AI is coming to the Arc browser — but probably not like you think
Sure, you could just stick a ChatGPT sidebar in your browser. But what do we really want AI to do for us as we use the web? That’s the much harder question.
At some point, if you’re a company doing pretty much anything in the year 2023, you have to have an AI strategy. It’s just business. You can make a ChatGPT plug-in. You can do a sidebar. You can bet your entire trillion-dollar company on AI being the future of how everyone does everything. But you have to do something, you know? These are just the times we live in.
The last one of these was crypto and the blockchain a couple of years ago, and Josh Miller, the CEO of The Browser Company, which makes the popular new Arc browser, says he’s very glad he missed that particular boat. “These waves have come and gone,” he says, “and sometimes — god, every time — we have to say, ‘Alright, are we doing this one?’ Historically, we’ve been good at saying no.” But with AI, Miller says, The Browser Company needed to jump in. Competitors were making noise about being “AI browsers.” The tech was getting better fast. There was clearly something to all this AI noise, even if most of it was just noise. Arc is all about innovation and speed and doing browsers better, and it suddenly felt like bailing on AI meant risking being left behind.
Miller and his team didn’t want to just, like, build a ChatGPT sidebar into Arc. And they certainly didn’t have the resources to build their own large language model. But they had to do something — ideally something that didn’t get in users’ way or seem like a waste of time for a company with an already-long road map or bankrupt The Browser Company through infinite complex AI processing. So they started experimenting and building prototypes for how AI might fit into your browser. They built more than two dozen of them, all of which Miller and Browser Company designer Nate Parrott recently showed me over Google Meet. They represent the full gamut of possibilities, from “What if AI was your browser?” to, well, basically a ChatGPT sidebar in Arc.
Ultimately, the team picked five things to ship — a collection of features they’re calling Arc Max that’s powered by a combination of GPT-3.5 and Anthropic. Here are the five:
→Ask ChatGPT: ask ChatGPT questions right from the Arc Command Line. Easily the least interesting feature of the bunch. →Tidy Tab Titles: when you pin a tab in Arc, the browser will automatically rename the tab — whether it’s a wordy Amazon listing or an unhelpfully vague page name — to make it easier to find in your sidebar. →Tidy Downloads: same as the tab titles, but for downloads. Instead of 3776x_L.jpg, Arc will rename your downloads with things that actually describe the download. →Five-Second Previews: hover over any link and press Shift, and Arc will fetch a short summary and preview of that webpage. →Ask on Page: when you Command-F to try and find something on a webpage, it’ll look for that keyword as normal. If it can’t find it, it’ll query Arc’s AI to get an answer to your query specific to the page you’re looking at.
All of these features, Parrott says, serve the same goal: they make your internet life easier and help you do stuff without screaming “Look at all this AI!” every time you encounter them. The Ask on Page feature is a good example, he says. “Say I’m reading about Mongolia.” (A lot of Parrot’s demos turn out to involve Mongolia. Even he’s not exactly sure why.) “I want to know, like, what are millennials doing in Mongolia? There’s not really a good way to answer that — but with LLMs, of course there is.” So he types Command-F on the news article he’s reading, types “millennials,” and once the on-page search fails, a little “Ask” button pops up. “You can click it,” Parrott says, “and we feed the model the page, and we answer it.”
(Ask on Page strikes me as the thing of thing every browser should do, by the way. It’s a simultaneously low-touch and useful way to bring in AI. I suspect every browser will do it, too; the SigmaOS browser launched a similar feature called Airis a few months ago, Opera’s Aria can do much the same, and Chrome and Edge and many others seem poised to implement similar tech in similar ways.)
These lower-touch features felt right to The Browser Company only after experimenting with a lot of much higher-touch ideas. One prototype Parrott showed me was basically a browser overrun by AI: he opened up a version of Arc with a huge text box as its home screen, typed in “I want to take a trip to Mongolia,” and the browser ran a series of AI queries that brought back travel tips, links to checks, follow-up questions about how you like to travel, and more. This might work if the only thing you ever did on the web was plan trips! But that’s not realistic.
Parrott and his team made a bunch of chat-first prototypes, along with a few “attempts to take the chat interface and make it a little more click-friendly.” One put an “Ask” button at the bottom of every page that you could click to either ask for information about the page or just query the model in general. Parrott went to an Amazon listing and asked for non-Amazon reviews and asked about the vibe of a neighborhood on an Airbnb listing. It all kind of worked in the way that so many AI models kind of work, but it felt overbearing.
One good reason for any browser maker to care about AI is that your browser knows more about your interests — through your history, your bookmarks, even the page you’re on right now — than practically any other software, so it can personalize an experience unlike almost anything else.
The prototypes kept coming: a version of Arc in which you could highlight text on a page, and the browser would take and automatically organize notes for you; a version that could automatically compile a table from a bunch of product pages, making your shopping easier. These were cool and worked pretty well, but they’re the sort of browser feature you’re likely to use sporadically, if at all. “A trap that we fall into sometimes,” Parrott says, “is just thinking the internet is about power information gathering when for most people, at most times, it really is not.”
One internal favorite prototype took over the Forward button in your browser and turned it into an AI-powered exploration engine. In Parrott’s prototype, you’d hit Forward, and suddenly, a starry animation would appear like you’re going to lightspeed in Star Wars. Then you’d be taken through a journey around the internet: if you were on IMDb, it might show you other shows and movies; if you were looking at Mongolia pictures, it might show you other travel spots. Miller loves this one. “We’ve forgotten how vast and wondrous the internet is,” he says, “and lost that part.” He even had a tagline in mind: never go backwards, go forward.
But that ran into a problem a lot of these prototypes did. Parrott puts it pretty bluntly: “People don’t want more internet. They want less internet.” If Arc’s AI tools were going to be actually useful, they had to help people spend less time dealing with their browser because, ultimately, most people would happily trade a delightful discovery engine for a few minutes of their life back.
The “Less Internet” path became the way forward for Arc’s AI work. The team prototyped a way to automatically help people make Boosts, Arc’s feature for quickly customizing a webpage, just by describing what they wanted to do. They made one that could help with data analytics and other basic scripting across all your web apps and another that was basically Google’s Smart Compose but in every text box on the web. They made a sidebar that would automatically organize all your tabs into categories — recipes with recipes, news with news. (Both Miller and Parrott love that one, actually — the only problem is that it’s currently a little slow.)
Ultimately, Miller says, the five shipping Arc Max features made the cut because they’re fast, they’re useful, and they actually help you use the internet better. And less. “People already have full lives,” he says, “so how do you make those little moments that are faster, fewer clicks, and not make them learn or invent anything new?” They may not be the most ambitious AI features ever, but Miller says that’s on purpose. The goal was to ship stuff that works.
Even after all this testing, Parrott and Miller say they’re not sure they’ve cracked it. So they’re asking users: the Arc Max features are technically only going to be live for 30 days, and users will get to vote on which features to keep. They’re all going to be opt-in, too, so many users might not encounter the AI features at all. Asking users to choose to use Max is also a privacy choice, Miller says — he hopes that by forcing people to turn on the features, and by telling them on the same page that using AI requires sending some data to these models’ owners, Arc can mitigate the creepy feeling of being watched by AI.
Even internally, Miller says, debates have been contentious about which features to keep and which to ditch. “With two of them specifically,” he says, “I can imagine a world where no one cares. I can imagine a world where a couple of them change the way you use the internet. I really don’t know. We don’t know.”
AI is coming for just about everything in your online life. Will any of it really change the way you use the internet, do your job, and live your life? Nobody knows, and for sure nobody knows what it’ll look like if it does. But Parrott says that’s why this moment is so much fun. “Some features do survive [these hype cycles], maybe 90 percent of them don’t. But out of these explorations, we’re going to find a handful of things that are just table stakes in every browser. And if we can do it first, we’ll be really happy about it.”