What do learners want from AI?
What does the Technical Report of Google's LearnLM tell us about expectations of learners from AI.
Google introduced LearnLM, a family of Large Language Models (LLM) fine tuned for learning at their Annual Developer event Google I/O. They have already started integrating these models with their flagship products like Search and YouTube and has received a wide range of reaction from users.
Over the weekend I came across
's post decoding the technical report of LearnLM which intrigued me and prompted me to read the actual paper. The technical report has a lot of interesting insights about generative AI in education and is a must read for anyone building in this space. Of all the interesting insights in that paper, today we will talk about only one "What do learners expect from AI models when it comes to tutoring"
Education research has notoriously struggled to keep pace with advancements in pedagogical practices, largely due to the difficulty in establishing universal best practices. This fragmentation across various disciplines leads to inconclusive evidence and makes translating research into practice unclear. Many studies highlight different, often non-overlapping interventions, further complicating the creation of a common set of recommendations. Additionally, much of the research is conducted on small, homogeneous populations, limiting the generalizability of the findings.
Then how do we say which AI models/products are a "good fit" for education/learning?
The technical report makes an attempt at providing a framework for evaluating the use of Generative AI in education/learning. As per this framework a model/product which is "good fit" for education must cater to the following 5 aspects of learning.
Encourage active learning: Learners prefer AI models that encourage active learning. This involves manipulating information through discussion, practice, and creation, rather than passively absorbing information. This active engagement helps learners better internalize and apply knowledge independently.
Manage cognitive load: Effective AI tutors manage cognitive load by presenting information in multiple modalities, structuring it well, and segmenting it into manageable chunks making it easier for learners to process and retain information.
Deepen metacognition: Learners also prefer AI models that help deepen their metacognition, or "thinking about thinking." This principle allows learners to generalize their skills beyond a single context enabling learners to better understand their own learning processes and apply their knowledge broadly.
Motivate and stimulate curiosity: AI tutors that motivate and stimulate curiosity are highly valued by learners. This leads to increased self-efficacy and promotes lifelong learning which are crucial for sustained engagement and continuous learning.
Adapt to learners’ goals and needs: Finally, learners prefer AI models that adapt to their individual goals and needs. This involves assessing the current state of the learner, understanding their goals, and creating a plan to bridge the gap personalizing the learning experience to meet individual learner requirements.
It is sure that learning as we know it is about to change forever, and the changes will be seen across a broad spectrum. What is interesting to note from the Technical Report that general purpose models are not exactly there yet, but they are not too far off even at GPT 3.5 equivalent levels in terms of their ability to help users "learn". But the state of the art models (GPT 4 equivalent) have caught up and can almost as effectively help learners with their intellectual needs.
I am personally quite excited about improving how we learn, if that is something you are interested in too, let's talk. Always happy to connect.