In this clip, Andrew asked ChatGPt to write a story about Marques. It’s a very impressive use of AI, but then they discuss the other not-so-impressive aspects of it as well which turns into a broader conversation about the state of using AI for creative reasons.
Watch full episode:
Shop the merch:
Twitters:
Waveform:
Marques:
Andrew:
Adam:
Ellis:
Instagram:
Join the Discord:
Music by 20syl:
Waveform is part of the Vox Media Podcast Network.
Foreign Something okay when I first started Seeing these I was very unimpressed okay Because I felt like it was basically Taking a few key words and then just Copying and pasting examples of popular Versions of that often from maybe just Wikipedia articles so I had people send Me like What is an armadillo or what is Ultimate Frisbee yeah and then it would just tell Me what ultimate frisbee is and it would Read exactly like the like the Wikipedia Article and I was like that’s not Impressive at all like I could ask Google that and it would tell me the Same thing So it really became much more impressive When the prompts got more creative I Think it’s similar to Dolly where like Dolly was impressive but clearly kind of Messed up but the more intense you got With the prompts yeah and how it still Was continued to be messed up Like errors here and there but it’s Still Could contextually figure out this two To three sentence long prompt and still Create something that made sense it’s The more so so there’s a a normal Accepted error rate let’s just call it Like 10 And when you just asked Dolly give me a Cat
And it gives you a cat but 10 of the Things look kind of weird you’re like Not impressed but when you ask Dolly Give me a cat with 12 legs in the style Of Monet On like canvas with like robots fighting In the background and it actually does 90 of that yeah it’s really impressive So I feel like it’s kind of along the Same lines here chat GTP I’ve had a Number of conversations with it I’ve Asked it lots of things I even tweeted What I asked for uh you know iOS versus Android essay I love Android versus iOS Essay and I at previous versions of that It got some things wrong like it was Telling me some things that weren’t Actually true but the fact that I was That detailed with my prompt and I was Like give me the style of the YouTube Video give me Android better than iOS Give me these things I was like pretty Impressed that it was actually able to Generate that much even though a lot of It still kind of does read like it’s Taking from the top line of an article It found or the second paragraph of the Wikipedia about Android it’s it’s sort Of throwing together a structure based On things that I asked for which is cool It’s blending them very well if it I Mean it pretty much is doing that I Guess the argument would be how much is It just copy and pasting from the exact
Things it finds or or how much is it Analyzing x amount of data and then just Through all of that data creating Something because if you search your Name Wikipedia is not going to be the Only one there’s going to be 400 pages That say like Marquez is an American Tech YouTuber with 16 million followers And then it’s going to be like oh 500 articles have said that that’s got To be important I want to take that and Put that in there some form of that in There Um but yeah the the way it Contextualizes different things and Actually make sure that it’s using them All correctly in like a story for the Most part you’re right there is still an Error there is still an error rate for Sure and some of them you read and You’re like that’s awesome and then I Think David was saying something about Fastest sea mammal it’s like oh that’s So cool and then looked it up it’s like Wait it’s not that like that’s not what It is yeah yeah and I also have like uh And then this will also I think be in The video there are some errors that are More Common than others like when you ask for These like AI paintings or portraits or Dolly your hands we’re gonna get into The the paintings in a minute the art I Guess can I do I I wrote I got made a
Story in Chad GPT that can I read this And then we’ll get into uh the more art Version of it all right Um in the holiday spirit I asked In the holiday spirit I asked chat GPT To write a story about how Tech YouTuber Marcus Brownlee had to take over the Role of Santa Claus Would you like to hear I would love to Hear what it came up with all right It was the night before Christmas and Marquez Brownlee better known as MKBHD On YouTube was feeling more than a Little stressed as a tech YouTuber he’s Used to filming and editing videos not Delivering presents to children all over The world but that’s exactly what he had To do thanks to a mix-up at the North Pole apparently Santa Claus had come Down with a terrible cold and was able To make the rounds and since Marquez has Been so good this year the elves had Chosen him to take over the job Marquez is a bit hesitant at first but He knew that it was his duty to help out So he quickly put on the red suit and Hopped into Santa’s sleigh which was Powered by the latest and greatest Technology The sleigh was equipped with a GPS System that showedmark has exactly where Children lived and it even had built-in Cameras so that he could capture all the Reactions as he delivered the presents
As Marquez flew from house to house he Couldn’t help but smile he loved seeing The excitement on the faces of the Children as they opened their gifts and Even got to try out some of the new tech Toys that he had helped design Before he knew it it was time to return To the North Pole Marquez was exhausted But he felt a sense of satisfaction Knowing that he helped to make Christmas Special for so many kids as he handed The reins back to Santa Marquez couldn’t Help but joke I think I might have to Retire from Tech and become a full-time Santa Santa just chuckled in reply I Think we might have to make you an elf Whoa he got burned at the end what that Was so good that was really good it Wasn’t really good what was the prompt Again give me that one more time it was Write a story about how Tech YouTuber Marquez Brownlee had to take over the Role of Santa Claus that was I kind of Missed the music already yeah that was Really good it had a punch line at the End it really So whoa yeah it’s a it’s a lot of fun to Play with I’m very interested to see Where this goes I can’t wait for like in Five years you just called this before This is kind of like the highest form of CH like conversational AI that we’re We’re seeing and in five years when we Look back on this and think about how
Terrible it was it’s going to be really Impressive to see where we go for real Okay now we should talk a little bit About the the visual versions The Visuals yes there’s been we talked we Did a whole video about Dolly we’ve Talked about Dolly it’s using you know Obviously diffusion and creating images From scratch using a natural language Text prompt which is sick uh there’s Another one uh that’s generating AI Avatars of people there’s a couple yeah And I think with this one basically you Feed it several images is that right so You give it some images of yourself yeah And it can spit out some cartoon version Of you with different situations it’s Very Dolly like except this time Focusing on humans although human that Are consenting by giving you their Pictures Um Prisma Labs did one for you and you Have a million photos of yourself online And it did a very very good job like Yours are insanely impressive and I Think what it’s doing is is similar to Dolly where it knows what a general Person looks like but the really cool Thing is that it confines the Stylization of other things and make you Into it like this is you it looks like You’re in a Chevron hoodie astronaut Suit Um and like cool more like Miami Vice uh
Here’s you in a purple suit which is Really awesome yeah um a lot of it is Really cool and yours are super Impressive you’re pretty good uh one Thing I’ve noticed that it never really Does well just like Dolly uh hands and Text so anytime there’s supposed to be Texting it like the one with the Astronaut suit supposed to have some Text on it right there doesn’t really Make a real word if you ask Dolly to Make a stop sign seven out of ten times Yeah some weird word and then uh I think If you there’s some other other YouTube Friends that have done this with hands But even look at the one with me in a Purple suit like that those hands don’t Quite match up there’s like a floating Thumb on the top of my hands there that Isn’t connected to either hand that just Happens when you have hands and photos With AI Um so hands are tough other than that Pretty good yeah yours yours is very Good Sarah dichi did one with Avatar AI That was incredible I mean but again Then there’s like Justine who did one on I forget if it was prism or Avatar hers Was bad yeah and she has a lot of really Good pictures out there and her face Looks nothing like her and some some of Them her eye is like almost off to the Side like off the side of her head it Seems to really like some people and
Really dislike some people myself who Doesn’t have a lot of good photos online Tried Avatar AI it is rough really I’m Gonna show you something I disagree These are fantastic these are I want to See these so I think these are really Bad to be fair the photos I fed into it Were not great although it still took About 30 photos Um and If you would like oh yeah Where are they yours as well there’s a Link in the dock oh it’s okay I’m Clicking on it now Some of the so like Some of these are very good some of These are very bad I also totally don’t Approve of all the categories that put Me in Um dominatrix as the first cat there’s Also like gym and for whatever reason Sunflower field none of them look like Me I look like Dr house and some of them Which I’m not opposed to like it’s funny So Andrew you did not tell me that you Starred and a mirac I don’t think we can show the gym one uh The gym ones are incredible they don’t So there’s a varying amount of Resemblance to your face because you can Tell it’s picking which photo it’s gonna Use and trying to make it look like you The Picasso one looks less like a Photo realistic version I would argue
The ones that look the most like me are Just the most standard ones that felt Like it’s pretty much just pasting the Photo that I put in there it like just Kind of looks the race car driver one You can just tell in specific ones it is Just my face on Michael Schumacher’s Face it’s a fashion model one of you in A turtleneck is very good this is this Is just a young Michael Schumacher with Kind of my facial details yeah so like You can tell where it’s pulling from on Some of these things and Um I don’t um this might be one of those Things where you should watch the clip Later if you can’t see all of it because It’s a little hard to explain I think I Think the thing is is it’s it’s nailing Very normal photos of me the fashion Model one the one of just like in a Business suit or the passport versions Of these are all just kind of normal Where photos I fed into this a lot of Them were from weddings which I had a Suit on so it’s basically just making Kind of a drawing version this side Profile of me as a fashion model is like I posted a side profile picture that was Exactly like this yeah and it just feels Like the photo of me okay that’s a good This is a really good transition point Because there’s a tough question behind All of this which is How do you credit what it’s taking
Inspiration from yes and so for the for The ones that you gave it the pictures Of it’s it’s fairly straightforward you Gave it a certain set of pictures so you Know what it’s drawing inspiration from And if you ever see any similarities It’s pretty obvious where it got that Stuff I think my whole thing is actually Pretty obvious in a lot of theirs as Where the ones that nailed it it’s Mostly taking it from the pictures I fed It the ones that didn’t nail it you can Kind of tell the photos it’s grabbing From yeah and that’s like you said it’s A it’s a discussion that I think we need To have because it’s a discussion on Twitter now is What part of where it’s scraping from is Okay I guess and that is a tough Question so you mentioned that other one Is clearly a picture of young Michael Schumacher it’s definitely just Michael Schumacher that photo where did that Come from exactly and we don’t know the Artist or the photographer or any of That behind that image and I think that Comes back to even like so dolly dolly From open AI is taking from an unknown Giant database of like all of the Information that they fed into it and Then if I ask it for a picture of a cat In space or whatever it’ll give me a Picture of a cat in space and maybe it’s Taken some photos from artists about
Cats and some photos from NASA about Space But if I ask it for a photo of a cat in Space in the style of An MKBHD video it might be taking frames From an MKBHD video that it has in its Database and should it be able to do That yes and that is a tougher question I think that is a very tough question And I’m gonna preface this by saying We’re just gonna say our opinions very Quickly on this and we’re looking at it So I don’t want anyone to yeah I just Want to preface this we’re not copyright Lawyers but there’s also just like I don’t know it’s a tough tough subject Yeah and I think there’s a lot of Conversation to be had about it Um but there are a lot of artists out There who seem to be seeing this and Saying that a lot of things look far far Too similar to art that is not Consenting into the into being part of These AI databases and we just don’t Know how much is a coincidence versus How much is is it literally dry Inspiration it’s very very hard so the Way here the weight Dolly works is Really interesting and I you can sort of Go back and watch the video again and Hear the explanations and there’s other Explainers of this but it’s like it has All this information it has all this Database it uses diffusion it has an
Image of you can you can listen to David Explain this on a previous episode but It has just like a static image of just Like black and white grain and then it Has a final image and it tries to use Diffusion to like add structure to this Static image on its way to the thing That you described to learn what you’re Making so if you ask it for a picture I’m just gonna go with a cat again it Knows because it’s learned from its Database what a picture of a cat looks Like and then it tries to create a Picture of a cat from nothing Essentially And so it’s not just like taking a piece Of someone else’s art and pasting it Into this new creation and calling it a New creation that would be a lot more Obvious what it’s more doing is learning From all these different sources what it Thinks you are trying to ask it for and Then creating a new version of that from Scratch and sometimes the level of Variation from that new thing is Pretty Tough like there’s probably not that Many pictures let’s say of a a cat with A purple sky behind it or something so If you’re an artist that made a picture Of a cat with a purple sky behind it and Then I ask dolly for a picture of a cat With a purple sky behind it it’ll Probably resemble A little bit more of that artist’s photo
Because it doesn’t have as many examples To draw from if that makes any sense so It’s it’s like if I make more specific More nuanced art I am more likely to Find issue with my art being drawn from Rather than if I just made like General Pictures of cats you know it’s tough I Also think depending on which AI you’re Using there’s probably ones that are Doing a better job because I I have seen People post that like some of these Avatars they’re getting have Old artists signatures on them which Means it’s definitely taking a little More than just completely learning and It’s more melding things together which Feels yeah if it literally has a Signature then it’s definitely not it’s Using the fusion thanks so much for Watching this clip we’re we’re gonna do Something kind of fun give you a little Sneak behind the scenes here um we Normally shoot thumbnails afterwards to Make it easier for Tim and the clips Outro is just going to be us shooting Thumbnails just gonna see how we do it Tim doesn’t have to use AI to cut us out Because we give ourselves a white Background yep so we do this white Background and then Adam is gonna bring Up a couple points that we talked about Today and Marquez has to try and make a Face about it yeah so we spoke about Knockoff watch bands
Foreign What’s his name Sam Sam Friedman oh Whatever exactly that guy shows up at Your house tonight He’s with people Now you’re interested Uh EV stick shift Nailed it perfect chef’s kiss yeah all Right now I really do think the EV stick shift is A face pump thumbnail yeah for sure Oh yeah All right Andrew Um There’s a AI story about you robbing Santa Claus You’re okay with that all right cool Okay how many gifts Wow good question it sounds like I’m Doing good uh turns out that your AI is Actually a super villain not a superhero Just because of the ones that showed up You you’re wearing like a black Turtleneck you look super creepy that’s True that’s my favorite one the black Turtleneck one is fantastic all right I Think we’re good Like And subscribe if you’re watching The clip like And subscribe Thank you
