Gross is every year’s top-grossing movie, since 1913, reviewed.
I have no good reason to be sharing this with you, except it's made me slightly weak at the knees. I uploaded my post about DeMille's The Greatest Show on Earth, from a couple of weeks ago, to Google's new AI, which is called NotebookLM. The AI is billed as a ‘study guide’ and one of its features is that it will take a text of any kind and turn it into a two-person, conversational podcast episode (Why? I hear you say). Two Americans who don’t exist - one female, one male - talk about the text. To be clear, they don’t read it out, like a lot of the AI tools pitched at creators right now. They talk about it. And it is, without question, my most head-spinning AI experience so far.
NotebookLM took less than ten minutes to generate this 8m 17s episode - complete with impeccable, idiomatic speech, exactly the correct degree of vocal fry, pauses, ums and ers, speaking over each other etc. (there are a few glitches too, that will trigger your uncanny valley detector, but really don’t make it seem any less human).
More to the point, Google’s AI has an opinion about my newsletter. Even - if I'm interpreting it right - disagreeing with me, or at least glossing my opinions, which is kind of nuts. Their disagreement is an ideological one. There’s a kind of toning down of the politics, a slightly disapproving tone. And, I suppose, of course there is. I mean these AIs won’t be cutting their ideologies from whole cloth will they? They’re going to have a position and it won’t be a politically radical one. It will be one inserted by its libertarian-scientologist-sovereign-individualist-ayahuasca-transhumanist inventors. So now I realise I have to accommodate the concern that allowing these systems access to our creative processes will necessarily contribute to a further flattening and dulling of political discourse. Another good reason to throw the whole thing into a volcano, I suppose.
Anyway, NotebookLM finds the structure in the review - structure that I didn’t even know was there. So it sets up a kind of dialectic and then tries to resolve it in the conclusion. NotebookLM also adds content that isn't in the post. It’s not just rewording or summarising it, it’s actually adding additional, researched content, including a reference to a scene from the film that isn't in my review. But maybe I’m just easily impressed.
Don’t worry. Messy humanity will be fully restored next time. And it’s Rear Window!
GROSS is free. No charge. Gratis. But if you’d like to support the project, then I’d be thrilled if you’d click this button for a 50% discount on a subscription - that’s £2/month instead of £4 and £20/year instead of £40.
To repeat, all I did was upload the post (I saved it as a PDF, which is one of the formats supported. This is exactly what I uploaded). I made no changes and didn’t have to provide any kind of additional prompt or guidance. The post is the prompt. And the AI seems to have been able to fillet the source like a post-doc - headlines, footnotes, quotes etc. - without confusion.
Brad DeLong thinks we ought to call the AI thing GPT-LLM-ML and not AI, and I’m not unsympathetic.
This is actually the third version that the AI generated. As we’ve come to expect, each time you prompt the model it generates a completely fresh response. This is the best one, slightly longer than the others.
I suspect the research element in the AI podcast reflects the fact that the thing was trained on Google’s store of ALL THE WORLD’S KNOWLEDGE, which must help.
Share this post