Instead of focusing just on bigger and better models, (which they are working on anyway with Gemini!) - Google is also coming up with interesting test cases for LLMs.

I have been looking at NotebookLM as an interesting idea – they allow me to set up a bunch of documents, text and PDF files, links etc as my sources and then ask questions on that knowledge. It provides answers with references to the right source with all other advantages of LLM like language style, variations of sentences, professional or casual tone etc. All of this remains personal in my own account. Really fantastic if you want to quickly find big ideas, common themes from a bunch of research papers, blog posts etc.

But recent update to NotebookLM is extremely intriguing! They can now generate audio deep-dive into the topic and create a lively conversation between two hosts generated by AI! I had to try it on one of my earlier blog posts and see how it will perform, especially given the fact that it is based on a Marathi album. It performed very well on some of these aspects:

• Liveliness of the conversation, voices feel natural enough and interesting with proper pauses in conversation etc. Really great!
• Good summary of the topic.

However it stumbled on some specific issues with my task:

• It could not pronounce the names and song names properly, although it did okay on some more famous words/people.
• It made a factual error - It said Ga. Di. Ma. is another name for P. Sawalaram! 🤦

It will be great if it could produce this conversation in native Marathi, but the model can only improve from here and I believe this is already a great start and an excellent use-case!

Listen to this conversation and decide for yourself!

Leave a comment below if possible!