Stripe Buys OpenRouter, Grok Bot Arrives, ChatGPT for Teens & AI Designs Proteins | Lightwork

0:00So basically it's a teammate and I think that's the potentially scary unlock for a lot of humanity is we've all been talking about AI being like so helpful and assisting us not yet getting to the point where it could just do everything that I do in a day. It's a little bit like I don't know the cash register for AI. It is very much so position less as a chatbot and almost like a roster of employees.

0:25Josh Machiz: It's like, I am Groot, but the AI version.

0:36Josh Machiz: We're back with another big week. Stripe reportedly bought three-year-old startup OpenRouter for more than$7 billion. SpaceX closed its$60 billion acquisition of Cursor and shipped GrokBot. OpenAI launched ChatGPT for teens. And Andropic ran a really cool experiment in the drug development space. Plus, two updates from across Lightspeed Universe. Let's get into it. Some big acquisition news. Stripe is reportedly acquiring Open Router for$7 to$8 billion. This is huge news for both venture returns and for this company. Yeah, just three months ago, I think they had raised$113 million Series B at a$1.3 billion valuation.

1:24Yeah, they've only raised$164 million total.

1:27Josh Machiz: Yeah, so this is a huge, huge, huge markup in just three months between their Series B and now a$7 billion exit. Yeah, it's funny. I think people thought that, you know, there weren't maybe venture scale huge opportunities like this in the new AI era without the models. But this is kind of showing that there are these big opportunities like this is only a three year old company. And now we have, you know, a seven or eight billion dollar exit. Yeah. And I think we're going to cover another one later in this episode as well. But it was awesome to see so much activity between this, Stripe, and OpenRouter, and SpaceX AI, and Cursor.

2:07Josh Machiz: Just lots of activity. And I think the founder of OpenRouter is a guy called Alex Atala. And he's actually also the founder of OpenSea. So this is sort of the second kind of marketplace play that he's built. It really goes to show those serial founders know what they're doing. And he probably learned a lot of lessons from OpenSea and that whole saga. Yeah. And for reference, OpenSea, for those who are less familiar, it is basically a marketplace for NFTs, but just awesome. And then it became something else and then it's back. Yeah, I haven't been tracking, but I think to your point, what's very clear is that great founders will build in important spaces and you can see OpenSea was his kind of crypto play.

2:49Josh Machiz: and now in AI, OpenRouter. So tell us about OpenRouter and why did Stripe acquire them? So the simplest way to think about OpenRouter is it's an AI model gateway. In this era, you can also sort of think of them as a toll booth. So if you think about the way that a traditional developer is able to access and serve their requests for model providers, it means building a lot of very different integrations. So say I want to plug into OpenAI, it means that you have to build for that API, I have to build a lot of unique integrations for that. All of these have different data structures, different token counting methods.

3:28Josh Machiz: And then I also want to plug into DeepSeq, and I also want to plug into Anthropic, and I want to plug into XAI. All of these are very different. And so OpenRouter came in as a sort of normalized access point where you just integrate with OpenRouter once and you can access 500 plus models and you point your request to OpenRouter and it routes it accordingly to whichever model based off of whatever parameters you have. So if you're optimizing for budget or you're optimizing for latency, you can actually optimize it even for workflow. So if you're running a coding task, it'll route you accordingly to whatever model is ranking highest in that function.

4:09Josh Machiz: And what's cool is as a CFO or someone who's managing the token budget spend, as we've discussed extensively on all things token maxing, you can actually put a hard budget and hard cap. Another cool feature is, for example, if a model goes down, they will automatically route you to another model so that if you're a user, which we know that sometimes a lot of these services go down. Yeah, models go down and it happens. And what's nice is if I'm a consumer or if I'm an app builder and I have a bunch of consumers, they don't feel that at all. It just automatically with almost seamlessly will switch to another model.

4:44Josh Machiz: So it's sort of this kind of model gateway, but also a toll booth. Yeah, kind of makes sense. Now I kind of get why Stripe acquired them, right? It's a little bit like, I don't know, the cash register for AI and, you know, just another place for Stripe to own a little piece of this like infrastructure. layer. Yeah. And we could talk extensively about why it makes sense for Stripe to lean in here or what I would assume, you know, we don't know, but what we can infer. But if you really do think of this as like the Expedia for the AI era, in the same way that if you're a consumer and you don't want to log into 50 different accounts to check your information, book flights, this is, again, that centralizer for developers where you can in real time have access to all of these different models.

5:33Josh Machiz: How they make money, though, is they take a 5 % to 5.5 % cut on all-inference traffic. So that means that if I'm spending a million dollars on Anthropics models or OpenAI's models, all of that will still funnel to the model providers accordingly. But OpenRouter will charge a 5 % fee of$50 ,000 on top of that. And so that's how they make money. So they're essentially a tollbooth. It's a pretty awesome business model. I mean, Stripe was already close to like where every consumer makes like a purchase, right? So it's just moving it down, realizing that agents are going to be making all of these different transactions and just another place to own that little tollbooth, I guess.

6:16Josh Machiz: Yeah, for sure. I mean, Stripe, as we know, is the payment processor for humans. And so, yeah, the internet. And this is a very natural extension where you have OpenRouter being the closest point to where AI agents are transacting and the ways that AI agents are plugging into different model sources. And that's really, it's really helpful to kind of have visibility into both. I will say I do have some questions about how this space will continue to evolve. I think one of the tailwinds for me is that, you know, what is increasingly becoming very clear is that we are moving away from this one model world and living in this multi-model world where if you are a startup or building an app, you're not going to default run every single query through one model.

7:04Josh Machiz: You are going to accordingly route them based off of the task. So for example, if I'm a company like an airline, I'm not going to route millions of customer service requests that I get every single day into OpenAI's best model or Meta's best model, what I'm going to do is route all those hard, difficult problems, throw them out the best models. But if it's something a little bit more routine, like summarizing or just simple Q &A, I'm going to want to maybe use something a little cheaper. And as we've seen with, you know, the advances in a lot of these Chinese models, I think what we're seeing is at the application layer, a lot more diversification.

7:41Josh Machiz: And OpenRouter is actually a really interesting place to track that. I think there was an interesting stat that the share of U.S. hosted models, developer traffic on open router went from about 70 % to 30 % over the last year. And so what you're seeing is also this mix of like close plus open, which is also another interesting phenomenon to track. So it'll be cool to see how it all plays out, but awesome to see, you know, great exits for great entrepreneurs. And I think what we're also really seeing and the takeaway for me here is that, yes, there's so much activity happening at the model layer.

8:15Josh Machiz: It's important. But you're also seeing all these players benefit from building the infrastructure around it to support all these models. And I also think in reaction to this will be interesting to see how the models and the labs continue to also iterate. I think you're also seeing them move a little bit upstream into really targeting building coding agents or really building the harness layer so that they can capture some of that value. They don't want to just be a commoditized model that anyone can switch easily between. mean, they also want to provide a product layer around that. It's a lot developing in real time, but it makes sense.

8:51Josh Machiz: And I'm excited to see how this partnership plays out. Yeah. I have a question for you. Do you think Stripe an independent enough player to be acquiring them? I always think about what happens after the acquisition. Are there potential buyers that would feel that it's not as independent as when OpenRotter was its own an independent company? I think that's a really good question. My sense of Stripe is relatively neutral. Like fairly neutral, right? Yes. Like if this is not meta, this is not one of the labs acquiring them, right? Like maybe there's a few payment processor companies that would be turned off open router, but probably not many.

9:30Josh Machiz: And I think with Stripe, they have been kind of building the stack. If you look at their history of acquisitions, this is sort of makes sense in that stack. It's not really them, I think, trying to have preference for any one model provider. They've really wanted to be the neutral player that, you know, I think they often thrive in these tollbooth-style businesses. And I think a big part of having a tollbooth-style business is you're not preferring one type of car or one type of vehicle. You are there to basically make money off of every transaction. And I think it is in your best interest to be neutral.

10:05Josh Machiz: I do think for For people and developers using OpenRouter, though, I think the space is changing very quickly in that, you know, for OpenAI, when they launched GPT-5, they took some of that routing part into their internalized that. And so, for example, for GPT-5, it'll actually base off of the conversation type, the complexity, the tool needs and the intent. It'll actually route you accordingly between, you know, a cheaper model, a more simple model versus something that is a little bit more computationally intense. And so you're kind of seeing some model providers also build some of that internal tooling.

10:43Josh Machiz: At the same time, I think at the application layer, the developers who are using on the other side of it, I also think they're figuring out how to think about... Use different models for different tasks and different parts of even one task. It still feels like Open Router has a pretty decent business model here. Yeah. And I think for these apps, though, I wouldn't say that routing is the most technologically complex thing to build. So my big question going into this is there are some open source options. Do apps as well want to internalize some of this routing logic? Because I do think one interesting thing to think about is if you are an application using OpenRouter, that means that Stripe and OpenRouter has visibility into a lot of your unit economics, which is also I think, you know, I think for certain companies who view that as a proprietary edge, you don't necessarily want to be giving away that freely, especially if it's not too hard to build some of the routing sacks.

11:41Josh Machiz: So yeah, I'll be interested to see. I think for now, OpenRouter serves a very clear purpose. And I think it makes a ton of sense why Stripe would embed them into their offering. But at the same time, I think there's still a lot happening at both sides where the models are building some of this. And at the application layer, they're also bringing some of this in-house. Still a very exciting acquisition, maybe moving topics from one acquisition to another. We had a rocket company finally close its acquisition of a code editor in Cursor. SpaceX acquiring Cursor feels like a very exciting moment. There's been some great products that they have already worked on.

12:21Do you want to talk to us a little bit about the deal?

12:23Josh Machiz: Yeah, we've discussed this, I think, quite a few times over the last few months. But for a rehash, SpaceX has evolved a lot in the last couple of months. It's embedded XAI into it. It's running compute clusters. It's a rockets company, but also has X. The social media platform also has XAI. So it's a very different animal than, you know, what the pure rocket company that existed. And you can sort of see how all of these have become this kind of vertically integrated AI company with a space business attached to. But basically, SpaceX announced that it would acquire Cursor. And back in June, I believe, it was still being teased.

13:11It was around the timing of the IPO, and it was basically a promise to purchase rather than a straight-up purchase.

13:19Josh Machiz: Exactly. And that's what had been teased. And so leading up to this, now that we finally have that being closed in the last week, it's exciting to finally see it formally become a partnership. But the logic, you might be wondering, is, okay, why in the world would a space company... Why did the space company acquire a coding company? There's a lot of reasons why those partnerships make sense. But if I could distill it into the most simple terms, it's because SpaceX AI as this massive entity with Grok, with XAI, X, has never really had that strong of an enterprise foothold. You know, most of its distribution was coming from consumers through X platform.

13:56Josh Machiz: People were using Grok as an extension of X, but it never really had massive usage in enterprise. On the flip side, you have Cursor, which has very strong brand and reputation in enterprises. And so in that sense, partnership makes sense. For Cursor, some of their big issues leading into this was, one, their dependency on... Lack of compute, lack of power, all the things that are the bottlenecks for these companies. Yeah. And one of the biggest questions at the time was, what is Cursor going to do in the era of Anthropic and OpenAI? also building coding agents and also having strong enterprise presences.

14:36So there was that

14:37Josh Machiz: question on dependency, also question on training its own models, needing compute. And obviously SpaceX AI had built Colossus, these two massive clusters that weren't necessarily being used at max capacity. And so sort of felt like there was a marriage where they both needed something from each other. And obviously for Cursor, the company that it is, it's an amazing exit for that team. I think it makes a lot of sense. I'd be curious to see how all of it comes together because to what you brought up, I think you'll probably have a lot of insight into the branding around mergers like this. The cursor brand, Elon usually does not, but it's seemingly sticking around and it's such a goaded team.

15:21I feel like he's actually doing a pretty good job.

15:24Josh Machiz: On the cursor side or on SpaceX's side? I think on the integration, it seems like the teams so far, especially with all the noise in the timeline this week around GrokBot and that being a really powerful innovation, lots of people, at least on Twitter, are saying that they're using it and it's incredible and it's saving hours of their lives and that it's truly like a teammate. Yeah, that's the kind of innovation that I think feels like it came out of, you know, we talk a lot about the value of the data, like cursor has all this data of software engineers training AI to be better at coding. like all of those learning loops translates to i'm sure building some you know better models better products and especially in this agentic future creating this new grok bot product even though it is under the grok brand and not the cursor brand that was my like marketers like why because i just think cursor is such a strong name like it it means what it says it feels so self-evident.

16:30Grok, you know, it has some rough edges to it as a brand name. And I was curious to see that, but it seems like they're doing a really good job.

16:39Josh Machiz: Yeah. What do you think reactions have been to GrokBot? Do you think they should have positioned it differently? Like if they had called it CursorBot, would I think it might have been a little bit more palatable? Yes. It's like, I am Groot. But the AI version. But the AI version. But also at the same time, Elon's proved time and time again that they will people will buy his, you know, what he's selling, whether it's a Tesla or X or, you know, people don't leave. People trust a good product and he makes a good product. I actually got in on the note of naming things. I did get in a little bit of a debate with someone on whether or not bot was the right terminology for this when you had agents, assistant.

17:24Josh Machiz: And I'm curious if you have thoughts on whether bot is the right positioning. But I actually kind of liked it because I think what we're seeing with GrokBot, and we'll get into the features and why it's huge. I do think actually bot is quite palatable for a everyday consumer because I do think at least my sense is for a lot of people, agents is still confusing terminology. And then assistant, I think, can mean a lot of things. And so I don't know. I actually don't mind the bot language. Yeah, I do think the XAI design team also cooked on the logo. I really liked the little ghost. It's both endearing, but also like looks like it's getting work done.

18:01And so far, people are loving it.

18:03Josh Machiz: To clarify, though, should we talk about what GrokBot actually is? Yes. So GrokBot is essentially an agentic platform that can do a lot of the work that you need done on a day-to-day basis. So it sits on your computer, and I think it costs$300 a month currently. You need to subscribe to GrokUltra to get access to it. And once you have access to it, it can kind of pick up a lot of those tasks. I saw a lot of people say I had Claude making me all these different skills that I don't need to necessarily pay as much attention to as I did because GrokBot is now just doing them on the back end. So basically it's a teammate.

18:47And I think that's the big, you know, potentially scary unlock for a lot of humanity is we've all been talking about AI being like so helpful and assisting us, not yet getting to the point where it could just do everything that I do in a day. and I'm seeing people say, you know, 25 % of the things I do in a day can now be done by CrockBot. And that starts to put it in the category of coworker and not assistant.

19:17Josh Machiz: One thing I thought was interesting is it is very much so positioned less as a chatbot and almost like a roster of employees. From a more technical angle, actually each bot gets their own persistent cloud computer with a real browser, its own file system and terminal, its own logins and tools it needs. And it also means that all these agents or bots can work even after you shut down your laptop, which is actually a big difference, at least for me. And what's cool is all several bots can run at once and they can actually coordinate, which I think was interesting. So I think, yes, we've had a lot of tools come out.

19:54Josh Machiz: You know, this is almost like the fourth big one to come out i think between claude co-work chat should bt work gemini spark is google's yes agentic assistant and then now you have grokbot but what i think here is that it's not just about completing multi-step tasks but rather kind of creating a roster of different agents that tackle different workflows i think another really cool feature i thought was interesting and you kind of see it with other uh the other players as well they've also built similar record a skill or watch me do my work features. But GrokBot also has a teach a task feature, which is basically where you perform the action only once.

20:33Josh Machiz: So for example, if you perform an action of any time, Josh, post something, give it a like or comment something, you can perform that once and then the agent will record that sequence and then turn it into a skill on its own. And if you use it and you tweak it, it'll actually update itself accordingly, which I think is really smart where we are increasingly, to your point, moving away from these systems that you have to handhold so much step by step. And correct them along the way. It's just watching you make the corrections. Exactly. I think that big step change going from reactive, taking step-by-step instructions to truly something that can watch and observe and infer based off of how you work, that I think, for me at least, is going to be huge.

21:17still worried about the runaway train effects like you know things sending hundreds of emails and all of those things are still so scary have you ever run into any of those issues of any of these i have i have not but i've had plenty of friends who have you know lived out their open claw lives

21:34Josh Machiz: get into a lot of trouble in my head it was like open claw is the most you know developer native you have to kind of know what you're doing there i think clawed co-work is like a little bit in this in-between yeah you can still get into trouble yeah and then gemini spark i felt like was the most traditional consumer friendly i'm curious to see where grothbot fits in this i would probably categorize in this half of the spectrum where it's somewhere between open claw co-work users uh and that being sort of the target demographic but i'd be curious to see how you know also if there are any everyday consumers being drawn to grothbot what how they're using it, what their value they're extracting from it.

22:16Yeah, I'm very excited to use it. So Claire, OpenAI announced ChatGPT for teens. I think this was a surprising announcement. Really? Yeah. I didn't expect them to be so fast at this. This feels like OpenAI for me is learning the lessons from the social media era where it took meta so long to understand that Facebook and Instagram might not be giving the kids the things that they need and might actually be causing harm. I'm actually surprised that they were so fast to innovate in this way and that they actually are saying, okay, yes, Silicon Valley did this wrong once before. We're learning our lessons and we're applying them with this announcement.

23:02Josh Machiz: That's interesting. I almost kind of expected them to have a kid or teen product. And maybe this is coming from because I've been tracking so many education and learning tools already and have had touch points of a lot of the team building out study mode. And I think with all of these chatbots, whether it's Cloud, Gemini, OpenAI, it's very natural to see them being used as learning tools. However, you are seeing that social friction where we know that in its current form function, these tools are not built for young humans. Yeah, they're not built for teenagers. They're not built for children.

23:39Josh Machiz: Yes, I think there's some big stat from Pew that says that a third of teens use AI chatbots daily, which is huge. You are seeing this first generation grow up on AI. So naturally, it makes sense to have teen mode. To go into what's different about teen mode. Yeah, what is different about it? It is going to be a distinct user experience built specifically for people age 13 to 17. They're going to be delivered to you if they know that one that's sated by the user directly. Or it'll actually try to infer your age through papers. So apparently they track more than 2 ,000 signals, like your login time, your account age, and they will flag if you seem to be a minor.

24:24Josh Machiz: So even if you are a minor pretending to be a 20-year-old or a 30-year-old, it'll just, based off of your conversations, start to infer and it'll flag your account accordingly. Or Minecraft tutorials or whatever. Yeah, but I was like, it got me thinking a little bit. I'm like, how old do you think Clotter or OpenAI or Meadow or any of these tools thinks I am? It's like those guess my age prompts from back in the day. But they actually have so much data that I think they can do a pretty good job of that. But basically, once it detects whether or not you are under 18 or if you've told it that you're under 18, it will introduce a couple of different features.

25:02One of the big differences is that if you are trying to shortcut an assignment, it will redirect you to study mode.

25:10Josh Machiz: So if you're trying to say, just give me the answer, it'll say, nope, why don't we try with some of these quizzes or I've built you this learning visualization. parents can specifically also schedule times when study mode is on by default and then there are all the natural things that you would expect like age-appropriate guardrails and safeguards like having conversations in high-risk topics like anything to do with self-harm violence eating disorders explicit sexual or graphic content so really building stricter guardrails around that And then I thought other behavioral things like you actually get nudged to take a break after 90 minutes of activity within a three hour window.

25:50Josh Machiz: It will also introduce cues and repeatedly mention that I am not an AI. I am not going to build emotional dependence with you, which I was I thought that was interesting because I was like, I feel like we should introduce that for adults to that. Like that shouldn't. Yeah, I feel like that is already happening with adults, right? There is this whole set of issues that affect all of us with AI that I don't think are being addressed. And maybe that is one of the unmet needs that is still in these products. You're seeing people start to be like, you know, especially in this new post-COVID era where people, I think, are further isolated with Zoom and all of these other ways where we kind of don't have to connect.

26:34and we're looking at our phones and like we can just chat to the AI bot, like a reminder that this isn't human and a replacement for human connection feels like a necessary one for adults as well.

26:45Josh Machiz: Yeah, I mean, I think obviously someone who maybe leans more wanting to build safety guard rails, like I would expect that ideally we are reminding users that these systems are AI, that they are not real and that they don't have feelings or consciousness. Like I believe that that should fundamentally just be baked into the ways that you interact with these models i mean the the incentives on the other side are to keep you logged in and making you know continue to use the platform and you know yeah but i also recognize that like i imagine there is a cohort of humans who do feel like okay if i have freedom of choice and ability and speech to engage with this system however I would like.

27:29Josh Machiz: I don't want some higher power or corporation to dictate how I can use these tools. For sure. But I don't know. My sense is, I think ultimately, there is always going to be a choice made here. And I'd be curious to see how these continue to develop. I think this is also coming at a time when there are also external regulations coming. So there's something called the GARD Act coming out of D.C., which would ban AI companions for minors and require age verification. So this feels like OpenAI just getting stuff ahead of that. And then in January of this year in California, something called SB243 came into effect, which basically, to what we were discussing earlier, requires chatbots to disclose that they're not human, which, again, feels like table stakes.

28:17Josh Machiz: But yeah, I think you're also seeing these external shifts as well. And I think companies are almost preemptively building in the way that I think YouTube decided to build YouTube kids, meta, Instagram teens, all of that. I think this feels like a product extension or surface area that makes sense. Yeah, it does feel like it's really good that they're getting ahead of this before there are congressional hearings like there were the last time. I also think there's something interesting to think about, which is like, if in the Instagram era, we were always worried about like what kids are viewing, what they might see that they shouldn't see.

28:58Now we're kind of worried what AI might be telling them to think, which I think is a little bit different, right? And it could be changing the way, you know, children feel on a topic or on a government. It could really be very interesting the way it shapes humanity. Not that AI hasn't already.

29:19Josh Machiz: my question is you have all those second order effects that you just laid out that i feel are very important to continue to have discussions on and and my hope is that we continue to do research on things like metacognition and whether there is cognitive offloading or the ways that engaging with ai changes our ability to have you know simple discourse like i think these are all important things to recognize but i think my fear coming from that is if you are one of these model providers yes you can build all these guardrails to your best ability but it's not like if i am a you know a teenager if i really want everywhere and there's so many different ones it feels like you've seen the ways that like i've seen people kids are so good at accessing the internet and now it's even easier yeah no you've seen people jailbreak chipotle's ai agent or ai assistant to start be used as like a regular chatbot or snapchat has snapchat ai baked into it so i think there probably needs to be some holistic discussion where it is like a full scale analysis rather than i think just all of these individual model providers or companies It sort of feels like it's a job for regulators.

30:37Josh Machiz: Yeah, and we'll see. And I think there is a question like how much you regulate, how much not, but it's all up in there. But my question is, yes, as much as you can enforce at an individual app layer, it does feel like there are some of these sort of second order effects that will play out regardless of building strict guardrails. It feels like every single week we have a new experiment coming out of the labs and new discoveries. This week we had one very cool experiment from Anthropic. Yes, Anthropic had a pretty amazing announcement. In one experiment, CLAW-designed protein binders were tested against 15 targets, and they produced binders against 14 of those 15.

31:22This is usually a very manual thing in, I think, most science labs, even with some AI tools. So this was pretty amazing. Yeah.

31:32Josh Machiz: And to distill it into why it matters for you and me as regular humans, the basic gist of it is that this is, as you mentioned, a really hard science job normally to be run by a group of humans. Specifically, the job is to design tiny proteins from scratch that grab onto specific targets in the body. This grabbing function is actually how a lot of drugs work, is that they latch onto something and they switch it on and off. That's how a lot of drugs work. normally as you mentioned this takes scientists weeks and weeks or months even per target and to what you mentioned it was able to do this on its own and outside labs physically built it and then they test it and then the results prove that they work and this isn't just like proving that they work in theory this is like real yeah about 27 percent of the designs worked 14 out of 15 that actually is really impressive because the human success rate in this field is usually 10 to 15%.

32:30Josh Machiz: So roughly double what humans are able to do. As we know, AI has been helping to design proteins for a very long time now. This is not completely new. We've had things like AlphaFold, RF diffusion. But the new thing here is that Claude was able to run the whole process itself. So it was actually almost like a workflow experiment as opposed to just whether or not we could do it. Was Claude using a bunch of those different models in concert? How did it do that? I'm not sure, but what Claude was able to do is, one, research a target, pick the strategy, pick the right tools to use, and then choose the winner.

33:07Josh Machiz: And all of this from one set of written instructions. So that's very different than, I think, AllFold is almost like one tool, which is amazing. But what was cool here is that it ran the process on its own. That being said, I think I want to be very clear that this is impressive on a technical level, but it doesn't mean that people are running around creating new drugs or that random people in their basins are solving diseases. Rather, it just shows that there's going to be a lot of efficiencies in the bottleneck. And this is important to scientists, to researchers, because often the bottleneck in science isn't the ideas part.

33:49Josh Machiz: It is the expert labor part. So there aren't that many people who can run a protein design campaign and there's a select few humans. And so because of that, there's so much scarcity in the amount of experts who can create new medicines. and that's why new medicines are typically very slow and expensive. So the idea here is less so that all of a sudden all of us are going to create our own drugs and you know all hell breaks loose rather it's that for the scientists who do want to create more specialized drugs or find new ways of binding proteins that process is much more efficient now and hopefully what that means is that for medicine providers you don't have to only focus on the biggest most profitable diseases you can go after some of these more unglamorous problems that are a little bit more rare disease that don't necessarily have massive billion dollar markets.

34:41Josh Machiz: Hopefully this is what that unlocks. So yeah, it's exciting. I love seeing experiments like this. As someone who is not as deep in the science and medicine space, I love being able to hear stories like this because it gives us such a tangible way of digesting the ways that AI can truly unlock massive step changes and paradigm changes in the ways healthcare evolves or the ways that we interact with the medical space. There's just so much exciting stuff. And these are little tidbits. And I love that on a weekly basis now we're getting some of these. Yeah, it's a game changer for biotech companies.

35:15It's nice to see AI accelerating something good, which I think we can all agree on that there are lots and lots of drugs that are still needed today. I also think just on an announcement side of things, just taking the marketing lens, I thought the anthropic announcement was really beautiful. They basically had all of these proteins kind of folding into the word anthropic in a really beautiful way that I stood out in the timeline. So I have to give them some props for that.

35:48Josh Machiz: Neko Health, the preventative health scanning company co-founded by Spotify's Daniel Ek, is officially coming to the U.S. They just announced their first U.S. location opening in New York this September. Neko says that more than 25 ,000 New Yorkers are already on the wait list for their 60-minute non-invasive full-body scan. This summer, Neko raised a$700 million Series C in a round led by Lightspeed, valuing the company at nearly$7 billion. Castilian announced a$1 billion Series C to scale production of Blackbeard, its low-cost hypersonic strike missile. The round values Castilian at$13 billion.

36:31The company says it has secured more than$500 million in U.S. military contracts over the past 18 months. And it's expanding its New Mexico manufacturing campus, which the company calls the largest dedicated hypersonic missile facility in the nation. What a big week. Venture Capital is really back with two big acquisitions, one closing, one getting announced. This is a compression of, you know, the value creation time. I feel like forever it was always 10 years to watch these companies exit. And now we've had, you know, two companies in quick succession where like they're two or three years old and they're being acquired for billions of dollars.

37:15Josh Machiz: Yeah, it's a really exciting time to be a builder. It's also with everything happening in the research and science space, exciting time to be a researcher. You're just seeing all of these amazing things compounding for every person in the ecosystem and things just feel like they're going faster and faster. And my hope is all of that gets built downstream in positive ways for users like us. And I think we are seeing delightful experiences like GrokBot, but also big questions around the shape of what this all looks like in application form, even in things like Chat to be a team. So lots to track and keep updated on.

37:54And we'll be distilling it down for you every week on Lightwork.

38:05you