# Ambition in AI — Jason Liu — session 2026-11-12T17:10:00.000Z → 2026-11-12T17:30:00.000Z _84 transcript lines · 40 slides_ ## Transcript You know, me too. These days, I don't think of myself as a developer. I'm mostly... I'm thinking about myself as someone who's trying to will things into existence. Right? And I have to, again, feel this ownership and describe what I like. In a past life, not many people know this, I used to be a machine learning engineer, I used to be a physicist, and I used to be an artist. And the the story of the camera is something I keep thinking about as we introduce more ways of doing this kind of replication, right, as as this way of creating things. When the camera first came out, painters who were Busy focusing on their ability to recreate images? Yeah. They were a little stressed out. But at that time, people weren't even thinking about whether or not photography could change the art world. We weren't thinking about what a movie meant. You see, like, a Charlie Chaplin film when it comes out. It's really easy to dismiss that as some, you know, silly thing. But at some point, you realize that not only is photography art, but it has changed the art world. Right? And now we have film and all these other mediums, and there's amazing things that you can explore. Walter Benjamin's Says this in a book, The Work of Art in the Age of Mechanical Reproduction, a book that came out right when the camera became popular. He says, photography freed the hand of the most important functions and brought them to the eyes and to the voice. Right? This is kind of the same with, like, AI now, because I just get to use voice mode on on chat and just say, you know, do this and do that in the same way that at some point, if you're a photographer or director, you can just say, you know, move that camera over there. I want a wider shot. I wanna capture some kind of feeling, and maybe have a team around me that helps me. For that vision. And now we live in a world where everyone has a camera. And sure, maybe I have a thousand photos in my in my background... In my my camera roll that I never look at. Maybe that is a slot. But it only matters whether or not I have captured a couple of really important moments in my life. Right? And and having ownership over those moments is the thing that's important. Like, for example, like, thirty years ago, like, I didn't really have childhood photos because we just grew up in a place where we couldn't afford a camera. Right? And now technology has allowed everyone the ability to have that kind of access. And I sure Like to have a camera back then even if it meant a little bit more slop. Alright. And so if you go back to this idea of this painter, I think the question you have to ask yourself is, do you identify yourself as the painter in this example or as the artist? Right? Maybe it means that painting photorealistically isn't the skill that's really valuable. But if you think about yourself as the artist, right, whether it is the paintbrush or the camera, whether it is code or clay or anything else, these are not just mediums that allow you to express yourself and to make things. Right? I don't make a table, and I say, like, the Walt made this table. Right? Just because a a director has thousands of people on set, it's still the director's movie. And even in open source, we might have thousands of contributors or even agents now running the system, but it's the core maintainers that help hold the responsibility for the quality of the work. Right? And I don't think the danger these days is whether or not, like, AI is gonna create more slop. Right? But instead, it's about self erasure. Right? Are you gonna be a passive observer of the work that Being created, which results in this slop, or are you gonna be someone who is an active participant in the act of creation? Right? I think many people have been sort of moving themselves away from this. Right? And the only difference between, like, a person who's making a thousand PRs versus another is whether or not they actually care about the work and whether they are an active participant. Know? And so as an engineer, what this means is you just have to, like, use your products more, notice when things break, and decide whether things need another pass. Like, a lot of the great examples you see online are not one shot exam. Right? These are things that you iterate over time because you feel proud to create this kind of work. You know, internally, we have a lot of amazing work with, like, our ownership over, like, the codex app. Right? No. Most people here have an incredible sense of ownership over the the product, and as soon as things slip as we ship more, we wanna bring them together. And, again, this is mostly because of the fact that we care. So in terms of worrying about slop, I think that the the medicine for slop is really just the fact that you you really did believe that you own this work. I think the second thing to address is whether or not you feel like you can actually describe what good looks like. And for that, you need to go out and be inspired. So the second question people are asking now is, like, do you think AI is gonna destroy our ability to think critically? Right? If I own the work, how do I know if something's good enough? If I just hand you a camera and this analogy, like, does it mean that it takes away all the effort of creation? I don't think the answer is yes. Right? Instead, it's now about having language to describe what's going on. So when I first started art school, you know, maybe I don't know how to develop film. I just use my iPhone. But I still need to understand, like, the elements of design and what does beautiful look like. Right? Space, value, shape, texture. You know, when I... I I would just spend a lot of time thinking, you know what? I don't think I understand what negative space means. Let me go do a lot of sketches and explore a little bit. It's the same thing with, like, development now. Right? I can make multiple apps, and I have to think about why I like something and why I don't. And most of my I'm in art school. I was just thinking, man. Like, I can make things, and I can look at them, and I just say, I don't I don't like how this looks. And what you... I was noticing was that my taste and my judgment exceeded my skill. I could not do the thing I liked. Right? And I can't just look in the mirror and tell myself, like, make it pop a 100 times and and know that it's gonna be better. I have to go out and develop that language. Right? People say, like, taste is the moat. But I think I I generally hate that line because this doesn't tell you what taste really means. Right? To me, taste is the ability to describe what you like. And predict what others will like. Right? And building this model of the world and of other people, that's the new thinking. Right? Winstein's the limits of my language means the limits of my world, and I think that's very much the case on pursuing ambitious projects. You know, it was a funny example of just, like, the use of language. I don't know if Steve was in the room, but he made a tweet a couple days ago that just says, find all the implicit state machines in my app. I had to go Google what that meant. Right? And all of a sudden, all the comments are just how, like, everyone's apps were, like, 20 Percent faster. These are now, like, really simple ways of just developing some language to understand what good looks like and how to make things better. If you think of, like, model progress, in machine learning, there's a really helpful concept called the generator discriminator gap. Right? You'll have two models. One grades the model, and then one actually creates the outputs. And what happens is is if the grader is too good, then the model can't learn. And if the generator is too good, the discriminator can't help the model improve. We are now that discriminator. Right? As model capabilities go up, at some point, maybe maybe the model's outputs hits your own taste levels. Then it's really hard to actually improve. And then to make things that are better, you have to go out, develop that language, and know what good looks like. And the failure mode is here just not being able to discern whether or not a model can improve. And again, what this means is we have to go out and learn again. We have to think again. This This is an example of a bunch of things I spent my past year learning, like, in just building software. Or, like, as I build these applications, I realized, oh, man. These websites look really ugly. Like, I need to learn about typography. Like, why is everything italic? Like, I don't understand. And I had to learn... I spent a lot of time learning about typography. Then it was about, like, how the information was presented to avoid looking like slop. And And then it was about, like, motion design and how can I make things move in a beautiful way? And I got to spend so much time learning about animation and then learning about, like, interaction patterns and what makes an app feel beautiful. And this is really fun. Right? And this is, again, spending more time learning in order to make great things. And most importantly, having the ability to describe things is what allows you to differentiate. Right? If you think about these purple apps, the reason these existed was because the people making these these applications, like, didn't have the site. They couldn't actually describe what was wrong, and they thought this was okay. Right? And this limiting in language is sort of the problem that we face. And then so the question now is, okay, how do we develop this new language? And honestly, the line that I've sort of been, like, punished to now is just like, if you want good taste, you have To eat. You have to go out and explore things. Right? People are, like, vibe coding video games. Maybe the answer isn't just to, like, tell the model to do better, but, like, play some video games and understand what feels good, what what animations feel good, what stories matter, what stories do I wanna tell. Right? Right? If you're building tools, try other people's tools and applications and understand what you like. Look at beautiful things. I've been doing a lot of, like, three d printing with AI, and everything just looks like rectangles glued together. And I just kept saying, like, make it more beautiful. Right? And that's a really frustrating experience. And then I started looking at, like, sports cars or, like, industrial design. I started going to stores and looking at things and going, well, let me describe and understand what, like, good curves look like. And, again, it's, like, a really fun experience. And now more of my designs are, like, incredibly beautiful for what I think is beautiful. Right? I've been helping with, like, marketing, and I realize that, like, I tell the marketing team that they should use, like, the codecs app more, but I'm not watching more ads. And as more, but I'm not watching more ads. And as I'm watching more movies, I realize there's great stories to tell, and everything becomes better again. Right? And this is kind of the really exciting part. Using these tools. Right? And, like, thinking about why I like things and why things are amazing and why things feel good now is how I spend my time. Right? So I'm not thinking less. I'm thinking more, and it's really enriching my world and making me see things differently. And if you're running a team, I think the the one takeaway here is that, like, if you really believe that AI is making you more productive, I think you owe it to yourself and the rest of the team to make space to go out and explore things and try new applications and and just actually go out and, like, The world. Right? And if you are someone that makes really great work, inspiring others as well. Right? And that's how you can come back with a point of view. So go out and be inspired. Learn to be a better critic. And recognize that you're not actually thinking less. You're thinking more. And the last thing is actually just being curious... Really, really curious to pursue new ideas. Like, one question I want you guys to think about as you are thinking about your role as software engineers is, like, what were the things you gave up or put aside to pursue? Software. Right? Where else could use your judgment and your experience you've developed as a software engineer? Because I think people worry about being replaced, but really, I think people are needed in way more places. If you think about when the, you know, invention of, like, the pencil or or sculpting became more easy, all these different industries in the creative world became really very much, multidisciplinary. Right? It completely made sense that if you were a furniture designer, you could be an architect, and you could, like, make a movie. And design kids' toys and playgrounds. Right? Like, Michelangelo is a sculptor and a painter and an architect and a poet. You know? It completely made sense that you could do all these things. Right? And this is a creative world. Like, you invent the pencil, you can do a lot of really amazing things. But with technology, it's become even more complicated to do so. I just can't get, like, four PhDs. And if I could, like, maybe I'll just be, like, 50. Right? And so being technical has become incredibly difficult. But when you do cross pollinate these different industries, amazing Happen. Like, the physicist was... Sorry. The founder of the molecular biology was a physicist. Right? A mathematician started leading the human genome project. Right? And the mathematics around magnets is what you... Is what is used to help understand neural networks. But this is now, like, super hard to do because you need so much information. Right? Even in our own world, like, we have a bunch of AI helping us design, like, chips and writing software, and these, again, are just very multidisciplinary pursuits. If you look at what it means to be ambitious Outside in the real world, there's a machine learning consultant using AI and molecular biology to, like, help treat their dog's cancer. There was this, public servant in Hokkaido who used, like, chat and, like, computer vision models to build, like, robots to automate their farm. Right? This cross pollination and this will to create these things and learn and cross pollinate across these different paths is what's really inspiring. Right? I I think earlier today, I saw folks are using AI to Like, make more robots and also, like, discover exoplanets. Right? And it's that curiosity and that enablement of what AI allows you to do in lowering that barrier is, what allows you to be ambitious in those fields. Right? And so when it comes to thinking about whether or not you can... You're gonna be replaced, a lot of it really is the fact that you now have the skills that everyone needs in every field. It's not just about, like, writing more apps, but actually thinking about, you know, what were the things I didn't pursue in college to pursue software engineering? What... How can I go back to those worlds? How can AI make those things easier? Right? You can, again, develop new language and understand what makes things amazing and then go go pursue those things. And the great joy of this and the reason why these, like, multidisciplinary artists are so successful is as they pursue new mediums, the act of being a kid again means that you have the naivety to ask for what it might be impossible. Right? The reason this cross pollination works is if you have the judgment you've built in software and you pursue another field, it means that you can A question that people might not think is possible, and you have the courage to use AI to help you get there. Right? And so there's great benefits in being able to feel new again in the field while all... Having all this experience in AI as well. Right? In my personal life, I've been, like, learning the drums. I've been building custom, like, drumming software. AI is helping me learn how to use, like, Ableton. I'm connecting my my drums at the MIDI and learning there. You know, I've been using computer use to teach me new software that I never really had time to learn. And And I've even been using AI to help me design, like, physical soft Physical engineering. So I'm three d printing parts, building chips. And, again, these are things I was always curious about but never had the time. And, again, part of having AI is not only getting my time back, everything else is accelerated. Right? And so being ambitious is about exploring those pursuits, not just spending more tokens like building an app. And so this is also what I want you guys to think about. Right? Again, what fields do you put aside? How can you pursue them now that you have more time with AI? And When I'm when I'm talking about, like, increasing ambition, I'm not just saying, like, become a biologist. Right? What I'm really saying is that, like, part of your job now to become more ambitious is to be really inspired, pursue these things, learn that language, and also for that next generation to be inspiring. When people used to ask me how to learn how to code, it was never a case where I said, you know what? Like, you should go read a textbook on Python. It was always build something. Right? It was always go build another app. Go build a note, like, note taking app. Have a real problem and solve. And now the scope of these problems are much bigger. And so as software developers, I think our job now is to not only build amazing things and share them, but also inspire the next generation to have more hope too. Right? Because maybe software is changing, but really, we're gonna be be able to contribute to so much more things. Right? And, like, we see this everywhere. Like, for example, the codecs app and sites, these are all things that were acts of love. Codecs app, the sites product were all built by people who really cared about these things, are talking to people online, and working through these things. Right? And now Our PMs can do engineering. Our engineering can do design. This cross pollination is inspiring others to do the same and take ownership of that work. Right? And most of the things here that exist are here because people are tenacious and are trying and want things to be beautiful. You know? And ultimately, our mission at OpenAI is to benefit all. And ultimately, what that means isn't just having OpenAI have all these AI to to control and, you know, use agents forms or whatever, but also making sure that everyone has this technology so they can contribute in their own And so that's basically it. You know? Think about the things that you own. Go out. Get inspired. Develop the language you need to describe what beautiful things look like. And then just make sure you're actually curious enough and take the time to pursue these new interests. And then that's how you become more ambitious. So I'm really excited to see what you guys are building, and, come by to the booth to chat. Thank you. From creating Flask and Jinja to exploring the future of AI powered software development, our next speaker has helped shape how developers build for the web. Please welcome Armin Ronaher, CTO of Arendelle. Hello. My name is Armin Ronaha, and I wanna share with you a little bit the lessons that we have learned from building software with agents. And in particular, I wanna make a strong case for, not so much building the factory as building the artiller or workshop. And, I wanna show some data with this from our own usage. I wanna explain what I mean by building the atelier or the workshop prior to building the soft factory. And, I think everybody here is being used... Has been using agents for quite a while and you have probably noticed that, it is... We're sort of an exponential curve in in what we can do with them. But the comparison that I wanna give here is that you can be an amazing artist and use the greatest and latest technology from CNC machines to three d printers, which is very good Our tools and it can increase the ambition of what it can do with your art. And I think we can also look a software a little bit in that sense. So less about building much more software and just building way more ambitious and high quality software. So I'm going to show you a little bit of what I mean by that. And, you might know that Arundel is working on a PyCoding agent, so we have obviously a little bit of a horse in a race here of how we're looking at this. But I think what is probably quite interest Thing is that the PyCoding agent has not really changed in a year on the core harness. It still has the same tools today as it did, in December last year. And I think this is interesting because it has become much more powerful purely as a effect of the models getting better. So I am much more capable today of what I can do with Pi without any changes landing in the core functionality of what the harness is doing. And I think this is an interesting thing to reflect on because it means that we are all simultaneously getting better. Without landing any significant changes in how we're doing this. And, you might have remembered throughout the last year and a half maybe that there were a lot of ideas of how you can prompt the machines better or how you can do some interesting tricks with the harness. Very few of them are still necessary. For me, a good example here is that I was in love with PyAutoResearch, which was a plugin that basically allowed an agent to drive towards a goal like improving performance. But if you use a present day state of the art model, you actually discover that the entire plugin is no longer necessary. So there's a lot that is just happening in the models that is dramatically improving how we're doing. So these machines are getting better. And I just want to show this from a couple of graphs from basically my PI sessions where I filtered them out to be have nothing but the default set. And the total amount of corrections, and I used Jeff to analyze this, the total amount of corrections that I had to do got down and down and down with every single model that got better and better. So I'm I'm pushing back a lot less against the ## Slides ### 00:00:09 # Ambition in AI ## Own what you ship You're still there after it ships. ### Jason Liu / Developer Experience **OpenAI** [QR code] ### 00:00:37 # Me too > No. I have zero technical ability. ## Ambition in AI Jason Liu / Developer Experience **OpenAI** [Meme image of a man with a long white ### 00:01:08 # Was photography art at all? The painters whose job was to recreate images were pretty stressed out. ## Ambition in AI Jason Liu / Developer Experience **OpenAI** [ ### 00:01:36 # Ambition in AI > "Photography freed the hand of the most important artistic functions." > > Walter Benjamin > The Work of Art in the Age of Mechanical Reproduction Jason Liu / Developer Experience **OpenAI** [Black and white portrait of Walter Benjamin] [QR code] ### 00:02:08 # Now everyone has a camera I can take 1000 photos I'll probably never look at them. Is that slop? Maybe. I just need to ### 00:02:40 # Painter, or artist? The camera, the paintbrush, clay, even code: all just mediums. We're the ones who make it. ## Ambition in AI Jason ### 00:03:08 ## Ambition in AI ### A table “I don’t say DeWalt made this table.” ### A movie A thousand people on set. It’s still the director’ ### 00:03:41 **AI Engineer Code Summit** Presented by Google DeepMind # Active participant, or passive observer? Somewhere, an engineer will merge a thousand PRs a day. The difference is ### 00:04:04 AI Engineer Code Summit PRESENTED BY Google DeepMind # Internally... ## Codex They're online talking to the people who use it. What broke, what felt weird ### 00:04:36 ## 02 Go out and get inspired Learn what you need to be a better critic. ### Ambition in AI Jason Liu / Developer Experience OpenAI [https://openai.com/ ### 00:05:09 # AI Engineer Code Summit Presented by Google DeepMind ## Ambition in AI - Line - Shape - Color - Form - Texture - Value - Space No. ### 00:05:36 # My taste exceeded my skill > You can't just look in the mirror and say, “Make it pop.” 100 times ### Ambition in AI - Jason Liu / Developer Experience OpenAI [QR code linking to https://openai.com/blog/ambition-in-ai] ### 00:06:09 # AI Engineer Code Summit Presented by Google DeepMind ## "The limits of my language mean the limits of my world." > "The limits of my language mean the limits of my world ### 00:06:40 # The generator-discriminator gap Failure: the model can generate far more than we can judge ## Ambition in AI Jason Liu / Developer Experience **OpenAI** [Line graph showing "Quality" over "Time". A blue line labeled "Your taste" and a white line labeled "Model capability" both increase over time, with "Your taste" initially higher, then "Model capability" catching up and surpassing it. A dashed line labeled "The work" follows "Your taste" then "Model capability". Key points are "Model launch" and "Model catches up".] [QR code for live translation] ### 00:07:07 # The generator-discriminator gap Failure: the model can generate far more than we can judge ## Ambition in AI Jason Liu / Developer Experience OpenAI [Line graph showing Quality over Time. A blue dashed line labeled "Your taste" rises, then flattens. A blue solid line labeled "The work" starts lower, rises to meet "Your taste", then flattens. A white solid line labeled "Model capability" starts low, rises sharply after "Model launch", and eventually surpasses "The work" and "Your taste", with an annotation "Model catches up" where it crosses "The work" line.] ### 00:07:41 # Words I started learning, so I can describe this better to AI ## Typography - Kerning - Tracking - Leading - X-height - Type scale - Line length - Tabular numbers ## Hierarchy & layout - Information hierarchy - Visual weight - Optical center - Negative space - Proximity - Alignment - Grid ## Motion - Easing - Duration - Stagger - Spring - ### 00:08:09 # Think back to the purple apps. - **Dayplan app:** A calendar view titled "My week" showing scheduled events like Design, Build, Sync, Focus, Demo, Lunch, and Review ### 00:08:36 # AI Engineer Code Summit Presented by Google DeepMind ## If you want taste, **you have to eat** - **Play games** What makes a game fun, ### 00:09:08 ## If you want taste, **you have to eat** ### Play games - What makes a game fun, and what kind of stories ### Try other tools - Spend an afternoon with other application ### Look at beautiful things - Learning industrial design ### Learn motion design - to make my application feel more beautiful ### Watch more movies - as I help with marketing And ask yourself: 'Am I able to ### 00:09:36 # Thinking about why I like things It's enriching my world, and inspiring me to build new tools and new systems ## Ambition in AI Jason Liu / Developer Experience OpenAI [QR code] ### 00:10:09 # Ambition in AI ## 02 Go out and get inspired Learn what you need to be a better critic. You're not thinking less. **Speaker:** Jason Liu / Developer Experience OpenAI **Resource:** [https://openai.com/blog/ambition-in-ai](https://openai ### 00:10:36 # AI Engineer Code Summit Presented by Google DeepMind ## Ambition in AI Jason Liu / Developer Experience OpenAI ### For creatives, working across fields was always the case - **Charles ### 00:11:08 # AI Engineer Code Summit Presented by Google DeepMind For creatives, working across fields was always the case - **Charles and Ray Eames**: Furniture, houses, movies, and toys. ### 00:11:40 # AI Engineer Code Summit Presented by Google DeepMind ## Working across fields changes fields, but was quite rare - **Max Delbrück**: A physicist who helped found molecular biology. ### 00:12:09 # What it could mean to be more **ambitious** - Paul Conyngham and Rosie: An ML consultant who learned molecular biology to help his dog. - Hiroki Tomiyasu: ### 00:12:42 # AI Engineer Code Summit ## Presented by Google DeepMind # People like you are needed everywhere The fun part is learning the languages of fields that weren't your own. ### Ambition ### 00:13:13 # You're a kid again, with the judgment you built in software Not limited by what you think is possible. ### Ambition in AI Jason Liu / Developer Experience **OpenAI** [https://codesummit.ai/](https://codesummit.ai/) ### 00:13:36 # You're a kid again, with the judgment you built in software Not limited by what you think is possible. ### Ambition in AI Jason Liu / Developer Experience **OpenAI** [https://codesummit.ai/](https://codesummit.ai/) [QR code linking to codesummit.ai] ### 00:14:09 ## I am a beginner again - Learning drums and music - Computer use to explore new software - Hardware Design and Physical Engineering ### Ambition in AI Jason Liu / Developer Experience **OpenAI** [https://openai.com/blog/ambition-in-ai](https://openai.com/blog/ambition-in-ai) ### 00:14:37 # Get inspired, # and be inspiring ## Ambition in AI Jason Liu / Developer Experience **OpenAI** [QR code] ### 00:15:09 # My advice was always '**solve a problem**' not 'read a textbook' ## Ambition in AI Jason Liu / Developer Experience **OpenAI** [https://www.aicode ### 00:15:41 # They're good because people cared And when you care, you inspire other people to care ### Ambition in AI Jason Liu / Developer Experience **OpenAI** [https ### 00:16:04 # When execution is free, your ambition is limited by - what you **own** - what you can **describe** - what you're curious enough to **pursue** ## ### 00:16:36 # AI Engineer Code Summit PRESENTED BY Google DeepMind [https://codesummit.ai/](https://codesummit.ai/) LIVE TRANSLATION [QR code linking to codesummit.ai] ### 00:17:06 # AI Engineer Code Summit Presented by Google DeepMind ### 00:17:38 # AI Engineer Code Summit - Event branding: AI Engineer Code Summit - Sponsor logos: AWS, Atlassian, Sonar, Databricks, Vercel, and another partially visible ### 00:18:09 # THE THESIS ## Before the factory, the atelier 1. Everyone is trying to build *software factories* 2. Most of us, me included, have not learned to use ### 00:18:37 # The Software Atelier ## THE THESIS ### Before the factory, the atelier 1. Everyone is trying to build software factories 2. Most of us, me included, ### 00:19:09 ### I · STABLE HARNESS ## The machine got better 1. Most prompting ideas were replaced by **better models** 2. Basic Pi did not change 3. Yet I am much ### 00:19:41 ## I • STABLE HARNESS ### The machine got better 1. Most prompting ideas were replaced by **better models** 2. Basic Pi did not change 3. Yet I am much more capable today [https://the