Google is improving its most advanced conversational AI model with the help of others, via new app with limited access called the AI Test Kitchen
Last year, Google unveiled LaMDA, an experimental natural language platform that’s designed to engage in natural, free-flowing conversations. On Wednesday, at Google I/O, the company showcased LaMDA 2. During the I/O keynote, Google CEO Sundar Pichai called it Google’s “most advanced conversational AI yet.”
The tech giant is working on LaMDA and other models to improve Search, Google Assistant and other tools, but “we need people to experience the technology and provide feedback,” Pichai said.
That’s why Google is opening up the AI Test Kitchen for others to explore AI technologies and capabilities under development. The app is “meant to give you a sense of what it might be like to have LaMDA in your hands,” Pichai said. In the coming months, Google will gradually open up the AI Test Kitchen to small groups of people.
When it opened up the first iteration of LaMDA to thousands of Googlers, it led to a reduction in inaccurate responses, Pichai said. The company has made efforts to limit the potentially shocking or harmful content that AI models deliver to users. However, AI models can still deliver inaccurate, inappropriate or offensive responses.
To work on this with the help of broader user feedback, Google will initially offer three LaMDA-powered experiences in the AI Test Kitchen app. The “experiences” are not previews of products — just demonstrations of the technology at work. Pichai demonstrated a couple on the I/O stage on Wednesday, including the “Imagine It” experience.
The “Imagine It” experience tests if the model can take a creative idea you give it and generate imaginative and relevant descriptions. Say you’re writing a story and you need some inspirational ideas, Pichai explained. He demonstrated asking LaMDA what it might feel like to explore the deep ocean. In response, LaMDA generated a scene in the Mariana Trench and generated follow-up questions on the fly.
The model wasn’t programmed for specific topics, Pichai noted. It synthesizes concepts from its training data, enabling users to ask about nearly any topic.
There’s also a demo to test how well LaMDA can stay on-topic during conversations, as well as a demo to test LaMDA’s ability to break down a complex goal into subtasks.
“These experiences show us the potential of language models to help us with things like planning or learning about the world,” Pichai said.
Over time, Google intends to incorporate other areas of AI into the AI Test Kitchen.