Google says it has developed a new family of generative AI models “tuned” for learning: LearnLM.
A collaboration between Google’s DeepMind AI research division and Google Research, LearnLM models, built on top of Google’s Gemini models, are designed to deliver “conversational” tutoring to students on a variety of topics, Google says.
LearnLM is already powering features across Google products, including YouTube, Google’s Gemini apps, Google Search, and Google Classroom.
“LearnLM is based on educational research, making learning experiences more personal and engaging,” James Manyika, senior vice president of research, technology and society at Google Research, on stage during a keynote speech at Google’s I/O 2024 developer conference on Tuesday.
Google says that, through a pilot program in Google Classroom, it is working with educators to see how LearnLM could simplify and improve the lesson planning process. LearnLM could help teachers discover new ideas, content and activities, Google says, or find materials tailored to the needs of specific student groups.
Separately, LearnLM is powering Circle to Search on Android, a feature that helps solve basic math and physics problems, and will soon understand problems including symbolic formulas, diagrams, and graphs. And LearnLM supports a tool on YouTube (for now only on Android in the US) that allows users watching academic videos to ask clarifying questions, get explanations, or take a quiz based on what they’re watching.
In Google’s Gemini apps in the coming months, LearnLM will allow users to create custom chatbots that can act as subject matter experts. These chatbots will provide study guidance and hands-on activities like quizzes and games, and will respect each student’s individual preferences, Google says.
Google also plans to partner with organizations like Columbia Teachers College, Arizona State University, NYU Tisch, and Khan Academy to see how LearnLM can extend beyond its own products.
“Today opens a new chapter for learning and education at Google,” Manyika said. “Generative AI is opening up new ways for us to make the world’s information and knowledge universally accessible and useful.”
LearnLM is an intriguing piece of work, no doubt. but a technical document Detailing the development of the model reveals that it suffers from some of the same problems as other generative AI models.
For example, LearnLM does not speak in a very “encouraging” tone and has trouble identifying when students answer practice questions correctly relative to one of the basic Gemini models (Gemini 1.0), although it is better at detecting mistakes. And he is not immune to hallucinations or making up facts and figures in response to the prompts he receives.
For these reasons, the paper’s co-authors caution against using LearnLM in applications “without additional evaluation and analysis of the specific harms of [the apps]” while at the same time suggesting that Google performed such evaluations and analyzes for its applications. For the sake of students and educators, let’s hope that’s true.
We’re launching an AI newsletter! Sign up here to start receiving it in your inboxes on June 5.
