Exploring Google's Notebook LM: Ideas for AI Interaction and Insights

I've just started playing with Google's Notebook LM tool and it's highly impressive! I find myself wanting to paste in content from Dovetail to interact with it in Notebook LM. Some specific ideas that would add value: -Have an AI chat sourced from 1 user - e.g. I have done 5 interviews with the user 'Bob Smith' - trained just on his responses can I interact and chat such as asking what he said or thinks about certain topics -Produce a podcast style audio where 2 hosts have a back and forth discussion about the insights within a particular project - to make an easy to digest and entertaining report summary of the insights

💡1

👀1

12 comments

armina
·
·
would love to hear more about your experiences with Notebook, if you ever wanna do a share-out (public or not)! 🤓
your first idea made me think immediately of synthetic users — i’ve been following their work for a while, am curious if this is something you would use? my only concern with the ‘Bob Smith’ chat bot is that it’s harder to spot-check for hallucinations, since the connection between the input and output is harder to follow 🤔 so it could be useful for ideation but, just like with synthetic users, it would still require validation with the real Bob Smith / real user types.
the podcast idea is something some of our leadership has actually been asking for! so definitely a lot of potential, very curious to hear how that’s worked for you.
Alex DeWitt
·
·
thanks for the reply, I've briefly looked into synthetic users but I am not convinced at all by it. What I meant was that answers come only from the source material so nothing is hypothetical. I have situations where I have had to do 5 or 6 discovery interviews with the same person as they are a domain expert and I need to trawl through it all to find out their answer to a specific question.
armina
·
·
aaah, that makes more sense and definitely easier to spot-check! 🌟
Alex DeWitt
·
·
I guess the feature would be to select which sources you want to train the AI chat on - they could be data notes, insights, projects, folders, users etc
🙌1
Nathalie T
·
·
thanks for posting your ideas with Notebook LM! I like your first idea a lot - feels like an easier way to interact with the transcript. i get a little worried because recently I was reviewing an interview and noticed a few errors in the transcript (e.g., user actually said "i do agree" but the transcript has "i don't agree"). i guess humans can mishear just as randomly as AI can mis-transcribe?
Alex DeWitt
·
·
indeed. I always review the video/transcript and correct typos, but yes an AI summary could be very misleading if that QA wasn't done
Sarah Köksal
·
·
A bit late to the thread, but I am really interested if you guys are doing some sort of sharing re. Notebook and Dovetail usage and combinations 🙏
Sarah Köksal
·
·
I use Notebook a lot for open exploration of data that we gather from customer-facing departments (success, sales etc.) for our UX topics.
Alex DeWitt
·
·
interestingly I pasted just relevant excerpts of several user interview transcripts into Notebook LM and asked for the common user needs. I then asked Dovetail AI search the same question based on a filter for the project containing the full interviews. The answers were virtually identical, almost word for word, so maybe Dovetail uses the same Google powered AI. It was reassuring.
Sarah Köksal
·
·
Very interesting!! I had a really different experience, Notebook was a lot more accurate than Dovetail which came up empty on a lot of questions that Notebook had detailed answers/summaries on. I thought it could be related to the language, most of our material is in German for which the Dovetail AI doesn't work perfectly (yet 😅)
Alex DeWitt
·
·
I also found Notebook to completely make stuff up. Like I gave it 4 documents and asked which info appears on all 4 and it just made it up completely. I told it the answer was wrong, and it said 'sorry I made a mistake, here's the actual answer...', and that was also wrong. Quite worrying as if it was a lot of data I wouldn't have been able to fact check it! SO maybe also do some spot checks on the answers it gives.
👍1
Sarah Köksal
·
·
That's true, it's so important to check (if possible). I also found that it has biases that are not explicable. For instance, I was asked to do an analysis comparing the best practices of customer success managers and couldn't understand why Notebook rated one specific manager as more competent. When I went back to manually sort through the material, I couldn't find any significant quality difference. Do you think the AI could be potentially biased in favor of AI? The only difference between the aforementioned employee and the others was that he used AI for his documentation more consistently than others.