A look at how Facebook's AI researchers are trying to build smart chatbots using unsupervised training of neural networks for natural language processing
Mark Sullivan / Fast Company :
Context & Ripple Effects
This 2019 Fast Company piece sits in the middle of a long arc of Facebook's conversational-AI bet. It began with the Wit.ai acquisition and the Bot Engine release for machine-learning-enabled developer bots, then moved into monetization with the plan to bring businesses and ads to Messenger through AI-enabled chatbots and the M assistant.
What changed by September 2019 was the training method: instead of hand-labeling conversations, Facebook AI turned to unsupervised neural-network training for natural language — an approach that paid off eight months later when the lab shipped Blender, its open source research chatbot trained at massive scale. The article matters because it documents the methodology pivot that made that model possible.
First-order effects
- Facebook's AI researchers gain a path to more capable chatbots without the cost of labeled training data, strengthening the same Applied Machine Learning infrastructure that had already made AI central to company strategy.
- Developers building on Facebook's bot tooling are the near-term beneficiaries: better underlying language models flow into the Messenger bot ecosystem the company has been cultivating since Bot Engine.
Second-order effects
- Rival labs racing on conversational AI face pressure to match unsupervised pretraining at scale — a race Blender's subsequent open-source release sharpened by putting a top-tier research model into everyone's hands.
- If the resulting chatbots get good enough, Facebook's plan to route businesses and ads through Messenger gains a credible interface, shifting conversational commerce toward whoever controls the messaging platform.
Third-order effects
- The pattern points toward self-supervised learning replacing curated labeled datasets as the default recipe for language AI — a structural shift in how big-platform labs train models and in who can afford to compete.
- Open-sourcing flagship models, as Facebook did with Blender, turns research leadership into ecosystem leverage: standards and talent cluster around the lab that gives its work away.
The trend: Conversational AI is consolidating around large-scale unsupervised training at big-platform labs, with open-source releases used to convert research wins into ecosystem control.