Google has unveiled a new series of generative AI models called LearnLM, specifically refined for educational purposes. This initiative is a collaboration between Google’s DeepMind division and Google Research, with LearnLM built on Google's Gemini models. The purpose of these models is to provide “conversational” tutoring across various subjects.
LearnLM is already enhancing multiple Google products, including YouTube, Google’s Gemini applications, Google Search, and Google Classroom. “LearnLM is grounded in educational research, making learning experiences more personal and engaging,” stated James Manyika, Senior Vice President of Research, Technology, and Society at Google Research, during the keynote at the Google I/O 2024 developer conference.
Through a pilot program in Google Classroom, Google is exploring how LearnLM can streamline lesson planning. It aims to assist educators by offering new ideas, content, and activities, as well as materials tailored to specific student needs.
Additionally, LearnLM powers the Circle to Search feature on Android, designed to help users solve basic math and physics problems. This feature will soon have the capability to understand complex prompts involving symbolic formulas, diagrams, and graphs. Furthermore, a tool on YouTube (currently available on Android in the U.S.) allows viewers of academic videos to ask clarifying questions, obtain explanations, or engage in quizzes based on the content.
In the coming months, Google’s Gemini apps will enable users to create custom chatbots acting as subject-matter experts, providing personalized study aids and interactive activities like quizzes and games that cater to individual learning styles.
Google plans to collaborate with institutions such as Columbia Teachers College, Arizona State University, NYU Tisch, and Khan Academy to explore extending LearnLM's capabilities beyond Google products. “Today marks a new chapter for learning and education at Google,” Manyika remarked. “Generative AI is unlocking new avenues for making the world’s information universally accessible and useful.”
While LearnLM presents exciting advancements, a technical paper on its development indicates it still faces challenges common to generative AI models. For instance, LearnLM tends to lack an “encouraging” tone and struggles with accurately assessing when students answer practice questions correctly compared to the original Gemini model (Gemini 1.0). However, it does demonstrate improved accuracy in identifying mistakes. Additionally, like other models, it is susceptible to generating inaccurate responses—known as hallucinations—when responding to prompts.
The paper’s co-authors have advised caution in deploying LearnLM in applications “without further evaluation and analysis of the harms specific to [those applications].” They also suggested that Google has conducted such evaluations for its own products. For the benefit of students and educators, we can only hope that's the case.
Stay tuned for more updates by signing up for our AI newsletter, launching on June 5.
Ranking
