Facebook's David Marcus on how humans work alongside AI automation to power Facebook M and whether such a service can scale
Time To Meet The Wizard: Facebook's Messenger Head Pulls Back The Curtain On “M” — Facebook's David Marcus says M is already a lot more than just people pretending to be robots. Tweets: @vanessaluvtini and @rossdawson . Thanks: @jmbooyah Tweets: Vanessa Martini / @vanessaluvtini : Last night my car drove me home on autopilot & today I read this article by @Kantrowitz via @buzzfeed #futureisnow http://twitter.com/... Ross Dawson / @rossdawson : Very interesting insights into Facebook's M: AI formulates response, vetted and improved by humans, bot learns http://www.buzzfeed.com/... Thanks: @jmbooyah
Context & Ripple Effects
Marcus's walkthrough closes the loop opened in August, when Facebook began testing M as a supervised virtual assistant that could complete purchases inside Messenger, and by November's hands-on reviews showing M outperforming pure AI rivals like Siri and Cortana at real-world tasks. His point — AI formulates the response, humans vet and improve it, and the bot learns — reframes those results: the magic reviewers saw was a hybrid system, not autonomous software.
First-order effects
- Facebook must staff and train the human vetting layer behind every M interaction, making each task completion a labor cost as well as an engineering one — the direct constraint behind Marcus's own scaling question.
Second-order effects
- Rivals building purely automated assistants face a quality benchmark set by human-vetted output they cannot match without adding people; meanwhile Facebook can mine the vetted exchanges as training data, so every human hour also buys future automation.
Third-order effects
- If the pattern holds, the human layer is transitional: the 2016 Bot Engine release for machine-learning-enabled developer bots and the plan to bring businesses and ads into Messenger show Facebook rebuilding M's capabilities as software it can distribute at scale rather than operate by hand.
The trend: Consumer assistants are passing through a human-supervised apprenticeship phase, where human labor bootstraps training data until automation can carry the workload alone.