Your Browser Is About to Drive Itself

0:00 Not putting it. All right. Shall we go ahead there? Yes. Yes. Yes. Cool. So, welcome to the Tesla Pod Arjun and Bhavani. Right. So, we are in a self-driving car now and then you're making browser self-driving. Right. So, that's pretty interesting, right? So, would love to hear more about Retriever, what you guys are doing, and then would love to dive in. Yeah. Thank you so much for the opportunity for the podcast. So saw the previous episodes looks super interesting and we even ourselves had like a video published where we were cloud code and driving like cloud coding and but yeah it's like our first ever tagline with retriever AI was make your browser self-driving. now so that's when we had a browser extension now we have a embeddible web agent. who now you're way past just browser agents, right? Yeah. Yeah. Yeah. Now it's we're more positioned as a website sellside agent like on inside a website like a agent that sells whatever you have to users agents visiting just makes it much more usable like you have people who can't figure it out. But yeah, that's a little diving into the product, getting

1:18 Our ahead of ourselves, but yeah, I'm Arjun. Yeah, I'm Bhavani. Yeah. So, Bhavani like now you're the CTO, right? And now the wheel is like you know turning on its own, right? So, and your specialty is basically like you know where you're differentiating from standard agents towards like you know you're mainly driving towards DOM, right? Yeah. So what is this differential? What does this unlock? Yeah. So before diving into like DOM what does a browser agent do right? The browser agent should understand the environment of like websites and need to take the actions just like your self-driving car need to understand the roads and take an action of you should go straight left right.

2:06 Yeah. The same way any browser agent understand the website what to click what to type what to do to achieve a user task. Okay. Now there are different approaches of doing it. The approach that we are taking is DOM based which means understanding the web pages. What is the you know semantic relationship between different elements. Okay. First you need to like click on this section to unlock the subsections inside it or to go to the next step just like how you should first like make a left and that's how you will get on to the next step of your journey. M and there's like another approach called kui agents or computer use agents which take screenshots of the page and try to identify I need to click this pixel coordinate I need to click at 100 120 and from what I read these vision based agents are the ones that's like you know mostly they are right and that's how you mainly differentiate right so the problem with the vision based agents are number one the LLM architecture is predicting the next token of like what should be the what is the highest probability of like the next token that fits well for the given task right the vision based approach does do not unlock this capabilities of like semantic understanding in the LLM architecture and also taking a screenshots the encoding of the screenshots is also inherently very expensive currently compared to like text based encoding and it's also

3:26 Extremely slow for example you want to subscribe to an email on a newsletter on like a website you need to like scroll down 10 times or take 10 screenshots to come back to like subscription part whereas in DOM based approach you can represent the whole page in a concise and compactful way for LLMs in one go itself. So a combination of things of like inefficiency in the representation and also vision based approach is not very good at structured data extraction. yeah so when you're trying to represent like lot of numbers and lot of data and trying to answer on top of it the current TLM architectures hallucinate from picking those numbers precisely from this vision based so these are all the gaps that we are trying to like solve by taking the other direction of like purely text based a graph representation of the pages. So you don't use vision at all or is it a combination that you take up? Yeah, we don't use vision at all. And that's because we started off like a series of hackathon projects where it's like we're trying to get a LLM to scrape a LinkedIn profile, right? And we were like, all right, just this should be straightforward. Just take a screenshot, ask it like, hey, get all these person's like information. Pretty straightforward job. But then it was

4:43 Just hallucinating every other line. And from that very beginning, we realized, hey, these are language models. they're still not good at like doing that conversion from vision to text and like using the screenshots very efficiently. So we never went with this vision based approach. We have so many friends in the web agent space. Everybody's saying the same thing. It's like you shouldn't use screenshots. It just confuses the model. They're all switching over to a DOM based. Yeah, there's a realization they are coming to terms these days. But right from like day one we realized that that's not the right approach with this architecture. So yeah, we take the full DOM approach of representing the pages in the text based graph for LLMs that yields the high success rate as well as extremely fast. Right. Right. So diving into bit of the story behind this, right? So you both were at Google before, right? And Bhavani, I think you've been at Adobe as well, right? So what made you I mean like you know being at these like you know big tech, right? And then they are I mean Google itself they are building quite a lot of things around web right I mean now Google is trying to make chrome kind of like you know self-driving right so how do you like you know what made the

6:00 Decision for both of you all okay now it is time for us to step out and then really start building because there is a clear gap. Yeah. So the backstory was me and Bhavani met at Google. we were doing a lot of hackathons like around the Bay Area like all of these companies you would know like I was actually there was a viral tweet that was like 2 years ago everybody was trying to set up these vector databases and like setting up like trying to fit everything within a chat GBT's 8K context window at the time and it was like oh yeah we did that exact thing. We did a hackathon where it was like tab was our first hackathon project of trying to create like a an agent of like all the tabs that you had opened on a browser so you can talk to Yeah. So that's but like we did a couple of these and the one that we felt had the most legs was like we had my roommate had a team in India to like basically collect for every municipality in California it's audit files and they had someone manually type and click in a browser to get this information. We're like, okay, this is like seems like very valuable and seems like something an agent should be able to do. And we built an agent that was able to do this to navigate the web, extract the information. And after that,

7:20 That's when we realized, hey, I think like let's explore this further. and this is kind of different than what Google and other the big labs are trying to do of building like a consumer first agent that's like helps you order DoorDash or does your groceries. So, we felt it was differentiated in that way. still after like publishing it, we grew to 30K users, 4 million workflows. Like the predominant use cases are like job applications are like LinkedIn outbound, social media outbound. but like we so if I'm looking for a job, I could straight away hand that task over to retriever and then it would apply on behalf of me. Yeah. Just give it a resume and be like, "Hey, apply to all the jobs on this page." Right. And we can the beauty of our approach was it could open the tabs in the background and like process them in parallel. And yeah but we saw that this is a nice feature in a platform not a like something users will like pay as a subscript. You kind of expect this if you have a cloud code subscription or codeex. So then we're like I think the value we provide is much more if we sell

8:39 To the website directly an agent that can convert those users that can't figure out your website and making your website consumable for AI as well. So that's what we pivoted this year to launching this embedible web agent and Rover, right? Yeah. Rover. Yeah, exactly. we so it's like you add this one line of code on your website and you have a agent bar on your website that can type click on your live site and take action. So this comes back to our approach of not using any screenshots. We do everything with HTML interaction and yeah it kind of like I had to repeat it every time in a pitch. It's like it's actually typing and clicking. It's like it's literally a DOM engine inside your website that can take actions because everyone else is used to chat bots that like you know they tell you click this link if you want to see the pricing or like hey this is how you can raise a ticket. They don't do actions in and of themselves right and we changed that paradigm with this new agent inside your site that can do the thing for you. You can kick it off, switch tabs, come back, and whatever you wanted is done. It's like seamless. The website owner is happy because like the

10:00 Users got what they wanted, got the value that they're deriving from the site. Yeah. Right. So, we inverted the paradigm. Right. Right. Right. And how would so basically what you're telling is right this allows any application, right, to have a complete fullyfledged self-s serve agent, right? that instead of instead of like you know you having to just interact on a text based basis right it gets the task done for you right but my question here is right so then how are you understanding what the application is all about right so then do you do some kind I mean is it purely based on documentation you understand and take the actions or what's the approach that you take yeah there are multiple layers to it so the number one kind of customers right now that are onboarded to Rover is smallmedium businesses, right, that are not very technical to know what is an AI agent, how to build an AI agent. All they have to do is just add one line of code on their website.

11:02 That's it. In one second, they go live. No need of any giving us any documentation, no need of giving us any material because our agent is trained on like nearly 4 million workflows across 30,000 users of Retriever ecosystem. The agent is capable of like navigating most of the websites either old website from your WordPress to like your modern Wix website, your modern like Shopify store whatever it is. So the agent work straight off the bat with a extreme high accuracy right now second layer of like now go to like B2B complex SAS websites. a lot of it is like besides UI there is also like a need for domain knowledge and the business knowledge needed. So for those kind of users, we provide an additional offering of like builds like these domain knowledge of your business where the agent like yeah understands the relevant documents and everything on the fly and uses it to put that knowledge together bake into its actions to complete the workflow. Yeah. Do you have any more thing to add? Yeah. just wanted to one of the selling points of our approach is all of these other chatbots you know you have to maintain a separate knowledge base of documentation and you'll always

12:18 Have like issues of stale knowledge whereas our approach lead reads the live website that is being served to the user. So it will always be live. This is information. There's no maintenance. It is just like reusing your existing infrastructure you sell to humans to sell to AI as well as like allowing those humans to use your site agentically. So that is a huge beauty of our approach that like people resonate with. Yeah. Yeah. There's this bread tailor obviously bread tailor is a lot of companies but now Sierra statement on like maybe it's easy to like plant sapling like do some plants once but the hardest part is like watering them maintaining them to create this beautiful garden. The same applies with any AI and agents right maybe you can if you have a small website you build an agent once or put some documents once but the difficulty arises to like maintaining it. your website changes all the time with especially with wipe coding these days people are shipping every second or every minute and if you are building this agent very tightly coupled with this documents or back end or even changes on your website you need to like go ahead and update it all the time make sure it's backward compatible all those things and the best part of like retriever is you can

13:35 Hand off all those things to us so rover understands your website real time at that second so no need of any maintenance no need of any like going updating it or anything. It works real time on the fly, right? And this is what is probably to most of the users and there's like lots of customizations of bringing your own tool. Say for example, some users be like book a demo. A robot does the demo, but it can also trigger maybe your Slack message. Hey, this user booked the demo right now. He's a warm lead. Go interact with them. Or you can call like another like tools through your HubSpot or whatever it is. So you can also do customizations of connect your tools based on intent of AI and humans visiting your sites and also like build your own knowledge bases everything.

14:19 Yeah. So you spoke about Sierra right? So Sierra Finn right? So all these guys they are in the customer support space right but then what I've seen is like you know most of their agents like you know they are still tex oriented only right? So even if you look at Finn as well as but then you're mostly it's a very actionoriented approach that you take right but where is the value coming from I mean if you're if your customer is a small business right so now sometimes like you know the visitors who are coming to these small business websites they simply want to like you know get the information right and then probably like you know I don't know like you know place an inquiry and then straight away like you know leave right But by rover providing actions within the website right where is the value coming in to the users of like you know the small business so websites this landscape I would split into like a 2x2 metrics one axis is autonomous level like how much autonomous taking actions versus read and stuff and the other axis is configuration or complexity let's say the technical complexity of the product now if you look at like Sieras and Fins, they're almost like

15:40 Consulting companies of like they take weeks and weeks of you onboarding integrating all of your docs APIs. So the complexity of like onboarding onto the platform is pretty on the higher side and that eliminates straight up like lot of SMBs to like adopt their platforms. That's the reason these they go after bigger enterprises mostly and the scale is like autonomous thing right the Sierra and finance stuff are able to provide some sort of autonomousity or autonomous actions of creating this Jira ticket or Zendesk ticket whatever it is through like API calls the tools connected all of it but again needs a really high technical configuration needed where does rover fits in is the complexity is really low for you to get onboarded and the autonomous autonomy is like extremely high but like how you add yeah I think to add on the Sierra in intercom and decagon they're what they're doing is solving for the customer support spend right what we're trying to do is the customer success spend where it's like making your product self-s serve the user finds the aha moment of what your the value from your product is And it's they so

16:57 They discover the value and they retain they engage they find activation much faster. So we're not focused on like handling support tickets or this kind of integrations that they all are like going after. It's so competitive. They're all doing really well by the way. They're doing fantastic. They're raising tons of money and we've just unlocked this new market, right? that is based on this browser agent technology figuring it out how to sell embedded and distributed to websites. So we see these both coexisting like Arjun mentioned there's customer support there's need for customer success of onboarding these users showing them the demos that work real time so your product coach marks or your loom videos get outdated right all like once rover is live on a website your sales team can be like simply show this user a demo of running a workflow or like you know setting up our platform and rover gives like a guided walk through as well as actually clicking typing doing everything in that demo real time on your interface rather than some like static UI that keeps changing. So we see these coexisting but at the same time the complexity the autonomous level is like yeah some of the differentiators right. Right. So here's an

18:15 Interesting question right. So now more and more the internet right is actually being browsed by agents themselves right. So you're currently helping the users or humans right that are browsing the web right so how do you like a as the agent traffic really keeps on increasing right and more and more agents starts to explore right how where would rover come into play right would you see yourself helping agents as well yeah so what is really resonating with our users so like for recap we launched two months ago, got number three on Product Hunt, already live on 2,000 plus websites. And what we're seeing the users are the website owners really love is that we launched a capability where from chat GPT your users are researching your product. They can directly book a meeting from chat GPT and you're like wait what like chat GPT only like visits sites. It like maybe links to your product. How's it taking actions? So we launched a A2W agent to web protocol.

19:23 We're just adding our script tag. We leave instructions for any visiting AI. It's like hey just visit this URL with the prompt encoded and we'll take care of the action like book a meeting with these details and you don't need to worry about like is this agent like what does it want? What is it here for? I'm just getting like a bunch of traffic and I don't know why. Right? There's a whole mushrooming of this GEO space and we're taking it to the next level of getting AI that is visiting your site to take actions on your site and this way you actually capture what is the intent of these agents that are visiting your site and you offer something right you're like hey this is a free agent on our site you can just hit it's already knows all our actions APIs and it will take like it will do this for free for you right so for the so the visiting agent it be much more cost effective for them to leverage this agent on your site. And for the publisher, the website owner, it's you get intent data. You get to control the access of the what? So you you can be like I'm not getting spammed. I'm not getting like scraped all of these things. Yeah. Yeah. So yes, we see the web users as two kinds. There will be humans still, but there will be also a lot more like AI on behalf of humans doing the work.

20:41 So with rover we serve both of them. We do the actions end to end for AI coming to your website as well as users through like technology and protocols arjun mentioned. Now we do to your questions of like yeah even cloud code or codeex they can launch browser do stuff right so there's two sides of ecosystem here. There's demand side and there's the supply side. See all these agent traffic is the demand side. They're coming with demand from the user to do a task build an app whatever it is. Now the website is the supply side. How does website effectively respond to them as well as learn from this so they can improve their product to make them better to serve these kind of demand right and our infrastructure sits on the supply side sits with these websites to understand these intent well do them do the better actions do successfully and also like in a cheaper way. We run with websites and also not all websites want to give access to all of their products for free to all agents. We sited this intersection of like we are like a gateway providing access to this AI identifying the AI providing access as well as doing all the things needed to the AI. So we take care of all of this infrastructure for the supply side and surface this data so websites can improve their product and surface right

21:57 Information to more and more AI visiting their product. Right. Right. And one of the most interesting things that I read was that the benchmarks that you all have posted. Right. So you have 81.4% on webbench right. So that's great right and then I saw the comparison like you know with all the other models as well right. So h how is I mean what is broken in this like you know still 18.6% that's there right now what is broken and what more room for improvement that's there for retriever. So this webbing benchmark history on it like has around 300 plus tasks I guess 320 or something of like doing various sorts of read actions write actions or what we call as crud on various websites right one of the examples would be like hey go to this Airbnb message a host and then delete the message all these things so in if you look at the 18% or 19% of the failure rate around 10 to 15% is what we call as like infrastructure failure what does it the agent is unable to access specific website. The website blocks your website has bot detection for JavaScript or things like that and why do they do that? Because the majority of websites don't want AI agent traffic because they cannot monetize that. that's

23:16 The reason and the rest four 5% is the agent the capabilities of agent to like recover try different things or figure out what is the best path to go ahead. Now if you with our current approach of having an agent inside the website the infrastructure problems are fully solved this 10 15%. Is because you're living in the website making work for the website. So you won't have the limitations of bot blocking or removing access for this rower agent. So what are the other things that we always continuously push for is the rest four to 5% problems with the agent. How do you effectively represent various elements pages trajectories learn from the failures and success? So the agent picks the shortest path, best path and selfheals to complete the task.

24:02 So the problem of we chose a different problem domain in a way than like the big labs of doing a vision agent, right? So there's it looks like on the output side, hey, it's a agent that can type, click, navigate the web. On the input side, it's like completely different problemscape, right? Like what we're trying to do is solve for the long tale of HTML interactions on the web. There's like anybody building a website knows there's so many different like edge cases and like there could be legacy websites. There could be like for some reason a type and click on LinkedIn could lead to double text instead of like just one like regular text. So it's like solving for these edge cases and handling them across the gamut of workflows. Whereas for the vision-based approaches that the big labs take, the problem is how do I train a bigger and bigger model that can handle more and more different kinds of workflows and it's like a model problem. So it's like different expertise you need to solve these two things and yeah where we see longterm of our agent is that like we can solve more of these edge cases and by partnering with websites we're no longer have a adversarial kind of thing baked into the business model itself right you know it's like our top use case in the with the browser extension

25:22 Was LinkedIn automation there that's a whole like mushroom of like a space and it's kind of adversarial LinkedIn doesn't want you to do that right and we are like trying to I pass LinkedIn. So it's like but now they here we're providing value to the website and it's go it can go much higher. Yeah. So double clicking onto this LinkedIn automations, right? So I think that's a pretty interesting space, right? Because there's like a ton of like you know ton of products that's there, right? and I mean some of them they work, some of them they partially work, right? Some of them they like, you know, get your LinkedIn blocked straight away, right? So I think we've all been there. We've all been there, right? Now what is the like now are you able to bypass LinkedIn right or because I mean LinkedIn does not in I mean like they do not actually allow automation that's one of their policies right but how do you comments no we do have comment we are very safe LinkedIn automation when I said I got my LinkedIn account like it was not blocked it was just I had to email them was like I just opened like 30 LinkedIn profiles all tabs in the same browser. Like I had manually just like command kept command clicking and like I was like oh my god my account like so it's like you know you don't even need an agent to like get yourself

26:39 Like flagged as suspicious. But so what we're doing the though is doing it responsibly and safely in your own browser on user's own device. Yeah. So most of the reason that these other LinkedIn automation solutions get blocked is they're like you give them your cookies or literally your username password and they run a cloud browser and then LinkedIn is detecting hey this is like not your usual IP address and then it gets blocked. This is like number one way all of this happens by us keeping it in your browser in your own regular device. It doesn't get flagged at all and we only open five to 10 tabs at a time. So even that LinkedIn is not like okay with. So it's like imagine I send a 100 LinkedIn connections in a background window. It's just the agent is cycling through them five at a time and it's like doing it and it's like completely fine with LinkedIn. Yeah. And also the traditional CDP style or pay puppeteer style like po using like ports for automation. We don't do all of those things. So you might see in the other Chrome extensions, they'll have this banner called like someone's debugging your browser and then a lot of websites just stop work. That's literally how literally every other browser agent is working. You claude for Chrome, Codex for Chrome, Manis, all of them are

27:59 Doing it this way and it's actually leads to a lot of vulnerabilities on your device because it leaves ports open. And it's like allows for untrusted execution and we're doing it safely and in a way that doesn't like it's less flaky. Yeah. Right. Right. So I think we spoke quite a bit on the product front. Right. So then diving into your co-founder life, right? Out of both of y'all, who's the optimist and who's a realist? Yeah. Yeah. Yeah, I would say Arjun is the most optimist because he we both are extremely technical but he believes in a lot of like some really cool technical features working out like we launch features new stuff all the time. He's definitely more optimistic of like oh people will adopt them like why not it's pretty straightforward. I'm bit more realist on the side of like one thing we learned in this AI world anything that's even remotely complex people will not use it. It's literally need to be baked inside the products like a magic tool. All they have to just talk and whatever optimizations, features, everything that you're doing should work under the hood and not that people have to discover that feature and click it. So that's one thing we definitely learned over time. I mean that kind of in a way is what we're

29:14 Doing like providing to the websites of like just giving it over like a command bar for your user to talk to your product and they don't need to figure out like the sequence of steps to like make it work and make it magical right yeah exactly and users don't have to figure out oh what new features you launched I need to learn them they just need to come up with what they want to do and rower on the website maximizes which feature is best to solve that and like does that thing end to end for user I think this is the future No one is like willing to put in an effort to see learn like we speak to a lot of our research friends apart from only software development. A lot of people running like RL training like day in and day out they are like laptops are going to die.

29:55 I don't need my laptop. I just want I speak to my AI and like I want my AI to send like a mobile notification or talk to me of like when the run is complete or when changes are like done. I don't even want to look at the changes being made by the AI. So yeah. Right. Right. So diving into hot takes right what is what's the hottest take in AI agent space right now right what's that like you know what's the thing that every founder in this space is like you know kind of too polite to say yeah I thought like I earlier this year I was going to my lead take was like hey Google is going to kill it at the foundation model layer and it's going to get killed at the application layer like there is just and then but like because they're just so slow with release cycles. It's just fundamentals of a big business. It's not even like something that like you can be against them. It's just like with the current pace that like you saw Codex and Claude Code.

30:51 They were iterating like every day multiple updates like they were the leads were tweeting like non-stop. Hey, we just launched this, launch this. And they were like fighting at each other like incorporating each other's things. And then you see like Google just like I don't know like just in the background everybody forgot about anti-gravity right yeah it's a just yeah so that was our hot take but then yeah just yesterday Google IO happened and I was my personally kind of disappointed with the not only the application layer releases I was also disappointed at their foundation model release I was like this model is three times as expensive and it's like similar performance as before and I was like, "Oh, okay. Maybe like I'm thinking even at the model layer like they might you know get Yeah. So, just to recap what happened at Google IO, they don't have any like strong foundation model updates yet. Just they've announced only like flash 3.5. but the biggest problem that we face and community on hackerne agrees with is they don't have the enough inference TPUs for everyone. So we you see a lot of like either 429s or 503 is unavailable even though they claim like the big Google they're like fast and

32:08 Stuff. I'm sure maybe if it's a dedicated enterprise they try to give like more inference right to them but even as big as Google with their own TPUs still doesn't have the model availability that's like needed by all these startups and stuff right another hot take I've got is probably on the hardware devices side I think all the all these hardware devices of like recording from like I don't want to name startups but like recording everything that's happening in your life on like Okay. A small gadget that you can hold or all sorts of like omi style keychains that you or pendants that you can wear. I think all of them are going to die, right? Wow. I didn't know that.

32:49 Right. You think they are way too overrated right now. Yeah. They are whatever job they're trying to like propose is like pretty much done by the existing your mobile, your laptop, which is all for someone to adopt a wholly new hardware device. it needs to like meet or like do lots more than just like recording or like taking notes because those are like table stakes that are done by your laptop and phone. So dedicated for that I don't think it's cut it's gonna cut it right like Google again they launched the these new glasses right like it's coming Google as part of Google IO because of their partnerships with like monster and other eyeear they're launching again new eyeear competing with like meta and it we'll have to see if it's going to succeed this time right so if there is a founder out there right now trying to scrape the web right say puppeteer right so what would you say to I think what we realized this year it really took off is we called it vibe hacking. So instead of you know writing a puppeteer script, writing a browser, you know models are gotten so advanced they could just reverse engineer the websit's APIs and we got the we had this problem cuz like customer like had a custom deal that we wanted to help them

34:06 With and we were getting our browser agent to actually type and click and like get all this data and then like we did it for like couple of pages we're like this like let's just get cloud code to write a script and so then that we realized we re-engineered ed our own harness to like reverse engineer the websit's APIs in your own browser and write code to just hit it at scale and that allows you to do a lot more scale of operations and I think that's everybody's going in that direction now so yeah so one thing that we and also the world of agents we see going ahead is like balancing this like tool calling of this GUI interact actions like clicking, typing, all those granular GUI things still needed because not all websites have APIs to do stuff plus whenever there is a possibility of APIs and stuff to do it something faster and cheaper like writing code snippets to balance it. So we see the research heading and ourel product both heading in a direction of like how can you have an agent harness that balances this GUI tool calling as well as like writing code to achieve like higher success rate as well as you know like the shortest path quicker things like to do things. Yeah. And it's like complex

35:25 Because a lot of foundation models are trained on like tool calling. they have affinity to like pick these GUI tools. maybe if not needed or even if they are like you know as a longer path and stuff sometimes code can be the optimized path. So yeah, the balance of both of this is where we see heading. Right. Right. So we touched base a little bit before, right? On this agentic web, right? It's a phrase that's really keep coming up more and more. Right. So where do you see like you know the internet being in 2030. So it's in that's insane. I didn't expect you know cloud code and codeex to be the killer super app like last year, right? I would have thought one of the many opening eyes offering they had so many things in the pipeline I thought one of them would stick instead of codecs but yeah I think the where the web is going in 2030 is judging from the users like what do they want right they want to be able to do a lot of things all at the same time and they want agents to go do those things for them so the end goal where we're going is like there's a new interface where you just talk to it and it will actually do the interacting with the websites, figure out like negotiate

36:44 Deals for you or like do handle your tasks, right? We got to like think back, right? Like think a step back. The browser itself was invented 30 years ago. And I think it is about time you have a new interface instead of the browser. But like I do think though the current way of distribution like how do you discover like a business right? You discover them via the URL and their online presence is still dictated by that URL and models are amazing at like encoding stuff into URLs and like going to that URL like it's like a really hacky way to do a lot of tasks is just to like put in query parameters and stuff it in. So I think websites and URLs will still be there. I think there will be a different interface where you can parallelize, get more done from the web, get more actions, research, all of this. Yeah. Right. Right. Yeah. I agree. Like probably AI traffic will be like at least like 80% on the web and 20% will still be some humans using it. And as Arjun mentioned maybe let's call it as like intelligent web. There's definitely lots more new kinds of GUI that's more intelligent powered by rower or powered by own themselves that will serve these AIs

38:03 Better but the browser like I think the browser is going to be there the web is going to be there websites are going to be there yeah right but it's up for quite a transformation now as well right so that's a good point right so now if there is a founder right so who is at Google right now who's thinking king of jumping the ship as you guys did, right? What would you say to them? I mean, like ju just do it. The only advice I've got is it's yeah, you've seen the announcements of meta layoffs today if you're watching this today is I don't know like May 20th. So, Meta announced started sacking people from like Singapore today. So, it's risky everywhere. Yes. But not a lot of times in life you get the opportunity to like double down and implement faster, do faster with just like one man army or two man army whatever it is to do stuff. So yeah, now is the time to like take the jump. you'll figure out I would say definitely do more planning get because like Google's ecosystem right there is a Google tech island where it's like if you've been there long enough you don't even know how the rest of the world works. It's kind of like what is next.js what is

39:21 Re like react like so it's like I think some of the learnings from our own approach is do a lot of hackathons build a lot of cool tech and see what is getting good feedback from like other people before doubling down on it and then like keep releasing build in public like showing other people and getting feedback instead of just like blindly going wherever you think the cool problems. Yeah. Yeah. Like I would really not recommend that like I think YC gives that as like release right now. Really like if you haven't released release it like so the just getting there's no secret sauces out there people will figure it out whatever you've done especially with the latest coding tools. So it's like just be in public build in the community and collect feedback. Right. Right. So I mean one thing that we learned at Founders Inc. as well, right? So, GT T F O L I think. Get the out of local host, right? So, yeah. I mean, just ship it, right? And then see like, you know, how it lands, right? That's kind of the best signal. Yeah. so where do you where do people find you and Retriever? Yeah, we are in all the social platforms. We publish a ton of YouTube

40:38 Videos. You can find it at Retriever AI. we also on a Yeah. Yeah. LinkedIn X. You can reach out to us anywhere. Our website is tricky but we got good SEO for you. It's RTRvr.ai like your golden retriever. You can like find our energy across all socials. Right. That's awesome. So these been great guys. Thanks. Yeah. And you know we should give a little bit background of the user like viewers. You know we met at Llama Lounge 24. Llama Lounge 24. Right. Was it? Yeah it was 20. I mean it was in San when we were pitching at Zoom, right? Yeah. Yeah. Yeah. Oh, 25. Yeah. Yeah. I mean it was in 2025, but like I think the number of the but like Yeah, that's how we met. We had booths next to each other. We got talking and yeah, that's how pretty fun. And afterwards I think like you know we just kept bumping at each other at all the SF events, right? Yeah. That's what you do when you become a founder. You have a founder community. Yeah.

41:40 Right. Right. So that's awesome guys. So thanks a lot for the time. Thanks for the opportunity. Thank you Tesla for driving. Yeah. All right. Thanks, guys. Thank you. That's a wrap. It was great. It was pretty smooth. All right.