Google announces LaMDA, a years-in-the-making language model trained on dialogue

Yesterday at Google I/O, the company announced its brand-new AI tool: a language model trained on dialogue, christened “LaMDA” (Language Model for Dialogue Applications).

Recap: Language models help computers process and produce natural, or human-like, language. They’re the reason AutoCorrect (sometimes) knows what you meant to type.

Meet the model

Language models are usually trained on text alone, sourced from large swaths of the internet. LaMDA, on the other hand, is trained on dialogue. That means it can pick up conversational nuances unique to open-ended conversation.

• The goal: Sound more like a human and less like a robot.

In practice: Google demonstrated LaMDA’s conversational chops by having it pose as a paper airplane and as Pluto during two separate conversations.

• When asked what the dwarf planet wished people knew about it, “Pluto” responded: “I wish people knew that I am not just a random ice ball. I am actually a beautiful planet.”

LaMDA is “designed to converse on any topic,” Google CEO Sundar Pichai said during the event, clarifying that the model is still in R&D but is being used internally. He said the company is working to ensure LaMDA meets its standards for fairness, accuracy, safety, and privacy, but Google’s ethical standards are under scrutiny due to the company's increasingly controversial AI ethics decisions in recent months. Pichai provided no details on training data, off-limits topics, or specific safeguards.

Bottom line: LaMDA could be a huge step forward for the conversational capacity of voice assistants and other AI language tools. —HF