Google's NotebookLM is by far the most actually useful AI product of the current generation, for people that actually need to augment their own capacity. Being able to process multiple large and complex sources and pull out key details (names, organizations, key themes, etc) is vital for anyone doing research involving large amounts of information.
Some ideas for future enhancements:
1. Source diversity – Being able to add sources of all kinds (video, audio, images, full websites) would be amazing. I have tools to convert a lot of these things into 'pasted text', but being able to skip those steps and stay in a flow state would be amazing.
2. PDF and OCR – The current OCR capability is BY FAR the best I've found. I deal a LOT with scanned historical documents, and you guys grind through them in ways nothing else can right now. Being able to deal with graphics, figures, logos, etc within the documents better would be amazing too. Likewise, sometimes certain text styles seem to throw off the OCR. But again, great job so far and love the idea this will only get better.
3. Promptable audio summary – The "deep dive" hosts are fun and do a decent job. But... sometimes less is more. Being able to prompt a specific audio query would be amazing (one or two+ voices, scope and focus, quick or detailed, snarky or professional, to name a few ideas.) Also being able to download as a MP4 video (with subtitles, or even talking head(s) could be awesome too) — for easier posting to YouTube, X, etc — and would help build audiences for NLM, too.
4. Combine notebooks – Notebooks are a good unit, and allow for separation of topic areas. Sometimes it would be amazing to combine a few different notebooks to gain a big picture analysis. I realize this is dependent on context windows etc... but this is a good long term goal.
5. Graph analysis – I frequently want to know "who are the people + orgs in this set of sources and what are the relationships between them." I have been able to do this successfully and also to produce graph data files (such as .gdf format) by including a source document that defines the file format. I can then visualize this file in software like Gephi. It would be next-level badass to do all of that inside of Notebook and to maintain flow. Many kinds of analysis could benefit from the ability to infer graphs from text or multimodal input sources.
I love where you're going with this. I would suggest focusing on how to maintain a flow state (add more sources, query, inspect, repeat) for the user, which the product does now. For me that means minimizing the fiddliness of "transforming" sources for processing and staying in a state of inquiry, discovery, and source augmentation.
Anyway, use me. I have started to build some pieces of this toolchain for my own use but you guys have way more resources to bring to bear. This is the way.
/cc @raiza_abubakar @stevenbjohnson