OpenAI's new browser is great for Google
After weeks of testing, Atlas has changed some of my core workflows.
Several weeks ago OpenAI released their ChatGPT-powered internet browser, Atlas. Atlas isn’t the first AI browser (Perplexity’s Comet gets that distinction), and it certainly won’t be the last (you can be sure Google will add similar AI features to Chrome, for example). I waited to try an AI browser because I didn’t see a use-case for it that was compelling enough to change my browsing habits. However, I do try to experiment with the latest technology as much as possible to keep abreast of current developments, so eventually I caved. Atlas surprised me in two ways: how useful it is for things OpenAI did not advertise, and how absolutely terrible it is for things they did. Over time the limitations will certainly disappear, but my final verdict is that it’s useful enough even now to change some of the ways I’ve been using computers for years. Ironically, Atlas has led me to preference Google’s apps!
When you start a new Atlas window you get what looks just like the ChatGPT webpage, or the stand-alone ChatGPT app. In the center of the screen there is the OpenAI logo, and below that is a text box where you can enter a prompt. There’s also a tiny button in the top-left corner of the screen that, when clicked, reveals a side bar on the left with your history of AI chats. For ChatGPT users this will all look very familiar.
Typing the address of a web page in the prompt window (or the menu bar), instead of a prompt, takes you to that page. From there the experience looks and feels much like any browser, e.g. Chrome or Safari. There are tabs, bookmarks, incognito mode, etc. So, right from the start, Atlas can either function like a conventional browser, or like the ChatGPT app. That’s already a very small win … until I started using Atlas I always had the ChatGPT app open as well as a conventional browser. One app is better than two, if that one app can really do everything.
What brings these two functions together (ChatGPT and browser) is a button on the top-right corner that says “Ask ChatGPT”. That opens up a sidebar on the right where you can have a conversation with ChatGPT about whatever is in the browser window.
OpenAI’s big selling point was a new “agent mode,” that allows ChatGPT to take control over the browser window. For example, if you select agent mode and type “Find me the nearest Mexican Restaurant that’s open for breakfast”, ChatGPT will get to work in the browser window and you’ll see the pointer moving and clicking on links, scrolling, typing stuff in text boxes, etc. On the surface that sounds amazing. You can ask it to shop for you, add stuff to your calendar (if you use a web-based calendar like Google’s), find flights, etc.
Unfortunately, it kinda sucks right now. Someone described it as like watching a 5-year old use a computer for the first time. I tried two queries. First, I asked it to find a store within 10 miles of my house that had inkjet photo paper in stock. It was amazing to watch it work autonomously, but in the end it got stuck and just gave me the obvious suggestions (e.g. Office Depot) without much confidence. About a week later I asked it to suggest a closing paragraph for a substack post I was working on and insert it in the right place. It had no problem reading what I wrote and crafting a final paragraph, but it had a lot of trouble scrolling to the bottom of the document and inserting the paragraph correctly. It would insert it in the middle of some earlier paragraph, then undo, then insert it between two paragraphs, then undo, etc. It was almost laughable how many times it tried and failed. I won’t be trying agent mode again any time soon.
With that said, I am increasingly finding that being able to ask ChatGPT about what’s in the browser window (without agent mode) is incredibly useful. For example, one of my most common use cases for ChatGPT is as a writing companion. I don’t use it for first drafts, but I do use it to clean up grammar, fact-check, suggest revisions for awkward wording, etc. When writing this substack I would have to copy and paste my entire post in to the ChatGPT window to ask for edits. However, it would copy as raw text. If I had any formatting, italics, boldface, hyperlinks, etc., none of that would copy over, and that would sometimes lead ChatGPT to make inappropriate suggestions. With Atlas I can ask ChatGPT about my writing in a sidebar, and it can see my original writing in the browser window. That not only saves me from having to copy & paste every time I revise my writing and want more feedback, it gives ChatGPT awareness of all formatting.
Interestingly, this has led me to change some very old ways of doing things on my computer. For example, if I were to write some document other than this substack, I would normally use Apple’s Pages, which is a stand-alone app. Now I’m finding I use Google’s Docs, so that whatever I’m writing is in a browser window directly accessible to ChatGPT. Similarly, when I write mathematics I used to use TeXShop on my mac, but I’m now more likely to use Overleaf in a browser window (especially because Atlas will be able to see the rendered PDF, instead of the raw latex I might have copy-pasted into a chat window before). I haven’t found ChatGPT’s suggested revisions to my presentations to be great so I won’t be switching away from Apple’s Keynote (which I like very much), but if that changes I’ll start using Google’s Slides. Similarly for Google’s Calendar app, Sheets app, etc.
Despite its rough edges, Atlas has quietly shifted how I work. Not because of its headline features, but because of how seamlessly it lets ChatGPT understand whatever I’m doing in the browser. It’s a small change in theory, but in practice it alters dozens of tiny habits I’ve accumulated over years. If a browser can continually do that it may turn out to be the most important application on my computer, even if agent mode still behaves like a toddler with a mouse.
Yes, the last paragraph was suggested by ChatGPT.