OpenAI's Will Depue on the Ghibli explosion, the Fiverr eval, and AI-powered personalized shopping
Apr 1, 2025 · Full transcript · This transcript is auto-generated and may contain errors.
Featuring Will Depue
let's see will how you doing how you guys doing yeah we're doing great uh congratulations on everything I mean everyone uh seems to be on a tear uh for those that don't know uh will created artificial intelligence yes that's right uh hand handbuilt he's on that he's an author of that paper artificial intelligence yeah how it's done invented the Transformer yeah yeah yeah the Transformer uh we appreciate you being here do you want to give a little bit of overview of like what you actually do day today I think that'd be interesting yeah yeah totally I mean um I joined open a like two years ago uh have like a startup background but um worked on like video gin for a while worked on Sora for about a year and then I work on like RL post training chbd stuff so cool uh how did you process the uh Studio gibl moment was that expected did and I really want to know did you how internally did you think that that images in chat GPT was going to go viral and you didn't know what the prompt would be or were you like there everyone's going to gibl I want to know I want to know where gibl is flying around for 48 hours ahead of ahead in slack give us some Secrets here yeah I mean I've honestly like I was very bullish on this I think we like definitely underprovision gpus I was surprised our team predicted like less traffic than we should have expected um yeah I think the thing that I was thinking about the whole time is that like I mean I was walking around SF and some of these like nail salons have like clearly doly 3 generated like image fronts or whatever that is right and I think it just clear to me that like pretty much every business in the world very soon is just going to use this to like you know make logos and you know we can do transparency we can do controllability I think like my favorite eval is the Fiverr eval just go on Fiverr and count how many of these things youve can automate and like this takes off like a lot of those things um and hopefully there's a lot of new things as well that this creates but I just think of like total utility to the world is just massive and yeah have you thought about why it was Studio gibl in like precisely because you could do Simpsons you could do South Park but there was something about the anime style of Studio gibl that was like it kept enough of the human in there it didn't just become a completely generic stick figure it was a magical experience you needed no you you needed to do nothing to prompt it just you could I I love that you could misspell gibli still get a perfect output great and then the output was just magical I would I would I would you know generate the same image with four different styles of Art and the gibli one every single time the yeah what was your take on it no yeah that sounds about right to me I don't think I predicted gibli specifically uh last time for the dolly 3 launch I made a bunch of Pepes and that was the thing that really went off but interesting yeah I I remember I mean for a while the the version one of uh image generation it felt very art station informed it felt very like sci-fi dystopia like it was really good at those but people weren't really I mean there were there there were certain like you know anime Styles but this one was just uh was just everywhere um what else can you talk about um like to me to me the most like the the thing that's most exciting for me is like uh the gibl ification of everything right it's like on a long enough time Horizon like everything is gibli right there just sort of abundant and like very free and uh if you thought even a year ago that would have this sort of like anybody could instantly basically for free get this sort of beautiful handrawn anime style artwork that historically and you know somebody else posted about this but you know there'd be like a 4C scene in Spirited Away that was like that took a year and a half to generate right and so like to take something that's just so time intensive so expensive and then make it free and abundant is so powerful and and to me that is that's why when when I think about everything that's happening in sort of politics and Tech drama and everything like that and then I and then I think about like we're moving towards like everything being G Leed it's like everything today feels like a a distraction is that sort of like how you guys like as a team stay focused within open AI which is just like hey there's going to be a lot of noise there's going to be like you know this Benchmark today that that release over here and it's like we're sort of moving towards this like future Target that's sort of a ramble but but I'm curious how you think about sort ofoc opening ey is extraordinarily focused on on like real progress I think you know um I mean like for so long like you know you see the diffusion models they can't spell and like it takes us like two years there's no Dolly update and then finally we come out with this and it's not for no reason right I think like um people who are focused on like really like what is what is pushing the frontier forward and like what's actually the next evolution of this technology um yeah I don't know I think there's a lot of potential here I think I'm actually not sure what the image gen team thinks about like what the next steps are here I think their view is mostly about like making this like truly useful like can we go from diffusions a fun to toy to like this is a truly useful tool in the world and like clearly it's already having that impact um yeah yeah I mean you said your your your backgrounds in startups can you talk a little bit about um just being a consumer tech company now and like the focus on product and some of the product decisions like uh images and chat GPT it's not a separate app like Sora that seems intentional um how are you thinking about product development what are the strategies how much feedback are you doing are you just testing stuff internally are there AB tests running is this secretly running in India six months before we get it here uh like how do you inform product decisions totally I mean I think it's hard to pre-launch models because they create so much noise but we really test them but I guess like a moved from being a research company to a deployment company over time and and we're still in the middle of that kind of you know we're doing both and and I think we're just doing a lot more deployment they ever had in the past I think people are very motivated around like how do you actually make this things useful how do you deploy them to the world um I think it's really cool to be in position where like we can launch new products that are completely unlike anything people use chbd today for and people are so receptive to them like it's so like you know I don't know I think it's hard to get product Market fit and get people to kind of try new products but to have this like kind of all-in-one AI thing where we can try deep research we throughout image gen and people are just excited to try it out and like see if it works for them it's just like so rare to find yeah are you excited one of the things that I felt like really worked with images in chat gbt was like I could go to my camera roll Click Share share it into chat type Studio G Style and boom it gets it and it takes the the it takes the the the pain of inspiration out of the process and I'm wondering if you're looking at ways to create those magical easy to generate moments I remember mid Journey was talking about this too where like part of why mid Journey worked was if you just give someone a text box they'll just be like cat and it'll be like okay that's a picture of a cat but if you show the mid journey and it's like the Discord and it's all these different ideas then they can spin off that um how do you think about giving the user really setting them up for success in the context of Sora yeah yeah I mean I think generally like it's just very difficult of a problem like I think one of the most biggest things we're thinking about right now which is like when you open chap you get dropped in a text box and like the person has no no idea how to use it um I think I mean we have a lot of views on how to do this I think personally the thing I'm most excited about is that like I think the if we're talking about like what is the like Dolly to image gen what is the like you know kind of whatever each evolution of of of like the next kind of order magnitude um I do think the the long-term thing that's extremely exciting is just like maximizing personalization super long context the model knows who you are the model can follow up being like hey by the way I've seen you doing this do you want to try a anime style image or you know like maybe it creat things for you it's proactive like I think there's a lot to do on like kind of yeah moving the agency away from you into the product as well yeah you call yourself a master of slop uh which is a great line but uh does slop even exist in 5 years when I when people were saying oh the timeline's all slop on the gibly moment I was like this is not slop everybody this is this is fantastic yeah in fact I would argue that the rage in politics was slop were free of slop it was the first day that the time was free of slop no but it but it but it seems like the trajectory that we're on it's like eventually like you have to try to intentionally prompt it and say like make me an image that's like the image gen outputs from 2022 and it's like yeah they've got six fingers and it's the hand is connected to the body I want vintage Dolly one give me the dolly one but like what happens like I I feel like we're in the slop the slop era feels like it's coming to an end you're not going to be you're not going to be an image model that's constantly producing slop that is getting funding going forward right you have like the new bar has been set depends what you mean by slop I feel like I went to the I went to the SF Moma recently and like half of it felt like slop and I think there like I mean half of it not to be clear but I think a friend of mine always quotes that like you know art is like the search for the periphery like you're kind of like constantly looking you know kind ofing things forward or whatever that is and you know I don't know like once things get played out they get boring like they get cringe they get SLO right like I don't know maybe the problem is that humans are SLO and not that the models are SLO right yeah that's a good point uh are you worried for college students today that have access to these tools and making sure they actually learn how to do things like writing which is as We Know associated with thinking you know if you can't write well then uh I I just you know the models are getting better it's less obvious that I'm seeing something that's out when I was in college if I had this tool I would have aced my studio gibl animation course yeah yeah and AP Bio AP Bio yeah which I did I'm curious like you know how how how do you think universities like and just curriculum teach learning yeah I'm extremely worried I think the world just gets a lot more extreme and I think this is just true with AI across the board where like I think the bottom percentile of students are just going to get a lot they're going to fall behind and I think the top percentile of students are just going to get like extremely overpowered um I think if I had this tool as a kid I would just be like vastly smarter than I am now like even now today like I just talk to jbd every day I was like you know asking at random Concepts like learning random things I'm like went super down like deep down like a bio Rabbit Hole yesterday and like rare supplements I could take from the internet you know and like I don't know this is like I think it's like an incredible tool for learning but yeah I think it's like it's kind of bodal right I think like a lot of people are going to a lot dum I think a lot of people are get smarter um I don't know what to do about that but yeah fine tuned on GW how's your uh how's your experiment where you just text a number to post going oh no it's gonna make me like way more radical because then I just like send something I forget about it and then I check again you would you recommend it to someone fighting brain rot yeah totally Ian I think also like it I think the problem with tweeting nowadays is it's not not as fun once you have more than like 3,000 followers because in this risk and then you're like oh I got to care formed me or something and even just seeing the numbers and being like Oh like this one didn't do as well as the you know you know from the first 60 impressions of it's a banger or not and then you have this desire I had a I had a post last night that like I like John and I both thought was like hilarious but like I I realized that like nobody like really got it you needed to know you needed to have like red Ben Thompson's like Friday article to like get the joke it this Tam was too small low Tam um but but anyways uh you got anything else John this fun congrats it was fun we you know for the record we we didn't plan this around the the fundraising news uh we just wanted to chat so congrats to to you and the whole team on a on a cool Milestone and job's not finished so job's not finished make get back to work please make the model better you know make it better this was this was just you talking to customers yeah this is customer we enjoyed it anyway okay sounds good see you guys cheers great see see you talk to you soon yeah little little mini open AI day we got a couple folks coming in we got Aiden coming in talking open AI today we'll see if he can join right now and I'm working on a surprise guest but I don't know if it's going to come together he hasn't read my messages yet but I'm hoping it happens we'll see cross double double message we got Aiden coming in we got a guest coming in how are you guys good to see you both we're great how are you good go ahead go to be on I love this show are you are you celebrating today and if so what uh over