I find NotebookLM fascinating.

The closer you look at the results, the more you see the patterns, the more you realize it's an uncreative process that is a cocktail of existing good podcasting tropes.

The closer you look at the outputs, the more you realize they are mediocre… but reliably mediocre, wrapped in a charismatic, superficially high-quality package.

Not unlike MidJourney results.

Masterfully applying useful tropes producing something that has superficial trappings of quality.

A checkbox can't produce upside, it can only avoid downside.

NotebookLM can produce a superficially high quality, engaging, charismatic podcast out of anything.

But it doesn't necessarily have anything interesting to say about anything.

Someone fed it a long file that just said "Poop" and "Fart" and it still generated compelling output.

That's how you know it's able to put a charismatic and engaging podcast-formatted wrapper around anything, without necessarily having anything to say.

An engaging dialogue between two open, worldly people is just inherently compelling to our lizard brains.

NotebookLM is clearly constructed by people who have a highly calibrated taste for compelling podcast content.

That gives a veneer of taste to every bit of output.

But a mechanistic, unchanging object can't produce output of true taste on its own.

True taste is sui generis; it adapts, it evolves.

NotebookLM is presumably a bundle of carefully tuned prompts and a few generic underlying foundation models.

Those mechanistic guts aren't alive, no matter how superficially captivating the results are.