Browserbase founder Paul Klein on building the web browser for AI agents

him in. He's ready. Welcome to the studio. How you doing? The basement. Hey guys, the basement. What's up? Looking good. Came prepared. Looking fantastic. Looking sharp, you know. Uh this is the big leagues. You know, you got to you got to dress up to impress. So this uh is this a green screen background? Whatever.

Oh, this is live. This is live. We're in the People are working. People are grinding. I love the whiteboard. I love the whiteboard action in the back. It's fantastic. Somebody's actively whiteboarding something integrals. Yeah, it's math. Yeah. Why why why why do you got a human doing that?

Can't you just uh wire up one of these uh computer use uh tools to do it for you? Well, that's that's that's all in the cloud. This is in real life, but yeah, you need an every hundreds of engineers running in the cloud right now. Fantastic. Yeah, those are just paid actors actors.

Uh anyway, uh can you introduce yourself, the company, uh kind of what you do for everyone? Yeah. Hey everybody, I'm Paul, founder of Browserbase. We power web browsing capabilities for AI agents and AI apps. So we're kind of building like a web browser for AI.

Basically, we think the future of software is software going to have to go out and do work on your behalf. But a lot of the work that we do is in a web browser on a website. So we want AI to do work on our behalf. We're going to have to have a AI use a web browser and we provide that web browser.

Now, we mostly sell to developers, people who are building these cutting edge AI features and applications, and they integrate our AI web browser into their applications to give them superpowers.

What that might be is like going out to a thousand websites and finding the right screw for your procurement team, or maybe it's going out and booking demos to help, you know, figure out what your customers are doing or, you know, trying to automate all the tedious work that people have to do every single day.

We do that with browser base. Um, I want to talk about MCP APIs and kind of the future of how agents and LLMs will interface with the internet. Um, there was a meme for a while that was just like the the front end is the API.

Just get good at web scraping because the the API might change, but the front end is going to stick around for much longer. And so if you can be really dynamic there and I and and I always wondered about, you know, are APIs just kind of going to go away as as LLMs get better and agents get better at computer use?

Uh what's your take on uh where how we will see companies make their content or their websites available to LLMs going forward? Yeah, we really believe that computers are going to be just as good at people of browsing the web.

And for a very long time, there's still going to be humans using the web doing the same task they do every single day. So why do we have to rebuild the internet for AI when AI can use the web the same way people do? So there's certainly going to still be APIs and MCP servers, especially for high volume stuff.

You know, if you're booking a flight, you probably should do that through the Delta MCP server, not through a website because that's just not the efficient way to do it.

But when we talk to companies that are not the cutting edge companies like our customers like Perplexity, but more cutting edge companies that are old school, have been around for a long time, there's a lot of websites that aren't going to have MCP servers.

The Nigerian immigration form won't have an MCP server for a long time. And AI needs to automate that just as much as automate kind of the more high volume tasks.

So we view the browser as kind of really the most primitive MCP server that can work on every single website without having to have, you know, first-party integrations.

Can you talk about uh how you thought about the opportunity early on because I imagine you and you you know you've made a tremendous amount of progress, raised a bunch of money uh working with some of the most you know important uh companies in the space but early on I imagine some people would have said hey like browser automation software has existed before like we don't think the market's that big and and clearly you didn't agree with them and uh I'd love to to hear you kind of speak to, you know, why you understood the opportunity is so much bigger than than people might have originally thought.

Yeah, there's there's a lot there. I mean, frankly, I just love this browser automation stuff. If if it wasn't, you know, as popular as it is, I still would be doing it because I just love the stack. I love the problem. It's it's really interesting to me.

Kind of it touches on the programming I did when I was a developer starting out, which was, you know, how can I automate something and help my life and make things easier for me? And that often was, you know, interacting with websites on my behalf. So, there's a bit of luck there because I just love the space.

Now, I'll reference one of our angel investors, Jeff Lawson, the founder of Twilio. When he was starting Twilio, people were saying, "Well, SMS is going to go away. " Like, why would you build a company on SMS? And his point was that it's a very long bridge.

I think it's the same thing here with browser automation is that sure, when we have Neurolink and all of our brains are talking to each other, software is going to be a lot different, but I think we have a really long time until software, especially software on the web, goes away.

And taking that perspective where browser race is going to build for that short medium-term future and really build great technology to help agents run and do tasks on the web.

That felt like an opportunity that was a bit contrarian and you know when you're contrarian you find alpha and that's what we've done at browser base and as you can see today the public markets are reacting you know after our Q1 results triple Q up 10%.

I don't know what happens but I think they may have gotten wind of what's happening at browser base and have really corrected accordingly. I love it. I love it. Uh can you talk about um you said you sell you said you sell to developers obviously perplexity it sounds like as a client and a very large company.

Uh do you have kind of self-s serve offering where someone can just kind of start import this in one line and start to get hacking?

Uh, and kind of how do you see the the market developing from like proumer, you know, just some noode developer is going to be using a browserbased like product or or you're going to be under the hood maybe um to someone who's, you know, building a business on top of browserbased.

Yeah, for us, you know, you could get started self-s serve. You can sign up. There's a free plan. You can upgrade accordingly. And then, you know, as you get to a certain volume, we want to give you pricing that's right for your use case. So, there is more of a contact sales option longer term. Yeah.

But you know what we think about most is our business model. We're a consumptionbased business model which means that we only win if you build something that's successful.

So our incentives are very much aligned with helping our customers actually build applications that use browsers a lot that are used by lots of customers.

And a lot of my job as founder and CEO is going out and helping our customers really think about how can they use AI in the best way to help impact their users because you can kind of map one to one the browser hours they're using on browser base to the time that they're saving for people.

And we get really excited seeing that number chart up every single month because that means there's hundreds of thousands of hours that are being saved by people instead using browser base. How do you uh was deep research the studio Gibli moment for agents or is there one on the horizon in your view?

It feels like you know we we've talked about this on the show over the last year there's been a lot of you know we were at YC demo day there was a lot of like AI agent infrastructure companies.

I don't remember getting pitched that many uh individual AI agents, but it does feel like as the underlying tech gets better and people figure out how to actually build quality agents that that will flip at some point. Um, but I'm curious like, you know, do you think that agents are even priced into the market yet?

I I imagine you're seeing a lot of stuff like, you know, in in, you know, using browserbased that like hasn't isn't really in the public domain yet. Mhm. Yeah. It's it's interesting because I really like to think about the vastness of our customers.

Like sure, we have cutting edge companies, but we also have a 55-year-old dairy trucking company in the middle of America.

And they hadn't hired a single engineer in all 55 years, but then they hired one and their first thing they got was browserbased to automate, hey, how do we know what the gas prices are along this route?

Previously, like an ops person calculating all the gas prices and they're like, oh, maybe we can hire an engineer to make this easier with AI.

So it's certainly AI has escaped Silicon Valley and every company in the world is thinking about how to use it and thankfully browser based since it's pretty easy to get integrated and get started with we have a framework called stage hand which makes it easy to tell AI how to control a browser stage hand team behind me uh those those guys uh it's much easier to on-ramp to use AI so we really try and think about like you don't have to know what the cutting edge models are you have to keep up with the research if you're using browserbased we try and bring it all to you in terms of it being priced in I think there's a lot people who maybe moved a little too early to AI agents.

I think the Clara stud story is a really good use case here where they went all in on AI agents. They threw away all the software and then I think they had to walk that back a little bit.

So it's still early days for agents and a lot of that's limited by great infrastructure and we think that if we can build good infrastructure that works best with the models, it can really help ease adoption.

How do you uh everybody's been really worried about uh AI agents and their their sort of potential to be used in nefarious ways. So like you know releasing agents across the web that are just like you know effectively uh spamming people. Uh is that like worry you at all?

I'm sure you can see if somebody's using browser base in a nefarious way, you can kind of identify it quickly and and eliminate access. But I feel like people have been just like worried about this idea of AI spam. But it hasn't seemingly had it isn't so prevalent yet that it's that we're in like a crisis. Yeah.

The only place that I've seen it is LinkedIn and we just disable LinkedIn on browser base. It's just not a business that we work with and we think the LinkedIn, you know, everyone at my LinkedIn inbox is crazy. It's just something we don't want to contribute to.

For us, we we really we really want to think about like how can browser base be an arbiter of good bots long term. I think there's this cat-and- mouse game and anti-bot blocks all bots, but there are good bots and bad bots.

And we're in this really great position where we can actually be advocates for our customers that can talk about how so and so customer has 20,000 employees on workday and the workday APIs aren't really cutting it for them and this is why they're using browserbased to automate it and build that relationship.

That's why in our last round we brought on octa ventures as an investor, the founders of clerk, the founders of stitch, the founders of work OS. You know, these are people who are really thinking about authentication. And my view is that the antibbot problem is probably solved at the authentication layer.

Antibbot is just proof of personhood, right? So if a agent can log in on your behalf, that's going to be really good to match their actions to the person that they're saying they're acting on behalf of and hopefully match the rate limits there accordingly. Can you talk a little bit about the llama evals for llama 4?

I saw that I I believe someone from your team posted that they had done some uh kind of I don't know I don't even know how you describe it like private evals stuff that you guys just look at directly and h how do you build that stack up? What are you actually testing for? And then what are your vibes around Llama 4?

Yeah, so what we did over the weekend of course whenever something comes out we want to move quick and try it out. We're all engineers here.

I'm an engineer and I love to try and play with new models and with stage hand it you can kind of imagine it like a burell a SDK just for building what's called a web agent or an agent that brows the web so we can slot in any sort of model into that and we have a standard eval set that we run stage against common tasks on the web that we've kind of built up to evaluate different performance like websites with dropdowns website with date pickers you know the many different types of elements you see on the web we want to see how our automation framework can stack up against that we spent a bunch of money this weekend evaluating every single model that we could get our hands on against that eval framework and we actually found that Gemini was one of the best performing models.

I think it was Gemini 20 flash and Llama 4 a little bit more disappointing. Now Llama 4 we used it with a inference provider and we know the different inference providers can have different performance based on these different open source models.

So I think it's still too early to call on Llama 4, but we're definitely excited about the stuff that Gemini is cooking up and they seem to have really figured out some of this interesting stuff with long context models. Yeah. And also Llama for they haven't released the behemoth model like their biggest and best.

So maybe that changes the conversation when that happens. And there was also a discussion about they might have used a different one on LM Arena is getting getting a little spicy on the timeline. What's your uh what's your Paul Doom? What's your P Doom? P Doom.

And specifically, I'm I'm curious how you you reacted to AI 2027.

Uh because I I just want to go out and say like if any of AI 2027 comes uh you know becomes true, I I hope that you'll turn into a defense tech company really quickly and you know help unleash a a an army of good browser agents to protect fighting on our behalf. No but fight the good fight. Yeah, exactly.

In terms of P Doom by Paul Doom right now very low. It still feels like there's a lot of you know capabilities that need to improve. Infrastructure needs to be built and everyone seems really willing to work together which is positive. I think I see a lot of partnerships happening in the ecosystem.

For us browser base will be a partnerships company. You know we need to work with the cloud flares of the world but also the open AIS the labs. When computers came out we were able to launch alongside open AI and kind of get early access to that model and give them feedback.

So it certainly seems like everyone is being collaborative to solve these problems. I mean we'll see what tariffs do to model prices. I don't know if that comes down one of these days, right?

Um, but in in terms of what's going to happen in the future, it's still too early to really tell, but I'm feeling pretty confident. Can I was going to say what are what are use cases that you want to see more of? Like maybe you see it happening on browser base today.

Like the the example of a dairy company like trying to more strategically uh route their you know basically like manage their logistics to drive down costs is cool and sort of random but I'm curious if there's other areas where you think there should be more people building um these sort of browser use type experiences.

Yeah, for us we really look to like all these these use cases that probably don't need to be standalone products on their own but really help individual businesses. We get really excited about companies that are kind of building things in house. One insight that I've had is that as code creation has become easier.

More and more companies are kind of opting to kind of semibu in house where they're pulling some primitives off the shelf like some infrastructure like browser base or maybe superbase for their database but then still building their own agent internally. I think it's kind of interesting.

It's like if you have all the primitives available, can you just stitch those together using cursor and actually build something that's more bespoke to you? What we found from some customers is when they buy blackbox solutions, they aren't able to really tune it to their specific use case.

And agents still need a lot of tuning for each individual use case. On the wider side of use cases I want to see, I really love seeing stuff around like every CEO should be looking at what do they have an operations team doing.

Is it going to an insurance website and doing a credential check on a provider or insurance verification? Is it doing some interesting procurement task? Is it doing some manual research? And how can they automate that? There's a lot more tools out there than there were before and it's easier to integrate than ever.

Can you talk a little bit about the actual tech decisions you've made? I could imagine one version of computer use on the tech tree is like take a screenshot, use a diffusion model to understand this. We're seeing this with images in chat GPT. It's getting very good at recognizing text.

It can recognize people within images. You could imagine taking a screenshot and just asking the the vision model. Where is the call to action? What like like define that in pixel terms so I can click that button.

At the same time, you could imagine just passing the entire you know view source like the entire DOM to uh to an LLM and just saying hey write some jQuery that clicks on the call to action uh and make an educated guess based on that. You could imagine hybrid approach. What has been most successful?

what are the paths that uh you are not looking at these days? Yeah, for our framework which really kind of abstracts all of this, we use everything. And I think that's the important thing is it's not one size fits all.

If it's just screenshots or just DOM or hybrid, you really need to kind of think about what's the context size of the website. It's a very long complex website with a lot of HTML, you probably want to go a vision approach. If it's Craigslist. org, DOM is probably sufficient.

So, I think every single website has different configurations that work well for it. And with our products, we really try and help developers make those right choices or not have to even think about that and just give us all the models they have access to.

We can, you know, route between that model based on the type of website that's important and help them build the reliable web automation.

For us internally, we do really love the DOM based approach or the HTML based approach just because it's frankly more cachable such that instead of having to send a screenshot to an LM every single time, if the website doesn't change, you could probably reuse the same button you used last time to perform a workflow.

So that DOMbased approach seems to be like the more performant and cachable longer term is also a little bit cheaper but we'll see as comput models keep getting better maybe there's going to be some you know model level caching on screenshots where they can say this screenshot we've seen before this is the prompt we've seen before we can say this is the same button so as comput models get a little bit better it might be on more equal footing with the DOMbased approach in terms of cost but it's going to be up to the labs to really do some innovation there.

Yeah. What's the team's take on vibe coding right now? Is it not a fit when you're needing to build reliable products that customers depend on for business processes or are you guys making it work internally?

Yeah, so I mean the team doesn't vibe code very much on critical infrastructure, you know, that goes under a lot of review, but it's made shipping integrations with browserbased and stage hand easier than ever and we've been able to integrate, you know, with lane chain or llama index or crew AI and those kind of come out a lot faster.

So this interconnected world that you're seeing with every tool integrating with every other tool that's a lot easier vibes wise, you know, something happened one day where the whole team switched from cursor to Windsurf overnight. I'm the only cursor user left. Maybe that means I'm slow to adopt.

But it seems like people are moving between these idees pretty damn fast and the next tool that comes out maybe could this like the great flipping, you know, between cursor and wind surf just kind of happened at browser base.

Yeah, it feels like there's this Cambrian explosion in like AI tooling and there's a lot of businesses that have built pretty solid businesses. We've seen like the crazy charts and the revenue ramp. Um, and in the past like cloud era, a lot of the companies actually went public.

You know, like Twilio and MongoDB, you know, are very point solution companies in my opinion, but they both built fantastic businesses eventually.

Um, but you could imagine a future where there's more competition and you're building into adjacencies or there's mergers or acquisitions and you kind of get some sort of like new age rollup of all the different AWS style tools in this infrastructure like how do you see the market evolving because uh at the same time I imagine the hyperscalers are thinking about should could browserbase be an option in AWS with the other 25,000 tools that they have up there.

Uh but but yeah, I mean obviously you're very partnerships driven right now. How does that evolve over the next couple years? Yeah, you know, we are seeing hyperscaler interest for sure. Amazon just launched Nova and the Nova SDK took a lot of inspiration from what we've done at Sage, right?

And we really appreciate that. That's the DX that we're all settling on. It's the right way to do things. Now, the model performance is a little bit subpar and I think trying to build a model while also building a great framework is just a really impossible mission for browserbased.

We think about how do we build best-in-class infrastructure and are able to go super deep on a single vertical in a way that a hyperscaler might not be able to do.

If you've used, you know, Amazon RDS versus maybe a more focused database provider like Planet Scale or Superbase, I think you'll find that the product just it's not as exciting for developers to use.

I think Stripe is a good example of a company that's gone very focused on something like payment infrastructure but then layered on this product platform on top of that. When we look at the expansion opportunities for browser base, I think there's a lot of interesting stuff.

You know, think of every Chrome extension you have. You have your password manager. You have your payment information saved in there. You have your browser history, which is like memory for an agent.

So, we can think about more areas within the browser to offer as agent primitives longer term, but we're really focused on just building best-in-class infrastructure that developers love.

Building a framework that really works well for everyone and just continue to move up the stack and find things that we can do to help developers build browser automations easier than ever before.

Can you talk a little bit about uh some of the optimization vectors that the various foundation model companies are kind of sprinting towards? Uh we saw this with Llama 4. They were claiming, hey, we got a 10 million token context window.

I think that was 10 times bigger than what Google was touting uh a year ago or even a couple months ago. There's obviously the the length of time a reasoning model can stay consistent and that's like doubling every couple months now.

Uh there's different things about just uh maybe it's not the highest IQ model, but it doesn't make stuff up because it's got a rag integration or something and it's actually pulling real facts instead of just hallucinating constantly.

And you could imagine there that if I'm just thinking about like what does a computer user user need? Uh well, I don't want them to lie, but I also don't necessarily need like a chess grandmaster. So maybe I'm not optimizing for IQ.

Um but what where are you most optimistic about where the foundation model companies will go and what kind of uh problems or vectors or the shape of that progress looks like? What what is it what does the shape of the progress need to look like for you to really succeed?

Yeah, for us I really like this paradigm of splitting up reasoning models to other types of models.

I think trying to make a generalistic model that's great at everything is going to be really really hard and you can fine-tune different types of models and different types of tasks if it's codegen or reasoning or even you know UI automation like a computer use model. So more models means more choice for our customers.

In the end, you know, our customers are trying to choose the best model for their task and that varies because the internet is so big. You know, there's billions of websites and billions of choices on what, you know, you can use or should and shouldn't use.

For us, like we do see the reasoning models being closer to the agent loop, whereas more like a 40 or a Gemini 20 flash being more close to the tool calling makes a lot of sense. you want the expensive model making the decisions and the cheaper faster model maybe making the individual tool calls.

That's been a great paradigm split and more of our customers than ever are actually using models from different providers. So it does seem like more choice means more innovation and more ways to kind of mix and match the best of every sort of eval. Very interesting.

I want to get your take on the latest from Microsoft uh today. They just uh this just leaked out or something that uh Microsoft said it would no longer move forward with its plans to build data centers in Licking County.

The company had planned to invest a billion dollars initially towards three data center campuses in New Albany, Heath, and Hibron. Uh Microsoft plans to ensure the land at two of the three sites will be available to be used for farming. And uh Nick on on X says, "Yep, it's over. " Uh so so how how are you feeling?

I mean, obviously, there's a lot of turmoil in the markets around around um around tariffs, but I feel like the whole narrative of like maybe we're getting over our skis on capex and hypers scaling.

Um does is all this just noise to you because you're just focused on product experience and you don't really need a 10x improvement in the foundation model performance that might come from a bigger data center buildout or are you still like inferencebound and you're like, "Yeah, it would actually be great if they move forward with that.

" Wh What's your take? Yeah, from what I've seen, the labs are still innovating and and I don't know if Microsoft is a lab that I've really think is going to be an innovator as much, you know, so I don't think OpenAI or Anthropic are holding back on investment.

Now, does that mean they're signing massive data center contracts right now? I I think people are kind of waiting to see what happens in the macro sense for us. I'm very downstream of that, right? And and of course, we're, you know, a private company, venturebacked.

Things that happen in public markets tend to trickle down to private companies. So being aware of it, but you know with browser base you know with the traction momentum so far uh we've had no shortage of inbound and it feels like there's a lot of room to run for us so not really concerned.

My last question uh and you've covered some of it before but can you talk about just general capability overhang? There's this like you know with with all the evals and the sort of pressure for the labs to constantly be number one or at least be you know at the top.

It feels like that's almost like a distraction at times from just like delivering all the capability that is now possible. Um, but I'm curious what your take is on that.

Yeah, maybe answering this in another way and we can go deeper, but for browser base to exist, there needs to be healthy competition at the model layer and we see that models are more funible than ever. People are switching between models all the time and people don't want to switch their infrastructure.

they want to have a piece of infrastructure that works they can kind of integrate with a little more deeper and if that means plugging in different models that's just going to help them long term I think there's a lot of innovation that needs to happen at the model level you know some stuff like multimodal tool calling being supported by all multiple providers and not just anthropic this is like a lack of capability but it seems like every time someone innovates that does trickle down to the other labs so it doesn't seem like there's going to be a shortage of choice from customers and there's still a lot of innovation happening especially long context like we think that long context is going to do a lot for DOM based web agents and there's still a lack of true performance there but we think with time all the stuff does just get better cheaper faster and it's inevitable.

I want to ask you about this this other story that kind of broke recently that uh OpenAI is considering acquiring a uh a a hardware device manufacturer that was started by Sam Alman and uh Johnny IV the exApple designer.

uh potential in designed includes a phone without a screen and an AI enabled house household devices and I and I'm curious to know your take about like generally we saw like the first wave of this with like the rabbit R1 the friend the humane and some of those didn't go so well some of them are plugging along but I could imagine that maybe the studio giblly moment or the touring test moment is when yeah I don't need a screen because I trust that I'm not just using this pin or this device without a screen as a knowledge engine.

Um, but I'm actually it's actually able to go use a web browser uh on the back end and just talk to me in conversational. Uh, I know that's a little like sci-fi and maybe like I don't know, maybe it's like three months out, who knows?

But I want I want to know your take on kind of devices and and how you could see browser base plugging into all that ecosystem.

Yeah, you know, for us like consumer web automation isn't really a focus just because that I think that is one on the device level and I think that the Google the Apple I mean Apple intelligence not really something I use every single day and oh I hope it gets better right but you can imagine the amount of data they have access to the integrations with your existing life they're really well positioned to offer more consumer web automation you can tell Siri hey go book that restaurant and it's going to have a lot more capabilities to do that maybe even on your device right but for our customers customers who are kind of building, like I said, these more traditional RPA or maybe they're using BO or maybe they're a company that's building something more verticalized to go sell to someone who uses RPA or BO.

Those people are still going to need infrastructure the way that they because they're building applications and and that's who we focus on. Yeah. I I wish OpenAI the best with that and I hope it works. I I want more cool devices. I I wanted Rabbit to work and Humane to work and I I don't have a friend Avi hook me up.

I'd love to get one, you know. Uh I I want all these things to work because I want this cool sci-fi future. So, I uh I try and stay super optimistic about that and yeah, hope it into existence. Yeah, it it doesn't feel like it's super in Apple's DNA.

Like, if they even if they just had, you know, an Apple intelligence query, oh, like summarize the news for me, you know, that they're going to do a bunch of partnerships and then they're going to pull that in from an API and it's going to be as uh as deterministic as possible.

Like they're I I feel like Apple does not have probabilistic computing in their DNA. And uh I'm sure you have to deal with a lot of a lot more rough edges.

Are are you are you seeing the thing that we're seeing with like agents and we heard about a little bit of this with like the dust up with like AI SDRs is just that um the like a hallucination rate of 1% is unacceptable. Um how are you thinking about like keeping a human in the loop?

Um and just kind of like even how do you message to a client the reliability of your model? Obviously you have benchmarks for all the LLMs that you're using, but what do what kind of claims are you making?

Because I think we all understand that like hey this stuff's new it's moving fast it's improving like you got to go you know fast and and and and just uh you know test this model and not put it in like a life critical situation yet and you can still get a lot of value out of it but how are you pitching browserbased to folks?

Yeah, you short term you are going to need some human supervision at some point, right? And the way that we allow this with browser base is we have this thing called a live view where you can actually embed an iframe that shows the browser working.

And with the browser working, a human can take over if the agent gets stuck.

And you can imagine maybe the future that's like very immediate is not a browser doing all your work for you, but maybe more of like a Devon experience where the browser is doing 90% of the work for you and maybe you have to come in and do the last 10% once in a while.

And when we talk about with our customers like how do we build human into the loop into the designs that you currently have in your application to make it just super easy to have that escape hatch into the human operator if something isn't working. Yeah. Yeah, that makes no sense.

Uh well, thanks so much for stopping by. Congrats. This is great. Congrats to the whole team. Would love to have you uh back on next time there's news. Yeah, this would be great, guys. We'll talk to you soon. Bye. Uh next up, we have Tracy Aloway coming in to the studio. Uh, but first, let me tell you about Bezel.

Shop over 24,500 luxury watches. Fully authenticated in-house by Bezel's team of experts. The market's up, baby. Take uh what do they call it? Uh uh take some gains or whatever. Realize some gains, right? And uh convert it into if you bought yesterday. Yeah. Uh you're doing pretty. Yeah.

If you bought the dip and you and you're up, you know, treat yourself. You've been working hard. You've been working hard. You deserve it. You deserve it. Go to go to Bezel. Uh download the app as well. And we have Tracy in the studio, I believe. Not yet. She's coming in in just a minute.

Um, she's been covering the chaos in the markets very closely. Oddlots has been on absolute tear. They had Ryan Peterson on. They've had a bunch of people uh dig into reactions to the tariffs. She also just put up a uh uh everything about the basis trade, which I want to dig into uh with her.

She writes, "It's worth taking a detour to explain just what the basis trade is and why it became so popular. To do that, we need to go back even further into the dark days of 2020, all the way back to the the aftermath of the 2008 financial crisis.

Scarred by the near collapse of the financial system, policy makers moved to makes bank safer.