NotebookLM is my top choice for knowledge curation and chat based on the knowledge, grounding in sources provided in a simple way that my team (non-coders) can manage, a tool I can share with my whole team without sharing to the outside world, a tool that interacts like that chatbots that they understand already at an entry-level end user perspective, NotebookLM's privacy terms (google enterprise), and it's pricing. Is it perfect?.... no. Do I trust google much? Not really. Are they ethical... lol.

I was surprised to conclude that NotebookLM has the knowledge curation features I was searching for. And it has content creation features based on that knowledge that are so end-user friendly that it's just overwhelmingly a win for team productivity. Those are the "studio"part... generating documents, flashcards, mind maps, reports, presentations, audio about the whole source set or specific sources.

I do not know what the creators of NotebookLM did to the google frontier model to make it behave this way, but whatever it is, I have not found any way to make gemini behave the same way. Give it documents, tell it to base output only using those, it still grabs things out of makebelieve in no time at all. I have yet to see that occur with NotebookLM. I can coax notebookLM to guess about things not provided, but so far not on subjects covered by the soruces.

My experience so far is that every frontier model chatbot lies to me constantly, despite every prompt-based attempt to ground it. I did this by asking questions about things I know well (homelessness, my own company, Salesforce). It makes up stuff that sounds good that's based on an average of what's out there. That can still be net-positive helpful in many circumstances, e.g., when you're trying to learn soemthing you don't know. It's fine when you're working with few details that you can name in your prompt and presenting basic information in a clear or persuasive common-sense format (probably the most frequent writing content creation scenario). And bringing in the net average of digested concepts actually improves this often, bringing in mainstrain ideas that are succeeding without you having to name them or know them specifically. (And there-in lie several ethical questions, which I cannot answer.)

Coding and agent task implementation are, of course, two other very different things apart from knowledge curation and query. I have experimented with and even done real work using those for months as well. Jury is still out, although I have some preliminary conclustions.

--Terry