Google Sets

» see labs.google.com/sets as it lived, on the wayback machine

~ the obit ~

Google Labs was a page I opened for no particular reason during university. It cost nothing and wanted no account, and it loaded from Ankara the same way it loaded from anywhere else, which in 2004 was not true of most interesting American web products. Sets was the item on that page that looked like a trick rather than a tool.

The layout was five empty text boxes and two buttons. You filled in a few examples of something, asked for a large set or a small one, and got back a page of things in the same category. Green and purple came back as the rest of the color names. Honda and Toyota came back as the car industry. It had no idea what a color or a car was, and the list was right anyway.

It launched in 2002, which makes it one of the first things Google Labs ever put out, from the period when Labs was where a Google engineer's side project went to be looked at by strangers. Sets outlived every one of its neighbors from that era and was still sitting at labs.google.com/sets nine years later, unchanged, when Google closed the building around it.

~ the rap sheet ~

~ the trick ~

The patent behind it was filed by Simon Tong and Jeff Dean, two people whose other work is all over the infrastructure that made Google what it is. The method was less exotic than the output made it look. The web is full of pages that list things, from product catalogs to fan pages enumerating every item in a franchise. If two terms keep appearing together in those lists, they are probably members of the same set, and the crawl has already collected the evidence for every category anyone has bothered to write down.

So Sets never learned what a category was. It reported back a pattern that was already sitting in the crawl, which is why it could answer for an obscure category about as readily as for the colors.

~ the next chapter ~

Google Squared came out of the same idea in 2009, taking the list and hanging attributes off it in a grid, and it lasted two years before dying in an earlier round of the same clean-up. The autofill in Google Spreadsheets was quietly running Sets underneath, so that dragging a cell down after typing two month names produced the rest of the months.

The longer inheritance landed outside Google entirely. Set expansion became its own small research area, and the behavior Sets sold as a novelty in 2002 is now the least surprising thing a language model does. You give one three examples and it continues the pattern, which is the same operation running against a different pile: the training text is also full of people who already made the list.

~ the funny part ~

On 20 July 2011 Google announced that Labs was closing, in the run of shutdown posts through Larry Page's first months back as chief executive. Sets went on 5 September along with nine other services. The oldest experiment in the building was killed by a memo about focus, having cost almost nothing to run and having been the reason a lot of people opened labs.google.com in the first place.

Then the joke landed. For a while after the shutdown, Sets was still in there doing the autofill in Spreadsheets, which means it was switched off as a product and kept as a feature nobody was told about. A nine-year-old experiment outlived its own obituary by hiding inside a keyboard shortcut in an office suite.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »