Ian Silber - What it's like designing at OpenAI

00:00:00I think some people might assume we'relike 2 years ahead thinking like that,but we're running very closely [music]with where all of these advancements aregoing. And so that's just like a verydifferent way of working. Things arechanging underneath your your feet allday long. And it's very exciting. It'sreally fun to be like, I don't know,we're we're going to figure this out aswe go. We're going to like we're goingto try it. We're going to turn thecrank. We're going to keep iterating.We're going to keep going. Welcome toDive Club. My name is Reid, and this iswhere designers never stop learning.Today's episode is with OpenAI's head ofproduct design, Ian Silver. [music] So,we're going to go deep into all of theways that they work, what it's likedesigning with AI as a material, Ian'sthoughts on how the role of designer isevolving, and a lot more. But before weget into all of that, I had to know howthe heck does somebody become the headof product design at OpenAI? I had beenat Instagram for like 8 years. [music]It was amazing. I mean, most everythingthat the big stuff that we'd worked on,I was able to be involved in in someway. Kevin and Mikey were starting a newcompany. I went and joined them.>> [music]>> It was Artifact. It was like this sortof news kind of AI thing. Ton of fun,great team, but I had some really goodfriends that had also left Instagramthat were like, "Hey, we're startingthis thing.

00:01:15We're kind of building agame. We don't totally know what we'redoing, but we'd love for you to comework with us on it, too." I don't know.I was at the point in my career where Iwas like, I should just try it. Somecrazy stuff. I'm not even like ahardcore gamer or I have zero backgroundin gaming. In that way, I was kind ofunqualified, but I mean, the team waslike that they were assembling was bothreally close friends, which is alwayslike fortunate if you ever get thatopportunity. But also some of the mosttalented people that I've had a chanceto work with in my career. Both fromdesign, but also like really greatengineers. I think we took a lot oflessons from Instagram, which is like alittle bit of a has some game typemechanics or behavior. So, we leaned alot into like the social side and andthat sort of thing. And it was a game,but it was it was a little bit ofeverything. The TLDR of what we builtwas Minecraft in the browser,essentially. But we were kind of tryingto combine Minecraft meets Roblox, whereRoblox is like anybody can create thesegames, and there's this like fullmarketplace. People are constantlycreating games, and then you know, youcan come [clears throat] and play them,and and people can actually like makemoney off making these games, and it'sthis whole sort of industry. But whenyou really go to make the games, it'slike pretty comp like it's not the mostapproachable experience. But we lovedhow approachable Minecraft was, whereyou're just like one-to-one buildingthings. And so, we wanted to kind ofcombine

00:02:30that together and say, "What ifyou could let anybody come into thisworld and build, but also let otherpeople participate in that?" And webuilt a lot of different mechanics intothe game, but one of the things that Ireally liked was the idea that anybodycould kind of create their own sort ofversion of it, and and then let otherpeople play that. It was ton of fun.Working on a game is very different. Onething that was interesting is like it'skind of hard to know like everything's agood idea in some ways. You feel like,"Oh, that that would probably be fun."And so, like the amount of scope you caneasily get into is quite wide.>> Yeah. And we realized why like athese AAA studios take 8 years to to torelease a game. Unless you're making,you know, a small mobile game. Like it'seasy to it's easy to kind of see howsprawling it can really be. So then,what's the story? Like how did you getto OpenAI? So, first of all, this waswhen ChatGPTor GPT-4, I think it kind of come outwhen we'relike a year into the the startup. Ourfounder went away and came back, andhe's like, "Guys, this is this ischanging everything. We need to likethink about what this means foreverything." It wasn't any specificdirection, but he was just like, "Thisis like the real deal." He studied AI inin college and

00:03:45has a background in allthat stuff, and he was like, "Yeah, thisis like a true kind of different pointin time now." We were using it to makethe game. And this was like so funny,but way back in the day, this was like Iguess we hadGitHub Copilot, so you could kind of dosome auto-complete stuff. And a lot ofcopying and pasting between ChatGPT andand and back into VS Code or whatever. Iwas doing a lot of front end at thetime. It was incredible, but it's funnyto think how far the work has come sincethen.>> like stone age, and it was only 2 yearsago. Stone age. Like I was basicallywriting all this front end, you know, byhand. And basically what happened was Imean, the true story is that me and thedesigner were both sort of had beenreached out to by the head of product atOpenAI at the time, who we had bothworked with in the past. He had recentlyjoined OpenAI. He was leading ChatGPT,and he was like, "Look,this thing's going great. We have agreat team, but we need to grow itreally fast, and we need people that canhelp us figure out how to like get tothis next scale." He didn't actuallyrealize that we were working togetherwhen he was kind of trying to sort ofhire both of us. And so, we talked, andhe also found out about the engineeringteam that we had, and one thing led toanother, we all ended up joining as ateam. There were eight of us. It was apretty funny like time where we wentfrom like talking about some randomspecific game mechanic on like a

00:05:00Friday,and then like that next Monday, youknow, we're we're like deep in theOpenAI trenches trying to figure outwhat we're launching next and and howthis whole place works. So.>> [music]>> Real quick message, and then we can jumpback into it. If you're like me, thenyou're prototyping a lot in code lately.But the problem is you kind of have thischoice between [music] a tool that isn'thooked up to your actual code base anddesign system, or a local host prototypethat's super annoying to share. That'swhy I love what Descript is doing. Inone click, you and everyone on your teamcan prototype directly in your code basewithout ever opening an IDE. Descript[music] extracts your design languageand gives you the perfect sandbox toexplore without any of the technicalhurdles. And when you're ready, there'sa nice little share button top right,and you can send it to anybody on yourteam. It's a pretty big deal, and youcan connect your code base and start[music] prototyping today. Just head todive.club/descript.That's [music] d e s s n. Never in amillion years did I think my designworkflow would change so drastically inthe last few months. And a big part ofit is Paper's new [music] snapshot tool.It's a Chrome extension that lets youcopy any component or element from yourlive website, and then paste it directlyinto Paper as editable layers.

00:06:15So, allthe time I'll grab something from prod,have Claude immediately spin it up sixdifferent ways on Paper's canvas. Andthen when I'm ready to send a conceptback to Claude, it's seamless becausePaper's canvas uses real HTML and CSS.That workflow feels a lot like thefuture of design to me, and you canstart doing this today. [music] Justhead to dive.club/paperto try it out.Now on to the episode. Before we get allinto the present day stuff, I want totalk a little bit about some of thosefirst few months even as you'reassimilating. Like what were some of thedefining characteristics that made theOpenAI trenches unique compared to pastexperience? From day one, it was just sosuch a different place. I remembervery distinctly my first all-hands. Youknow, they they have these all-hands,and they kind of show you, "Here's wherewe're at we're at today, andhere's what we're seeing is possiblewith, you know, the future. Like wherewhere where these models are going." Andit really just like sinks in like thethe rate of progress and what is comingfrom that. Moment I was like, "Wow, thisplace is different." I think the biggestdifference is working in a research-ledenvironment. OpenAI started as aresearch lab, and that DNA is very muchthere. It's very different than

00:07:30something like, you know, Instagram,which I was very familiar with, or maybea more kind of like company that startedas a consumer product. You know, OpenAIstarted as a research lab, and like justtrying to figure out what they're goingto build. And in ChatGPT launched aslike a low-key research preview. Andwhen you think about it through thatlens, it's just like a lot of differentthere's just a a different way that thethe entire company approaches it. Youknow, the company's super mission-drivenand really thinking about how we developreally far forward into the future. Canwe go a little deeper into what it lookslike to thrive as a designer when you'reworking closely with a research lab?Like how does that shift the practice ofdesign on a daily weekly basis? So muchof our work is figuring out what themodels are good at, and then trying towrap that in a product that people canunderstand and can use. We like ourdesigners to be really close to themodel side, and really thinking deeplyabout and playing with and and trying toand seeing where it breaks, seeing whereit falls down, understanding the kind ofmechanics we can use to tweak thebehavior. But a lot of it is really justhaving a ton of curiosity, I think, iswhat what it really comes down to. Isyou don't have to be like technical towork here, but I think you have to bereally curious and really interested tosee like how might somebody actually

00:08:45usethis? A lot of times you basically justget kind of handed this new thing, andwe have to figure out how we canproductize that, how we can like getthat into athing that people can now go to theircomputer or their phone and be like,"Wow, look at this thing it can do."Yeah, it's really just a lot of thinkingabout the model as the product. I thinkone thing we we try to spend more timeon is honestly like a lot of times it'slike, "What can we do this withoutpixels? Can we do this with tokens? Canwe do this with the model itself?" Andalways trying to push to to do more ofthat. What can happen directly in theconversation and and in the flow of ofhow you're using the experience versustrying to think about what new kind ofbespoke UI to do we need for this?There's obviously a balance, and I thinkthere's, you know, sometimes where youreally want that, and sometimes whereyou don't. It's just a new tool or newkind ofmaterial that I think you have to thinkabout working with. I think for mostpeople listening, they probably don'thave as much of a grid for what it lookslike to design outside of the pixels.So, like how deep can we go on to that?Like maybe is there an example of someof the design decisions or explorationsthat you all are doing internally thatare happening more at the model level?I'll give you one one example that we'rethinking about right now which

00:10:01we're still just honestly experimentingwith. If you think about how youintroduce chat GPT to a new user,there's like this very traditionalonboarding flow that lots of productsdo. There's kind of this tour andaccount creation and maybe trying to askyou some questions to learn about you.That's been effective and people areused to that but you know, you couldthink about that totally different,right? You could say, well actually wehave this like super intelligent modelthat could probably do a much better jobtrying to understand what this person'sgoals are, what they're trying to learnor what you know, what questions theymay have, what functionality we mightwant to tell them about that will helpthem with whatever task they're tryingtocarry out. We're really like strippingback a lot of maybe kind of what youmight traditionally do and trying tosay, well actually what let's thinkabout like how we should give thiscontext to the model that this person isbrand new and they might need somehand-holding or they might need to learnabout a specific feature or or honestlyunderstand what this thing even isbecause it's so different than whatmaybe you've traditionally used beforeto you know, get information or to learnor whatever. And so how can we let themodel do the work versus trying to kindof write a bunch of static kind ofexplanation of what this thing is. Andso there

00:11:16you start to get into thinkingabout you know, the system prompt andyou know, the model behavior and andthat sort of thing which is just like atotally fun new way of working and withthat you can quickly prototype likehere's what it would typically output ifwe did nothing. Well, what if we youknow, gave it a different context orwhat if we tried telling it this or thator whatever and how does that actuallychange behavior? Does that make it morefriendly or easier to use or make helppeople understand better what you can dowith this thing. And so I think that'slike a one example where designersworking on this arehopefully spending a lot less time inFigma or whatever like tool you use todraw pixels and more time reallythinking about how you interact withthis thing and the fact that the modelreally is like the core product. Isthere like a set of guiding principlesthat help you think about when to reachfor interface level solutions and whento outsource it to the model because Iwould imagine there's probablya lot of potential for all kinds ofdifferent bespoke interfaces. I know youdid the learning experiences now likehow do you maintain like the simplicityof chat as an interaction paradigm whilealso you know, reaching for the fullpotential of what this experience canbe? That's a really good question. Wedon't have principles. We probablyshould.

00:12:32I think it's more I guess atthis point a little bit intuition.You're totally right that just text isdefinitely not the end-all and we'reseeing that, right? So we have all sortsof things that you can do with chat now,right? You can you ask it for somethingand it and if you're asking to writesomething instead of it just yeah, sureand then it gives you a bunch of text,you know, do you want me to changeanything and you ask and you get a newturn. What we did there was we looked atthat and we just thought critically. Wesaw that like one of our number one usecases for chat GPT and this was like a afun really fun design led exercise. It'sfun fun one to talk about. We looked atthe data of how people use chat GPT. Ahuge percentage of people use chat GPTto write to help them write. And thenwithin that there's like lots ofdifferent use cases for how you do that.There's a bunch of things that are justkind of funky with it. So for example, Imean there's the classic likeyou know, you see somebody copy theentire thing and they they paste thatand it's like an email to somebody andat the bottom there's like let me knowif you want this shorter or something,you know. [laughter] That's one thingbut but also just the interaction looplike right? If you want to just likechange one word, it was kind of tediousbecause you would then have to okay,change the second paragraph to say thisor that or whatever. And so we wanted tokind of lean into more directmanipulation and so with that

00:13:47now whenyou ask it to write and we're stillrolling this out so for some writingcases you'll see it others you won't butwe want to get this out to all writinguse cases but now you get this likecontainer where it puts what you werewriting into the container and you canstill use chat to just say, oh actuallymake this longer, make it shorter andand it'll that'll work kind of the thetraditional way but you can now likeselect a certain part of text. You candelete it. You can ask for a specificchange and it's just targeting that onething. And so it's just a way moreergonomic way of working. And so that'san example of where it's a mix of modelbehavior and working with the model tosay when should it show this, whenshould it not and then how do youmanipulate stuff directly in this inthis writing block? And and then thenext thing we think about too is likehow do we make sure we're building asystem so that anytime we add one ofthese kind of things that feels a littlebit different than like the pure text,how can we make that a system so thatwe're kind of building these differentbuilding blocks so that eventually themodel I think will be able to composeand kind of pull these together in indifferent like bespoke ways specific forwhat that user's task is at hand. Whatdo you think are some of the traits ofthe best systems thinkers at OpenAI?It's easy to kind of think about the oneuse case that you're working on rightnow. If you think about how people usechat GPT, it's very fluid, right?

00:15:02Youmight be like working on one thing andthen flowing into something else or onemoment you're asking a question aboutlike what to pack for a trip coming upand then next you're helping it write anemail to your boss and then you know,next thing you know, you're doing someresearch. And so people move betweenthese contexts in very different ways.And so I think the the best systemsthinkers are thinking not just abouttheir feature but how does this featurelike extend the system? One thing we'rewe're really trying to think a lot aboutis kind of what are the underlying sortof primitives that these products needin order to serve the the user's needsand so that if you build this once, it'sgoing to make every other thing thatwe're trying to do better. I think youcan see some of these examples that arestarting to crystallize around forexample like skills is like this newprimitive that I think these productsare starting to embrace. And the more wecan kind of like find these the sort ofdeepest abstraction that that kind ofencapsulates what it is you're trying todo so then we can use that for for otherfeatures or other tasks that the modelmight want to do. And eventually we wantto like have a system that not onlyhuman can understand but a model canreason about and know when and how touse this. I kind of want to talk aboutwhat gets shipped and how because you'rein many ways sitting on

00:16:17top of this likecauldron of research led innovation andall these new capabilities are poppingup. It's a very generalizable productthat is being used by literally everyonein the world.And so you can go in almost anydirection and so you have all of thisexperimentation that's happening. Talkto me a little bit about what it lookslike to effectively steward an idea as adesigner at OpenAI. How do you go fromraw experiment to actually gettingsomething like that direct manipulationor maybe the new learning experiencesout the door?So much of it starts with a prototype ordesign. It's not just a designer ofcourse like anybody has these greatideas, an engineer, a designer, a PM, aresearcher. But from a designer'sperspective as everybody knows like it'sbecome much easier to kind of build aworking version of something. The bestversions of these are like a designerwill have this idea and now with Codexor whatever tool you want to use, youcan build real versions of this thataren't just clickable prototypes but areactually like live model responses andand are working with the model. Youknow, so I think that opens the door andso for example like these math ones,that was a designer who was justthinking about like you know, if you askfor something and it gives you back likethis LaTeX like the math

00:17:33you know,expressions and all that, it seemspretty archaic. Like that is not the waythat you should be able to learn. Youshould be able to interact with thesethings. And so that was a designer justlike kind of observed that, put togethera version but then actually just liketried it with a bunch of things wherethey said, hey, you know, you get themodel and you tell the model if[clears throat] you would typicallyrespond with this, tryrespond with this now. I think it wasjust so clearly valuable when we sharedthat around that you know, the teamrallied around it and eventually likehelped harden it, helped expand the usecases and and and ship it. So I thinkthere's a lot of pockets like that.There's a lot of bottoms up just ingeneral. That that's another thingthat's like I think different at OpenAIis just tons of bottoms up stuff. And Ithink this is probably maybe the waysome of the industry is going becauseanybody can build an idea now. You canof course just like fire up Codex or orwhatever tool and ship it basically.That's something that we think a lotabout is like we want to empowereverybody to kind of have these ideasand then it becomes like how do you editand make sure that it all fits togetheras a system and we're shipping the rightthings. Looking at when you joined, youmentioned like you were copying andpasting from chat GPT and now everythingjust feels like shooting lightning boltsout of your fingertips in many ways. Howhas that changed the way that teamscollaborate,

00:18:48the way that ideas getpresented? How have you seen thepractice of design shift as a result ofthis new technology? So when I joined, Imean two and a half years ago, somedesigners were using Origami. Somedesigners were I think very technicaland were like really working with thewith the API. For example, they'd belike in in the playground tryingdifferent stuff. I think it wasn't asaccessible to designers that were morefamiliar with like traditional designtools. And then you fast forward today,obviously you now have and this reallyhappened in the last like few monthswhich is crazy to think. First you hadlike cursor and I think that was coolbut I think even that was like a littlebit inaccessible to to some some partsof the team. And now with Codex and justtheI think the future where these productsare going, you have these tools that canreally go and do real work for you andit's more about getting your idea outand into it and expressing it properly.There's a whole kind of effort goingaround to figure out what tooling do weneed to make this like super effective.So for example, how How hook it up to adesign system so that it's not just likespitting out random UI but somethingthat feels like it would fit within oursystem. And so, you know, how do youlike make the connection betweenprototyping, production work, and Figmaand and all of this. But

00:20:03I think we'reseeing more and more of our designersembracing this. It's been really cool tosee like all of a sudden instead of aFigma prototypeorstatic thing or even like a recording ofa video, it's like very interactive. Ican go and and just tap around and I canplay with it. And And actually in manycases I can start asking questions or Ican like see what the the actual likemodel behavior would be. And it's alsonot just about design, which I think isone one thing that's kind of cool islike for example, we have an internalagent that we use which is like a datascientist. Any designer working onsomething can ask questions. Well, howdo people actually use this? Give me thedata on how this is used, how this isused, what are the use cases, whateveryour question might be. Now you can beway more informed with a kind of whatdecisions you're you're you're making.And when you're working at scale, that'slike really valuable. I love thatbecause it's it speaks to thislightening of reliance that we have on abunch of other people to be able to likeexplore and and put something out intothe world. Like everything you'redescribing is like, man, the amount oftimes I've been stuck in like, I don'tknow, Metalab or something trying toremember how to do a SQL query. It'slike I just want this answer, you know?I mean, it's going to be wild to thinkonce these things continue to developand everybody has these. Like I thinkyou just everyone will be will be ableto do so much more. And And the other

00:21:18thing like that is interesting to thinkabout is just understanding when to turnto what toolkit. I think we're stillkind of evolving as an industry, right?In some casesyou should definitely be coding a liveprototype or using a tool to code ato build a live prototype. I think inother cases a paper sketch is alsovaluable and right part of the process.And I I think thatthe discourse has been interesting. Ithink it's easy to swing one way or theother, but I think like sometimes youwant a whiteboard, sometimes you want aa paper sketch, sometimes you want awireframe, sometimes you want Figma, andsometimes you want these liveprototypes. And I think that the bestdesigners I think will understand andintuit when to go to to which one.Because sometimes you want to go superwide, you want to try a thousand ideasor a hundred ideas. And I don't think wewe have great tools yet to that help usdo that part, the the kind of wideexploration. I think that's one area I'minterested in seeingI think we'll we'll probably evolvequite a bit. Cuz you don't want to likejust take one prototype. And you canit's pretty easy to kind of like getobsessed with that and just be like, howdo we like, okay, I want to refine this.And your idea might be way off. I thinkthere's like this depth and breadththing that we need to figure out and andnot swing too far one way

00:22:34and just be Ithink as a designer you just need to bereally flexible understanding when to gowhere. Especially as a designer whodoesn't have a rich front end backgroundand all of a sudden you can make thisthing and it feels amazing and you'relike, this is the greatest thing I'veever made in my life because it's fullyfunctional, but it might be completelywrong. 100% and like, you know, Istarted as a hybrid designer engineer. Imean, honestly like dating myself, Ithink most designers code like we'rewriting code the front end code at thetime. You know, I was like really intothe 37signals crew at the time and thatkind of stuff where you're just likereally close to building this liveexperience. I think most a lot of peoplewere like much closer to code and werekind of using it as a material. You werekind of like a lot of designer engineerhybrids. Um you of course you hadspecialists and like iconographers andvisual designers and and that sort ofthing, but like the traditional kind oflike software design I think was likereallykind of like playing both roles. Andthen as the industry grew, I thinkpeople started to specialize and you hadthis new kind of like product design,which was definitely more on the justpure design and then we're going to havespecialist engineers that are going togo and build that. I think that likethat's now changing, it's kind ofswinging back where I think both canoverlap way

00:23:49more. And so there's just abig question around what skill set andwhere when do you work where and how doyou work? It's just like shifting quitea bit. Let's pull on that a little bitthen because as somebody in a leadershipposition, you're having to think alittle bit about like, where is thisgoing? How do we want to work in thefuture? If you do extrapolate, what doesthis kind of look like? So when youthink about the skills that become morevaluable, you mentioned curiosity. Isthere anything else that's top of mindfor you that you're looking for indesigners today? Yeah, I you know, insome ways that I think nothing haschanged as far as like what a greatdesigner is. Our team I think does a lotof this is just thinking really deeplyabout like, what problem are we tryingto solve? Who are we building this for?And let's try a bunch of different ideasand like, how do you kind of likeconverge on on an idea? And then ofcourse craft and taste and you know,understanding of interaction patternsand and all of that. Like none of thathas changed and I think is just asimportant. I think it's just all of thisunderstanding of working with somethingthat is changing very quickly that isgoing to evolve. It's not like set instone. It's like it's changing every daythat can kind of shapeshift and do manydifferent things. And then just likeyeah, you have these new tools and soyou can you can express your ideas verydifferently. Going back to my time atInstagram, which

00:25:04was part of Facebook, Ithink one thing that was but that reallydrew me to that to that company wastheir deep investment in in designtools. At the time when I joined it wasQuartz Composer. They had like all theselike custom kind of stuff they'd builton top of it that like Mike Matas andand and a few other people had kind oflike I think led the charge on. Theybuilt Origami. That was like BrandonWalkin. And that was like so cool to seethat there was this like new tool thatwhen a designer went from sketch to thenusing Origami, like you could totallysee their ideas come across much better.They had to think through not just howit looks but how does it work and howdoes it feel and how do you move betweenthese states? And I think it at the timeit was we were shifting from web tomobile and I think that was reallyimportant cuz so much of it was aboutlike how you interact with this thing.And so that tool I think did a reallygood job for that time that point intime. What I saw is the people thatembraced it I think were able to uplevel their their work and their craftand their impact as a designer. I seethat same shift happening now with likethese new tools that are emerging. Ithink Codex or Cursor or anything canreally unlock like a new way ofexpressing yourself for a a designer.[clears throat] Yeah, cuz it's not evenjust about the visuals, it's also abouthow it feels, but in your case it's alsolike the content

00:26:20itself is design, whichis really interesting. Like I've nevereven worked in a place where that is asmuch of the design as the corner radiusthat it sits in. Yeah, and it's sointeresting because a lot of us haveprobably worked in places where likeuser generated content is kind of whatpeople see. So like Instagramis kind of a shell, right? You're kindof like when you're designing, you'resort of designing the shell and then youdon't know what content's going to go inthere. You don't have control over that.But this is like something in betweenbecause we do have some control actuallyover what goes in, but not really. Like,you know, you don't know what theexactly the way the user is going tokind of use this or approach or ask. Butyou can you can play a little bit withhow the model should behave and how itshould respond. And so there's like it'ssomewhere in between this kind of likeyou have no control and you have somecontrol and you know, and so I don'tknow. It's just totally different.[music]One thing that Dive Club has madeabundantly clear to me over the lastyear is that the practice of design ischanging. And the old process of gettingfeedback just doesn't quite [music] cutit in today's world. That's why I'mexcited to announce that Inflight isofficially in open beta. It's thefeedback tool that I've always wantedand it's built for a world that moves atthe speed of AI. So I can share myprototypes, give context in video

00:27:35walkthroughs, and Inflight makes it easyto get the exact feedback that I need tomove forward, whether it's voting ondirections or maybe even getting thegreen light to ship a new idea. And allof this is available in a single linkthat I can [music] drop into Slack ormaybe even share with power users totest out a new prototype. I use Inflightevery day and it's totally [music]transformed the way that I share work.So I'm excited for you to try theproduct. And if you ever want to jamabout [music] it, just email me atrid@inflight.co.

00:28:06So you mentioned to me that the firsttwo years were kind of holding on fordear life. You're seeing absurd scale,like the level of scale that most peopledon't get to experience. Now that you'rekind of on the other end of this, likewhat are some of the more intentionalshifts that you're trying to make as adesign leader? I'm still holding on fordear life. Yeah, I figured.>> [laughter]>> But I'm trying to not just let it allkind of come at us as a design team andtry to figure out how we can be moreintentional about some of the thingsthat we want to work on. I mean,I think there is the classic stuff. Likefor example, we didn't really ever havea design systems team because we weremoving so fast and we're establishingthat now. I think what's cool about thatis we're really trying to think of itthough from first principles, how youprototype ideas and how does it workwith the model? So we have a wholesystem called the dynamic user interfacelibrary, which allows us to designthings that the model can theninterpret. And so that's like a new wayof thinking about this. And so I thinkthere's that whole thing. There's kindof figuring out like, what are thesystems and tools that you need as adesigner so that you can just like hitthe ground running and and and do greatwork in this new way of working. Andthen of course there's what is ourprocess? Honestly, like it changes

00:29:22day-to-day. Sometimes we're off in Figmaland and sometimes we're off like, youknow, prototyping things and we'refiring off things over Slack and andgiving feedback and moving reallyquickly. But I think we're trying tofigure out a slightly more intentionalwhile keeping ourselves honest that yougot to just be willing to just like rollup your sleeves and and try thingsreally fast and Can we add a little bitof clarity there? Like let us be a flyon the wall for a week as a designer atOpen AI. Like how does it work?>> we have a lot of traditional ritualsthat of course we're trying we're alwaysfiguring out. A lot of designers willjust have an idea and explore it and wewe have a channel for example that wecall PD whip which is like designersjust work in progress and you just throwstuff in there. It's got to be aprototype or video or something. Like wetry to make it like something thatpeople can like really easily react to.That's just like a a place we've had fora really long time like an open place tojust kind of share stuff and just likeget ideas out and and riff. We do docrits and I think that they aregenerally valuable to like just get abunch of people to kind of like riff onideas and that's really about just kindof like building ideas. We're stillfiguring out do we do design reviews andif we do exactly how and then likeday-to-day I mean I guess it depends onwhat you're working on. We do have atraditional kind of structure where youhave like a PM that you're working withand an engineer

00:30:38and honestly a lot ofthis is like building and iterating. WeI think we we try to embrace thatprocess of like we're going to have tofeel this out. We definitely don't spendour time I think like trying to craftthe exact perfect solution until we knowthat it even works and so I think tryingto get to an early version that we canplay with in the product I think it'slike a really good milestone and theniterating from there and spending abunch of time talking about well like isthere another way we could do this? Havewe looked at how this fits with this andand connecting all the dots. Of coursenow like cuz when we ship stuff we haveto be thoughtful about is this somethingwe want to ship to all of our users andwill this extend the system nicely andif not are there other ways or otherthings that we can do to kind of likebuild out the system so that it all kindof comes together cohesively which issomething we're trying to do more of tobe totally honest. You've talked aboutthe importance of systems thinkingmultiple times now. So it's obviouslylike a core tenant of what it looks liketo design at Open AI. We talked aboutwhat it looks like when it's good andwhen it's working. Are there pitfallsthat you're trying to avoid where maybeit's like what's the opposite of goodsystems thinking look like? You know? Ithink what we're trying to balanceI would say is how do we put out thingsthat are experimental and early andgoing to change?

00:31:53That's like theresearch lab nature but then how do youdo that in a way that for the thingsthat truly matter feel cohesive. So Ithink it's more about just understandingthe balance of like okay this is new wedon't totally know versus like this issomething that we feel like we reallyneed to harden and we're going to getright. I guess what what is what is notgood systems thinking look like? I youknow honestly I think it's just ifyou're a little bit too blinders on andjust trying to say well how do we getthis thing out as quickly as possibleand not totally understanding that wehave a bunch of other things going onthat are actually pretty similar and ifwe all all came together and that'shonestly my job. That's why I thinkprobably why I think about this a lot islike my job is to kind of try to connectall this stuff and say well what if wedid less here actually? What if what ifthere's a way that like we actually didfewer of these things but we did it thisway maybe that pulls it all together. AsI think about the evolution of ChatGPTthat's something that we're trying tofigure out is just like how do wetighten up a lot of the new things thatwe're kind of experimenting with andturn it into something that feels likeyou understand when to go to which toolor how to expose what you should whatyou could do with this thing in a waythat is clear as we have new kind offeatures or new advancements like weunderstand that that could maybe fitinto this system all while recognizing

00:33:08that honestly we don't know because likethere could be something tomorrow thatcomes out that completely changes theway you might want to interact with thisthing. So you have to be super flexible.Can we talk about that through the lensof the dynamic interface library then?What are some of the design challengesor opportunities that you all arewrestling with there? How do you designthis system so that it can rendereverywhere natively? How do you figureout how to make it interactive? How doyou make sure that it's truly addingvalue to the experience and not justlike doing what was done before buttrying to think of like the AI kind ofAGI pill version of whatever you mightbe trying to kind of interact with orwork with. And then I think the bigthing looking forward is how do thesethings come together and and like howdoes the model eventually kind of havean understanding of it so that maybe itcan start to help us compose these withwithout having to have a designer handcraft all of these. We're not quitethere yet but I do think that that willbe possible very soon. So when you thinkabout these you're thinking about theselike individual components that kind ofstack up to a system that a designer canuse an engineer can use and the modelcan use and how does that all kind ofcome together? It begs some interestingquestions about what the future role ofa designer even is in that world whereyou're you're kind of loosening

00:34:24all gripon what the interface can be. I mean Ithink it's an important question. I dothink that as all these tools let peoplemake more things like you know you'regoing to need an editor. I think a goodexample like a good analogy of course islike before the camera if you wanted astill of somebody you had to sit downand paint that thing. You had to be ableto paint a portrait and not everybodycould do that. If you wanted to it wasyou know time consuming and expensive.Now all of a sudden anybody can take aphoto. There are still excellentphotographers and people that likemaster the craft and people that havetaste and then there's you know peoplethat use photography for their personallife and it doesn't matter like thecraft doesn't really matter there. It'sabout the memories and and moments andall of that. There's a similar way youcould think about that with the factthat now anybody can write software foranything. I think that just becauseanybody can do it doesn't necessarilymean it's going to be good or the rightthing to build or the right instinct. Ithink we will as designers or or anyfunction I think like your job's goingto be more and more about kind ofhelping edit and direct and curate. Idon't think like the the job of adesigner is going away anytime soon. Ithink that if anything I think we'll beable to do more but I think that thecore skills will still be just asimportant cuz I think people canrecognize good software from badsoftware right?

00:35:39And we all know whatthat feels like and it's not just aboutdoes it work but like how it works andhow it feels and is it solving the rightproblems for me. You've talked aboutthis tension between designing for likepresent capabilities or maybe the mostrecent thing that the models haveproduced versus you know the super AGIpill part of it. So like how do you dealwith that tension and what are some ofthe more futuristic things that arerattling around your brain right now?One thing we think a lot about is thecapability gap. The capability gap issaying that the models actually are nowat a point where they can do a lotright? If you look at Codex and what itcan do sometimes it's spending a lot oftokens it's taking a long time and thenwhat people use let's say ChatGPT-4today right? There's actually now a gapbetween what the models are actuallycapable of and that actually justhappened very recently. I think asdesigners we have to think a lot aboutwell how do you actually expose that ina way that people understand what theycan do with it and how do you give themthe tools so that they can go out and doreal meaningful work right in ChatGPT.And so that's I think that's somethingwe think about a ton. How do we shapethe product give you the tools so youcan really take advantage of everythingthat these models are able to do if theyreally sit and and spend the time to towork for you. And we just haven't done

00:36:54that yet right? Like you know there'sthere's limited amount of functionalitythat you can do inside of ChatGPT unlikeif you think about Codex Codex can goand like use your computer and spend abunch of time and use the differentskills and all of that. And again youalways have to remind yourself so muchhappened just in three months. What'sthe next three months going to looklike? What's the next like I also try toremind myself of the um you know we'reworking on this thing is three threeyears or whatever it is into itsexistence. When computers first came outand they were three or four years intotheir existence or five whatever youknow and then where they are now andjust each year how do you have thatadvancement and then more and morethings were able to be done. It's funnyside note but there's this really funpodcast 30 for 30 which is like thesports podcast.Yeah okay cool. So so 30 [clears throat]for 30's like the the show but they alsohave a podcast. Have you heard the oneabout Madden? No. Okay there's thisreally great episode about Maddenfootball the game. This is a pillar ofmy childhood by the way so I'm I'm edgeof my seat. I cannot wait to see how youloop this back in. I just remember thestory from that where they they'd madethis like game on NES I think or maybeit was even before NES I don't remember.They wanted Madden to be thespokesperson and Madden was super intoit but he looked at it and he was likethere's only

00:38:09eight people on the fieldor whatever five people on the fieldlike football has how many is it? 11? Idon't know. Mhm. Uh they're like they'rethey're not enough bits to do that. It'stechnically not possible to put22 players on the field. We can only put10 total and he was like come back to mewhen that's possible.>> [laughter]>> Wow. You know and then like a year ortwo later like these the chips advancedand then all of a sudden they were ableto make it. And that was like the verybasic version of Madden where theyfinally had the 11 players and then hewas like cool I'll endorse it like I Iwant to be part of this I love it andthen it's a really fun podcast cuz youhear him like in the booth recordinglike his sound bites and stuff likethat. But it was a cool story of justlike also thinking about how muchtechnology advances like just on thisregular basis and I mean we all knowthis as you're working with thesemodels. They're not perfect like you runinto these issues you know like why itfalls down in a certain case or you knowit's really good at this but it'sactually like really bad at some otherthing andand you just have to have faith that ayear from now or even less six months oryou know and and you extrapolate outwhere that's going. It's like kind ofhard to imagine but I think a lot aboutlike that and and what it must have beenlike to design software or buildsoftware 20 30 years ago when there waslike all these such like limitations youwere constantly running into and you'reworking with these

00:39:24details. I mean we'reyou know we talk about like contextmanagement and context window and thatsort of thing and like I don't know isthat going to like feel archaic in in 510 years you know? And so Even theconcept of prompt engineering I wasthinking about the other day. Like weput so much emphasis on exactly how toformat and structure and you had allthose like copy and paste graphics thatwould be all over Twitter with thedifferent color coding and now I'm like,I don't even think about that anymore.Yeah, it's all changing so quickly. Soit's our job to kind of try to thinkabout going back to like, well you dohave to be deal with the realities ofwhat it is now because if you go too farahead, you can't put 11 people on thefield. So like you can't design the gamethat So trying to bring it back.So you have to kind of find the balanceand that's where it like reallyunderstanding what it's capable of nowand then also like being able to push itand say, well we actually do want to beable to do this. So So how can we likemake sure we're trying to help set somevision for that? Is there anything thatyou are doing intentionally as a leaderto help the design org push on thatfuture or be more generative or morerisk-taking in what they're exploring? Ithink it's a good question and probablysomething we need to be more clear aboutwhen to do it. I think that's the bigquestion is finding when the right timeis to do that. I think

00:40:40in some cases youwant to do it, in other cases you don'tand I think what I'd like to do is findmore time where you can kind offlip-flop [snorts] between doing both. Ithink that's something we're kind ofhonestly trying to figure out as a as ateam. Like how much time should I spendbe spending doing that kind of workversus more execution and I think youhave to be able to be very fluid. Wecovered a lot of ground. What have wenot talked about yet that you thinkpaints a picture or shines an accuratelight on what it looks like to design atOpenAI and the culture that you have andjust the way that you all operate? Thereality is that it isvery fast-paced. We are figuring thingsout quickly. We kind of try to updateour own thinking very quicklyand we have to evolve quickly with thetechnology and and as a team. And so Ithink some people might assume we'relike 2 years ahead thinking like that,but we're running very closely withwhere all of these advancements aregoing. And so that's just like a verydifferent way of working. You know, likethings like things are changingunderneath your your feet all day longand it's very exciting. It's really funto be like, I don't know, we're going tofigure this out as we go. We're going tolike We're going to try it. We're goingto turn the crank. We're going to keepiterating. We're going to

00:41:55keep going. Ikind of want to push on something thatyou said earlier. You talked about howin many ways, you know, being a greatdesigner hasn't changed. You're solvingproblems. But yet a lot of the thingsthat you are talking about do feelalmost more reactive to the technologyrather than user behaviors. And I got toimagine that that throws a little bit ofa wrench into the design practice thatmaybe you would have at Instagram orsomething like that.>> Sure. So what I should say, I I think tomake it clear, I think that greatdesigners can do both. Great designerscan both say, oh wow, we have this newtechnology. Okay, cool. How do we How dowe package that up? But then I thinkother designers or other times ourdesigners are thinking about, wellactually it needs to do this. It needsto be really good at this or here's howthis could work. And it's just a blendnow and you just have to think about onemore thing, honestly. It's like you haveto think not just about the problems,the user problems and all that, you haveto think about what the technology can'tdo that it should be able to do. So youhave to be able to do all that. So Ithink it's basically just extending onepart of the process and I think the bestprocess for designers is thinking like,well how do we push this forward? How dowe make this be more useful if it couldjust do X, Y, and Z or here's kind of anideal way this might work and then wecan like work with

00:43:10engineering orresearch or whatever to like get itthere. Okay, so for somebody listeningwho's inspired by the conversation, theywant to join the team. You'll be hiringthroughout this year, I'm sure.Yeah.>> What are some of the main signals thatyou would be hunting for in a designcandidate? And what would you be doingto figure out if they're present and ifthey're the type of person that wouldthrive in this environment? I mean, Ithink there's just a few candidate typesthat we always look for. I'm always veryexcited about the kind of moreup-and-coming people that just have tonsof energy. I think that It's funny. Iused to say like when when I firstjoined two two and a half years ago, youdon't need a background in AI to likecome work here. You have to be curiousabout the technology. And I And that'sstill generally true, but I think thatthere's enough now there's been enoughtime where finding people that havestarted to experiment and play with thisstuff and understand what it's good at,what it's bad at, where it needs to goand how we need to like push on these onthe tools and the technology and theproducts that we're all building.I think there's like enough there toplay with. I've always respected peoplethat will go deep on some side projector get really passionate about someidea. You need the fundamentals. Youneed to be great at just kind of likedoing, you know, general product design,but I think that that other piece justlike being truly curious and

00:44:25andinterested in the idea and ideallyyou've really spent time playing with itand and understand like and hopefullyyou have lots of ideas for for where wecan push things. We're at this fortunateplace where you're not just reacting tothe technology, but you're you'rehopefully helping shape where this isgoing and like helping to expand thecapabilities that we can put into theproduct into people's hands so that theycan do more with it. And so people thatare excited about about that kind of wayof working. Well, Ian, I appreciate youcoming on and sharing this with ustoday. You all have had a a massiveimpact on state of design and eveninteracting like just creating theseparadigms that we're all building on topof and so it's been fun to hear a littlebit of the behind the scenes and how youall operate today.>> [music]>> Yeah, this is super fun. Thanks forhaving me.