Facebook releases the Blender chatbot

Facebook AI Research published “Recipes for Building an Open-Domain Chatbot,” describing BlenderBot, a family of dialogue models — released at 90 million, 2.7 billion and 9.4 billion parameters — built by blending techniques the paper’s authors argued were each necessary but individually insufficient: demonstrating knowledge, showing empathy, and maintaining a consistent persona across a conversation.

The paper reported that human evaluators rated its largest model as more engaging and more human-sounding in multi-turn conversation than existing open-domain systems, including Google’s Meena, which had been published earlier that year. Unlike Meena, which Google kept internal, Facebook released BlenderBot’s model weights and code publicly, making a large, competitive dialogue system available for outside researchers to run and inspect rather than only query through a demo.

The paper was candid about what human-likeness on these evaluations did not capture: the models still lost coherence over long conversations, could contradict themselves, and had no mechanism for factual grounding beyond what was encoded in their training data or a retrieved passage. BlenderBot nonetheless set a marker other labs’ open-domain chatbots were compared against for the next two years, and its open release anticipated an argument — whether large dialogue models should be published as weights or kept behind a controlled interface — that would recur across the industry as the models grew larger.