Google unveils LaMDA at I/O

Background

LaMDA grew out of Meena, a chatbot built by the Google Brain team ^[1]. The move from a chatbot to a headline announcement was small in name and large in framing: Google presented the new system as a language model aimed specifically at conversation. A dialogue model is trained on large amounts of text so that it can carry a conversation forward turn by turn, keeping track of what was said earlier, rather than answering one question and stopping. That goal explains both the name and the way the system was introduced.

The announcement

Google announced LaMDA during its I/O keynote on 18 May 2021. The name is short for "Language Model for Dialogue Applications" ^[1]. It spells out the two halves of the project: a language model on one side, dialogue as the application it was built for on the other.

The same day, Google published a blog post introducing LaMDA, describing it in the same terms as the keynote did ^[2]. The post was written by Eli Collins and Zoubin Ghahramani ^[2]. Announcing a research system at a keynote and explaining it in a written post on the same day is the usual pattern for Google's large model work: the talk gives the headline, the post gives the description people can read afterwards.

LaMDA 2 and the AI Test Kitchen

LaMDA 2 followed on 11 May 2022, announced together with an app called the AI Test Kitchen ^[1][3]. Google described LaMDA 2 as its most advanced conversational AI at the time ^[3]. Presenting the model and the app side by side put the technology in front of people outside the company rather than only in a description. Google said it intends to keep adding other emerging areas of AI to the AI Test Kitchen over time ^[3], which framed the app as something that would keep growing instead of a one-off demo.

A claim of sentience

In June 2022, an engineer's claim that LaMDA was sentient drew wide attention ^[1]. Google fired him in July 2022 ^[1]. The episode carried the model's name well beyond the audience that normally follows language model research, and the argument turned on whether the behaviour of a dialogue system can be read as something more than generated text.

What came after

Bard, announced on 6 February 2023, was powered by LaMDA ^[1]. That placed the model first shown at the 2021 keynote inside a product rather than a demonstration. The line of work that began with a chatbot from the Google Brain team and continued through LaMDA 2 and the AI Test Kitchen ended up as the engine of Google's first widely announced conversational product.