279 comments

[ 2.2 ms ] story [ 270 ms ] thread
I'm surprised how real Rogan sounds and how Jobs does not. Why is that?
Presumably because we have hours, days, weeks of Joe Rogan speaking - not just on his podcast but as a sports announcer as well. Steve Jobs... we have a few speeches and presentations, but we don't have much data on how he spoke by comparison.
More data to train on?
Exponentially more training data for Joe than Steve, and infinity more training data in a podcast episode setting.
Probably training set size. Joe Rogan talks for a living.
Maybe there's more training data available for Rogan. The guy pumps out hundreds of hours of content a year in which he's recorded discussing every topic under the sun. I can't imagine there's a similar quantity of recordings of Jobs's voice - or of almost anyone's voice for that matter.

Edit: four other people replied in the time it took me to type two sentences. I guess the answer is that obvious.

Probably significantly more training data for Rogan than Jobs and a much wider range thanks to his long running pod cast. I am not super familiar with Steve Jobs so I can't think of anything other than his keynotes and some interviews that you would be able to use for him.

Unrelated point...that laugh was incredibly bad and repetitive to the point it felt like they were playing laugh.wav file each time they wanted a laugh instead of generating a new laugh of variable pitch and length.

The issue is Job's training data is likely 99% his public "presentation voice" audio -- cadence, inflection, emphasis from public remarks at Apple events, commencement addresses, shareholder meetings, etc -- which OF COURSE sounds unnatural in regular conversation.

Meanwhile Rogan has million hours of regular conversation audio to learn from.

Would the fact that Joe's data is more standardized and produced the same way. Job's data is likely a mix of different volumes, echo levels, processing have an effect
one could hire a _really good_ steve jobs voice actor to generate more training data for the AI algorithm?
At that point using them to create the exact audio would be easier
Yeah but what VC is interested in funding /that/?
Humans are expensive though. If you have a lot of speech to record, it might be cheaper to use the human to train the AI and then let the AI finish the rest.
Then you could just hire the actor to read Jobs' part directly?

Hiring people to train their replacements seems off to me.

ethically questionable, but financially it makes some sense
Then you'd need to hire the actor for every part. After enough training with the actor, you won't need to hire the actor anymore.
There's also Respeecher, which lets you realistically "puppet" someone else's voice.
What non-presentation source material do Steve Jobs voice actors train with? Seems like that same source material can be used to train the AI voice.
One million hours = 114.2 years
Not sure if you meant "million hours" as hyperbole; but that'd be about 114 years of non-stop conversation.

If there's ~2000 episodes of his podcast and he's talked in a bunch of other place too, it's probably less than 5000 hours.

lol, everyone calls it training data. Here I was thinking I was in the right practice.
(comment deleted)
(comment deleted)
This really feels like next level stuff--as Jobs would love to note, that's the magic of integration. But, it's really not novel--just someone putting together the building blocks of foundational ML models already out there.

I'm really excited to see version 2 and so on of this.

I think this is novel.

If being "novel" required being made of things that didn't exist before, would there be anything that meets that bar?

(comment deleted)
I think what they're saying is that this artifact could have been produced a long time ago with very similar techniques, and the novelty is in the ecosystem and tooling which lowers the barriers to creating it.
Exactly! My comment is not a ding on the product, but amazement at what a couple engineers with the right mindset can achieve with existing tools in this day and age.
Is this scripted by an AI or scripted by a human and simply voiced by an AI?
If you can't tell the difference, then the AI scriptwriter just passed a Turing test.
Human scriptwriter just failed it.
"The episodes are rendered using play.ht's ultra-realistic voices, and transcripts are generated with fine-tuned language models. For example, the Steve Jobs episode was trained on his biography and all recordings of him we could find online so the AI could accurately bring him back to life." - https://podcast.ai/about
Feels like both written and voiced by AI after listening to it. There were some tangents that felt a bit off. The Gizmodo bit was probably about the iPhone that was left at the bar but they never actually addressed the subject matter which was strange.
(comment deleted)
Some of it is very convincing! The first parts about spirituality especially. But for me it really fell apart and became an obvious GPT text later on — for instance the part about what Adobe should do is completely nonsensical.
Wow Rogan sounds very realistic. Is the voice completely generated or are some words/phrases "borrowed" from real sample data portions?
There's a huge amount of data to base this on, in the very context that it's presented in.
This is mindblowing. Like this could be real, and I'm learning stuff from it: There's an Indian epic that's 10 times as long as ...

There's some audio distortion (sounds like clips cut together, little "notches" in the soundscape) but apart from that, and some weirdness in sensing "the spatial location" where this audio was recorded...the concepts and the dialog are amazing.

Some parts are weird...but people can be weird. It you tidied this up, and added the right sounds affects and audio processing to this, without the cue that this is AI generated...holy fuck, I think people would believe it. Particularly if you cut it together as a "highlights reel". Jobs does sound a bit off tho, a bit thin...there should be enough data on him to do a sparse reconstruction of his voice to a level of accuracy beyond human discernment tho.

The thing this got wrong about Job's voice cadence, tho is: Jobs speaks a lot more slowly and deliberately, and with a lot more pauses, than here. I suspect the cadence / timing is not so emphatically modelled by this AI.

I think also they're missing some emotional trajectory coherence in both their voices. Like the emotional register of the voice does not sound or transition as naturally, and is less diverse.

Incredible PoC. AI folks are the new dark wizards. WTF can they not do? That list is shorter

> I'm learning stuff from it: There's an Indian epic that's 10 times as long as ...

From a cursory websearch:

> At about 1.8 million words in total, the Mahābhārata is roughly ten times the length of the Iliad and the Odyssey combined

https://en.wikipedia.org/wiki/Mahabharata

> The thing this got wrong about Job's voice cadence, tho is: Jobs speaks a lot more slowly and deliberately, and with a lot more pauses, than here. I suspect the cadence / timing is not so emphatically modelled by this AI.

That's because AI Broe Jogan got AI Jeve Stobs high before the show started.

>> That's because AI Broe Jogan got AI Jeve Stobs high before the show started.

He does this for every interviewee! I was shocked, but pleased when the Substack vp of comms wrote about her experience whilst colloquially escorting the Substack CEO for the interview. [1] [2]

>> At a long desk in the big main room sits an attractive nurse. She offers us an enhancer of B12 or NAD+, through a shot or an IV.

>> I get a shot of NAD+, which is supposed to be good for energy and metabolism. NAD stands for Nicotinamide Adenine Dinucleotide, and I don’t know what the + is. (Actually, I don’t know what any of it is, but the nurse said she takes it, and if you saw this woman, you too would ask for a shot of whatever she’s on.)

Ultimately, this is pretty crazy to me. Those along with the interviewees get “high” too and since Steve Jobs, this is the first time it’s been written about it.

I wish people knew more about this! It’s been a while since Steve Jobs has been interviewed. 1000+ shots later, Rogan has grown and is stronger than ever.

[1] https://substack.com/profile/39686168-lulu-cheng-meservey

[2] https://news.ycombinator.com/item?id=32649123

Neither of those substances are used to get what is colloquially known as "high" (one is a vitamin), you don't take either as a recreational drug.

I think what GP is referring to is Joe Rogan smoking weed with some guests before, during and after the show, rather than some vitamin and pro-metabolism shot.

> colloquially

What word did you mean here?

Is it legal to offer casually offer injections to passersby, even if you are a nurse?

There are IV clinics popping up all over the place, so there has to be some sort of streamlined process. No way they are full of MDs overseeing everything.
B vitamin IV injections are great for hangovers.
That might just be lack of inputs for Jobs compared to Rogan. Like we have tons of Rogan speaking in a podcast setting. We have very few Jobs interviews in comparison and the ones we do have take place over a much longer stretch of his life (so speech patterns will have changed). I might also guess that a decent amount of Jobs' audio might be from non-conversational contexts like keynotes and lectures.
It was really hard to find good quality of Steve Jobs' voice, most of his speeches are keynotes on stage, I honestly was surprised that we managed to get to that quality from such poor quality recordings.
You can hear the "space" when Steve Jobs is talking that a lot of the training material are from keynotes on a stage in a big space. That's why it sounds a bit "thin" compared to Joe Rogans voice which is probably mostly based on recordings in a studio with Joe being close to a microphone (and a proper microphone for recording)
I work in audio post production and would be happy to take a crack at cleaning up your source audio. There's a lot we can do to isolate the dialogue from the space and restore presence.
I agree it's a fantastic PoC. Jobs' voice sounds very heavily derived from keynote speeches. People speak with different intonation and cadence in different situations and the uncanny valley effect here is due to having him speak casually on a podcast but in keynote voice. It'd be like training a Paul McCartney AI voice solely on Beatles songs - it wouldn't sound anything like McCartney's real life speech.
Agreed

Joe Roegan I can guarantee you just had a MUCH better amount of background data to feed this thing - I doubt we have nearly enough high quality audio of Steve Jobs speaking casually to make a convincing facsimile.

A future Steve Jobs though where everyone has a smartphone and more is recorded.....that is kind of scary.

Stuff like this kind of makes me want to secretly record myself in very high quality in tons of situations and then put it all in a safe with a clause in my will to have it all donated to some AI research firm.

AI Jobs' comments on the emergence of spirituality from the Indian subcontinent and his belief that intuitive knowledge is as real as scientific knowledge sounds like something he absolutely would have said.
I wonder if people on their deathbeds will start recording themselves so that AI can immortalize them for their loved ones after they're gone.
This is a common trope in sci-fi, but I personally don't get the appeal. It's not talking to my dead loved one, it's just a weird technological puppet of them. An actual recording of their voice is much more meaningful, even if it's not interactive.
It actually already exists. There's companies that train AI on a deceased loved one's messages and then have them send texts back. They act as if they're alive and well.
That honestly sounds like it encourages a very unhealthy way to cope with grief.
It'd be interesting if it weren't for the current tech industry that would bend over backwards to find a way to use it for advertising purposes.
(comment deleted)
Ray Kurzweil discussed something like this on the Lex Fridman podcast. He fed all his late father's writing into an AI system and it's able to ask it questions and get answers that seem believable to agree with what his dad would have said.
Putting the ML conversation to one side, this is a fantastic piece of marketing.

It's perfect for getting attention from a wide audience and is a great demo of their product.

Only the voice is AI right? Is the whole conversation flow generated by AI too? That'd be crazy.
No, the text is AI too. It is crazy.
I wonder how much more effective this would be if the prompt wasn't "this was generated by AI." It would be really interesting if someone with a popular podcast injected an AI conversation into their feed and didn't tell their audience for a couple weeks. A real life Turing Test.
Entirely generated by artificial intelligence? So this text was generated by GPT3?
Fantastic marketing execution. Most surprising was the gravitation of the conversation. I'm really curious about the script behind the voices. Was it written by a person or ai as well?
Most mindblowing thing I've seen in recent memory. With the amount of available training data, isn't it now just a small step to AI-generate new episodes of existing podcasts? Exciting and scary.
This sounds like if you spliced together tape of Rogan and Jobs with scotch tape. An interesting concept but I never felt like I was listening to a conversation between two people. Jobs sounds weirdly far away, he also was never a "fast talker". Listen to his interview at MIT where he takes questions. He always stopped for long pauses, stares off into space for a minute. He didn't rush into responses and tended to nail timing really well.

Rogan sounds more believable but I suspect you'd get the same results sitting down at Audacity and cutting together an MP3. You just have better quality samples of him talking to people in a podcast setting.

It's an interesting demo but I'm not sure why I would want this to exist. It seems like technology which you could only abuse, either through the generation of podcast spam or through the production of fake audio. Podcasts are legion and are already cheap to make and based on open standards.

I'd say it is highly unethical. Imagine thousands of people listening to fake podcasts as if they were real. Not only are they being lied to, but if they find out many are sure to become very angry on learning they wasted their time.

Fortunately, it's not there yet, since it's obviously fake.

> wasted their time

That’s the best case scenario. It gets a lot worse than that.

I saw a (scam) video on YouTube that had Elon Musk saying “if you want to make money in crypto, the best way is doing xxx, for example using the website yyy.com”.

I had to do a spit take because it was very convincing for a layman. Both the audio/video were deepfaked, but you couldn’t really tell other than the fact that it was uploaded in 240p.

If my mother had seen that, I’m sure she would have believed it since it’s something Elon would plausibly say. We’re entering a scary time.

Well he only promotes scams so it was extra believable!
I remember that. I reported it to YouTube and never heard anything back. Days later it was still running.

I remember looking at the chain and seeing over 6 figures sent in to the address. It was also timed to line up with one of Musk's "big announcements", which helped it be plausible.

As you said, it was very convincing to a layman - but YouTube aren't laymen, and they must have had dozens if not hundreds of reports about that scam.

Thanks for reminding me of that.

Eh, celebs promote scams all the time. This new way just saves the cost of the bribe, which is a net positive.

More of a concern is putting lies into the mouth of someone who actually is legitimately credible.

If you prohibit it then it will only be done secretly by governments and then used against naive populations. At least this way the population can learn how to identify and defend against it.

If we had transparent governance then maybe this wouldn't be a problem.

Why wouldn't people learn to spot it when the government does it? And I don't know if your argument follows - people clearly either don't know or don't care about the rise of Facebook false reporting.
How do people learn better: lots and lots of practice with varying levels of quality and published examples, or against a single powerful adversary that's trying to conceal what they are doing?

I.e. your question is stupid.

I don’t think it’s worse than how people already believe headlines and even pictures of headlines (popular on Twitter, Reddit) while reacting to it much less demanding to read the actual article, if there even was one.

You can even just make up a caption next to an image like “this is Todd from Finland who raised $100k for the war in Ukraine” and this pic might spread virally as everyone pats Todd on the back “omg faith in humanity restored”.

And these scenarios are even less work and easier to consume than your doomsday. It’s already here and you don’t even have to listen to ab audio clip.

True. But listening to an hour-long or half-hour long podcast is considerable more investment than a simple headline. In terms of the investment in time the listener makes.
> many are sure to become very angry on learning they wasted their time

That is the best case outcome. Given ample real world evidence, however, I think most people will not get to this stage. They will be angry, yes, but at you for trying to introduce facts into their reality.

> Dennett told Motherboard that these sorts of ethical considerations will be important in the future, when natural language processing systems become more available. “There are very dangerous prospects lying in the near future of this technology,” he said. “Copyright doesn’t come close to dealing with all of them. GPT-3 is a sort of automatic plagiarist, and unless great care is taken in how it is used, it can do great damage!”

Dennett after none of the experts familiar with his work was able to select the real reply from GPT-generated ones. Seems like we are there, or at least, started on the path.

I'd say it is highly unethical if this technology is used for deception, or if permission is not asked. Not unthinkable to soon be able to publish a high-quality book, by finetuning on the books another author already wrote.

I think the reason Jobs sounds weird is because it seems like the AI was trained on his public presentations, like maybe the iPhone introduction and his AllThingsD interview with Kara Swisher and Bill Gates - there's a noticeable amount of reverb baked in, he projects a bit more, sounds a bit more prepared - it's not how he'd speak in a 1-on-1 conversation without an audience. I would love to hear a version trained on his more intimate, 1-on-1 interviews, which he's done plenty of.
Style transfer is a pretty well solved problem, at least for visuals. Could be applied to audio too, to fix reverb and such.
What I find funny is that both Jobs and Rogan sounds like they're doing stilted readings off of teleprompters at some points. The AI must have been trained on canned sponsor ad reads for the latter.
Exactly. I don't think it would be too difficult to do a little editing and end up with much more natural sounding speech from both of them.
> you'd get the same results sitting down at Audacity and cutting together an MP3

Yes, but it would take you weeks to do it.

> It's an interesting demo but I'm not sure why I would want this to exist. It seems like technology which you could only abuse, either through the generation of podcast spam or through the production of fake audio.

Generally agree. There seems to be the kernel of some really cool tech here, and I think the non-abusive applications may be obvious only in hindsight.

next AI website will be 'idontwantthistoexist.com'
How much did you listen to? The first few minutes with Jobs is rough but once there are questions presented, there is somewhat of a real conversation. Very interesting.
"My hope is someday, when the next Aristotle is alive, we can capture the underlying worldview of that Aristotle - in a computer. And someday, some student will be able not only to read the words Aristotle wrote, but ask Aristotle a question - and get an answer!"- Steve Jobs, 1985
Maybe he should have tried reading some of Aristotle's books.
A slight tangent, but there's something extremely interesting on the topic of Ancient philosophers related to this Jobs' quote. Socrates wrote nothing down, to the point that some have claimed he was illiterate. And that is a possibility, though improbable.

But the reason he claimed to not want to write anything down is because, in a nutshell, 'books cannot defend themselves' : words can be taken out of context, meanings misconstrued, and text made to mean what the interpreter wants to make it mean instead of what the author meant for it to mean.

It's quite fortunate for the world that many of his students disagreed, but it's interesting nonetheless.

There's no way to know if Socrates' absence would have led to a net benefit or not

Edit: to expand on this, we can say it is fortunate that his words were written down, but in a Socrates-less world or a world where his thoughts evaporated, you don't know which philosophers would have taken up the mantle instead and what the consequences would have been

Yes, no one can ever no anything about a counterfactual. So does nothing matter beyond whatever each individual chooses to matter? (Nihilism / Hedonism.)
I meant more in the sense that we are primed to take one side of the issue by the default, the one we know. In the counter-factual Socrates-less world I'm describing, we are the unprovable counter-factual and people there are posting about how grateful they are that Eulogothenes' words were put to papyrus.

Perhaps a better way to illustrate this is to point out that the Mona Lisa's cultural prominence essentially ballooned once it was stolen.

A tangent on a tangent.

This is the first time I've read why Socrates didn't leave any writings. Which incidentally reminds me of a saying in Zen Buddhism of "do not establish words and letters" (see eg. https://en.wikipedia.org/wiki/Zen_scriptures ) If you squint hard enough, Zen's preferred method of teaching via teacher-student interactions very vaguely resemble the Socratic method.

So I wonder whether there's any more-than-incidental connection between the two, given https://en.wikipedia.org/wiki/Greco-Buddhism -- i.e. was Zen Buddhism a very distant offshoot of the Socratic philosophy and methods?

(Sorry, might be too much of a tangent. But hey, Steve Jobs probably held Zen in high regard :D)

It's a good point. Do any of his dialogues include the other side of the debate?

Did Plato ever speak about how we wants to communicate ideas to later generations? Maybe he didn't want to, and he trusted his students to preserve and improve his important ideas? How did he expect people far away in time and space to study philosophy without access to great thinkers? Did he not care? Did he expect every niche to reinvent philosophy from scratch?

Thomas Jefferson wanted the US Constitution to be rewritten by 1810, but we are still stuck with it today (with amendments, granted).

>It's a good point. Do any of his dialogues include the other side of the debate?

oh sure, they include the other side, but amusingly enough the other side are always idiots.

It’s also quite unfortunate that the world needs to keep learning the basic lessons of media with every new communications technology.

The epistemological issues that Socrates had with the written word are not too dissimilar from the epistemological issues of contemporary social media.

Socbae dropping bombs about context collapse
This is probably the best use case for this technology. Although I think there's a lot of potential for very realistic fake news
You dont need fake news to be realistic to be believable. You just need to repeat them over and over again. Works wonders.
Not even that. Just need it to fit a narrative people already had.
That's good for extension and motivation, but not persuasion.
A motivated person will persuade his close friend/family.

That's why this scales well

This concept is explored in-depth in the Hyperion book series. Highly recommended if you like the genre :)
"What bothers me is, nothing does." - Dixie Flatline in William Gibson's _Neuromancer_, 1984

Wonder if Jobs had read that by 1985.

Really? We just made a version of GPT3 finetuned on the complete works of Plato. It produced some solid new dialogues. About, for instance, the relationship between beauty and the good.
the mind of a person is ever changing. you cant capture it across all their works.
GPT was trained on the works of Dan Dennett. People deeply familiar were asked to select the real reply from a multiple choice of generated replies.

Even the experts were only 50% accurate. I was close to random.

And it is not like Dennett or Plato are going to change their minds at the end of their careers of changing the minds of others.

https://www.vice.com/en/article/epzx3m/in-experiment-ai-succ...

Out of that whole gang (Pinker, et el), he seems the easiest to GPT3.
Did you just put Steven Pinker and Daniel Dennett in "a gang" ?

Wash out your mouth with soap, this instant!

Seriously, is there anything that connects these two other than having written some reasonably accessible books on matters in the realm of cognitive science?

Is there some feud here? I see a lot of crossover in the 2 seconds of googling I just did but know nothing about these people.
Here’s my guess: Pinker writes sometimes on evolutionary psychology. A controversy I heard around the atheism plus days was that evopsych should be banned. Therefore it’s likely that Pinker has been declared cancelled. Thus, to mention his name is to sin, and a sinner must repent, so wash your mouth out with soap and divert yourself from this path of heresy lest you join him.
Completely and utterly wrong.

I just don't see Pinker and Dennett as remotely connected. They don't work in the same field, they don't have much to say about the things the other is interested, and they were not even really on the same page w.r.t. "new atheism".

Hm, let us see. 10 seconds on google and wikipedia revealed:

"Steven Arthur Pinker is a Canadian-American cognitive psychologist, psycholinguist, popular science author and public intellectual. He is an advocate of evolutionary psychology and the computational theory of mind."

"Daniel Clement Dennett III is an American philosopher, writer, and cognitive scientist whose research centers on the philosophy of mind, philosophy of science, and philosophy of biology, particularly as those fields relate to evolutionary biology and cognitive science."

Not even remotely connected, you say?

Curious, how are Pinker and Dennett opposed wrt. "new atheism"?
Pinker and Dennett were quite chummy at least 15 years ago (along with Richard Dawkins and Sam Harris) as part of the "New Atheist" movement in addition to their shared interest in cognitive matters. I don't know what they feel about each other today -- there's been quite a lot of infighting and "New Atheism" as a movement has sort of splintered.
And how was GPT3's Attic Greek? Did you run it by classicists?
That’s the goal is to run a Human GAN, powered by Classicists.
> We just made a version of GPT3 finetuned on the complete works of Plato. It produced some solid new dialogues.

I highly highly doubt that.

New dialogues aren't so hard. They are rehashes of old dialogues. New ideas are hard.
* Is beauty wise, socrates? * Or rather, I will put the question more clearly: We know, do we not, that which is wise must also be good? * Certainly. * And what is good must also be beautiful? * True. * Then where there is wisdom there is goodness and beauty. * Certainly. * But where there is no wisdom there is no goodness. And where there is no goodness there is no beauty. Nor is this only true of the matter which we were discussing, but of all others as well. * I agree. * And therefore, where there is no music, we shall find neither wisdom nor goodness, nor beauty. * I dare say. * And that which has no music is unpleasing to the ear. But the unpleasing is the disagreeable, and the disagreeable is the evil, and therefore in music there is the greatness of goodness, and beauty, and wisdom, and lastly, in all that knowledge which relates to the music and harmony of the soul, which we have escaped to the end of our argument, in that, I fancy, you would place justice. * 19. "And so, glaucon," I said, the cause of every thing, which you want to find, has been found ?" * Certainly," he said. * "And is not that cause beauty?" * "I do not understand you," he said. * "The cause of all things has the name of beauty.”
There are some awkward jumps there to the topic of music, and I suspect it lacks coherence but I didn't want to re-read to the point of fatigue.

That said, the passage and its idea sure are beautiful!

Plato would never write this crap.
So a series of sentences with no underlying meaning, written in the style of a socratic dialogue? x)
This is such a great idea. And for me it goes even beyond being able to exploit the intellect of deceased geniuses. It could be the path to generating believable AI personas out of everyday people. If sufficiently accurate know I would love to have an AI with my father's advice and sense of humor once he is no longer with us.
> Rogan sounds more believable but I suspect you'd get the same results sitting down at Audacity and cutting together an MP3.

I am sure you would. The interesting bit is not having to do that.

> An interesting concept but

The ol' "it's never gonna be as good as..." keeps failing us, and increasingly so. If the recent AI trajectories are any indicator, I'll give it a solid 6 month tops before it's good enough to generate something that fools absolutely everyone (granted, having an interviewee that is, you know, alive would probably be a precondition).

He sounds far away because the samples are probably from him speaking on stage
(comment deleted)
It reminds me of those fake interviews they used to do on radio - where the artists would record one side of an interview and it would be sent out to hundreds of radio stations and the morning DJ from some podunk town would be interviewing Bono or whatever.
Weird Al Yankovic also did some of those in video form. I thought it was pretty clever and at times even convincing, if a bit ridiculous.
Let's also not deny that this is hilarious.

If you view it as parody rather than looking at what it lacks in realism, I'd say it did a very good job.

> He always stopped for long pauses, stares off into space for a minute. He didn't rush into responses and tended to nail timing really well.

The first part makes me wonder if Jobs had ADHD, the second part screams NO.

I agree. But give it some years and we'll get a proper AI "Joe Rogan & Steve Jobs" podcast (although I'm personally more interested in hearing more about Wozniak than Jobs).
The really bizarre thing is that I found myself enjoying this podcast, perhaps even more than the average Joe Rogan podcast. This didn't feel quite as hollow as the average AI-generated content.
The flow of Jobs' voice is a little off, but wow this is incredible. But also incredibly concerning. Its cool tech, and I guess nothing can stop the inevitability of new technology, but the ability to replicate people's voice, speech patterns, and how they look to a pretty accurate degree is mega sketchy. I guess to future proof yourself, just minimize your audio and visual data online...
Or, you know, if you want to be "resurrected" after the singularity from your digital web-crumbs then make sure you leave plenty of them. People who left the best data will be first to be uploaded.
The impression I got is that his tone sounded like someone giving a speech or presentation. Rogan sounded more natural.

I suspect a large part of the training data was Jobs giving presentations in front of an audience so his AI voice sounds like public speaking. Whereas Rogan's data would be most likely be conversations from his podcast so his AI voice sounds more conversational.

Am I going crazy? This sounds awful.

I've never listened to Joe Rogan - but does he really sound like his voice has passed through a square wave? He has almost no intonation. The AI can't even pronounce "Swayze" correctly.

Job's laugh sounds psychotic and out of place. And his voice, again, sounds robotic.

The dialogue is utterly nonsense.

What's going on with the hype here?

Everyone knows it's terrible, but it's cool that they could do this at all. Like, it's a small preview of what's possible. Imagine this 10 years from now.
10 years? Imagine it 2 years from now. Or even tomorrow, because it's unlikely this is the apex of what's possible now.
I imagine in 10 years it'll be the same. Firstly because if you can't get "enough" good content to model, you will have to manually shape the models to make it sound authentic - and that isn't AI, that's editing. But secondly because humans are very good at recognizing authentically human interactions, and algorithms simply aren't good enough to generate all the nuances we identify as authentic.

If you want to fake someone out, use a regular person, and clean it up in post.

Yeah. The first few seconds of Rogan talking were pretty decent, but Jobs doesn't sound right at all, mainly because of the dialogue specifically.

But I do kind of get the hype.

The first few minutes sounded bad but the last half was quite impressive.
The first minute or two of Jobs is pretty bad, but after that it is not that bad. Not totally convincing, but it isn't bad. I suspect most people saying this did not listen past that part.
It's state of the art. It sounds amazing compared to anything that's existed before. What alternative are you comparing this to to say that it's awful?
A contrarian outburst seems to be the default on HN these days. Perhaps it has seeped in via the rampant infestation of politics. The ‘hackers’ get blasted when sharing projects.

Personally I’m in awe of this tech. But I also detest its inevitable weaponisation. But the tech! Wow.

If you are only looking at the current product and not the potential that comes with it then you are missing the point.
The problem is - can the potential ever be realised?

Self-driving cars are, what, 90% of the way there? And have basically been stuck at "almost there" for quite some time.

AR had huge potential when I first played with it in the early 2000s. Still basically nowhere.

So we've now got systems which can generate text which fools people into thinking it was written by a human - well, Eliza got there first.

Text to speech? The Voder was doing that in 1939! This stuff is better, sure, but pretty far from being realistic.

Perhaps this will succeed. But potential doesn't always mean achievable.

Hi everyone, cofounder of play.ht (the startup behind this podcast) here. let me know if you have any questions.

To give more context, the podcast was totally AI generated, the content itself was generated from a finetuned GPT3 on SteveJobs' biography, the voices were cloned from few hours of both Joe and Steve voices, even though it was tough to get good content for Steve Jobs. And the podcast artwork was generated by SD.

We will be releasing more episodes soon which will be even more mind-blowing!

How much audio do you need to build a model?
Can you elaborate more on what the prompts for the text content were?
This was the prompt: " Podcast.AI Great people, great interviews with our host Joe Rogan. Episode 1 - Steve Jobs Summary: " Then GPT3 generated the summary, then we added: " Transcript: Joe: " That is all.
I still have a hard time believing that the opening of the script was auto-generated. Did GPT-3 really generate the part about the movie Ghost? If so, it has a surprising amount of understanding of what it's generating.
It's still unclear to me. Did you then feed the entire above blurb into GPT3 on a continuous basis to get the actual interview? To be more specific:

In: "Podcast.AI Great people, great interviews with our host Joe Rogan. Episode 1 - Steve Jobs Summary:"

Out: [the summary]

In: [all the above] + " Transcript: Joe: "

Out: the entire podcast?

Do you have Joe Rogan's permission to use his like-ness?

Do you have Jobs'?

This probably falls under the same umbrella as remixing samples into a song or parodying a public figure.
heyy felfel, Mustafa here, nice work man! so proud to see you on the orange website.
Wow, I was sure, from listening to the first few minutes, that the script was written by a human trying to be funny. The part about the NeXT Computer and the three applications was just too funny. Edit: Not to mention the reference to the movie Ghost.
How do you intend to monetize your startup? If you're planning on advertising, do you expect to pay the people you are cloning?
We will never use any cloned voice in any commercial way without consent and compensation, we only wanted to show the community what is possible and what generative AI models can do.
Aren't you using the cloned voices to generate marketing material for your company to make profit? Indirectly profiting without consent or compensation seems like a difficult ethical line. It also seems like something you want to have a very good stance on when a law could be passed that completely shuts down your ability to operate.
Please take the naysayers with a grain of salt, this is a fantastic demo of what's possible. While the voices aren't 100% convincing on good speakers, if this was over the phone it would be indistinguishable, the content of the podcast is also spot on with Joe's sort of rambling and Steve always acting like he's making a point or telling a story.

Great work.

Honestly I expected Joe to ramble a lot more, but I understand this demo is meant for people to listen to Jobs so they probably had to cut out a lot Joe's ramblings.
>...the content itself was generated from a finetuned GPT3 on SteveJobs' biography

Was the actual dialog generated by GPT-3, or just Jobs' responses? If the former, was any of the dialog human edited/spliced together or was the entire script generated as a single output?

No, All was GPT3 generated, check the other comment about how we prompted GPT3 to start the conversation
I know the reaction here is mixed and tbh that's what makes HN so interesting for me. But FWIW I love this podcast! It's a great demonstration of what AI can do. I am going to share it with my students before the next class so that they see what can be possible.
Thank you :) that is the main point of it, to show people what is possible and inspire them to create.
How much human editing was done on the output from GPT3?
Have you reached out to Joe Rogan's team?

It seriously seems like something he would be interested in putting on his show feed, and accompany it with a real interview about the state of AI.

Is the audio entirely generated from scratch along with their voices or was it cut and paste from existing audio?
Everything, the content and the voices are generated by AI. check my other comments for how we did that.
Absolutely incredible! I've shared this with everyone I know and they are all gobsmacked!
Just wanted to say this is amazing!
Is your TTS model a decoder-only language model using discrete audio features inspired by the likes of Tortoise-TTS? The results are impressive.
Does this have a textual transcription? Please publish it!!

I found it so interesting when he was talking about how glad he was that his wife was born, and the purpose of life and such.

Great work! How did you fine-tune GPT3? Did you convert the bio into prompt/completion format?
It sounds good on a first glance, but then the topic switches are totally absurd.

Like Steve talking about LSD all of a sudden, then Joe Rogan asks about Newton (which Jobs didn’t make) and Steve starts talking about… product and iteration?

It sounds like someone is superficially trying to act like Jobs would act, which … it kind of is!

I needed to turn it off after 10 minutes as it obviously goes nowhere. But really interesting experiment nonetheless, and better than I would expect.

I found it pretty funny, as if Jogan was making the joke that ‘the Newton was so bad that you must've been on drugs while you made it’.
The first 20 secs of dialog from Jobs is not believable. He never did chit-chat. He wouldn't give a toot about some podcast. His mission would be to get to the point - himself, his company, or a big idea that intersects one or the other. He wouldn't try to flatter the the host.

Bro Jogan, on the other hand, sounds very believable.

This might have something to do with the ratio of Bro/Jobs training material.

Still, this technology has huge potential.

Mind is blown.. it's verry good. Although there's no real coherence in the 'ainterview', it's very convincing. In terms of TTS, there are some minor glitches here and there, but overall I'm super impressed.

one of the things was the "sell you a quarter of a car" part. I couldn't find a google hit for that.

The opening minute or two sounded a little off, but either Jobs' cadence sped up or I got used to it, and it got it better as it went on. The last few minutes were astounding - Jobs riffing about the good and bad potential of technology, and the final bit about throwing computers - if that was not either heavily curated or copied from pre-existing material I am quite impressed.
Steve Wozniak famously said to "never trust a computer you can't throw out the window."
This has some obvious rough edges but still it's a harbinger of things to come. Seems this was trained on recordings of Jobs' on-stage voice rather than his casual conversation voice. The rough edges in this will be smoothed out over time by new tricks of AI trade. This will get crazy when combined with an AI Jobs that paces around a virtual stage in a black turtleneck hyping AI-invented products.
I too throw my MacBook out the window every now and then. Not to check if it still works, but to get it to work when it stops working