A Box of Many Inputs

One of the interesting questions when designing AI-enabled software is, “What is the intent of search input?”

This was once a simple question: if a user entered “squish” in a search box, it would of course return entries that contained the string “squish”. Over time, computers’ capabilities grew to the point where a kind of universal “omnibox” is now possible. Today software can:

1. Return results with your input’s typo fixed
2. Return results with a synonym to your input
3. Autocomplete recent or popular results
4. Directly retrieve an answer to your input
5. Jump to an action based on your input
6. Generate a conversational response to your input

Ideally, users shouldn’t need to think about which mode they want. They should just be able to type enough characters to make their intent clear, then press Enter.

While this is great in theory, it creates a very crowded design space for input boxes. While most kinds of software will end up with an omnibox-like input – especially with the rise of universal command bars – one of the most fiendish cases is found in the web browser.

The Omnibox

At birth, browsers’ address bar required precise input. To get to Wikipedia, you would have had to type http://www.wikipedia.org precisely. Now, you can mangle it as wiikipo and still get there, which is impressive in its own right.

However, with modern technology there is even more we can do with the address bar. The rise of “AI browsers” like Dia, Atlas, and Comet all strive to layer more functionality into this input field.

For example, in Dia if you type how much did atlassian buy the browser company for in the address bar, it will forego Google and use the in-browser chat to correctly answer $610 million. If it sees a question, it sends it to chat. Seems maybe helpful, right?

Except sometimes you’re not asking a question, you’re searching for a question. Check out what happens when you search for the movie “Who Framed Roger Rabbit?” below.

This is basically never what the user wants when they type a film title into the browser omnibox.

Now, if you typed “who framed roger rabbit” into ChatGPT, a chat response would not be surprising. It’s a chat app. But people expect browsers to be able to browse to things, even if the thing’s name starts with who.

Admittedly, Dia’s algorithm for question detection is more complex than prefix-matching. They’ve bundled a little classifier that detects if a phrase has “question vibes”.

Phrase triggers search	Question starts chat
may june july	may rutabagas be eaten
leather shoes white soles	help me pick shoes

By doing question classification locally with a very small model¹, they can identify questions within milliseconds, with no effect on privacy. Smart!

Except. There are a lot of cases where it’s difficult to classify a phrase by whether it should get an answer or a search result. Consider these results from Dia’s question classifier, all of which should do a search:

Phrase triggers search	Question starts chat
roger rabbit movie	who framed roger rabbit wikipedia
a can of beans recipe	can of beans recipe2
steppenwolf lead singer	guess who lead singer
gladiator you not entertained	are you not entertained

So it’s imperfect at identifying questions. But more critically, even if they hill-climbed to 100% on “is this a question” accuracy, that’s still a dead end.

The actual classification we want is “is this a question they are asking, or is it a phrase they want to search for?” Unfortunately, that’s a much harder training problem! Currently, it’s beyond the ken of a fast local classifier.

As I’ve been weighing alternative browsers, I expected other “AI browser” contenders to have the same problem. To my surprise, no. ChatGPT Atlas avoids it entirely by keeping their heuristic very simple: if a query has less than 10 words, send it to search – otherwise chat.³ If you want to explicitly send a short query to chat, you can press ⌘ + return.

Meanwhile, Perplexity’s Comet and Google’s Chrome send every query to their respective search engines, which use larger server-side models to determine whether a given result should get a more chat-forward response or a more results-centric one. As of today, Google errs on the side of returning web results, and Perplexity errs on the side of giving an AI-generated answer, but both can do either.

In time, these different search and chat engines will converge toward a design that gives users a good experience almost every time, without the need for modes. Already Google is testing out merging their “AI mode” into the AI overviews that appear on results. And presumably OpenAI will evolve to give more search-like responses to pedestrian search queries, in their ever-growing quest for world domination.

Of the four leading contenders for the “AI browser” crown, only Dia tries to interpret who framed roger rabbit as a question.

The Path for Dia

From here, Dia can go one of two ways. They can either go all-in on their own answer engine, competing with Perplexity and Google, so they can send every query there. Or, they can concede on the local classifier and do as ChatGPT Atlas has done – just route all short queries to search.

Earlier this month, Atlassian CEO and Dia acquirer Mike Cannon-Brookes gave an interview where he mentioned the acquisition and the future vision. Reading the tea leaves, it seems Dia could fulfill Atlassian’s vision of building the browser for professional productivity without needing to become a full-on search engine.

On the other hand, Cannon-Brookes mentioned the supposedly-retired Arc⁴ browser three times and didn’t use the phrase Dia once. So… it’s hard to tell exactly what the path is going to look like.

Regardless, with The Browser Company’s talent and Atlassian’s reach, it would be rash to count them out in this one. They’ll fix the Roger Rabbit thing soon enough.

But while browsers encountered it first, time will bring more and more ambitious text boxes. Many of them will initially struggle to deliver on our intent. But, despite some growing pains, the holy grail is within reach. People should be able to type what they want their software to do, and it should do it. And it should do so without any unpleasant surprises. Quickly.

It won’t be easy! But it’ll be great.