For commenters who didn't read the text below the audio, I'll repost it here:
This was created by Google’s NotebookLM by a subscriber who was bored on the Slack channel this morning. All he did, as far as I’m aware, was point NotebookLM to the Substack url and click “go.”
It’s absurd.
And if Google didn’t pay NPR for using their Fresh Air, Markteplace, and All Things Considered material to train this against, I hope NPR has a great lawsuit warmed up.
Has someone tried to reproduce this? What happens if you do it a second time? Is it identical? Similar? Dissimilar? Completely different? I tried following the instructions below with Chat GPT and I got this:
"It seems I can't access the Substack link directly. However, if you can give me a quick overview or some key points of "HWFO" content, I'd be glad to help create an NPR-style interview based on it."
The AI seems to have accurately assessed your writing as much as the simpering affect of NPR talking heads. The only thing more cringe than the faux "banter" was the repeated use of plural pronouns for the author who I believe is one male person.
The only thing really missing was the condescension and scorn "real" NPR hosts would heap upon you if they ever broadcast an assessment of your work.
It doesn't appear to. After you have the AI digest the text material there is literally one button that will generate the audio, and it doesn't appear to have any settings. The cool thing is that it also provides a LLM-like chat interface you can use to query about the subject material. I'm not sure if you've been given the link to the notebook; if you haven't, you're missing out on an analysis of your own work which might or might not be valuable to you.
You say it refers to the author with plural pronouns, I didn’t notice anything like “we” or “us” just they, which is sometimes plural but pretty frequently used to refer to one person whose gender is unknown
Oh my god that's amazing. If you hadn't said it was AI, I'd never have known, short of the fact there wasn't the sound of heads exploding about thirty seconds in. :D
Huh, Talk AI more fair and accurate than Human Talk. What a world. Funniest tihng I've listened to all year. They even have the sense to diss "common sense'' which is so uncommon.
I will put on my pedant hat for a moment though and point out that Kai Ryssdal is not an NPR employee. He works for American Public Media (APM), whose shows air on many public radio stations. A distinction without a difference? Yes, most likely.
Waitaminute...the entire 8 minutes of audio was AI generated!?? How? I get that NotebookLM was involved, but does that have audio generation in its capabilities?
I write a small special-interest blog about politics in my Detroit area suburb. I fed my blog's URL to it and it generated some very interesting insights about the stuff I've written over the past decade or so. It's working on generating the fake radio talk show right now.
It appears to have understood two distinct nodes, with very little fudging to the mean, AND the vector of how one would speak about the other... is the time of the LLM egregores upon us?
If something is free OTA, is there the expectation of payment? I'm just trying to think through the chain of events here. DMCA issues? Which would be kind of amusing given how much google takes down on their various platforms for "DMCA" requests.
For commenters who didn't read the text below the audio, I'll repost it here:
This was created by Google’s NotebookLM by a subscriber who was bored on the Slack channel this morning. All he did, as far as I’m aware, was point NotebookLM to the Substack url and click “go.”
It’s absurd.
And if Google didn’t pay NPR for using their Fresh Air, Markteplace, and All Things Considered material to train this against, I hope NPR has a great lawsuit warmed up.
J. E., I literally posted https://hwfo.substack.com/ and clicked the generate button. This was the result.
Has someone tried to reproduce this? What happens if you do it a second time? Is it identical? Similar? Dissimilar? Completely different? I tried following the instructions below with Chat GPT and I got this:
"It seems I can't access the Substack link directly. However, if you can give me a quick overview or some key points of "HWFO" content, I'd be glad to help create an NPR-style interview based on it."
It just did it's thing, which sounded curiously like NPR. The guess is they used thousands of hours of NPR recordings as their training data.
Wow, on multiple levels.
The AI seems to have accurately assessed your writing as much as the simpering affect of NPR talking heads. The only thing more cringe than the faux "banter" was the repeated use of plural pronouns for the author who I believe is one male person.
The only thing really missing was the condescension and scorn "real" NPR hosts would heap upon you if they ever broadcast an assessment of your work.
I'm not sure if the AI has a prompt tree which could be used to include that scorn. It would be interesting if it did, I'd like to hear the results.
It doesn't appear to. After you have the AI digest the text material there is literally one button that will generate the audio, and it doesn't appear to have any settings. The cool thing is that it also provides a LLM-like chat interface you can use to query about the subject material. I'm not sure if you've been given the link to the notebook; if you haven't, you're missing out on an analysis of your own work which might or might not be valuable to you.
By plural pronouns do you mean “they” which has been used to refer to single individuals of unknown gender for centuries?
I’m not sure I understand your question, or what you’re actually asking.
You say it refers to the author with plural pronouns, I didn’t notice anything like “we” or “us” just they, which is sometimes plural but pretty frequently used to refer to one person whose gender is unknown
Amazing how hard it is to tell whether or not this is real or AI.
Oh my god that's amazing. If you hadn't said it was AI, I'd never have known, short of the fact there wasn't the sound of heads exploding about thirty seconds in. :D
Huh, Talk AI more fair and accurate than Human Talk. What a world. Funniest tihng I've listened to all year. They even have the sense to diss "common sense'' which is so uncommon.
Impressive.
I will put on my pedant hat for a moment though and point out that Kai Ryssdal is not an NPR employee. He works for American Public Media (APM), whose shows air on many public radio stations. A distinction without a difference? Yes, most likely.
You are correct and the voice sounds nothing like Ryssdal.
I refuse to believe this is AI - no language model could conjure this level of intellectual pretension and pomposity.
“See if you can turn the tables”
Gold.
This is how we know they tuned it off of NPR.
I just ran this on our blog, and I am so impressed by this.
Handwaving Freakoutery is one of my echo chambers. That and Ecosophia.
Fake AI NPR (absurd PMC/regime propaganda 95% of the time) claims that HWFO uses facts and logic, but is "cringe". lolololololol
The irony.
What a blast. I’m fairly amazed at the content and repartee. What an agreeable post cast to boot.
It seems like they treated you about as fairly as one could hope NPR to treat you.
Sometimes you have to trust the machines more than the people. :)
I'd say it treated you far more fairly than the actual hosts would have. IMO. This was pretty nuts though.
Waitaminute...the entire 8 minutes of audio was AI generated!?? How? I get that NotebookLM was involved, but does that have audio generation in its capabilities?
Get on Slack and ask them, they made it in about two minutes. I don't even know how they made it.
*Wow*. Gobsmacked.
https://blog.google/technology/ai/notebooklm-audio-overviews/
I write a small special-interest blog about politics in my Detroit area suburb. I fed my blog's URL to it and it generated some very interesting insights about the stuff I've written over the past decade or so. It's working on generating the fake radio talk show right now.
NPR NPC vs. AI NPC
Hard to tell the difference!
I was so confused listening to this! If only we lived in a world where this could be real!
It appears to have understood two distinct nodes, with very little fudging to the mean, AND the vector of how one would speak about the other... is the time of the LLM egregores upon us?
If something is free OTA, is there the expectation of payment? I'm just trying to think through the chain of events here. DMCA issues? Which would be kind of amusing given how much google takes down on their various platforms for "DMCA" requests.
I would presume it's still got copyright? Maybe not. I don't pay much attention to NPR, honestly.