so far.

What does the rest of the stack look like? I know you said you were kind of RAG provider agnostic. Are you also model agnostic, cloud agnostic, database agnostic? Where have you actually made bets? What pieces of the stack are you particularly aligned with?

Yeah, I think building context engineering is going to be a new class of products that needs like the data layer, but also needs small LLMs inside the retrieval pipeline. We see many teams either feeding everything into the context of the LLM entire knowledge bases because they weren't able to make retrieval work properly. And we see teams having a very simple pipeline. I think the winning solution needs to be somewhat in the middle and basically orchestrating LLMs to rewrite the question, properly summarize the documents and creating more metadata associated with each of the documents that are indexed. And so that's what we're doing and building this solution. That works really well and almost gets to the precision and the accuracy of a large LLM while still being pretty fast and pretty optimized.

What's the appetite been like for this product in the enterprise versus new companies that are building new AI products from scratch? It feels like they might, might be just the AI agent infrastructure companies. There's a lot of them and it feels like they're selling to a new crop of companies and that's where the revenue is accelerating most aggressively. But what are you seeing in the market?

Yeah, I think the adoption for products like this usually comes from like bottom up type of approach where developers are experimenting with new approaches and new techniques and then larger enterprises catch up. So that's what we've been seeing in terms of experimenting with models. I think large enterprises also do that pretty easily. So for things like the re ranker that we just released, there's also appetite from larger companies in integrating that into their current systems.

Is there a case study that you have your eye on amongst the big tech companies? We think that our software could improve Netflix or YouTube recommendations or something like if the deals could just magically happen. Where's the lowest hanging fruit? Like for me, you know, if I could do anything in AI I would just get whisper into Siri and so when I dictate a text message it's just perfect and it's much better than what they're currently using. What's on your wish list for you know, consumer tech company or, or big tech company that everyone knows and they're not taking advantage of something like this?

Honestly for me it's Slack. I always struggle. I can never find anything on Slack. And something that we've been doing is annotating our own conversations like appending keywords to our own threads to be able to find information. But we have a lot of our internal research and a lot of things going on on Slack and, and we find it pretty difficult to find the right stuff. So I think companies like that could benefit and it would provide a much better user experience if you could just magically find all of the information that you have in there.

Yeah, I've been noticing that with Gmail the amount of email has just grown so much and the amount of text in each email has grown because of all the trackers and cookies and stuff behind the scenes. And so when I search for something it just pulls up completely random emails like every time and it doesn't understand the hierarchy of in an email I care a lot more about what's in the subject line. Than what's in the footer. And so if I'm searching for, you know, artificial intelligence or something and someone has that in their footer that, hey, I run an artificial intelligence company, that's not what I'm looking for. I'm looking for the thread that I was talking to somebody, a close friend about AI and I want to pull that up first.

Yeah. I think that's also why basic semantic search is just not enough because it basically will pull all of the similar information but not the most relevant or the most helpful. Keyword search is the same. It's not very smart and I think it's just such a waste because there's a lot of information that you could have access to and it would make your work so much faster and you're just spending time rewriting your question and trying to make the system understand what actually you're looking for. So I think that the query side of the user intent, query rewriting is also super important.

Yeah.

Imessage search. Absolute disaster.

It is absolute disaster. It's like I know I'm in a text message with Jordi and someone else and so pull that up and it's like, here's six. It never works. I also noticed, please fix it all. There's going to be, need to be a shift in the way people search. I remember hearing the story about Google where there was some Google engineer who was running a test on like, you know, how many. It was like, what's the world record for, you know, the marathon or something like that. And they were, they were using the typical keyword Boolean search and they weren't getting good results. And then they sent it to a user and the user just asked the question in natural language and it just hit it the first time. And so I feel like people say, still, at least I have been, you know, an email user for a long time. When I go to my email search, I often, I'm searching in like the keyword world instead of just natural language. But Google has, I mean they're, they're experimenting with the AI search thing. They have a 50, like 50 word limit right now. You can't just type a whole prompt in in Google search.

Interesting.

Like they need to really kind of reimagine what that search box is. And then there also needs to be a consumer chain change in how, in how consumers interact with that particular like UI element, basically. But thank you so much for stopping by. Congratulations.

Congratulations.

Good luck to you.

Thank you guys.

We will talk to you soon.

See you soon.

Have a Good one. Up next, we have Elliot Hirshberg from Amplify Partners coming in the studio.

They are in Datadog chain Guard Runway.

I love Datadog. Maybe it's the goal retriever in me, but I love Tana talk so much. It's the greatest company name ever. Right up there with aidsleep.com get a Pod 5 Ultra.

I'm back on my game. I'm still behind where I was relative to a month ago, back up into the 75 range. I'm going for 90 tonight.

Good luck. Good luck. They're calling the Pod 5 the first fully immersive sleep system that works with any bed. Pod 5 actively adjusts your temperature, elevates your body, and plays integrated soundscapes to improve. I have some new copy.

Simply too good.

Welcome to the stream, Elliot. Hopefully you're here. How you doing?

What's going on?

Hey, what's going on, guys? It's a pleasure to be here.

Thanks so much. And I love the suit. You are dressed fantastically.

Sign of great respect for culture.

I feel like there was a culture, you know, there was like a period where people would actually sort of match the vibes and have the suit for the technology brothers. And I feel like it's dropped off, so I want to bring it back, you know.

I appreciate it. I appreciate it.

You're in the boardroom, you know, it is. I'm glad to be here.

Yeah, it looks great.

Yeah, proper uniform.

I wanted to get a state of the union from you on a few things, but why don't you just kick us off with an introduction on you?

Yeah, for sure. My name is Elliot Hirshberg. I started my career as an experimental biologist. So I was in the lab trying to make new treatments for cancer. Got super frustrated, decided to retrain as a computer scientist. So I became a computational biologist and was obsessed with that as sort of a practitioner for about a decade and then got really obsessed with writing about it. So I started writing a newsletter called the Century of Biology, writing about companies in the space, data on the frontier. And then that sort of was a rabbit hole into investing with a friend of the network with none other than Paki McCormick. So spent some time at Not Boring, where I was writing and investing, and then recently joined Amplify. We just closed 900 million in new capital, including 200 million for a dedicated.

Are you announcing the new. The new capital, the new. The new fund today, or was that a little bit ago?

It was a little bit ago. 900 million, including 200 million specifically for bio that I'm helping to build.

Okay.

Yeah, you guys were not loud enough about that.

No, well, that's part of the thing. I feel like it's, it's, it's a really quiet fund where for three of the four funds for Amplify, they've been in the top 5% of venture returns. Not top decile, but like top 5%, like really good at what they do. And I feel like it's just not thought of as much and just like very quiet and stealthily doing really phenomenal work. And so, yeah, excited to talk more about it.

Okay, State of the Union on Bio. I want to know about where we are in artificial intelligence and technology helping advance bio. We've seen Alphafold, we've seen kind of tools and amazing breakthroughs. I think everyone has a really concrete idea of the impact AI is having on software engineering. Whether it's like amazing autocomplete, you have cursor now you have agents. Where are we in the deployment of AI tooling in Bio? A lot of the narrative just jumps straight to we're going to one shot cancer. And I love that. I'm optimistic. It happens eventually. It doesn't feel like we're there, but where are we actually in terms of the impact on productivity in AI With Bio?

Yeah. So you guys know like the Gartner hype cycle, right? Where you have these huge swings for a new technology where I've been working in this field for like a decade and there's been a bunch of companies, you know, Amplify invested in Recursion, which was one of the early leaders of this. Right. Like there's been this general sentiment that you can make a huge amount of progress with new data, new tools, new technology and life sciences. And it just turns out that it's like a really hard problem. Right. And it takes time, it takes some time for like the market to ingest that actually figure out the right business models and strategies. And so there was like a really strong wave of early adopters and then there was some disillusionment and disappointment where it's like, oh shit. It actually turns out that it's really hard to one shot a cure for cancer. And then as that sort of happens, there was just a bunch of breakthroughs in the technology. Right. So it became consensus that this is making a huge impact on hard problems in biology where we had the Nobel Prize for Alphafold.

Right.

So like a Nobel Prize going to an AI lab to Demis, in part to David Baker at the University of Washington, because it's actually starting to make real meaningful impacts on hard problems in biology. And so that's true for sort of molecular machine learning where you're thinking about designing new molecules and proteins. The virtual cell research where you're trying to actually model how cells behave. And so you're seeing this sort of step change now where there are a couple sort of, you know, actually faster than Moore's law curves, DNA sequencing is decreasing in cost faster than Moore's law. And so you get this huge data tailwind plus the tailwinds in machine learning and modeling. People are actually starting to scale these models and it's just like getting pretty impressive pretty fast.

Where are we on new drug discovery? Companies that were targeting a single thing. We're going to build a drug to solve a problem versus we're going to start a company that's SaaS, it's tool, it's going to help with all different drug companies. What's, what's working? What's more overhyped, under hyped where. What's your take on like picks and shovels versus drugs?

Basically, yeah. So like, short answer is we do both. So you guys had Jake on the other day at Centevax, like absolutely insane. Founder was one of the early computational immunologists at Stanford and he's making a medicine that you just couldn't otherwise make without these technologies where it's all these impacts within biotech and within modeling to make these just really incredible drugs you couldn't otherwise make. And so that's people just making these singular things that are really phenomenal. And there's also a real platform opportunity there for other things beyond the universal flu vaccine. And then I think if we take a little trip down history lane and think about how hard it was to actually sell software in the life sciences. One of the early companies in the space is Schrodinger, and they've been around for about 25 years. They're a molecular dynamics company and it just took an extraordinary amount of time and effort to actually saturate and get people to adopt the technology and get people to pay for it. And you know, they're a phenomenal business. They're a public company. They're actually vertically integrating into making their own drugs. But the thing that we're hearing consistently like from CEOs of top pharma companies is just like there's huge demand for new infrastructure. People realize this technology is here now and that they need to adopt it. They're hearing this from their shareholders, they're hearing this from the Scientists at their companies. And so there's just a very different moment post alphafold and post even chatgpt where like they're using it, their kids are using these models and they're just like oh, I really actually need to adopt this. And so I think opportunity if you're both like, like fundamentally new picks and shovels where you sort of replace experiments with compute and then also just fundamentally new drug products.

Jordy, last question.

No, I mean I think we should have you back on as new news hits because yeah, I think there's.

We'll do the TVPN bio drop in. You know I would love.

Feels like there's. Within the traditional lab labs bio has been used as almost like marketing. Right. Or even Elon yesterday saying like we're gonna discover new physics. And I don't think that that's obviously not where like a lot of the true innovation is happening. So yeah, let's make it a regular thing.

Yeah, this is great. Thank you so much.

When you think about that, right. Like there's just been a couple things like I had an investor who made a joke. Like there's just a couple things that have consistently delivered venture returns and that's like software and also drugs. And so like as far as a physical prediction goes for it being something that's super valuable, if you can make your inference be a billion dollar drug product, that's a pretty good spot to be. We're excited about it. So see you guys.

There's who. I forget who we had on when. When Trump had the executive order around drug prices. We talked to a few different people that were saying biotech has like on average been a terrible asset class. And there's some amazing outliers maybe. Yeah, like basically there's all these amazing outliers that that do deliver returns. But if you just index the market, you were going to underperform dramatically. Specifically yeah. In the venture category. Feels like this could be a massive shift where suddenly you know, the next five years become the golden age of venture bio investing.

I mean there have been some massive companies before, right. So you have like breakouts like the genentechs of the world, huge companies. There actually is. You guys should have Bruce Booth on who's an OG biotech investor at Atlas Ventures. He's done a bunch of analysis and the fact that like there's actually some interesting just return data for biotech versus tech where it's like not as gloomy as you would as you would think. But I think in general biotech's hurting right now we want to make a world where it's actually like engineering. Right. And that you're actually just getting these really scalable, amazing medicines. And I think that's where we're headed.

Incredible. Thank you for joining. Great to have you on.

Yeah.

Awesome.

Thanks, guys.

Cheers. All right, bye.

See ya.

And next up we have Kareem from Ramp. The man himself coming in to talk about the launch.

Ramp's new agent Launch. Is he in the waiting room?

We will bring in Kareem from Ramp to chat. Second time in the show he hopped on at Hill and Valley.

That's right.

First time as a remote guest.

Great to see you.

How you doing, Kareem? Hello. Great to see you guys. Can you hear me okay?

Yes, loud and clear.

I don't think you need much of an introduction, so why don't you just kick it off with the announcement and break down the launch today and then we'll have a bunch of questions. Yeah, of course.

I mean, it's been a very exciting day for us at Ramp. We finally announced our. It's our first agent. We're going to be announcing a lot more agents soon, so it's hard to keep track sometimes. We've been playing with a lot of tech internally. We think we're in a very interesting space where maybe, maybe to think about it is a lot of people from the outside look at Ramp and think of, maybe visualize the card. They think about the fintech aspects. But, but at the end of the day, what we're really trying to do is help reduce the drag on companies that happens when there's just a lot of bullshit work in between teams and a lot of papers being passed around, questions being asked, the things that really get in the way of doing work. And that first agent that we're building is really just that it operates in the messy middle between finance teams and every other team trying to spend to move the business forward. And yeah, that's basically what we launched today. So it's an agent for controllers. It knows a lot more about the expense policy of a company, the rules that are in place that govern spend, than any single employee. And it knows a lot more about every single transaction than any single person on the finance team. So it can operate in the middle and automate all the little decisions and the extra work that needs to get done to figure out what's in policy.

Or not, and it's immediately available. Like that's part of the power, right? Is that if you're, if you're in the sense of, if an employee wants to decide whether or not they can buy something or if something's in policy. You no longer have to be slacking somebody. You know, it could be the night or something like that, or off hours where there's creating that drag, that delay. Right.

100%. That's certainly one part of it. Like you can ask questions about your policy and ask questions about specific transactions, live to figure out whether they would be in or out of policy. But more interestingly, once you make a transaction, it's already doing work to go and figure out like, well, that transaction that you made at that restaurant, it looks big, but if that was a dinner with 10 people, maybe it's not as bad as initially thought. And that's actually impossible. So while that information is in your calendar, it's in your email, it's in sometimes outside of just the immediate context of the transaction. So the agent will go out on the Internet, in some cases contact vendors or pull data from APIs on your behalf to really gather all that context and make better decisions on behalf of the company.

I want to talk about the word agent and the decisions to how agents are fitting in the different stack of a tech company these days. Because there is kind of always, there's kind of always been an agent behind the scenes working. We think of these as like cron jobs before. It's like there is a long run running process that when a receipt comes in it gets tagged and there has been for I think years. I don't want to share anything, you can't. But like there's been an LLM interacting with receipt data for a long time, but it's been fully agentic in the sense that it was behind the scenes. And so I've been thinking about this in the context of like meta and like some of the value that Zuck is going to be getting from having a frontier AI model. It's like there's, there's so many workloads inside a business that has billions of users. That just happens behind the scenes. And these are agents, but they're like almost internal agents. And so I'm wondering about your decision to position an agent as like this is a user facing agent versus something that we're just going to have a process that's running behind the scenes entirely.

Well, there's a bit of a difference, right? Because when you think about these processes running behind the scenes, for the most part the code is pretty deterministic. The tools are the same, it's built for accuracy and auditability and you have a High confidence. You can trace back the path that the old school agent, let's call it that went through. Exactly. And in this case, it's less deterministic. You give the agent a set of tools, you could tell the agent, or you can essentially give it access to, let's say ability to call, ability to email. And it can be like, go figure out a way to get the receipt. That's what you know about the restaurant. And it'll browse the web and figure out that that's the phone number of the restaurant and then try to call the restaurant. And if that doesn't work, then we'll try to email the restaurant until they achieve that goal of getting you the receipt or it fails and you can then interact with it. In this case, the instructions that we are giving the agent as we're building it are very high level. You're just giving it high level instruction and access to tools. And that's very different from the old way of building these processes in which you had to be very specific about all these paths. It would take a lot longer to build these systems, to debug them, to update them, etc.

We lost you. Your, your zoom is turning like, like a ghost. It's very funny.

Super intelligence.

I think you just need a little bit more light on your face.

I think we actually, I, I actually lost power.

But wait, you lost power?

There we go. I'm back.

Okay, there we go. That's wild. Much better. I want to talk to you about the data walls that are going up and some of the battles that are playing out in like the enterprise world. Because when I read stories about companies that want to do enterprise search, you can see that, well, maybe Google doesn't want you to be taking. Maybe they want that for themselves. Ramp's in a very different position. But at the same time, there's just evolving policies about how friendly. This is a classic with Amazon not sending the itemized reception receipts to Gmail because they just didn't want to give Google the data. But as a ramp customer, I want the Amazon details pulled in via Gmail, via the ramp integration. So talk to me about, like, how's the broader trend playing out? And then how do you go to big companies and say, hey, like, you know, work with us? Our clients want to be able to pull data from your service and we're not going to build a delivery network Amazon, so we're not a competitor for you.

Yeah, for sure. Most of the data that we need at the end of the day is data that is quote Unquote, owned by our users, the businesses that are on ramp their employees. I think it's a little bit easier to operate in the B2B space because like those, I guess what governs who owns the data and whose data it is is a lot clearer than in a lot of consumer applications. So in our case it's like what data do we really need to know in the case of the agents that we just launched to figure out whether something's in policy or not? It's metadata about the transaction. Right? Like what's in the receipt at the end of the day, like stores all your receipt, that's your receipt. Right. We get that information. You have information that we get through the networks through Visa. The metadata about the geographical location of the transaction, maybe whether it was an in person transaction or not. There's data that's in your inbox, in your email, which again like that information is also owned by the company. We haven't really encountered a lot of pushback and challenges. I found most of the challenges in getting the data to be more like technical. How do you make sure you get it quickly, clean it up and get it accurately as opposed to ones where there are third parties that are trying to make it harder and harder for us to access the data.

We kind of been in that, in the previous company, that was kind of the stuff. Previous company, yeah.

Of course we had a lot of these problems. I mean there were lots of funny moments at Paribus or previous companies where we were. I mean, we're really building an agent for consumers to help them save money on their online shopping. Right. And we're trying to log in on their behalf to Amazon accounts and Walmart accounts, et cetera. And of course they'll put blocks, they'll put captchas. And today those captchas seem like a joke. I think any version of any half recent version of ChatGPT or Claude is able to solve those captchas very easily.

That's one of the ways the Internet's getting worse right now is the captchas are actually getting so hard in a while.

When we go to the gym in the morning, Jordy has to log in like two minutes to get through the captchas for this gym.

It has the most military grade security to get to a gym login and.

It just gives you a barcode that you just scan.

But on that I'm actually interested because I can imagine Ramp has tens of thousands of customers, like high value business customers and other people that are building agents. I'm sure would love to actually be able to make actions on the ramp platform, but at the same time you guys are trusted to handle the, basically the finance back office for these companies. You don't want like an agent, like hallucinating, like saying, you know, based on and taking actions on the ramp platform. So I'm curious how you see that dynamic playing out, because I'm sure you've been approached by a lot of companies saying, hey, we're building this agent to do this thing. We'd love to be able to, you know, get authorization.

Of course, I mean, we're thinking through that a lot right now. I think there are good ways of exposing the right information to the right agent as long as our customers are very aware of what they're exposing. And there are lots of interesting applications for us to work on. Like in the case of any large purchase at a company, there are multiple parties within that company that need to review it or approve. Like you want to review a certain vendor and look at their data protection policies, you want to look at the legal agreements, in some cases you want to negotiate the price. And you can imagine a day in the future where a lot of our customers have an agentic tool that they trust or agents that they trust for legal work, agents that they trust for IT work, etc. And we're very interested in actually working with some of these companies. But we got to figure out on our end how we expose the right interface so that we're ensuring really the security of the data of our customers. So it's an ongoing work stream.

Last question for me, the Grok 4 launch was very benchmark heavy. It seems like the consensus is that it's a good model. And so if as soon as I see that it's now about cost per token. And so I want to hear from your perspective, what drives decision making? How big of a line item roughly? Or how much time is spent thinking about LLM inference optimization at your scale? Like roughly how big of a deal is it? And then what is the workflow to decide? Can we use a cheaper model? Do you have internal benchmarks? Are you just checking these things? How are you making decisions about which model to use for what problem?

Yeah, that's a great question. I mean, I'm a lot more paranoid about being too slow to try the newest model and the latest and greatest tool than I am by maybe overspending a little bit in one area. I mean, the amount of time and money wasted at companies doing BS work is just insane. That if we're Debating whether you can make something faster by spending an extra dollar or half dollar. The value that we're able to create is so big that I don't worry about it too much. But we do have internally somewhat imprecise like stack ranking of the different places where we need to make inference calls. And in some cases they're very simple, high volume, kind of low risk.

Right.

Like you're trying to normalize or clean up some like merchant data to figure out the appropriate spelling and maybe the right like photo to use. It's not the end of the world if it's not like perfect. We're doing it at high volume. It better be cheap. So we have a kind of stack ranking of like this is something high volume where we need to be cheap. This is something that's low volume and high stakes where you need to be accurate. And we'll generally try the newest and greatest models in the places where we think will make the biggest difference. And over time we'll break up some workflows and some parts of it will become cheaper, more repeatable with, with smaller versions or cheaper versions of the model and it will just evolve. I remember micro optimizing every single thing on our AWS account back in 2014. It was a lot harder back then. I think. We also pride ourselves in being the time and money company. So we do care a lot about making sure that we don't waste our own money and our own time. But. But I would say that the TLDR is like our time and engineering. Time is the most valuable thing here and I'm a lot more focused on that than anything else.

Yeah. On the time issue, what do you think about the various latency trade offs? I'm sure if an employee wants to know, is this in policy? And you hit O3 Pro and it waits 10 minutes, they're probably just going to slack their manager and ask them, but you're going to get a really accurate answer that's really detailed. And so how do you think about those trade offs in latency?

Yeah, I mean it really depends on where in the workflow are we making the inference call. If it's live in the interface and the user expects a quick answer, we'll be using some of the faster models. But the reality is a lot of these agentic workflows that are being kicked off at ramp happen behind the scenes. Make a transaction, you maybe get a very quick question from Ramps AI to gather a little bit more context, like that's enough. And then from there we'll kick up another, kick up, kick off another task that can be a little bit slower. They'll happen in the background and by the time it reaches a bottleneck or it'll reach a place where it needs like additional feedback, it'll be in someone else's like notifications or on someone else's slack. You could take a little bit of time when the work is going from one person to another person, but less when it's the same person interacting with the interface live. That's generally the thinking Cool. Yeah, I have. I tried some of the newest browsers today and I tried Comet today, I tried DI a couple of weeks ago and I think what they're trying to do is incredibly cool, but I often find myself thinking like, damn, I wish this was a little bit faster. I know it's coming, but I think unlike some of the browser agentic calls, you want it to be really fast.

Yeah, I was thinking about that in the context of the OpenAI browser and unless they figure out something that makes it basically 10 times as fast, I'm still going to default to Chrome if I have both of them open just because I'm like, well, I just really need a fast answer here.

I mean that was Chrome innovation, right? Chrome won on speed. They just optimized the code and they nailed speed and it was enough to leapfrog and so you could see. I mean that's like the bull case for like Apple coming from behind is like, yeah, they like, it feels like if X, AI and Anthropic and OpenAI are all kind of like Gemini are like roughly at the frontier. If you can just get something that's at that frontier, not any new innovations, but hyper optimized and it runs locally on your phone and it's spitting out like tons of tokens every second. Like you have a product that would be very, very. It would be very rapidly adopted. It's exciting.

It matters a lot. I mean, I think one of the weirdest ux patterns on ChatGPT now is that I have to do the work to figure out whether to use O3.03.

Or 4.0 every time. Do I have 10 minutes or do I want the. And 4.0 is always so good that I usually don't need to, but then I'm just like, well, I want the best. Of course I'll come back to it. And it's such a weird paradigm. It's going to be something that dates us and I just know our kids are going to be like, what did you have to do back Then you had to rewind the VCR tape. You had to put the help in the X. You had to pick which model to use. This is insane. It's so legacy and it's going away, but we're just in this weird. Like, we don't have a model router solved. And it feels like the easiest thing is, like, which model should we use for this? I don't know. We'll see if.

I mean, I don't know if you guys. I grew up in Lebanon. I still remember the days of dial up where you would have to.

Kick everyone else off the phone line.

Exactly.

Well, select the phone line in our case, like, okay, which phone line am I going to use? I don't know. Can't you tell me which one is free and pick it for me? No, I still have to, right?

Yeah, it seems like the easiest thing to do. And also, I mean, this is just complaining about the app that I use 30 minutes a day at least chatgpt. But I almost wish I could just define it in the prompt and just say, hey, use O3 Pro. And then here's the prompt as opposed to needing to click the ui, change it, switch it and then pick instead of just being able to go back and forth. And I don't know, I mean, it's a good sign because like, like, people are using the stuff so much that they're frustrated by these, like, niche UI things things. So it's exciting.

There's a lot of. I forgot who it was who posted on X. I think it was like a couple of weeks ago that every company is one great UX breakthrough away from something amazing. And I think that will be true for a long time. There's a lot of alpha right now and just great UX and good patterns. We haven't figured it out. We're still in the maybe terminal phase of personal computers. Like, when is the mouse going to come out? When are the right GUI is going to come out? Like, there's a lot of that happening right now. And yeah, it's a fun time to be building.

One last question for me. On Monday, Dwarkesh released an article and then came on the show, kind of talking about his timelines around when an AI agent would be able to do his taxes. Right? Sort of like basically fully agentic experience being like, I want to do my 2025 taxes. And then it just sort of autonomously runs. How do like big. How are like, you know, Fortune 500 CFO, like, what are, what are their timelines around? Maybe you just Tell them what the timelines are like. Okay. By 2028, you know we're going to be able to do this for you. But how is the sort of finance arm of the C suite kind of anticipating, like the rate of advancement? Obviously, the agent today is a step towards that future, but you'll obviously need a variety of different agents.

Well, I think in terms of capabilities of LLMs, we're there, we have the capabilities. The bottleneck on being able to do this today is having the right context.

Right.

It's, it's. Well, some of that context is in my head. So the AI needs to know to, like, ask me the right questions efficiently so I can answer those right. Even when I'm working with my accountant. Like, pick the best accountant in the world for your personal taxes. If you just tell them like, file my taxes, they can't do anything. Maybe you tell them, file my taxes and here's access to my email. They could do a little bit more, but they can't get it fully. Just tell them like, file my taxes. Here's access to my email. You can call my wife as much as you want. You can look through my drawers and you give it more and more of these things. Maybe it could do it, but it's going to get lost, it's going to take forever. And really what we need to do, even for businesses, is what are the right patterns for us to extract context that's in people's head, organize it, get them comfortable with connecting different rules like your inbox and things of that nature. And I think in terms of tech and capabilities, we're there, we're not really missing anything. So there's a lot of UX and.

Classic, like product building agents that can email me a question and put it in my inbox, which is effectively my to do list. And that's what my accountant does. When that access happened, they email me.

And say, well, that's why this, you.

Got to do this.

You can just, you can take a picture of a product and ask it if I buy this, you know?

Yeah. Is it in policy?

Yeah, 100%.

Yeah. Well, thank you so much for stopping by. Do you have anything else? This is great. We'll definitely see you soon. Kareem, this is great.

Congrats to the whole team.

Thank you.

Talk soon.

Talk to you soon. Bye. And that is the rest of our guests. We are through that. In other news, Periodic Labs, there's this, a scoop from Natasha Mascarenas. The startup being co founded by Liam Fetus and Eric Dogus Kubuk Great Names is in talk to raise hundreds of millions of dollars in funding at above a one billion dollar valuation. The two month old startup is looking to apply AI to physical science, starting with discovering novel materials.

Let's give it up to the two month old year unicorn. We got to have these guys on.

The show extremely fast. I also like this post from David Perel where we're getting him back on the show asap. We had a lot of fun talking to him a couple months ago. He said, I'm touring apartments in New York and just about every new build has the same soulless aesthetic. Flat walls, white paint, no cornices, no ornamentation, just a room in a box. Only one real estate agent said to me, if you want something with character, you're going to have to stick to pre war buildings. Look, look, I'm all for some efficiency gains, but we've created a world where new things are soulless things. And that's how a society as modern as ours, and that's not how a society as modern as ours should function. Intuitively, you'd think that a wealthier society would build more beautiful things, but not ours. And I completely agree. What's crazy is that this isn't. I mean, I don't know, these apartments look nice, but this continues on all the way to $20 million. Houses that are still bland. And I think it's mostly because maybe time and all the difficulties with permitting because if you even have the resources to build something from scratch, creating, okay, I want these ornaments and I want this and I want something that's really expressive of my personality. Well, now you, if you want that, no one else wants that. So you have to build it and you have to, and you have to, you know, underwrite it and you're going.

To be underwritten code.

You make sure it's to code and then get it built and, and then, and then the secondary market value is gonna be less because not everyone wants Hearst Castle. Whereas if you build, if everyone builds the exact same thing, they're perfectly, it's perfectly liquid market because every, every apartment is interchangeable with every others.

Yeah, so good point.

It's kind of a, it's kind of a function of just like modernity, but it's more a function of people not, you know, just risking it ever on building a disaster project, making their forever home. People learn the lesson of William Randolph Hearst too much. They should have just like never learned that lesson. Send it, ripped it and just send it and just build something that no one else will want to buy and we'll take decades to build. That's always the best.

Well, I have a good place to end it. Rob Petrozo says the original Hermes Birkin bag prototype just sold for $10 million in Sotheby's. There was a two minute standing ovation. He says bull market confirmed. And a God, we love a bull market.

The original prototype. Fascinating.

That's wild. Makes sense. Very cool. Its incredible lore and I would be excited for a bull market and alternative assets such as Birkins. Be great and you should be too. But that's a great show, folks. We will be back tomorrow. I cannot wait.

We will talk to you tomorrow. Have a good day.

Cheers.

Bye.