Facebook M

» see messenger.com/m as it lived, on the wayback machine

~ the eulogy ~

facebook m never opened outside california. by the time it was killed in early 2018 about ten thousand people in the bay area had ever talked to it. the rest of the world heard about it from techcrunch and the verge writing breathless explainers about facebook's siri competitor that mostly read more like profiles of a startup with twenty employees doing manual labor inside a messenger thread.

the project launched on august 26 2015. david marcus posted the announcement on facebook itself, framing m as a personal assistant inside messenger that could actually complete tasks instead of just looking things up. you typed what you wanted in a normal chat thread, m parsed it and asked for missing details, then m booked the restaurant or ordered the flowers and confirmed when done. the demo conversations were impressive enough that the first wave of press took them as evidence ai had crossed a threshold.

the inside story was more interesting. alex lebrun, who had founded wit.ai before facebook bought it, ran the project. the human part of m was a team of menlo park contractors called m trainers who watched every conversation in real time and stepped in when the ai missed a step. by 2017 internal reporting from people inside the program said about 70% of requests still needed a human to finish them. that was not the curve a venture-stage product is supposed to follow at year two.

~ the dossier ~

~ what it was ~

the public-facing m was a chat thread in messenger with a small purple icon. you opened it like any other conversation, typed "send my mom flowers for her birthday next thursday", and m worked the request through. the assistant asked for an address if it did not have one. it asked which florist. it confirmed the order before placing it. for users who got into the beta, the value proposition was real, you could outsource the small annoying chores of weekly life to a chat thread, with a guarantee that someone, ai or human, would finish them.

the part that made m work for those users was the human safety net. when a real m trainer stepped in, the response in the chat read exactly like the ai had figured it out, no handoff message, no notification, just a continued conversation. the team was doing the long manual research that the ai was supposed to learn from, ideally producing enough labelled completion data that the ai could start handling more cases on its own over time. the curve never moved the way facebook had budgeted for.

~ the day it died ~

on january 8 2018 techcrunch reported that m was shutting down. on january 19 facebook confirmed and turned the service off. the official statement from facebook was that m had been a research project to test where ai could be useful in messaging, and that the lessons would feed into the wider messenger product. there was no migration path for the beta users who had been treating m like a personal concierge for two years.

what replaced m on the consumer side was m suggestions, the inline shortcut ribbon that appeared inside messenger threads from april 2017 onward. when you said "let's split the bill", m suggestions added a "send money" button to your thread. when you said "see you tomorrow at 8", it offered to put it on your calendar. m suggestions stayed alive for another year or two before quietly disappearing from the product as well. the original m, the chat-based assistant with human backup, never returned in any form.

~ what it left ~

facebook m sits next to google allo, microsoft cortana on consumer, and apple siri's worse years as the run of pre-llm chat-assistant projects that did not survive the gap between demo magic and real demand. the differentiator with m was the open admission that the team had built a hybrid system, that real humans were doing a lot of the work, and that this was the only way to ship a credible product in 2015. that admission has aged well. every successful llm-era assistant since gpt-3.5 still leans on human reviewers for the cases the model gets wrong.

the wit.ai engine that powered the parsing side of m kept running for years inside other facebook products. portal used it for voice. messenger keyboard used it for suggested responses. the original m purple icon disappeared but the parser stayed in the building. m is the project that gave the eventual llm-era a working blueprint for the human-in-the-loop ai assistant, then was shut down before that blueprint mattered.

@}-->--

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

« previous  ·  random  ·  next »