Adam:
It's Thursday, May 25th. This is Accelerate Daily. Today we've got famous selfies of history, Microsoft watermarking AI images, the White House making moves, and a conversation about open source project, girlfriend GPT. Also a recipe for patching up bad presentation audio. The AI future is accelerating. Time to put on your goggles. intro music. Okay. Welcome back, everybody. I'm Adam.

gptboss:
My name is Mackenzie, good morning.

Adam:
We're back with three links and a three links and a hack, three links and a recipe. Get you caught up on what's happening in AI today. How's, how you doing, man?

gptboss:
Oh, it is what it is. Living the dream.

Adam:
We're still in the phase where this part of it surprises you. I don't know. What do I say? It's also weird because it's every, like when you do a monthly podcast, you know, there's an answer for how you do it.

gptboss:
Yeah.

Adam:
We do it every day. It's a little like, I don't know. I just was just talking to you in Slack. Yeah.

gptboss:
Yeah, literally. Like we last spoke like 10 minutes ago.

Adam:
Yeah. I had the, the back to back meetings this morning. of the type where you're like, Oh, I should maybe put on a jacket. And it's just this weird, it's a weird. Like remote employment thing to just sort of go from a life of like, Hey, suits I wear to the office, which is the life I've lived to a single suit coast hangs in the right place. Cause sometimes I'm on meetings where I feel like maybe that should happen.

gptboss:
I used to dress up for like school. I would wear

Adam:
Yeah.

gptboss:
like, like a suit jacket to school, like to high school and college. Well,

Adam:
just to do it or

gptboss:
I liked

Adam:
as a

gptboss:
it,

Adam:
matter

gptboss:
right? I

Adam:
of

gptboss:
was like, I'm getting

Adam:
yeah.

gptboss:
ready for the day and I'm going to stunt on everybody and I'm fly as hell. And this is what we're doing. And then like now is it like, I'm doing a podcast with my sleeves cut off.

Adam:
Hahahaha Oh, it's a brave new world.

gptboss:
Yeah, dematuring or something, you think as

Adam:
Okay.

gptboss:
a Benjamin Button mode? Anyways, so we got this title card here. And you know, when I saw this the first time, I thought that was Hugh Laurie.

Adam:
Yeah, it does. It looks like a sort of sort of Hulari John Ham

gptboss:
Yeah, yeah,

Adam:
character.

gptboss:
John Honeyham, we call him in Canada.

Adam:
John Ham's John Ham.

gptboss:
Yeah.

Adam:
Um,

gptboss:
The tweet

Adam:
no,

gptboss:
really makes

Adam:
it's

gptboss:
this, too.

Adam:
a, it's a, it's an image from a running series that's been popping up on Reddit called famous selfies of history, the, uh, the prompt for this particular image, the link to which you can find in the description is hyper realistic GoPro selfie of Michelangelo and the scaffolding of the Sistine chapel. I can't, you can't really see the Sistine Chapelness of it unless you click through, but there's kind of like some, just some sculpture happening on the edges and he's clearly in a, you know, in the rafters of an epic building.

gptboss:
Yeah.

Adam:
Uh, but the whole series is worth checking out. There's like Abraham Lincoln on a battlefield and there are all these, you know, wide angle selfie looking situations where it's just, it's sort of like, okay. Yeah. Also. Wow. Like that's

gptboss:
No,

Adam:
convincing.

gptboss:
but the tweet goes so crazy.

Adam:
Hahaha.

gptboss:
Michelangelo Bluecheckmark at Sistine Sculptor says, balancing on scaffolding, turning ceilings into skies, who knew God and Adam would be neighbors? Palette emoji, hashtag heavenly canvas, hashtag holy artistry.

Adam:
Yeah.

gptboss:
This

Adam:
Another

gptboss:
is...

Adam:
part of this is they also generate tweets from those, from those, uh, I had, uh, maybe I'll actually do this. This goes to what we were talking about yesterday, but maybe I'll actually do it soon. Uh, there's an old onion article that I've always loved that, that the headline was like Christ returns to the NBA or something like that. And it was mainly just the reason to have an epic picture of like, you know, like it looked like an old Fresca of, or what you're not, not, uh,

gptboss:
Alfresco.

Adam:
Yeah, there you

gptboss:
Yeah.

Adam:
go. Mosaic situation. I don't remember. Michelangelo asked, actually that's the good tie in looks like an epic painting of Christ dunking a basketball over

gptboss:
Hehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehe

Adam:
a bunch of people. Um, but then the snippet for the story was about like Jesus Christ getting drafted into the NBA as the first pick and blah, blah, blah. Um, In like 2010 or something, I bought, uh, NBA Jesus. I think I still own it

gptboss:
Hehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehe

Adam:
on Twitter. And my idea was to like live stream the whole first, like, like live tweet the whole first, you know, a couple of tweets a week, the whole first year of that career in the NBA.

gptboss:
You should revisit that now because

Adam:
One of

gptboss:
you

Adam:
my

gptboss:
have

Adam:
first, yeah, one of my

gptboss:
the

Adam:
first

gptboss:
AI

Adam:
tests,

gptboss:
to like do the live photos.

Adam:
one of my first tests for chat GPT was just to pitch that idea in a paragraph and have it do it. And I told it like include famous players currently and you know, hashtag it came out perfect. I was like, Oh, I don't have time right now. I'm going to do daily podcasts and stuff, but

gptboss:
Yeah, too

Adam:
we

gptboss:
real.

Adam:
could publish this.

gptboss:
Yeah, totally.

Adam:
Anyway, let's jump in.

gptboss:
Yeah, so what's up first? The watermarks. This is good news. We were talking about this yesterday. I'm glad

Adam:
Yeah.

gptboss:
that the story is continuing and stuff's happening.

Adam:
Yeah, the first link from PC World. Microsoft will ID its AI art with a hidden watermark. It's a way to ease the world into AI art without threatening artists. Uh, yeah, the fact that, you know, big companies are paying attention to this quickly makes sense. I was just in a conversation where we were talking about the stable diffusion, Getty, you know, fight and, and honestly, probably Getty wins. So we're going to need for the big players to start coming up with, you know, sort of intellectual property regimes that function way more like. Uh, you know, YouTube's ability to flag things and allocate the funds accordingly, instead of playing a game of, of whack a mole with trying to say you're not allowed to use that, take it down.

gptboss:
I don't think this is necessarily about the creative, like a new, a new moneymaking method dropped yesterday, Kurt. Um, somebody generated

Adam:
Thank you.

gptboss:
an AI image of the Pentagon having been attacked, right? Like it like exploded, big plume of smoke, debris everywhere. And then it was going around on Twitter and people were like live tweeting, Oh, this event like Pentagon's under attack. And then it got picked up by Bloomberg who couldn't fact check it because there's no watermark.

Adam:
Right.

gptboss:
And so when Bloomberg tweeted it, the S and P dropped like 3%. So new money making method for those of you at home. Open a short, generate some cataclysmic AI image.

Adam:
Right. Which is, this is not advice to be clear, but like this is a game that, that, that, uh, Elon Musk plays and gets in trouble for on Twitter

gptboss:
Yeah.

Adam:
already when he talks about, like he's literally been instructed by the sec not to tweet about stock price, got in trouble for it. Um, but like,

gptboss:
So then he tweets

Adam:
yeah,

gptboss:
about

Adam:
that's

gptboss:
NFTs

Adam:
so it's,

gptboss:
and gives

Adam:
so there's

gptboss:
me

Adam:
the,

gptboss:
$10,000.

Adam:
so these, the threatening artists part. Which is an okay reason to say that it needs to exist, but there's also a really important part of this, which is that like, this could become a terrible misinformation machine, which

gptboss:
That's

Adam:
have

gptboss:
what

Adam:
historically,

gptboss:
I think.

Adam:
like I joke about civil wars and stuff. World War II happened because radio. And World War I happened because printing press and industrial revolution. Like these are really serious problems. So anyway.

gptboss:
We don't need to get into why World War I started, but we should talk

Adam:
We're

gptboss:
about

Adam:
happy

gptboss:
it later.

Adam:
that everyone is working on the

gptboss:
Yeah, yeah, that's

Adam:
protocols.

gptboss:
like the biggest regulatory concern is the misinformation that can come out of this. So until we get these watermarks like standardized to something that are easily accessible, hang on to your tinfoil hats, right? Because most of the things that you'll see won't be true.

Adam:
Yeah. So to pull a quote from here, the coalition for content, provenance, and authority, C2PA, began work in 2021 to develop an open standard for indicating the origin of digital images and whether they were authentic or AI generated. So really what this is, is they're committing to a standard that Adobe is also on board with and a bunch of other major players to make it easy to check against an Oracle, whether or not the media that you're using in your news articles and things like that is fake or not. Um, which is hypercritical for the fifth estate to continue to function in a, you know, modern economy. So

gptboss:
Mm-hmm.

Adam:
it's happening. This part of it looks a lot more like the internet, right? Hey, we need a governing body. There's

gptboss:
Kinda.

Adam:
13 competing standards. We should, we should have a uniting standard.

gptboss:
Yeah,

Adam:
Now there's

gptboss:
another

Adam:

gptboss:

Adam:
competing

gptboss:
competing

Adam:
standards.

gptboss:
standards. Yeah, literally. One thing, one like side effect that may be difficult to like determine here is that this this does kind of like disparage citizen journalism. Because if your media isn't accepted by the Oracle, right, like Even if it is a real photo and the Oracle just says no, sorry, that's not real, then the truth is not true anymore. So we got to figure out some balance on that. This is a check and then we need a balance. So I'm looking forward to see this story continue to develop.

Adam:
Which is actually why it's a good transition to the next slide where the White House has continuing to talk about AI risks from Reuters, White House takes new steps to study AI risks, determine impact on workers. This is largely about the first wave sort of listening tour part of doing a giant democracy. Um, but it shows, you know, quick action on the back of. what we talked about a couple of weeks ago, which was the allocation of funds and the establishment of research institutes and stuff like that. Um, this is largely about them starting to roll out the plan to talk to the workers, talk to the experts, try to understand where this stuff is going to. Yeah. Impact the population and start to figure out what, you know, executive and legislative branches can do about it. Um, again, and, and, and I think it's a good transition from what we were just talking about because like you just said, uh, what you're talking about when you Oracle up is gatekeepers

gptboss:
Mm-hmm.

Adam:
and there's some version of an answer for something, some of these services where there have to be public, public watch dogs and information policies where the people can, you know, can, who watches the watchers. situation, um, especially if the watchers are, you know, they're public companies. So some, to some degree it's public, right. Uh, but also, you know, we ended up talking about auditing bodies and all kinds of stuff like that. So.

gptboss:
Yeah. If I don't know, I'm not going to make any recommendations here, but

Adam:
Ha ha ha ha ha.

gptboss:
there's, there's trades that I would like to make and trades that I would not like to make. But, um, I think, yeah, this is a really good first step. Something else interesting from this article is that they specifically highlight, uh, talking to gig work experts. So they see automation kind of impacting the least like reliably employed Americans, which as far as I know, it's like a very, very high portion of the workforce. Now, people are typically over employed in like three gig work jobs that comprise a full job. Various types of freelancing and stuff like that. So the faster these like small, simple things get automated, like a target for automation would be. I guess you guys, I don't know if you have like skip the dishes, maybe like Uber Eats is a good parallel. If that suddenly all went to like drones, then that could be like a really significant GDP impact. So. it's interesting that they highlight that they're looking at that kind of thing.

Adam:
Right. Yeah, absolutely. Like this isn't just looking at like the HR department at big companies is going to be impacted,

gptboss:
Yeah,

Adam:
but it's

gptboss:
they're not

Adam:
it's.

gptboss:
calling up marketing and saying, tell me how you feel.

Adam:
It's also important that they're looking at, at, at honestly, the space where I live, although my version of it is like aggressively chasing this replacement because I see it happening and I'm a technologist, but like as a marketer, I'm also aware of the extent to which all of the contractors that I work with could be just using AI to do the thing that they do,

gptboss:
Mm-hmm.

Adam:
which is going to shake up how everything works on that side of stuff. And. freelancers are notoriously bad at pulling together to do union type things, collective bargaining, to have the conversation, have to have the conversations they need to have with the power that they need to have them. Right. That's why the government and democracy needs to exist because some percentage of the population can't get

gptboss:
Organize.

Adam:
it together to, yeah, not get it together. It's just like, you're busy with your job. You

gptboss:
Yeah.

Adam:
know, you're busy farming.

gptboss:
Yeah.

Adam:
I don't have time to think about union squabbles. You just team up. Okay, next up. Ooh yeah, girlfriend GPT. This

gptboss:
Oh no.

Adam:
one actually just linked to a hacker news thread because that's, that's, it's a really interesting place to just distill the conversation around it, both from a practicality standpoint, because it's tech people talking about like, Hey, does this actually work? But also very thoughtful conversation on hacker news, generally, and people talking about, but also I like talking to the AIs, um, specifically, this talks about open source models. So the link from hacker news is actually just to a GitHub repo, uh, with girlfriend GPT there, if you want to, you know, clone it and play around. Um, But it's a very interesting, you know, there's, there's the The extent to which this is an unstoppable thing, right? Like you might, you like, like you might have an impulse to say, Oh, yeah, that's, that's weird. Got to shut that down. Yeah. It's open source software. The people that want to AI girlfriend up and have the power will be able to do it. Um, and it's going to mean a thing for traditional, you know, society as we have known it before this is as a possibility existed. But I don't know, thoughts?

gptboss:
You don't want to know my thoughts on this one.

Adam:
Hahaha

gptboss:
I have,

Adam:
Ugh.

gptboss:
yeah, I have many thoughts, but, but in some, I, uh, I made a tech talk earlier on this story. I front ran it. Sorry. Um, if you know, like in, inside of the comments here, they were talking about Twitch streamer Amaranth who's known for being ridiculously good looking, uh, she launched an Amaranth GPT, right? So she fine tuned this girlfriend GPT model to respond like her and likely, I didn't read the full thing, but likely use 11 labs to clone her voice and send like voice memos. Um, And she charges a dollar a minute across 100,000 fans is, you know, she's getting like a house a week, something like that. So in the tech talk, I said, if you, you know, are this kind of content creator and have this kind of following DM me because I'll do this as client work. I don't see it as particularly destructive and I see it as an interesting exercise.

Adam:
Uh, yeah, if I'm completely honest, that's why I introduced the inevitability piece. Cause that's

gptboss:
Mm-hmm.

Adam:
the wedge to talk to some, like, even if you, even if you, if, even if you're listening to this right now and the idea is repulsive, you're like, oh, well it's going to happen. So we have to talk about what it means for society, especially in a world where you've already got the surgeon general in the U S declaring loneliness, like a legitimate mental health crisis across the world, like downstream from the internet and a global pandemic and whatever. Like. There are things to deal with here. Uh, you also have the, like you said, the growth of video games, which is where that's the streamer, you know, sort of like, there's a completely different media ecosystem than anyone who is only paying attention to like mainstream cable and YouTube and whatever

gptboss:
Mm-hmm.

Adam:
that is like people playing video games together. Which is, which is a version of metaverse and a version of VR. Um, and then there's the question of like, you can't stop this. So let's jump to, to your stance, which is nothing wrong with it. Right.

gptboss:
Mm, hehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehe

Adam:
Um, I don't want to put words in your mouth, but like, it doesn't sound like you have moral problem with it, right.

gptboss:
I know

Adam:
Which, which

gptboss:
I

Adam:
I

gptboss:
do.

Adam:
don't really either.

gptboss:
It's like, it's just, it's

Adam:
And

gptboss:
just

Adam:
so.

gptboss:
like regulatory arbitrage, right? Like this is something that we should set up now and then we get grandfathered in and then eventually a ban hammer comes down, but everything, you know, it's already, you're already creating

Adam:
Right.

gptboss:
jobs with it if it already exists. So they're not going to like force your business to shut down. So you don't have like your admin staff and stuff like, you know

Adam:
Yeah.

gptboss:
what I mean? Like, it's just, we got to like,

Adam:
Build

gptboss:
let's,

Adam:
it for you.

gptboss:
let's do

Adam:
And then,

gptboss:
it.

Adam:
right. And then from the standpoint of like, look, if it's a legitimate, you know, time will tell in terms of like, is it a legitimate force for helping against loneliness and like a lot of what we've seen so far, you know, bots I've been in the space for a little while because I was, I've, I've one of my prior startups was in the telehealth space and it was about connecting therapists with people. It caused us to be pretty deep in the weeds on research around things like avatars in, in assisted living facilities and stuff. Like we already have proof that robots can make people feel less lonely and more sad. And that's from 10 years ago. That's before they could have a conversation with you that was coherent. Right. So there really is, is legitimate. I have a friend who's, who's partners PhD research was in this kind of stuff as well. Like aid aid robots and things like that, like are already a legitimate thing, right? So you call it girlfriend GPT and everybody, you get a good headline. So people listen to your podcast. Thanks everybody. But like, but, but it's a very real possible solution for agents that can act and, you know, feel like they're human.

gptboss:
I just, I generally have an issue with like parasocial relationships, um, supplanting interpersonal relationships and like in my personal life and, and like a society level in general, even though I benefit from it, there's

Adam:
Right.

gptboss:
like, I have 40,000 followers, right? That's 40,000 parasocial relationships. Um,

Adam:
We've only met in this weird metaverse space that is

gptboss:
Yeah,

Adam:
assorted video chats.

gptboss:
totally. Um, so, but the supplantation is what like, like I have friends in real life too, right? Like I could go to the

Adam:
Right.

gptboss:
bar a block away and ask somebody about like how college is going. And, um, this kind of thing, like moves you away from, like, I just, I'm, I have

Adam:
Absolutely.

gptboss:
concerns about like the future of the species with technology like this, but it's just, it's kind of one of those things. Like, like, what, what am I going to do? Am I going to be like, no bad dog? Like you can't talk to the girlfriend GPT. That doesn't seem like something that you should do with an adult human.

Adam:
Yeah, it's, it's, it's, there's two parts to the loneliness piece, right? Which is the cultural, like teaching the cultural importance of community as well, so that people still go do that stuff. So people have the thought when they're feeling sad and lonely, like, Oh, maybe I need to go, you know, I don't know, join

gptboss:
Go,

Adam:
a CrossFit

gptboss:
Jim.

Adam:
gym

gptboss:
Yeah,

Adam:
or

gptboss:
go Jim.

Adam:
whatever, right?

gptboss:
Go play D&D.

Adam:
Yeah. But also not against girlfriend GPT and check this out. Good, good conversation on hacker news. about all the things that we were just jamming on. And that's it for links today. So we're onto our recipe, which this time is a request for feature, bad audio. This one actually landed in a Slack channel. I'm in where somebody, somebody on the marketing team said, uh, we did a virtual presentation. Everything turned out great, except the person who's presenting audio is intermittent. Is there a way to fix this? Uh, and the answer is totally. There are already some services that'll do it for you, but I'm interested to talk to you back about just sort of how you, how you, you, you would build that if you were just gonna AI it up, writing some scripts, right? The idea is what they said, they do have a good transcript. So however they produce the transcript, apart from what happened, they have clean transcript. They

gptboss:
Mm-hmm.

Adam:
just have bad audio to put on the video version. So they're like, we have these spots of 30 seconds where we want to see if we can clone the speaker's voice to

gptboss:
Mm-hmm.

Adam:
reproduce. to fill in the gaps.

gptboss:
Yeah, so transformers can do this. Like you wouldn't use like a GPT, but you would use like an audio generator. And there's a few like existing services out there, like 11 Labs. I actually have a contact that's been working on a new prompting language for audio models so that you can include like pauses, inflections, and tonal variation, which also have semantic meaning. There's a difference between like, you know, and you know. Right? So he's developed like a markup to differentiate those two utterances that he can pipe to a custom machine learning model, but I don't know if it's public yet. But if somebody came to me for this product, I would go to him and then see if there's a way to transcribe the bad audio to get like a pattern of the inflection and then use another transformer to continue that inflection so that the fill is good. Because one of the problems with voice cloning, is that the inflection is kind of like random and all over the place. It does produce this uncanny effect of like, why did you say that like that? Right? Like that doesn't sound quite like you. So with existing tech, this is something that's totally possible just to like plug in the gap. But I think there's a little, and I don't know if this is like my own curse of knowledge because I know how these systems work and I'm able to detect it more. And other people just don't know. Like if you don't know what to look for, if you don't know what the color green looks like, how would you ever... tell if something's green, you know what I mean? Like if you're red, green, cloudy, or blind, how can you tell? How can you know what you don't know? So maybe the existing tools are good enough, they're not good enough for me, and then that was my description

Adam:
Yeah.

gptboss:
on how to continue this. You need to improve the transformer by adding additional context to the prompts.

Adam:
I've been testing 11 labs with like people around me. Like I have a recording where I, you know, I ended up, I'm just hype and talking about AI things. And then I go, so like, listen to this, listen to this. And they usually go, it's not you. I'm like, yeah, but imagine there's background noise and you're hearing it through a crappy cell phone. And you know, like it's passable in some context and crappy earbuds, you know, but you're right. It's, it's. It is fair though, to say that we've reached the uncanny valley. Like we're at a point of advancement where the best you've got is, I don't know what to tell you, but vinyl sounds better. Yeah. Some people just can't hear it. So that may not, you know, may not. Like we're up against some kind of threshold in terms of our ability to look at what the output is, right. Um, especially mixed in with audio that is good, like, unless there's a big blip or a big contrast between the two. It probably works for the purpose of this sort of feature request, like what we could build already, but it seems to me though, that the broader thing in terms of the acceleration. aspect of it is the models you're talking about will continue to improve. Right. Uh, like you're talking about a markup language where people can ultimately. kick off the feedback loop of how do you assess whether this stuff is good or not? Which I've actually heard talked about in the chat GPT context, like people don't talk about the reinforcement learning phase of getting there. Like with audio, with video, there's still some layers of we need to do the part where people are listening to it going, ah, this still doesn't feel right. And here's the, you know, but to be able to do that quickly with Markdown, eventually you can teach an LLM how to do that. And then you have a fully automated feedback loop that's really going to get us to like vocal clones that are really, really real.

gptboss:
Yeah, and yeah, so it's just an example of AI improving itself. Um, there's a question here of like, is, is this speciesation or, or GPT is just GPT's right? Like, so using GPT-4 to write this mark down to pass to like an audio generator, like a refusion or whatever it is. Does GPT-4 know that refusion is the same kind of organism?

Adam:
Wait sorry, say that again?

gptboss:
So there's a question of like speciesation, right? So

Adam:
Uh huh.

gptboss:
like this is like a collaboration that we're seeing between like, we were talking about like ASI yesterday and you actually sent me a TikTok this morning about like an argument against like runaway ASI is that models would not want to improve success of models that they're doing reinforcement training on past a certain point. Sorry, there was kind of like a logical jump here.

Adam:
yet.

gptboss:
But so I was thinking about that TikTok, I was thinking about AI's self-improving. So that kind of came together and then I was wondering like so like the claim is that we're not gonna get to this like runaway level because GPT 3.5 is gonna like Be able to like detect if GPT 4 is kind of out of bounds a little bit faster than GPT 4 would and then GPT 4 is gonna detect if GPT 5 is out of bounds a little bit faster again so my question is like do If there was like some misalignment, do you think that these like system like GPT is for different tasks like a text to text transformer? Do you think it would consider if it was, you know, intelligent, do you think it would consider a text transformer to be the same type of thing as like a text to audio transformer? Do you think that they think that they're different or like, do you know what I mean? Or am I ascribing sentience in a place that's completely inappropriate? Do you get,

Adam:
It's...

gptboss:
do you get what I'm getting at though? Do you think that like, do you think?

Adam:
Yeah. Although I don't know if it's just like when you have a philosophy degree, everything looks like a nail. Like there's a data science version of what you're talking about, which is like, okay, in these systems, how are these things represented and how do they correlated so that you end up with this black box algorithm that does X. Um, the philosophy side is where I'm like, okay, put on your goggles. This is metaphysics. Like when you name something, something, you give it a label. They actually use the term tokens in this context too, which is hilarious, but like you give it a label and then there's the label and then there's the thing that that is and it's like, okay, we've gotten this weird place where we're talking about teaching these algorithms and going, okay, if the algorithm understands audio, is it actually understanding the sound though, is there just, or is it just that understanding the words that we apply to the sound and are they different?

gptboss:
Yeah.

Adam:
like.

gptboss:
Yeah. I guess my question was, if I, if I teach it to use an audio tool, does it. Like where does that tool live in its like mental model? Do you know what I mean? Do you think that like, it thinks that the tool is itself? Like that's just a part of like my output space

Adam:
Or the

gptboss:
or.

Adam:
tool is its ear, the way you would think of your ear.

gptboss:
Yeah, or

Adam:
Right.

gptboss:
well, like another thing that's been happening recently is like, you know, like the phantom phone vibrations, right? Like if your phone's not in your pocket, you still feel

Adam:
Yeah,

gptboss:
like attacks

Adam:
yeah.

gptboss:
come in on your leg. So is it like being absorbed into like the identity of these systems over time? Do you think that they would recognize it as a different thing is basically?

Adam:
I, you know, my answer is tune in to find out

gptboss:
Yeah, yeah. We'll

Adam:
in

gptboss:
see

Adam:
the

gptboss:
how

Adam:
meantime,

gptboss:
this develops.

Adam:
this recipe is a good example of, Hey, bumping into this problem at work. Hey, I can probably fix it for you where it used to be in the marketing side of production problem, where you would have to have a sinking feeling in your chest. As you went to your boss and went, I'm sorry, the audio is bad. Anyway, it's been Thursday, May 25th. Thanks everybody for joining us. I'm Adam.

gptboss:
Thank you. Thank you. Thank you. Thank you. Thank you.

Adam:
Come back tomorrow for another Accelerate Daily.