Keen

» see staykeen.com as it lived, on the wayback machine

~ the obituary ~

Keen was a machine learning demonstration wearing a product, and I mean that as a description rather than an insult. I work on models for a living, so the internal logic of this one is obvious to me from the outside: somebody had a recommender that got sharper the more labeled examples you fed it, and the fastest way to get a person to label things by hand is to convince them they are making a collection. That is the whole design. The collection is the training signal.

It launched on 19 June 2020 out of Area 120, the incubator Google ran for employees with an idea. It was built together with the PAIR team, the group inside Google working on human-facing machine learning. A "keen" was a board on a subject. You made one about gardening or bread or whatever it was and dropped links and notes into it. Google Search then fed you more things it thought belonged there. You could invite people to collaborate on a board and you could follow other people's.

CJ Adams, one of its founders, described it as curating the content you love and finding new content based on what you had saved. The pitch had a second half. Keen was presented as an answer to passive algorithmic feeds, a way to steer your own recommendations by hand rather than letting an engagement model pick for you.

That is an interesting premise, and it is also a premise that requires the user to do work. Pinterest had solved the same problem by being a shopping catalog that felt like a hobby. Keen asked you to curate for the sake of curating, on a web app and an Android app in the middle of 2020. It offered no reason to believe Google would still be running it in three years. Very few people took that deal.

~ the case file ~

~ the frozen years ~

The last update that meant anything shipped in December 2021, eighteen months after launch. Everything after that was a running server with nobody on the other end, which is a state a lot of Area 120 output ended up in. The incubator's deal was that a project either found traction fast enough to be adopted by a real product group or it sat there.

Google cut Area 120 roughly in half in the January 2023 layoffs and then wound the whole thing down. Once that happened, Keen was not a product with a bad quarter. It was an orphan on a server that somebody would eventually have to pay for.

~ the day it died ~

Users got an email on 14 March 2024. Keen would close on the 24th, and everything saved to it would be deleted. There was an export tool at staykeen.com that would package your saved links and text posts into a zip file, which is more than Google has offered on some shutdowns and less than ten days to act on it.

Ten days for four years of collections is a strange amount of notice. It is enough to be defensible and short enough that most of the people who had a board there never read the email in time.

~ what it was really for ~

Keen never announced a user count, and companies do not withhold numbers they are pleased with. What it did produce, for four years, was a corpus of human-curated topic collections, which is exactly the kind of clean training data that a recommendation system is otherwise very expensive to obtain.

I do not think anyone at Google was being cynical about this. The team clearly believed in the anti-feed argument they were making. But the thing that made Keen worth building inside Google was never going to be the same thing that made it worth keeping, and once the incubator that housed it stopped existing, there was no one left to make the second argument.

@}-->--

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

« previous  ·  random  ·  next »