สำเนาบทสนทนา 🔗
speaker-0: No, the internet that we use today was designed for humans, ⁓ browsing pages, clicking links, ⁓ but but increasingly AI agents will be the ones that are interacting with with ⁓ with the web. And so we have three folks in this panel today that are building the core infrastructure for that world. So I'd love to have you all introduce yourselves.
speaker-2: My name is Cleel. I run a company called Natural, not sure why it says a non, but we build payments infrastructure for agents.
speaker-3: ⁓ my name is Paul. ⁓
speaker-1: ⁓ I'm the founder of Browser Base. We provide real browsers for your AI agent to do real work on your behalf. We're an infrastructure on to developers who were about two years old. And people who are calling us the browser tool for AI.
speaker-3: Yeah.
speaker-4: And my name is Sudhish Nair. I'm the co-founder and CEO of TinyFish. We are building the entire superhighway for agents to do whatever they want on the web.
speaker-0: Awesome, awesome. Great. Well thank you. I have a I have a great list of questions from my clipboard here, and we're gonna jump in and make it pretty fun. ⁓ so so I think you know the internet was designed for humans browsing pages. What changes when the primary user of the web becomes an AI agent?
speaker-1: Yeah, maybe I'll take I'll take it away. ⁓ I've talked a lot about this publicly. ⁓ and and generally when we think about how cars are driving on the roads we have right now, especially AI controlled cars, we didn't have to read roads for for Waymo. Way no just learn how to drive on the roads we already have. Interact with the, you ⁓ know.
speaker-3: We rebuild all the roads. Way more learned how to drive. And I think the same thing with AI using the internet. We don't need to rebuild the internet. AI can learn to interact with millions and millions of websites.
speaker-1: already exists. And I think about like the bitter lesson here where good that you won't have to pull the internet for AI to work with it. I think a lot of people are thinking very short term CPs or APIs for all the software we already have.
speaker-3: You should just expect the models to get so you don't have to retool. We might need ⁓ a bunch of MC And we should do that for high volume use cases, but for the majority of the long tail, agents are gonna be able to use the browser to accomplish all these tasks.
speaker-1: Yeah.
speaker-0: Super interest.
speaker-4: Think ⁓ for me the question always is like why? Why would people do anything? I the answer is that by twenty twenty eight. I don't trust a lot of things Gardner says, but one of those things that self-serving, I will trust, is that approximately 90% of B2B transactions will be identified or disintermediated by agents. That we are talking about is approximately $15 trillion worth of value. If you think of ⁓ web primary usage today, it is page view ad commercialization. And what does agents not do? They don't experience, they don't do page views. They are focused on executing. The task. So what is happening is that it's breaking the bone of what makes web commerce work. And it is going to have profound implications, which means that if I want to know what is March Capital about, I won't go to March Capital website. If I want to know what are the latest coupons on Best Buy, I don't go to Best Buy website. I'm going to ChatGPT, I'm going Cloud, I'm going to Gemini. And what is Gemini doing? It is doing WebFetch in the back. And that execution is what we need to think about. All of the website that you have put so much time and effort building is not going to be visited by humans. They are going to be indirectly visited through agents. So you have to think: is bot detection, all of that bullshit that we are currently dealing with, is that going to live? No. Today, Cloudflare announced they're going to crawl the website, they were supposed to be the guardians. If you don't see this happening, I think we are ⁓ not watching what's happening.
speaker-0: Fascinating. I I think you know agents feel incredibly viable in the last year. Things have been changing so fast in all of our in industry. ⁓ And so what do you think changed recently to enable this shift?
speaker-2: Sure. you know what's interesting is I actually don't get that answer quite yet. The so we're two hundred and eleven days old today as a company. We've publicly announced ten million dollars in in funding. ⁓ and when we went out to raise our round in August, a lot of the questions I got were When do you think agentic payments is gonna materialize as a market? And like the world we're building for is one where agents execute a majority of global payment volume. And my belief is that is within the next 10 years. ⁓ of course, there's the point we're in today, and there's the point we're in in the future. And the point in the future requires a lot of agent interoperability where agents are just interacting with other agents and there's not a lot of human interaction. But I feel like the place we've seen even between August and today is like ⁓ you've gone from just internal code gen agents or very, very basic agent workflows to long run agents with many steps, many tool calls. And that allows agents to take on increasingly wider tasks. And a lot of those are culminating in well, there needs to be a payment interaction. And so that's where we fit in. But there's just been a lot of progression in how long an agent can run without ⁓ Yeah.
speaker-1: Yeah, it's the models and the infrastructure are getting better, right? I mean, in over the holiday break, it seemed like every developer in San Francisco was vibe coding using a a plethora of different models and and seeing the progression that previously was
speaker-3: The mob. In in finding that the infrastructure that sort of limited to them in the past now exists. We now have companies.
speaker-1: You know, browser based is two years old, right? You know, two years of RD to build the infrastructure of the
Load full transcript
HS actually be productive. So I think, you know, people have said every single year has been the year of the age.
speaker-3: You need to add but we might finally have the model capabilities and the supplemental information payment actually make actual automation of work happen online.
speaker-1: infrastructure, maybe plans, maybe browsing, to really
speaker-4: I think look, if I were to add ⁓ 2024 is the first time humans became minority on the web. That is the first time we crossed the Rubicon, so to speak, 51% of the traffic was ⁓ bots, non-human, let's say. Out of that, around 40% was bad bots and 15 or so percent was good bots. Today that is probably much bigger than that. But if you look within that, the AI agents right now is less than, I think it's one basis point or something. But here is the difference. It is growing almost 500 times. We haven't had this kind of growth in the web for a long time, which means that there are approximately 3 to 3.5 billion humans in the whole world getting some sort of value out of the web. We are expecting 500 billion to a trillion agents. And this is not just AI agents that you think of software agents. We are talking about devices. Like home automation. If your laundry is going to be done by a robot, you are not going to buy laundry detergent. And if you're not buying the lot detection, do you think your robot is actually going to think about Tide as a brand? No freaking way. It's going to figure out what is the chemical composition that makes it smell exactly like what Paul's family wants, and they want organic, they will go find it. Do you think it'll care about ⁓ first page, sponsor result, and click? No. For a single answer, an agent currently reads approximately 70,000 pages. It has unlimited context, unlimited memory, hundred page, five hundred page. Which means that companies that have put so much money, like Coca-Cola, for example, $300 billion valuation, $100 billion is on the market, the brand. That brand is no longer useful. Because it's not like humans. We don't go by walking to ⁓ a Walmart and the best shelf space for tide and then pick it up and run away. It doesn't overbuy, it doesn't underbuy, it is not sentimental, it doesn't care. People who bought that also bought that. It doesn't care, only two items left. This is the transformation that we are talking about in the next few months, if not quarters. What happens when a car is driving autonomously? You don't go to gas station to buy impulse buy. What are the implications? How much of the business? Freaking love it.
speaker-0: Yeah.
speaker-1: I think we have a few more longer than a few more laundry's boxes detergent for us. But ⁓ he's all I I mean in being pragmatic, I I maybe ⁓ the most pessimistic founder who takes revenue out of the investor report, you know, as opposed to bumping it up more. ⁓ I really think it's gonna be a long time before AGI gets here and don't worry, we're all gonna have your jobs for a little bit longer, us included, hopefully. ⁓
speaker-3: until the find that it's I'm I'm messy guy in AI. I I'm but A long time. But what's gonna happen is that you're going to see legacy systems that nobody wanted to invest in, no one wanted to touch, those can be automated by I. It's gonna be a natural extension. The busy work going
speaker-1: doing, you know, good thing I'll think about we're founders. Our Delaware franchise tax was due on March 1st. Did you file yours?
speaker-4: I don't know about that, but hopefully we're
speaker-3: Well, my agent will file my own cap table and go to the and account and calculate that. And that's the kind of stuff that we can see AI payment. And then make a payment, the minimum payment. ⁓ that's the kind of stuff you'll see AI automating right now and it's automating right now. The use cases that we have gentic automation in the short term, it's the boring stuff.
speaker-1: Because they will go to a browser, we'll get my go to the Delaware Franchise Tax website, take my share. You see at browser base. ⁓ stuff.
speaker-3: Drey itch, it's taking the the container from the ship, putting on the truck and using a browser to check the inventory. I think that's a pragmatic real automation. We try and avoid putting our head in the
speaker-1: software the scheduling. Stuff that we like to focus on on browser base though. in the clouds and thinking too far out because it's very distracting to our customers and our business.
speaker-2: Yeah.
speaker-0: Super, super interesting. I I think what I'm hearing is like there's kind of these two approaches that are that are ⁓ surfacing. One is this more probabilistic approach we used with ⁓ kind of browser use and intuiting things based on the rails that exist as they are today and ⁓ enabling agents to do that. Then you also have this other side which is more deterministic, which is more code-based and more interacting with the tide ⁓ based on the chemicals, not on the branding. So how do these things ⁓ Do they persist independently on their own over time? Do they converge? ⁓ is this some kind of a hybrid that evolves over time? Like how do you see the future?
speaker-4: I mean I I can start. ⁓ I actually kinda disagree with Paul on this because the last point at this point are connected for me. Because I I'll give you an example. Deterministic execution at scale should not become non-deterministic. In the sense that if you start with an off-degree variation on one task, and if you have to do that 10,000 times, which is what enterprise always struggles with. And this is where models struggle, because models are so super intelligent that if you have to let's say click a calendar, if you want to do 10,000. Thousand times you want exact same answers. So you have to figure out how to use models for the right use cases, but at the same time, you should also figure out how to build the infrastructure to turn them into deterministic execution. The reason I said I disagree with ⁓ Paul a little bit is
that, yes, there are routine automation and there think of it like RPA 2.0. You can definitely do that. There is significant value there. I'll give you one use case example. There is a beverage company wanted to know the other beverage companies' products out there that's being sold. And long story short, it actually found all the agents were able to find all the restaurants, but 60% of the restaurants didn't have websites. And those where they don't have a websites actually went to Yelp and others to see the photos and found whether they're selling product one or product two based on the photos that others have uploaded. And within that, they found if the brand is being misrepresented. We did not ask for that. happening is think of it like mechanical execution, reasoning, outcome, and then pipe that into the next workflow. And then if you keep piping, what happens is you get a lot more than mechanical execution.
speaker-1: Yeah, you maybe I'll I'll queue it up for payments and I I think to respond, like where where we really see in like two years of learning this lesson at browser base is you have to build the right primitives for AI due to implementation next iteration model. Thanks to this guy, actually. It's just me, actually. It's just you in the lab, right?
speaker-3: Anything you try and do work around the current models limit, it's gonna be fixed in the next you you really have to focus on building the present pieces of
speaker-1: Primitive piece infrastructure. When you look at browser-based, when we started with browser infrastructure for AI, but today we launched a fetch API.
speaker-3: Because we know that AI and models can call more tools. And we're continuing to think about what's the next additional tool that your AI model is going to want to do when that's where some of the capabilities that are missing, but now payments and authentication are really the next things that AI is gonna think about. And whether it's deterministic code mode or MCPs, having the right primitives to enable AI to do the work that you do on the web is and I think that there's still a lot of
speaker-1: browsing the internet. I think like actual's working on is like Yeah. important part. gaps for us to fill out primitives. It's really the way
speaker-3: Bill, in in focusing on those problems, how do you solve problems and create the most value right now? I'm not sure if you
speaker-1: You agree.
speaker-2: Yep, Paul's a value add investor. He's shouting us out, I don't even need to be here. I think one of the interesting things that we think quite a lot about that I'll add to this is a lot of the internet and payments in specific is built around like how do humans interact with these systems. And there was a recent Mintlify post, I think it was that 60% of browser, 60% of documentation viewing is now agents. And so as we build out natural, one of the things that we think quite a lot about is how do we make sure that when agents introspect and they go, how do
I execute payments? that they on their Their own come to the conclusion of I need to introduce natural, and they can reliably pass us information about their quote unquote owner or the legal entity that they're tied to. And then they can implement all of natural's APIs and start executing. And that is a world where humans are not making the buying decision, it is agents making the buying decision, and they come to one buying decision, and that buying decision is beneficial for us.
speaker-0: You talk about humans making the buying decision. I I think it's interesting, like we s in many of these production, ⁓ in many of these agents ⁓ and agent workflows that are in production, there is still a human in the loop. ⁓ and how do you see that evolving over time? Do you see the when when do you see the human being removed from the process ⁓ in in each of your workflows?
speaker-2: This is a great one for payments. I got that question a lot too. The I think the what ⁓ my belief, and of course, like take my belief with a grain of salt, I'm biased, is that the feeling is it's slow and then all at once. Like today, you might say anything above 250 or 500 or $1,000, I want to human escalate and manually approve. And that's fine. At some point you'll see it, okay, natural runs in production, there are no mistakes. I I reliably approve and deny the transactions that I like. And then you're like, well, why don't we try like $2,500? We're $5,000. And then you're like, okay, I see that run in production for a little while longer. And then you're like, okay, let's try $10,000. And then I think genuinely, by the time we're at this conference next year, it's like, okay, agentic payments is real. It doesn't make mistakes. I trust it. It's slightly bounded still, but that's just for like my absolute base case downside protection protection. ⁓ but broadly it is like we've gone from seemingly nominal nominal payments with a lot of oversight to basically no oversight.
speaker-1: Yeah, I mean Carpathy has this really smart idea about the autonomy slider. And I think about this a lot when building products for AI at browser base. Because our customers, they might be from the heart of San Francisco.
speaker-3: And they trust the AI a ton and they want to give it the most autonomy, like go go out and buy, go out and do, I don't care. And then I have customers.
speaker-1: points where customer can
speaker-3: So you have to build the right product entry. Maybe a customer can be a lot more deterministic and build their trust in the AI before giving it more autonomy. Because in the end, you know, while browser base does support human-in-the-loop type browsing where you can see what happens in a
speaker-1: our time. For a lot of our customers that doesn't scale. You know, if you're doing a
speaker-3: a ⁓ you know insurance claim verification at night at two a ⁓ you don't want to wake someone up to watch the agent so you kind have to gradually add a lot and that's where we're seeing the most successful enterprise AI companies. They start with a lot less autonomy and as they prove out their POCs they scale up volume.
speaker-1: Okay. Yeah. go bigger and bigger and also scale volume bigger and bigger.
speaker-4: I think my view is that it depends on the use case. For example, what we have done with TinyFish is our thesis is that everybody wants to do AI projects, but they're not sure how to measure ROI on it. And last year it was okay to do science projects, but this year they want to see return. And the challenge is, just like any other project, the numerator needs to be really big and the denominator needs to be small. The risk has to be low and the value has to be high. So we have intentionally focused on use cases where the outcome is different. As long as the data is accurate and verifiable and auditable, the worst case, if we stop working, they will just go back to the way things were. That is where we have uncovered value. And in that case, human doesn't have to be in the loop as long as it is evaluator framework really works and the data quality is high. And we intentionally are, like we are happily seeing customers saying that human needs to be, like there are use cases. If you want to watch a cat video, for sure your agent is not gonna do that. But if you want to know supply chain and competitive pricing at massive scale, there is no freaking way you can do that with humans, right? So we are intentionally using ⁓ going after use cases where humans, superhuman scale is the value at very low risk. And what is happening is once they get used to the idea. They are actually getting used to the idea of human doesn't have to be in the loop. Look at a cursor. Our company used to be almost 90% cursor. Today there is no cursor. Why is that? Because we are sort of deconstructing the idea that you are gonna sit there and watch the IDE. You are actually letting cloud code cook. And you are going to trust the code, and maybe first few times you don't even trust the code, you verify the code. Now review is automated. What's happening? The crab is boiled, it doesn't know it has been boiled, right? So human is already being outlooped by codecs and ⁓ claw code. This is the same thing that is gonna happen to everything on the web as well, and we are seeing it.
speaker-0: Speaking of crabs, ⁓ we saw ⁓ you know Maltbook and OpenClaw and all these things. Nice! ⁓ actually thank you. And ⁓ and and but I ⁓ we saw recently through Maltbook what agent to agent an AI to AI interaction looks like, right? We got a sneak peek at that. ⁓ what do you think that looks like ⁓ at scale? And you know, you each agent needs ⁓ needs needs boundaries. needs guards, rails, security, ⁓ it may be carrying PII, it has all kinds of different ⁓ compliance needs. And so, you know, that that might have been what what it looks like today, which I don't think is ⁓ scalable for the long term. What what what kind of future do you see for agent-to-agent
interaction in the next few years?
speaker-4: Why don't you choose start? I mean you guys bought the company.
speaker-1: Yeah, the ⁓
speaker-0: I can't comment on it.
speaker-4: Look, I think my view is that we should not freak out about security vulnerabilities right now because this is early stage. But what I will see is that just like every other generation, things actually get fixed very fast. Web people didn't trust, cloud people didn't trust, mobile people didn't trust. Now we do. What will happen here differently is how fast it will happen. So I'm pretty optimistic about all of that.
speaker-1: I completely disagree. I mean, I think browser base cares the most about security in the industry. We are SOC2, Type 2, HIPAA. We have real PII, PHI, you know, sensitive information running through our browsers. And like security is our biggest CTO or CISO, your your developer says, run a browser in our AWS account.
speaker-3: Value add when you talk to a and hey, I want to run account, and AI is going to control the browser, can go to any website and download any file. You should see what CTOs, their head explodes when they get that developer access. They come to browser base to isolate and ensure their agents. I think OpenClaw and Moltbook is a prime example of what can go wrong, especially if you let Shadow IT run.
speaker-1: So, we can browse securely. For AI, AI within your organization. So for our enterprise customers, we think a lot about, hey, your customers or sorry, push your internal users to like a remote browser on a browser base, structure safe, everything is secure the whole time. Security is our number one focus in browser base.
speaker-3: How do you push your to using trusted systems? Keep your infrastructure safe and make sure everything stays.
speaker-4: ⁓ payment you better be
speaker-2: Yeah, I'm as I obviously we have a lot of sensitive data as well, and we execute transactions that move real money, and that's important to to do correctly. The company was started based off of a memo that I wrote on agentic payments in the summer of last year. And it's public on our website if you want to read it. It's fifteen pages. And a lot of that talks about The agent-to-agent payments use case. ⁓ and so obviously I have a strong belief that that is where the world eventually gets to. I don't think it's this year, and a lot of the customers that we serve this year won't be doing that. They'll be doing real-world things, automating legacy workflows and things that are very B2B embedded looking. as we think about how do we scope natural and build like domain modeling and data constructs that make sense for that world, ⁓ it is important that we do that correctly and safely and I don't think agent to agent becomes real unless we do our job well. So
speaker-0: Cool. I I think ⁓ I want to go back to something you said previously about ⁓ like tide and like you know taste and and user preference. You know, it in this in this future where agents are
handling a lot of the workflows for consumers, prosumers, and businesses, ⁓ what what is how does do you think that changes ⁓ today's like advertising and search models, not not not like LLMs, but like the model of how search works today.
speaker-4: I think this is a question that I'm very ill-equipped to answer because this is a complex question that Google will have to figure out. Because the fundamental fight that is going to happen that will redefine all of this is between what will happen with your company and Google on this. Because as big as the company is and everything else, Google still approximately, I might be off by a few points, but close to 80% of their revenue still comes from ads. There's a reason why they're not opening up semantic search for agents directly, as you know, other semantic search. Companies are doing it because it's a cash cow. We are sitting in this bubble talking about all of this, and when you step out, people still don't know Claude. They still know just Chat GPT's AI for them. So, what I want to make sure is that the point of contact is already changing, which means that if I want to know about a product, a review, a person, something, research, it is all going to chat, which means OpenAI will have to figure out how to monetize it. Because when you have access to data of users' first point of contact, Contact, you are the most powerful person. And you will figure out how to monetize it. And Google knows that, OpenAI knows that. These are the only two players of consequence in this space on this space. And this battle is going to be brutal. Whatever happens there will define. My sense is that Google will have to evolve because the old model will change, and when shift happens, it will be not that gradual.
speaker-2: Uh-oh I'll give you the self-serving answer, but I genuinely believe it, which is ⁓ and it's the pitch we give to customers. I think there are a lot of companies today that build agents that charge some some pricing that they determine whether it's a success-based fee or this is a subscription fee or a platform fee or whatever. ⁓ I think all those companies should be payments companies. I think they should be in the flow of funds. I think they should be taking a take rate on GMV or TPV or whatever you wanna call it. I think it's basically like What would Airbnb's business be if they weren't essentially a payments company? I think that is the future of how agents charge, and and I think that ⁓ when it looks that way, their their outcome is one or two or three orders of magnitude larger.
speaker-1: I don't think Google's going away at all. So I mean, I think the search model's gonna change, but search APIs are very powerful. I think companies like Profound are driving a ton of traffic to websites through GEO. ⁓ we work with search partners like EXA and Parallel, and those are just really incredible businesses that are really blowing up right now. I can imagine a sponsorship model there at some point.
What I think about the time is like how often of our customers are referred to us from
speaker-3: one of the things that I all the time often, you know, we see almost twenty percent of AI. ⁓ and you know, we don't do very much.
speaker-1: Search advertising at all. So really I think you have to think about how you're investing your content, how it's being propagated to the weights of the models and models are using search APIs, reading comparisons and saying browser base is the right choice for building AI infrastructure. So I think we're gonna see more of that. I don't think it's gonna be as disruptive in the short term, but maybe five years from now we'll we'll come back and talk about it.
speaker-3: Hopefully it's showing up in the right places, but a lot of times these going to pages, reading sheets.
speaker-0: Awesome, fascinating. I I think ⁓ there's there's a lot of new companies that are possible because of this technology. And you all are building at the frontier and the very edges of it. If you were starting a company today and ⁓ you wanted to leverage these technologies, you know. What do you think, what new businesses are possible ⁓ as a result of the agentic infrastructure being here that were not possible before? And this question is like for the founders in the audience as well as the investors in the audience for like what what type of things to be on the lookout for?
speaker-1: Maybe take it first. I I put this out on a tweet recently. I think post-AGI there will be like three to four types of businesses. There will be AI labs, there will be infrastructure companies, there'll be systems of records, which might collapse into infra. And then there'll be services businesses. And I think the majority of businesses in the world will be, you know, services probably enabled by AI. The numbers that work at labs will not scale with the size of the population. But I think there's just a lot of opportunity in AI-enabled services. I think there's so
speaker-3: I think a lot of interesting opportunities in so many special
speaker-1: Areas of the market that are underserved where there aren't enough people or the services are too expensive. ⁓ And there's, you know, the single person dollar revenue services business is gonna be started in the next year or two using all the tools that we have here. And looking at these underserved populations, ⁓ and that could be in professional services or something along those lines, there's a lot of opportunity there to go build a great business. Might not be as sexy to invest in, but you can build a really cool business there.
speaker-3: The next single hundred million I think looking at ⁓ It may not be venture scalable, you might be to invest in
speaker-2: You know, I really I think I spend a lot of my day thinking about payments, not a lot about other people's companies, except if there are customers. I think the thing though I am genuinely excited about is I think at the tail end of last year you see this like proliferation of companies that are built on top of the fact that voice infrastructure is very good or browser infrastructure is very good. And I think that will be true of payments infrastructure this year. And so if you ask me like who does the most TPV in June, I think that answer is just very different than who does the most TPV in December. I don't necessarily have the creativity or or will to come up with those ideas, but I'm excited for those founders who do.
speaker-4: I ⁓ will echo that. And the way I will say that is 1849, the gold rush happened, and people were mining, kneeling down, clothing was not good enough. Somebody invented jeans as a material and he's still wearing that 200 years later. This is literally one of those moments. We don't know how many will be created and what will be relevant 50 years from now. You're all most of your finance people, modern banking came out of gold rush because people were up in the mountains making money, they want to send it to the new world. Elevators, artists discovered that as we were. So this is one of those moments there will be Levi's where 1849 and 49ers will be still playing in the same place. Some of them will become iconic, some of them will not. But ⁓ it'll be very difficult to predict. I think Paul's point of view around the categorization, I agree with that. How many of them are built with the right kind of founders to be a Levi? I don't know.
speaker-0: It's a great time to be building ⁓ right now. I feel like the the the models, the LLMs that have enabled so much growth and and reasoning ⁓ capabilities for long running task horizon, ⁓ long horizon tasks, ⁓ and then the infrastructure that you are building enables the whole new wave of folks to come and build on top of it and create things that are brand new and accretive to the to the whole
speaker-4: I'll just say one last thing. I love San Francisco, and what OpenAI did to San Francisco is fantastic. So I just want to say thank you. No problem.
speaker-0: No problem.
speaker-3: Okay.
speaker-1: How many credits did you for that?
speaker-4: You wish.
speaker-0: ⁓ we'll end it there and thank you so much for coming. It was awesome being here.