Using Google’s NotebookLM for summarization

Stefan Probst of Innovationsbeirat sent me a NotebookLM summary of my research methods course “Nailing Your Thesis” (NYT). The summary comes in the form of a 30 minute podcast with two hosts talking to each other, asking and answering questions, and making the content more palatable. To create the podcast, NotebookLM simple loaded my slides from uni1.de, which is hosting the course, and created the podcast from it.

First off, the technical quality is amazing. You can’t hear it is artificial people. The questions and answers make sense and clearly reflect my material.

To avoid any wrong expectations: While eloquent, the reception is also shallow: Everything bad you’d expect is present, from simple factual errors (“did you know a nobel prize was awarded for research on how to make cows produce more milk”) through poor assessment (overemphasizing less important aspects, ommitting critical insights) to simply not understanding the meaning of some important concepts. I’m also not enthusiastic about some of the examples the AI added. If the podcast was a student homework, they would have passed, but not with a good grade.

As a consequence, I could stop providing my materials to students and the public and disallow any recordings. It would also be rather pointless.

Given my experience with students, I intend not to fight AIs like NotebookLM but work with them. My assumption is that students will use these tools and will skip the course and will assume that 15 sessions of 90 minutes each can be replaced or at least sufficiently summarized by a podcast like this. As a consequence, I will look into how AIs parse my content and prepare to increase the probability that any derived AI-generated materials are a more accurate reflection of what I’m teaching. I expect to be adjusting my slide structure and how things are emphasized or deemphasized. For kicks, I’ll probably try some prompt hacking.

We now have two audiences: Humans, and machines filling in for humans. Whether they ultimately end up as one abstraction in my head remains to be seen.

My discussion so far ignored what may be well be the main win of this approach: A more accessible presentation. Me droning on in a monotonous voice in a lecture (“Anyone? Anyone?”) for 90 minutes is wasted time. The interaction of the two artificial hosts is engaging, somewhat witty, and fast-paced. This particular setup is probably preordained in NotebookLM, but even in this specific form, I could probably improve my lectures by embedding AI generated podcast snippets. Assuming that in a next step, AI assistants will be significantly more versatile gives me hope for creating more engaging lecture content, as one-way audio and video, and also for chat-based engagement.

Subscribe!

Comments

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Navigation

Share the content

Share on LinkedIn

Share by email

Share on X (Twitter)

Share on WhatsApp

Research projects

Open data, easy and social
Engineering intelligence unleashed
Open source in products, easy and safe