Google announces LaMDA, a language model for dialogue applications that it says represents a “breakthrough” for having natural conversations with AI
The VergeJames Vincent
Context & Ripple Effects
LaMDA lands as Google's flagship dialogue model at I/O 2021, arriving on top of a string of applied-language experiments like Look to Speak, which let users with speech impairments select phrases with their eyes. Within weeks of the announcement, reporting surfaced on Tivoli, an internal project to teach foreign languages through Google Search — an early sign Google intended LaMDA-style conversation to live inside its existing surfaces rather than a standalone chatbot.
First-order effects
LaMDA immediately becomes Google's reusable conversational substrate: it later powers Apprentice Bard when the company asks employees to test potential ChatGPT competitors, and it becomes public enough that a Google engineer is placed on paid leave after claiming it had become sentient.
Second-order effects
ChatGPT's arrival forces Google to retrofit LaMDA into defensive chatbot tests rather than shipping conversation on its own timeline, while parallel efforts push sibling models outward — PaLM is used to help Everyday Robots understand complex human commands, and Google unveils a speech model trained on 400+ languages with a stated goal of supporting the 1,000 most spoken.
Third-order effects
If the pattern holds, dialogue models stop being single-product features and become infrastructure threaded through Search, robotics, and accessibility tools — with the notable structural lesson that a 'breakthrough' announced in 2021 sat mostly unshipped until a competitor forced productization.
The trend: Conversational AI at Google is migrating from research-demo status to being embedded across Search, Assistant-scale surfaces, and hardware, with OpenAI's ChatGPT accelerating the shift from lab to product.
This blogpost co-authored by @ZoubinGhahrama1, who's taken over for Megan Kacholia, says “Responsibility first.” Its a joke right? You informed our team of @mmitchell_ai's firing right after @JeffDean wrote an email about “retention” or whatever. https://blog.google/...
Google has published a new language model that understands “Related Search Activity” and “Possible Search Activities” better. It is similar to the Google patents that I have read from Dear @bill_slawski. #seo https://blog.google/...
Google has announced MUM (https://blog.google/...) “1000 times more powerful than BERT and can generate as well” First thing coming into mind: how is it different than T5 then? (an encoder-decoder model) Answer: It is trained on downstream tasks on 75 non-English languages!
Google's MUM: A new AI milestone for understanding information: “Like BERT, MUM is built on a Transformer architecture, but it's 1,000 times more powerful. MUM not only understands language, but also generates it. It's trained across 75 languages” https://blog.google/...
#GoogleIO hyping the new LaMDA chatbot on the heels of the controversy behind the Stochastic Parrots paper. The LaMDA press release has this about AI responsibility https://blog.google/... https://twitter.com/...
Odd irony: Google have created a new language system (LaMDA) &, using expert knowledge of language, they've given it basically the same name as a tool from their main cloud competitor. (Google Lamda, Amazon Lambda) https://blog.google/... https://aws.amazon.com/lambda/