I used the tool on its default setting, leaving it to do its thing after pointing it at the URL of the cookbook. It only took three or four minutes to generate this seriously impressive 15 minute podcast chat between two affable American AI podcast personalities.
Positives
It sounds incredibly natural. I listen to a lot of podcasts and I think I'd be hard pushed to say that this was produced using artificial intelligence. It sounds just like any corporate-leaning podcast discussing a technical topic in a light hearted way.
The hosts um and err, you can hear them breathing between words, they vary the cadence of their words to place emphasis in the right places. I cannot get over just how real it sounds
There's proper banter between the hosts: a realistic back and forth that comes across as two people actually speaking to each other rather than at each other.
The hosts are very good at 'yes and-ing' each other which makes the format of their conversation feel so very familiar and plays to the tropes of this kind of podcast genre.
This is probably the first piece of AI audio generated content I've felt genuinely engaged by. I've heard various AI songs and quote generators of famous people saying things that other famous people have said, but they always seemed like tech toys or proof of concepts. Even though NotebookLM is in beta it's seriously impressive and immersive.
No AI generated podcast ads (yet!) that are twice as loud as the hosts' conversation.
Negatives
Because it's so natural, it's even harder to spot standard LLM problems like hallucinations, important omissions or faulty 'reasoning' than it is when using a text-based LLM.
It highlighted relatively unimportant parts of the cookbook and didn't mention, at least from my perspective, much more important aspects of the guide. I probably could have tinkered with the prompt through to solve that.
I now have no doubt that I can be fooled by (and even, shock horror, entertained by) AI generated audio. I suppose this is a sad realisation in some ways but also a good time to know this as part of being an informed consumer of content and collector of scam threat vectors.
At 1:25 I think it trips up on my cookbook metaphor and cites 'internal stakeholders' and 'restaurant critics' as two groups of people for product managers to try to please! I have to say I did miss this the first time I listened though because it sounded so smooth and authoritative.
Podcasts and the future of AI parasocial relationships
I once saw podcasts described as 'friendship simulators'. You spend hours of your life listening to and being invested in what your favourite podcast hosts have to say about a topic you're really interested in, so it's inevitable you'll develop some kind of relationship with them (and typically, a unreciprocated parasocial one at that). What does a realistic sounding podcast-based friendship simulator mean for the future of our already complex relationships with our mass media performers? Well strap in Dorothy because we're about to find out!
What is this useful for?
This could be a novel and engaging way of learning new things or digesting lots of complex text but I think I'm too concerned about the much too effective masking of hallucinations and omissions to use it for anything really important. So for me it's utility is going to be limited to infotainment, and I feel like there are already enough good 'real' options for that without having to put in the effort to direct my own bespoke version of Stuff You Should Know.
From a business perspective though, if I was running a company like Gimlet Media that made branded podcasts for other companies, I would be more than a little bit worried right now.