Darin Fisher is a little older than the fresh-faced, newly minted PhD types you see roaming the well-appointed floors at OpenAI’s second location in San Francisco’s Mission Bay district.
Before arriving at the AI super-startup, he spent 25 years working on some of the most important web browsers in the history of the web: He worked on Netscape Navigator, which helped define the early consumer internet. He worked on the popular Firefox browser at Mozilla, then went to Google, where he was a member of the Chrome team. After Google, he wanted to explore alternative browsers; he did so first at Neeva (which offered an ad-free experience), then at the Browser Company, which built the influential Arc browser. “The opportunity to come to OpenAI and infuse the AI model into all of this and to think about how that can really transform the experience was all kinds of super interesting to me,” Fisher says.
In OpenAI’s new ChatGPT Atlas browser, all tasks start with a prompt to the AI models working in the background. As the user accesses the web, the chatbot, which rides along at the right of the screen, can see the content of each webpage, answer questions about it, or take actions on it. An agent mode allows the AI to perform complex, multi-step tasks like filling out forms or shopping on the user’s behalf.
We asked Fisher about the choices, trade-offs, and innovations that went into designing OpenAI’s AI-first browser. The interview has been edited for length and clarity.
What’s the core philosophy behind Atlas?
One of the most important features of Atlas is really that chat is at the heart of it. You should start your journey with ChatGPT. Not because we work at OpenAI, but because I actually realize so many times, I kick myself “why didn’t I just ask the model first?” It would save me time. It should be the thing that’s there on autocomplete in the browser. It should be there so conveniently, effortlessly. This is the feedback we hear from people who’ve gotten to try Atlas.
How did you approach the design challenge of getting people to adopt new habits?
There’s this dance of the familiar as well as moments when you can optimize some things. I worked on browsers that have tried much more radical takes on browser UI, for good reasons, because there’s frustrations with a traditional browser UI. When you have all these little tabs at the top, at some point that starts to break down and you’re like, “Well, I guess I have to clean up now.” We really tried to get the basics right because people live in this thing. Not only are they used to it, but you’re working. You don’t want to be frustrated by it.
One of the big challenges of browser design is tab management. Can you explain Atlas’s scrolling tabs feature?
We built a classic tabs mode which works just like Chrome, but we also have a scrolling tabs mode in Atlas. What it does is it makes the tabs stay a fixed width and they start to scroll. But importantly, for this to work, new tabs enter at the left, which can be very disorienting for people. But once you get used to it, it’s kind of neat because what it means is new tabs are always opening on the left, old tabs are going off to the right. Your area of focus and what you’re working on stays on the left and all the tabs are nice and visible. What’s really cool is once you get used to this, you start accumulating a bunch of tabs and you can use the tab search feature Command+Shift+A. It’ll find your tab, and then you can zoom it back to the left.
How did you approach building on Chromium while maintaining design freedom? [Chromium is a Google-developed open-source browser project that provides the foundation of several major browsers, including Google Chrome itself.]
When people build on top of Chrome, you’re sort of constrained in some ways to the shape that it takes and the structure that it has. Not because you couldn’t change a lot of stuff, but because the more you change, the harder it is to update Chrome. What we really wanted to do was have our cake and eat it too. We came up with this clever way of essentially running Chrome almost unmodified, projecting the contents of the web pages into a Swift app [Swift is an Apple-developed programming language for building apps for Apple devices.] So Atlas could just be a pure Swift app, a relatively small app actually. It means we had this blank canvas on which to make anything look like anything we want. We were very free from a design perspective to rethink how so many things work. The main constraint is what people think a browser is and how they think it should work.
Can you walk me through the side chat feature?
When you click on links inside of ChatGPT inside of Atlas, it does this transition of moving the chat into side chat and opening the web page. This side chat . . . is now connected to the site that you’re on. You can ask questions about the site. [While shopping for couches at a retail site, for example] it could be things like, “What is the price range of your couches?” or “Who else sells something like this?” The model can go and look at the internet and tell you about all that stuff and you can just ask it a very simple question. Sometimes these websites are so cluttered too. A good example is a recipe site. You might say, “Can you just tell me the recipe?” You just ask the model to do that for you. Even on the recipe, you can be like, “I actually want to make this for four people, not six. Adjust the amounts of each portion.”
When you open a new tab, you don’t immediately see the chat sidebar. You see a chat window right in the middle of the page. Why did you do it like that?
We really went with this idea of one box. You should feel like it’s a simpler, cleaner experience. When we did this, we got a lot of feedback where people were like, “Where’s my URL bar?” We discovered that we could put one there—you just have to hover and then it will be there. Or if you’re a keyboard user and you did Command+L, it would activate. One of the innovations with Chrome was one box where you can enter URLs or searches. But if you go to Chrome and you open a new tab, you’ll actually see there are two boxes—one at the top and one in the middle. We’re like, “Well, can we have one instead of two?” It was remarkable that it didn’t have to actually be in your face for that to be true. Everybody was happy. Everybody has no problem finding it.
Using “agent mode,” you might, for example, ask Atlas to go out and find the best deal on a plane ticket, or even go further toward a purchase.
You’re definitely asking this thing to go do stuff on your behalf, but you want to feel in control too. There’s a stop button that’s very prominent. There’s a take control button. The model is tuned to understand that it should present you with results at a certain point and now you can take it to the next step, review its work, see what it’s doing. The model can open multiple web pages in the background to do its work. Also prominently, you can choose, “Do I want this to use my authentication, my cookies, or no?” [This might be your username and password at Google, and your preferences stored as cookies.] This is actually huge because maybe you just don’t trust it yet and you want to develop some trust. You want to see what it’s going to do and you want to try it out in a safe way.
How do you think about ChatGPT search versus traditional search engines?
Google, full stop, is amazing. It’s an amazing tool. But at the same time, it works the way it works. It works a certain way. People are used to it as the tool that it is. This AI stuff is different and it’s a different kind of interaction. What we’ve done to improve ChatGPT search capabilities inside of Atlas was not just because it was important for ChatGPT, but also because it’s kind of essential when you’re approaching it from a browser lens. Sometimes you’re typing in that box because you have high navigational intent. Like, “I want DNS for the web” or “I want to go look at a product on Amazon.” I don’t need anything else other than just get me there. Google’s outstanding at that, but ChatGPT with these capabilities does a great job at that, too.
What did Sam Altman and the leadership team tell you about wanting to build a browser when they brought you in?
The focus was squarely on bringing ChatGPT to more people more readily and having it be just core to that experience—realizing that in a browser traditionally you’re like Open tab, Go to ChatGPT.com. There’s just an extra clunkiness to that. The time’s right, you want a more streamlined experience, let’s go build it. Sam was super clear about that, which actually I really appreciate. This is a company that is rapidly putting out features. We’re often in the mode of “what can we do this month?” But this is an investment, because it’s not just what we’re going to do in January—it was what are we going to do in January to unlock what we could do in November? We spent all that time building a foundation and now we’re on a weekly cadence of putting out new features and new things and building on this foundation.
How did you approach testing and getting feedback during development?
We are an opinionated bunch, that’s for sure. But you want to validate your ideas. We had the internal population of OpenAI, which is not the most representative sample—these are really heavy tech folks. The kind of features that people are asking for internally would be things that we know from past experience might not be what everybody wants. We also brought in trusted testers. We brought in friends and family. There was this cohort of students and other cohorts—we got feedback from different folks to help inform and understand just how people experience this. We had this guiding principle that people can only learn so many things differently. You’ve got to start out with the familiar.
What’s been the feedback since launch?
I think we’ve gotten an enormous amount of positive feedback. There’s general excitement around these agentic browsers. Maybe some degree of people wanting to just kick the tires and see what it’s all about and maybe some level of skepticism too, but also enthusiasm, generally . . . The feedback really hasn’t been surprising at all. For example, we know that it’s a lot to ask for people’s habits to change on anything. Bringing ChatGPT front and center in the experience is a big change actually. For many people, early tech adopters, this feels very natural and they actually remark at how comfortable it feels that they’re [the browser and the AI] side by side. But I think for most people, they’re still early on that journey to be honest.
What features are you working on for the future?
Every week we put out a new version addressing feedback. We’ve already addressed some of the top feedback that we got. There’s a lot of stuff that was in the hopper that we paired back because we didn’t want to just go out—we wanted to go out with as polished of an experience as we could. Adding a model picker in the side chat was one—they already added that. Vertical tabs is another feature that we’re getting a lot of requests for.
What was the biggest design challenge in Atlas?
Thematically, probably one of the biggest arguments was just, “Is this in or out? How do you keep it simple?” Pretty soon if you’re not careful, you have the kitchen sink, you have the Swiss Army knife. You’re trying to satisfy everybody, but you satisfy nobody. You don’t want to overwhelm at the start. You want it to be familiar, easy. They [users] can do things that maybe aren’t the most efficient, but that’s okay. Then they can start to learn how to leverage more efficient ways. As a user, it keeps you in control. When I want to use chat, it’s there for me. But if I don’t want to be using chat on a web page, I don’t have to. I think that’s very important. You should feel empowered.
How do you design a product that has a deep set of features but still looks simple?
There’s this idea of progressive disclosure in design. [We] have a new product that can do all these things and maybe you’re really eager to try to tell everybody about it, but if when they open the product the first time, it’s telling you about all the things you can do, suddenly you’re like, “I don’t know what to even do.” Progressive disclosure can mean that as you use the product it might advertise progressively different features, but I also think of it as there’s a bunch of Easter eggs for you to discover. What do they do? They give you superpowers and help you feel like you can do better, but they aren’t in your face so you still get a product that’s approachable.
Looking ahead, what excites you most about the future of Atlas?
This is the beginning of a journey. It’s going to continue evolving. I think it’s also going to continue evolving as ChatGPT itself evolves. All these things are being built in tandem by different teams at OpenAI. A lot of ChatGPT’s best features, including things like deep research and study mode—all of those are in Atlas too. There’s all these modes—model picker, tools that you can invoke. At some level, those are powerful tools—but do you want to try them out? But over time, the model absorbs some of that. There’s a natural tension there. You give people a palette of things, but you also want to keep it simple. Ultimately, it’s just meant to be “ask the model to do stuff for you.”
