🎙️The browser as a body for AGI?

Now to the show: What has the browser unlocked for us so far? First, it let anyone navigate the digital world by clicking. Then it became the home for our apps, our work, our life. Ben Goodger believes it is about to unlock something else: a world where machines understand what we want in our own language, inside the environment where we already spend most of our time.

Ben Goodger, head of engineering for ChatGPT Atlas at OpenAI (and former Firefox and Chrome builder – can you imagine, building three browsers in three different eras!), explains why this moment feels like a fresh Netscape 1.0 moment and why the browser is becoming the body for an AI brain.

In this episode of Inference, we get into:

• How AI inside the browser shifts user cognition and sparks new curiosity

• The real bottlenecks of building an AI browser in 2025

• The “lethal trifecta” of private data, untrusted content, and external actions - How Atlas manages autonomy with transparency, guardrails, and that literal red stop button

• The best use cases for AI browsers right now

• OpenAI’s strategic view: AGI needs a body, and the browser is the first one

• And why East of Eden – a book about family – matters when building human-computer relationships

We also discuss the new browser wars, the long-term strategy for OpenAI, and what early users still struggle to understand about “an AI inside your tabs.” This is a conversation about the interface where AI will actually live, the hidden engineering behind agentic browsing, and the next chapter of how humans communicate with machines. Watch it!

We prepared a transcript (edited by Genspark Super Agent) for your convenience. But as always – please watch the full video, subscribe, like and leave your feedback. It helps us grow on YouTube and bring you more insights ⬇️

Ksenia Se: Today on Inference, I'm talking to Ben, someone who helped build three browsers through three different eras: early Firefox, Google Chrome, and now ChatGPT Atlas. Each time the web changed direction, Ben was there rewriting the foundations. That's a lot to carry, Ben.

Ben Goodger: Well, thanks for that intro! It's a pretty incredible journey that the web has been on over all these years. I'm very impressed with how the web has remained extremely relevant, even over such a long period of time. It's sort of kept up with the times. And it is, in fact, now powering this latest wave of AI, which makes it an extremely exciting place to be building.

The Cognitive Shift: From Search to Conversation

Ksenia: This era of AI browsers is so very new right now. How do you think about what's changing at the cognitive level for users who now will interact with AI in this browsing flow?

Ben: I think one of the things that bringing AI assistance directly into the browser helps with is just lowering the friction associated with asking questions of this technology. One of the amazing things about this technology is that it's so easy to ask questions in whatever way you want to ask them, and it will just come up with a good response for you. So just making it easier to access that capability is really important.

When I've used Atlas, I've found it's made me more curious. I find myself asking more questions, even things that I wouldn't have before. Because I think in the previous era, when search was primary, I would have to think for some time about what the search query I'd want to run would be. And sometimes a question doesn't even fit the mold of a search query. It might just be some piece of information, and then you'd have to think about: How do I get this information? Where should I go? What should I do with it?

Instead, with these AI assistants like ChatGPT, I can just ask the question and it will give me a response. Sometimes it will search and sometimes it will just know. So just being able to do that from wherever I am on the web is really fun.

Ksenia: I remember the times when people would compete in better Google search phrasing. The better way you phrase it, the better results you get. It's a little bit like that with prompting right now, but I totally agree with the question part. You can be a child again. You can ask as many questions as you want. That's fun.

Ben: Right. It's also extremely helpful, both to help you navigate around the web, helping you find the most effective site to go to, the best deal on a product. If you've been shopping recently, doing holiday shopping – that's something I've been up to that's been super useful.

And then just helping keep on top of all the work that you have to do as well. Actually, I think we're on the cusp of these agents beginning to take action for you on your behalf, which is also really exciting.

The Messy Web vs. The Agentic Web

Ksenia: It is very exciting, but also the internet is super, super messy, right? And we're seeing the split between the human web and the agentic web. Humans are used to navigating this messy environment, but agents do not want this. They want a clean state. They don't want the soup of all this ancient code that we live in. So what does the web look like from the agent's point of view? And as an engineer, how do you build for this new paradigm?

Ben: That's a great question. I think it depends on how the assistant is operating the web. There's a bunch of what we've seen with the development of this agentic web – a bunch of different technologies used to get stuff done across websites. And we see this in ChatGPT.

We see features like connectors, which have sort of API access directly to a service's content. That is very fast and very direct because the developer of the service has exposed that data directly to ChatGPT through that connector API.

Then there's something that really requires no participation at all from the developer. If you think of ChatGPT Atlas, we have Agent Mode. Agent uses a variety of tools – either to work directly with the site's DOM code or to work using computer use vision aspects. In that latter mode, it really can operate without any participation at all from the developer and still get many things right.

I've been impressed by – I was updating a spreadsheet recently with a hobby of mine, going and extracting a bunch of data from various websites. The agent was able to do that quite effectively.

Ksenia: But do you think we will need to rebuild the internet in a sense to make it easier for agents?

Ben: I think there will be some amount of that that occurs. There's a lot of interest from developers in making their services available in these new ways that users are exploring the world. So I think some of that will happen as a byproduct of that. Then there'll be some cases where people want to upgrade their site or make certain things easier.

What we're telling folks right now is: just making your site more accessible is actually the best thing. All of the tools and technologies that you would use – web accessibility – to make your site more accessible, for example, to a screen reader, actually helps make it more accessible to an agent. And that's just a good thing to do anyway, because we have people who are vision-impaired or have no vision who would use a screen reader to use your site.

So I think it's good just to support that technology anyway. And some of that stuff will make it much easier for agents to operate as well.

Building a Miniature Operating System

Ksenia: What are the main bottlenecks for you right now building the AI browser?

Ben: Well, there's just a lot that we want to do! We're trying to do many different things. A browser is – I think of it like a miniature operating system that runs inside of your desktop operating system. It's where a lot of your favorite apps run. They run in tabs.

So there's a lot of functionality inherent in that. We started this project at the beginning of this year. We've been building out a lot of that functionality. We shipped the browser on macOS just over a month ago, and that was a bunch of work. We're working right now to bring it to Windows and mobile.

There's a lot of that work for us to do. We're also digesting a huge amount of feedback that we've received from the user community, which is really exciting to get and work through and add improvements. So the team is – even though we're working on Windows and mobile and stuff like that – we're also improving the Mac app on a weekly basis.

There's a huge appetite from us to improve the product, to make it available to more people. And as always, there's more work than we can do. So we're trying to scale with that.

Ksenia: I wonder about the feedback. What is the main feedback? How did it change your priorities?

Ben: Somewhat predictably, putting ChatGPT at the heart of the browser is a net new capability that people don't fully understand yet what that means. And I think we don't either. One of the reasons we're so excited to get feedback is we hope to learn from folks as they use it.

We certainly have some use cases ourselves that we like. We want to hear from the user community as well when they discover things that work really well for them.

But I think for a huge number of people, just thinking what question to ask is sometimes hard. From our perspective, we want to figure out ways to make the value-add of having ChatGPT there more obvious to people as they browse. I think that's something people are still figuring out: How do I use this new capability?

And then of course, because it is a browser and people are passionate about browsers – they spend a lot of time on them – they just have a lot of feedback on the browser itself. We launched very early. We didn't have all of the traditional browser functionality, things like multi-profile, things like tab groups, stuff like that.

We're in the process of adding a bunch of that stuff, or we have added some features already that people have asked for. So we're also just working on some of the basic stuff too. Just trying to keep that steady continuous improvement going.

The Lethal Trifecta: Safety by Design

Ksenia: Right, browser becomes your little home on the internet where everything happens for you. It's so important. I was also wondering about some tension around building web browsers, specifically AI browsers. Because with more autonomy comes more responsibility, of course.

If you read Simon Willison's recent blog post where he writes about the "lethal trifecta" for AI agents: private data, untrusted content, and external communication. I heard the DIA team recently was also operating with this trifecta, thinking about building an AI browser.

So my question is: How do you think about these concerns? And again, as an engineer, how do you address this problem – these three very important parts to prevent attackers from extracting your private data and just ruining your experience?

Ben: It's a huge responsibility. For more than a decade now, we've known as the browser community that safety is sort of a top concern.

But the way I think about this in the world of agents – a truism that I've always felt for products is: transparency and control for users first and foremost. And so this is infused in the way that we have designed the ChatGPT Atlas agent, for example.

It's something that you're always opting into. It's never just going to run or happen. It's never going to run out of control without your knowledge.

The first time you run it, we give you this nice disclosure that explains how it works and tells you what it can and can't do so that you get some information.

Then when it does start running, you actually have choice as a user on whether or not you want – for the task that you've given it – do you want it to run on your private data, or do you want to run signed out? So you can choose to go either way. If it's more of a generic public web research task, you might get it to run signed out. If the goal of the run, the thing you're asking it to do, is to have it work on your private data, then you may let it do that, but you can choose.

And then when it is running, there is knowledge that it has about whether or not you're doing certain types of sensitive tasks. Some of those tasks will require you to attend it. So at this stage, if it's in your email, for example, you can't switch away from the tab. If you switch away from the tab, it will stop, because it doesn't want to accidentally do something that you might be unhappy about. So you're going to sit there and you're going to watch it.

And while it is running, there is a little bar at the bottom of the screen with a big red stop button. If you've ever been in a machine shop and there's a machine that spins or cuts or something like that, there's always that big red stop button that you can hit. So if something starts to go in a direction that you're not comfortable with, you just hit that button and everything stops.

That's kind of the mindset that we have right now. I think over the course of time, we imagine evolving and improving the set of capabilities that we have so that we can make this more seamless. But that's where we're at right now. That's reflective of the state of the technology right now.

We think nonetheless that this is incredibly useful and people can use it to do some good work, but we want to make sure that people understand the current state of things and give them some good controls around that.

Use Cases: From Curiosity to Commerce

Ksenia: What would you say are the best use cases for an AI browser?

Ben: Right now, I think where this product shines is really helping fuel your curiosity around just anything that you might be interested in online. Whether your passion is a specific hobby where you're learning more and you want to boost your knowledge by asking questions about something without having to necessarily repeat yourself or type the name of the thing that you're looking at over and over.

I think we have this experience with ChatGPT in a pre-AI browsing context where maybe you have a ChatGPT tab and then you have to go and tell it, "I am looking at this thing about something or other," and you have to sort of repeat yourself over and over every time you're asking about that.

Whereas in Atlas, you don't need to do that. You can just pop ChatGPT open when you're looking at something and just ask questions directly about it.

I think it is especially good at commerce. Over this Black Friday, Cyber Monday weekend, I've been doing a lot of shopping. It can help you find good coupon codes. It can help you find better prices on things. It's really good at researchy tasks like that that involve the web. So I think that's super helpful.

And then, like I said, it's early days for Agent as well. But there are certainly some tasks that I've found that Agent has been really exceptional at – just automating a lot of the drudgery.

I've had it make survey forms for me using Google Forms. I find the Google Forms UI a bit complicated. I'd have to go in and craft a survey form, and I've just told it what I want. And it's gone and made one for me and handled all the clicking and just given me something that I can send out with relatively little effort. So there's stuff like that that I think is super useful too.

Ksenia: How does access to ChatGPT correlate with what you have in Atlas? For example, you have Team subscription or you can have Pro subscription. Does it – if you use Atlas, does it move what you have, what you're paying for, for ChatGPT?

Ben: It just inherits the ChatGPT subscription level that you have, and so you get all of the access capabilities.

Ksenia: If you want to do deep research, will it do deep research if you have this?

Ben: Right, yeah. So you get the access limits that you would have under your plan – that's what you would get in Atlas.

The New Browser Wars

Ksenia: That was just a practical clarification. Let's talk about – we recently saw a launch of different AI browsers. It was Comet from Perplexity, DIA by the Browser Company, Chrome is integrating Gemini everywhere. And you lived through at least two browser wars, right? So what does the new war look like to you, now that OpenAI also entered this space? And what's the real strategic point for OpenAI launching the browser?

Ben: OpenAI's mission is to make sure that AGI benefits all of humanity. And I think part of that is both reach and then also AGI itself – what that looks like.

We have an incredible team here at OpenAI building cutting-edge models. I am a product builder myself. I am not a researcher. I love to work with researchers. Researchers are building some of this incredible brain that will power so much of the future of technology.

But a brain needs a body to operate in. And so I think AGI is something that can take action for you. The browser itself is an environment where that can happen. That's why I'm so excited about the browser.

I think the browser is a stepping stone to AI being able to orchestrate more of computing for you. It's a great environment for us to feel out how that works.

And I think one of the huge opportunities for us in building this product: we weren't starting from a baseline of something else. We weren't starting from a lot of legacy where we had to deal with a certain amount of expectation. I mean, I think there is expectation inherent in a browser. People kind of know what it is and how it works. So in some places, we've tried to push the envelope a little bit and then realized we went too far and had to back off a little bit. But that's something we can kind of control ourselves.

So what this product has allowed us to do is to really go back to first principles and say: If we have an AI assistant at the core of the app, what opportunities does that give us? We've built a few of them into this initial version. I think this is something where we're excited to iterate with our user community.

I think long-term, where I see this product going is it will become kind of a task management console for your life. Really, it's helped make me more curious. I think the next step is to help me self-actualize more in different areas because it is helping break through some of the ambiguity and bring me concrete actions that I can take, and even maybe pre-run some of them if they're easily reversible and that sort of thing.

So I see that as the direction for this technology that I would love to see it take. I think we're really at the earliest stages. As you mentioned, I've worked on a few browsers before. I view this as kind of a – if I use an archaic analogy – I could say this is like a Netscape 1.0 moment for this new era of the web. So I'm super excited to see how it plays out over the next few years.

From Idea to Reality

Ksenia: You wrote in your LinkedIn post that you joined OpenAI a year ago with just an idea of how this new browser could be built. How far are you from – or how close are you to – this idea that you initially had?

Ben: Like I said, I think we're just at the early days of building it. It's hard for me to predict when, for two reasons.

One is I've been pretty stunned over the past few years by how rapid progress has been in this field, as I think we all have been. So what I've learned is not to underestimate how quickly these models will improve and thus what capabilities will be unlocked by new iterations of them. So that's one.

But two, there's this iteration loop that needs to run where we learn from users, we improve the product, models improve, we generate new iterations of the technology and so on. So there's some ambiguity inherent in that.

But I would say that the next few years are going to be very interesting. I think we're going to make some major progress. And the team here is really fired up to get a lot of stuff done. So yeah, super exciting.

Building Trust Through Control

Ksenia: How do you build trust?

Ben: I think trust is built over time. It's built through giving users the sense that they have control, that they understand what's happening, that the system is being transparent with them about what it's doing.

I see this kind of progression playing out. We want to make sure that we give users that sense of: I can keep my hand on this and I can keep some control. But eventually, we hope that the accuracy will improve to the point that they feel the ability to just let go and let it do really valuable things for them.

AGI and Superintelligence

Ksenia: You mentioned AGI, and I always ask this question to people who I talk to. What is your definition of AGI, and is the browser part of this path to AGI? And as a clarifying question, how does it relate to superintelligence? Because it's a little bit of a mix of terms.

Ben: I don't want to get out ahead or differ from other perspectives at OpenAI on this, but what I'd say I'm excited to work towards is something where this technology can take really meaningful, valuable action for you and help you sift through all of the challenges and opportunities in your life and get stuff done for you in a really meaningful way. So that's the kind of product that I'm interested in building.

I think in the future, we will see the computing platform that supports all of this evolve in massive ways where these technologies can do things way faster than you can. What that means is that you can do much more.

So I'm very bullish on productivity increasing massively as a result. There are certain things that I do by hand today where I might spend an hour or more doing some information retrieval drudgery, whereas I can imagine a system that is massively parallel doing that in minutes or seconds even.

Whatever definition that fits, I'm not sure, but that's something that I'm excited to see happen and I'm excited to build.

A Practical Guide for the Skeptical

Ksenia: I always try to leave my readers and viewers with some practical guidance. So if we imagine a person who distrusts AI browsers, what would your guide be – step by step – for how to approach it and try it?

Ben: I would say with ChatGPT Atlas, you can just use it like a normal browser. I think that undersells the benefit in it, but certainly you can just use it as a browser.

You can ask questions with the Ask ChatGPT sidebar. There's a button at the right of the toolbar that you can pop open and just ask a question about whatever you're looking at. That is, I think, about as low-friction as it can be.

And you can start simple. You don't have to jump into Agent Mode right away. Just ask simple questions about pages you're looking at. See how it helps you understand things better or find what you need faster. Build up from there as you get more comfortable.

Planning in Uncertainty

Ksenia: Let me ask you one more question about how you plan the development. We know that the models – apparently you operate on OpenAI infrastructure, so you depend on the models that are launched and released. But how is your planning going? How far ahead do you plan the development, the engineering plan for Atlas?

Ben: Well, right now we're working on bringing Atlas to Windows and mobile and keeping up with the feedback that we've received from our Mac launch. Those are pretty concrete milestones that we're working toward.

We organize in a fairly agile way where we're constantly reassessing priorities based on user feedback, based on model capabilities, based on what we're learning. So we have short-term concrete goals, and then longer-term, it's more like we're building capacity in our team over the course of time to be able to do more ambitious things.

But for right now, we're pretty busy just dealing with the feedback that we have.

East of Eden and Human Connection

Ksenia: My last question is always about books, and if you can share the books that shaped you or maybe influenced you recently.

Ben: So a book that I read recently that really moved me was a recommendation from my wife. It was East of Eden by John Steinbeck. I think it was probably the most beautiful, poignant writing that I've read in that sense. It felt cinematographic – I don't know if that's a word – in a way. Just the feeling of the Salinas Valley and how that felt at that time, at that point in time.

It was really quite – of course, a gripping tale of the personalities involved. This is less about the world of technology and more just about the relationships that we have as people with each other and some of the oldest types of relationships: the relationships between siblings and family and all that kind of stuff. But I think that those relationships are really important. So I was really moved by that.

Ksenia: That's a wonderful book. Do you think a lot about the relationship that we are building in this new world with technology?

Ben: I think one of the things that's always motivated me to be in technology – originally I was someone that didn't like technology because I thought it was too hard. You know, I had to learn all this stuff that didn't seem that relevant. Like, should I learn this arcane command-line syntax or something?

Drag-and-drop was a big unlock for a lot of people. And then the web was another big unlock because it feels like with the web, you can just click around and serendipitously explore the way you would explore a new city.

I've been to Amsterdam a couple of times over a long period of time with gaps in between. But when I go back, I can kind of remember the maze-like layout of the streets and how to get from one place to another, which is fun. I think the web is a bit like that.

And so then I think of AI as kind of the next step in that progression of technology, where people can truly convey what they want to the machine in their own terms.

And I think now our obligation as technology builders is to make sure that that machine can do the thing that they want. But it can certainly communicate back in a way that feels very relevant, very meaningful, empathetic, which is just a massive evolution in how technology works.

Ksenia: Let's finish on this beautiful note. Thank you so much.

Ben: Cool, thank you. Great to be here.