Lex Fridman Interviews Richard Feynman
podcast.aiNow that these deepfake interviews are getting posted regularly, they should have something in the title indicating that they are deepfakes / AI-generated. I thought this was an old interview with Richard Feynman that someone had found and posted, which would've been pretty cool, but instead it's just another one of these auto-generated interviews. (It being from podcast.ai was what gave it away since I saw the last one with Steve Jobs and Joe Rogan, but really it should be obvious from the title.)
> should have something in the title indicating that they are deepfakes
What gave it away for me is that there is only a small overlap in time when Feynman and Fridman were both alive. Feynman died in 1988 and Fridman was born in 1983. :)
Fridman is around my age and Feynman was always a historical figure.
You can also check the url. `.ai` subdomains are a pretty good indicator of the content you are about to consume.
Lex Fridman was born in 1983 and Feynman died in 1988, that's more obvious than Steve Jobs and Joe Rogan since although Rogan wasn't doing much in interviewing during Jobs' life, it's plausible since they are somewhat contemporaries.
This assumes everyone knows who Lex Fridman is. It seems reasonable to me that deepfake interviews be labeled as such.
I was interested when Feynman pushed back on the laziness of the opening question - seems like a very Feynman thing to do - so I was interested.
The follow up question, though, snuck in a lie - "you went to UC Berkeley for two years." That property of these interviews seems dangerous, and likely to implant false factoids in people's heads.
Yeah I commented on this on the last thread by this company on here, but it's like literally a demo of a fake news / misinformation machine. Content that sounds stylistically like the parroted speakers but is just factually incorrect or misrepresentative of what the speakers would likely say.
I guess if the point is to demo the pure technical achievement it's maybe cool but there needs to be a huge 'misinfo' disclaimer plastered literally everywhere that it's for entertainment, although even then people could just rip the audio and repost it and pretend it was real.
Fun fact, factoids are always false.
Necessarily? I'd assume a factoid has the potential to be either true or false.
By some original definition of the word they are necessarily false. Usage seems to have shifted to ‘small fact’ though
The common misunderstanding of the word `factoid` to mean "a small fact" is itself a factoid.
Oxford English Dictionary (via Google) gives:
Factoid
- a brief or trivial item of news or information - an assumption or speculation that is reported and repeated so often that it becomes accepted as fact
So both uses are reasonable?
Let the dead be dead. Read what they left us. Watch what they left us. Listen to what they left us. Touch what they left us.
But if they weren't aware of this future or didn't leave any instructions on how to perceive them if such a future came about, then they should be left alone.
Can you imagine seeing the first photograph of a person? Like all you'd ever known was paintings and then one day someone shows you a glossy, pixel-perfect photograph of yourself.
Photos didn't start perfect and neither is this black magic of recreating a person's speech patterns.
Hard to imagine a world without photographs.
This is not a photograph. This is a drawing by someone who never saw the subject but has had the subject's appearance described to them, and the artist is not even a person to properly understand the descriptions, being presented as a photograph.
a bit too real and now i m learning all the false factoids about feynman. I hope i never remember them. i guess if i let it play too long, it will be forever a memory. next time i watch a video of feynman i will see him as a very ramblingman
this is fucked up
manufacturing content from the voices/faces of prominent people that have died is no bueno
fuck podcast.ai
I also think it's kind of pointless as anything other than a demonstration of the speech synthesis (which is getting pretty good). The Feynman character especially shows that while these models can pick up a bunch of the details about the surface style, they certainly can't be clever or insightful, which is the appeal of the people being imitated.
With respect to using the voices or faces of dead people, I do think it's interesting that the estates of the deceased can have some control over commercial uses of personal images. The estate of MLK I think still controls use of the whole audio from some of his famous speeches. But my understanding is we haven't yet settled whether synthetic / derived artifacts which rely heavily on existing photos or recordings should themselves be controlled in the same way.
In all earnestness - why?
It's not like anyone would be fooled into thinking this was real. What's the harm?
> It's not like anyone would be fooled into thinking this was real. What's the harm?
You are grossly overestimating the collective intelligence of people on the Internet and yes this also includes the "wittier" Lex Fridman audience.
~~also~~ especially
>not like anyone would be fooled into thinking this was real
my god, are you asleep?
Clickbait. I was fooled by the title and would have listened.
Dan, if there is a way to demote deepfake fiction, a way that doesn't cause other problems, that would be great.
> i hate Lex Friedman and Richard Feynnman
That's pretty harsh.
You never get to close pandora's box.
The only choice you have is how to decide to adapt to the new reality.
Whining into the void about the new reality isn't adaptation.
It's weakness.
This is such an insane worldview. Lots of things are technically possible but socially or legally discouraged. You could infringe copyright. You could steal. You could walk around naked. It's not whining to say that something is outside of social norms.
When it's trivially easy to break the law and there are no obvious victims, and it can be done on computers, those rules go out the window.
Almost everyone who shared music on Napster would have never shoplifted a CD from a record store. Almost everyone who shares TV episodes on Bittorrent would never have stolen a DVD from a store or climbed up a power pole to illegally tap into cable.
Your single counter example is "people do media piracy". I personally do torrent media, and I'm pretty clear eyed that I do it because there are no negative consequences for me. It has nothing to do with computers or victims. There's no enforcement. If ISPs started fining people or cutting off access for pirating media I would stop because those consequences are worse than the benefit I get.
Who benefits from making deepfakes, and how do we respond to them? By simply throwing our hands up in the air and upvoting you're encouraging this behaviour by normalizing it. If people recognize it as deception that violates consent it will be discouraged.
As I said in another comment, deepfakes benefit those who want to manipulate the public into believing someone did/said something that they didn't.
Do you really not see society-ending possibilities if every dictator or troll on earth had that power?
Sufficiently realistic deepfakes will erase trust in any and all recordings of people doing things, whether those things happened or not.
I have not encouraged, upvoted, or defended this behavior in any way. Nor will I condemn it. I'm just telling you how the real world works.
When it becomes trivial to 3d print gray goo (https://en.wikipedia.org/wiki/Gray_goo) I wonder if you'll sing the same tune to humanity's destruction or if you'd admit that there's a spectrum and there's some threshold past which a technology becomes condemnable. Because right now your statement basically implies there is no such threshold (since clearly the person you're replying to thinks podcast.ai is past it).
You can "condemn" it all you want, but if people can easily create gray goo we will all die, and no laws or social norms will save us.
Humanity had a good run. If 3D printing and AI podcasts is what ends us all, then so be it.
Imagine an enemy nation-state of [insert your country] decides to create a "recording" of a political candidate admitting to pedophilia.
Yes, those of us in tech would be skeptical. But 99.9% of people don't read technology or political news and don't know this is possible. They haven't been inoculated against it.
What would happen? You can't make that stigma go away.
One of Trump's only drops in polling during his first presidential campaign was due to an audio recording. In his 2023 campaign, any recording that comes out could easily be AI-generated.
If that doesn't terrify you, I don't think you've been paying enough attention to the tactics of modern authoritarian regimes.
It will likely eventually be possible to create bioweapons like https://en.wikipedia.org/wiki/Gray_goo, but we would still condemn someone for doing so, your statement makes literally no sense.
Even the fake Feynman is smarter than most real people Lex has interviewed.
I get the interest in Richard Feynman, but why lex?
A huge set of easily accessible training data?
It just seems sad to use the ghost of a guy who desperately chases pop culture/mainstream media trends to talk with feynman's ghost, and not even have it be part of some bit or social commentary
I liked it a lot and am very surprised about much negative comments here. The episode of Rogan vs. Jobs is interesting as well.
So did I. When you apply critical thinking, I don't think it matters where the insights came from, they are still valuable.
Is this direct text-to-speech or is someone reading the text and transforming the audio?
Listen to the audio between 4:00 and 4:15 - maybe in a quiet room with headphones. That's either TTS, or a Lex Fridman impersonator with some fucked up vocal cords. You tell me what's more likely.
There's also the mispronunciations for common words - words that you can easily verify Lex Fridman has no difficulty with.
I think it's important that this makes it clear that this is fake. The title here is 100% a lie.
At least, it should be "AI Generated Fake Lex Fridman interviews Fake Richard Feynman."
Have some self-respect.
This is the problem with "fake news" and increasingly AI generated content. From an information theoretic perspective this is a perfect title. It sounds 100% improbable, it is therefore 100% surprising, yet it (apparently) delivers what it says it does.
From a receivers point of view, this message is a super-carrier of information, unless (or until) they can discern that it's fake.
Clearly, they are not contemporaries so it’s obvious.
Is there any value to these? like its a cool tech demo, but they are mostly saying non sense.
They mostly don't say nonsense. I listened to the whole thing and I found it very convincing. I've never been able to listen all the way through AI generated content before. It's hard for me to believe this was 100% AI generated.
> I listened to the whole thing and I found it very convincing.
Eh? "Lex Fridman's" voice is shaky as fuck in the first 30 seconds and apparently he can't pronounce science now? - it says sci-in. (It continues to be flawed - he sounds sick when he says UC Berkeley). If I heard this knowing it was really him I would be concerned he was having a stroke.
His voice is defective enough that it is actually irritating to listen to.
Also, the opening monologue doesn't sound much like Feynman's style - he was known for being irreverent, flippant, intense and maybe even occasionally dismissive - but not this aggressive - certainly in a sitdown interview he agreed to. Also hard to believe he would fuck up the word "interviewee".
He wants to "walk"? and listen?.
https://www.youtube.com/watch?v=36GT2zI8lVA
This would only fool someone that has never heard Feynman speak for an appreciable period - but then you might as well get a decent impersonator. And really the content doesn't hold up to even basic critical listening scrutiny.
> They mostly don't say nonsense.
Oh god dude, most of the material is false and easily verified as false from basic sources. It's mostly nonsense.
EDIT: Ok listening to segment between 3:30 and 4:30 I know I'm being trolled.
We must have different frames of reference. I'm surprised that anyone would scoff at this.
I think the beginning and end of the value of these is as a tech demo. At this point in the lifecycle of the technology, I don't think anyone is trying to achieve novel, valuable content.
It's marketing for a tech demo.
Interesting. Very weird.
This is amazing.