Builders Unscripted: Ep. 4 - Pietro Schirano

Ciao Pietro! Hey, what's up Roman? How are you writing the theme song for Builders Unscripted? Maybe, yeah, like maybe we call it like tokens or something, I need more. Let's hear it. Tokens! Oh, I need more! That sounds like a great idea. We'll save that for later. Well, I'm so excited about this one. You're a man of many talents. Thank you. You're a musician, as we can tell. You're also an engineer, a designer, founder. I would say, AI whisperer, early testers of our models. So I'm super excited to chat. But to get started, I'd love to hear, how does your brain work? How do you have so much creativity across all of these different fields? Yeah. Well, first of all, thanks so much for having me. I feel like you and I have been talking for a while, so this is amazing to just be here and have this chat. For me, I feel like there is this quote that I love, which is like, we're born too late for the sea, but too early for the stars. And with AI right now, I feel like there's this new field that you can explore, like the latent space, right? It's like you can get into the mind of this incredible new technology and start to explore, almost like truly like an explorer back in the days, like a Magellan or Marco Polo, and just trying to figure out what are the things I can do with this technology. And I think that, to me, it's really exciting. Love that quote. That's super cool. I mean, we've seen your work from afar for a long, long time, and then you've been one of our early testers, kind of having access to these models, because you always have an interesting lens. You're poking at these models to find the little edges, and what can they do that the previous one could not do? What was your first impression of GPT-5.5, for instance? It was a big step change. I feel like there was a lot of the things that were very hard to do in the past, and now, all of a sudden, it just works. It's just like... complete sort of like wave that just shocked me when I first started interacting with the model in the beta testing. It's things like, you know, like I have all this set of evils, personal evils that I have every time a new model comes out. And those are things from, you know, agentic behaviors, like understanding like how can I run multiple tools at the same time or multimodal capabilities. I'll give you an example, in the case of GPT Vision, at the time, in order for the model to understand where to look at, I have to provide the image with an overlay grid on top of squares so the model will know, oh, the bottom is in this square and whatnot. And now you don't have to do that, right? Like 5.5, it's so good at vision capability, you can just reasoning around where to click, where to look, and stuff like that, right? Ultimately, what I really care is what kind of capabilities those models have that my people may miss. And so those are things around creativity and fun, like how good the model is at acting as a designer, or how good is the model acting as a creative director to direct other design agents and models. And actually, there's a really funny demo. The first thing I built with 5.5 was like, what if I could create something where 5.5 takes an image and deconstruct the image in a sound that is harmonic? based on the picture and then I can use that sound to reconstruct the image. Almost sort of like a secret communication game language so that two people can communicate with sound in this secret way. And I was like, there's no way it's gonna get this right because it's a very complex, there's cryptography involved, there's how good the music needs to be, how are you compagning the basic 4th of the image to sound. And I was like, there's no way it's gonna get it right. And then I did it and it just worked. It was crazy. Why don't we take a look at your laptop and you show me a little bit what you've been building with these new models. Yes. So the first thing that I built with 5.5, so going back to this idea of like, how can we poke the model for creativity and sort of like, you know, fun ideas, like out of the ordinary ideas. So what I did, I basically built this app, right? And so what this app can do... it's pretty crazy so it can convert an image into a sound and then from a sound back to an image almost sort of like a secret communication way between uh different people and so what i'm going to do right now i'm going to take the openai logo here so i'm going to pop it right here and then i'm going to ask to create a sound and and codex plus 5.5 was basically able to do this like in one shot and i was like mind blown because like there's so much going on here there is cryptography going on there is like understanding you know what does even mean to make a beautiful sound and you know there's this movie um i think it's called uh contact from the third kind it's like an old steven spielberg movie in the 1980s and there is at the very end of the movie i don't know if you remember this the scientists are communicating with the alien yes And the way the aliens communicated is actually very similar to this, which I was kind of impressed. It sounds really good, too. It sounds really good, right? And so what I can do now, I can download this sound, and then I'm going to pretend now that I sent that you are accessing this app on your computer, right? And what I'm going to do, I'm going to take the sound, drag it back here, and I'm going to ask to recover the image. Boom. And now I have the image there. So we convert, yeah, basically like an image to sound and from sound back. How is that possible? I know. For me, this is what's important when you think about AI. It's like, what can I do? No idea is too crazy. I think there's an element of... just your imagination. Your imagination shouldn't be your constraint. Just feel free. And also, I think there is an element where as the models get better, you really need to start thinking about what are the things I can do if I wasn't afraid, in a way. Because I think as human, we always have these walls a little bit. It's like, oh, can this be done? But just try. And now 5.5 is so good that it gets those things right. That's really awesome because you have your... musician background that can help you have these ideas. This is so cool. And you're talking about evals too. I'm curious, so when the new model comes your way, like GPT Image 2 or GPT 5.5, what's your process to poke at a new model and kind of find what it's good at? Yes. So I basically look for three different things. One is instruction following. So it's like, how good is the model? Are following this instruction? And you might be surprised actually sometimes. a modern model doesn't necessarily mean it's better instruction following because they might get maybe like a little more secured or whatever. And so what I love about the GPT-5 series is how consistent has been in just being helpful, basically, just being helpful assistant. So there's like the instruction following part. Then there is like, how was the task? executed? Is it actually being able to do what I ask? And then there's an element of how good is the model actually directing other agent? Because that's actually a very complex task. If I say to Codex, actually, this is really fun. So I have all this set of shortcuts that I use with Codex. And so, for example, if I say spawn... it does this. So I have all this text replacement that I have for myself. Because Codes can run multi-agents, and the reason why I have shortcuts is because by now I know what are the things that I want the model to be good at, and then I see how the model performs. Interesting. So for example, I have this one. When I say agent, I say, hey, remember you're an agent, and you have to complete the goal. Until you complete the goal, don't stop. And one thing about 5.5 is really good at that, which now you guys launch the goal. With the slash goal command. Exactly. So I was doing this before goal. And so this is another thing. It's like one, I think, of the most important lever that we've seen as the model get better is how long of a task they can do. And I actually believe the true unlock and distribution of wealth for everybody is going to be when the model is going to be able to just basically run for hours, which Code is already capable of. But in order to do that, you kind of have to nudge it a little bit. And so that's something that I look for when a new model comes up. What about subagents? Sub-agent, I have spawn, so I say spawn, and it spawn multiple agent. I have things like PR, so this is actually something I do a lot. So it's like, oh, push this PR based on this repository. Or another thing, actually, that Codex does extremely well is debug. So every time a session ends, I basically say, okay, review all the code that we've done. Is it any potential bug, regression, edge cases, all this kind of stuff, right? And so by now, I know when the model performs well on this task, like instruction following, multi-aging capability. And then there's an entire element of multimodality, right? It's like, okay, like... how good are you at taking this image and converting it to code, right? And 4.5 is actually amazing. In fact, I think a really fun flow that I recommend people to do with 5.5 is generating an image and then asking to make something around that, which it does really well. That's why we build the Build Web Apps plugin, where we package up in one plugin the idea of using GPT Image 2 to generate a design and then get 5.5 to implement that design, because usually it's like... it's pretty flawless to take a design in and like actually implement it out and i feel like those stuff truly were impossible like you know even like four or five months ago right and like you know they were in some category but they're just better now what about beyond games and pixel art like yeah you're a designer you designed icons in the past like how do you look at icon design for instance now so i built this app which i showed you in a second it's really cool it's an autocomplete writing tool and i call it odysseo and odysseo is basically the protagonist of like you know the odyssey and ulysses and the reason why i call it like that is because because you're writing with AI in this app, you're kind of in this infinite sea of possibility and writing that you can do. I like that. And so I asked Kolek, I was like, can you give me an app icon that has a Greek chip and a quill? And he just did this. This is incredible. And in one shot, right? And then he added it to DMG. And so now, you know, like my app, it's right here. I mean, the Pietro of like 15 years ago, how many hours would you have spent trying to design this? Oh, don't even understand. This was like Photoshop. multiple layer. I can even tell you, there is a gradient effect applied here, there is a border with gradient applied, there's all this texture. I mean, it's actually insane. It would be many, many hours. Yeah, to get it to this level, it would be hours of work. And you just one-shoted it. Yeah, exactly. But yeah, so maybe we can take a look at this app. I'm curious to hear more about it, because you also play with multi-modal models, right? And we just launched... New real-time audio models in the API and sounds like it's also something you you try to poke at already Yeah, so like like I think the team that I always go back to it is this idea of human to AI mind connection, right? It's like what can I do so that this tool like codex or certain AI model becomes basically just like an Sort of like an extension of my mind right and you'd be surprised like I feel like we as human we're an incredible input machine, with the vision, the sound, the smell, all those things, but we kind of suck at output. And so I was like, okay, what if I can make an app where as I write, it complete what could be the next text, and then I can also talk to it. So in this case, we're saying, today I'm going to the OpenAI office. to record a new episode of Building an Escape with Roman, and I'm thinking about demoing, and then I can just press tab, and now a GPD model is generating, you know, like, you know, okay, this is the last interaction of my agentic workflow, and, like, what if I say, you know, what if I say, like, oh, maybe I want to add something in the middle, right, so I can go here, and it's going to autocomplete that sentence right there, but then I was like, well... what if I do this mind extension thing? So I can start GPT real-time, the new version, and now I can just talk. I can talk right here. And you can see in real-time, it's basically writing what I'm saying, which is pretty insane. And then I can stop it, and then I can autocomplete whatever I say. That's pretty impressive. Impressive. Almost like having a second brain capture every straight hour in real time. That's kind of crazy. It really helps how far we've gone in making these interfaces feel like a natural extension of your own internal model. That's insane. That's what you wanted to say. That's insane. That's what we're talking about. That's wild. That's awesome. I love also how you're going to bring these modalities together, right? Not just like, well, obviously, I'm a writer. I'm typing. But also, I can just talk to the computer, capture that in real time, and then finish my thoughts in writing. This interplay is pretty magical. Yeah. And honestly, the UI that you build is pretty nice. There's a dark mode. There is all these commands. If I want to bold this, I can just command B. I mean, it's pretty crazy. And it's just talking to the model. Just talk to the model. And I think this is a perfect example of what is the half of your dream, right? That you had, you left in the drawer in your mind and you never built because you were busy or you thought you couldn't done it, like all those things, right? And I feel like now basically everybody can do this. And like the analogy that I have with Codex, that I was thinking about the other day, was like, it's like in the past, not a lot of people were great painters, right? But everybody can take a photo. Great, not everybody's a great photographer, but everybody can take a photo. And to me, this is like, the taking a photo of coding right it's like it's like it's just a snapshot of an idea that i had and now it's in this new sort of basically medium slash material that coding is becoming i feel like coding now is becoming this very malleable clay like type of thing yeah you can just build it you can just build it yeah exactly that's that's pretty awesome thank you for sharing me like all of these projects really cool to see your your creativity in fact i'm curious like you playing with so many modalities from image gen to these voice and real-time models to the frontier of 5.5. When a new OpenAI model drops and you get early access, how do you go about that? What's your process or your ritual to go about poking at those models? Yes, I love that question because it's something that I basically think about all the time. Sometimes I don't sleep because I think, what are the things I can do? There's almost like this... quote-unquote, building anxiety that I have because I just want to make sure I have time to do all those ideas. But what's interesting, it's almost like the models are getting so good at coding software and, you know, just computer software that it feels like they're most saturated in a way. And so I'm always like, what can I do that I haven't tried? And is there something that is really cool that I, maybe an idea that I had that I couldn't do before because I tried in the past and it just didn't work out. And an interesting side effect of 5.5 is that I basically surround myself with all these old gadgets that I had in my drawers for years. And I don't speak the language of these devices. I don't know what they're coded on. This might be coded in C, this might be coded in some proprietary language and whatnot. And so I was like, would it be cool if Codex in 5.5 could understand the nature of these devices and then build it out for it? That's pretty cool. I love the idea of you on one hand... poking at the very frontier of AI, but on the other hand... Coming back. Coming back to these old devices and see what you can potentially do with them. Yeah, and it's interesting because it's almost a way to revive this obsolete tech. There's almost an environmental element that comes from, oh, now I have all those things that are in my drawer, they're broken, and I don't know what to do with it. I can just... repurpose of something else which i think is really interesting is that what uh kind of pulls you back to them it's like well i have these drawers of like old devices sure enough at some point i can do something creative with them exactly it's like i can do something that i could maybe use it for my work or you know to have fun or whatnot right so tell me about this one what is it yeah so this is basically just like uh it's a device to do like security control on like so it's kind of like a working device if you want to actually to think about it and uh i was like wouldn't be fun to make a game for this But then I was like, you know, if I was making a game for this, it's kind of like weird. Like there's like a, you know, the screen is weird. Like I don't know how to play. But then I was like, hmm, but if I play it this way, maybe it's more interesting. And so I asked Codex to basically understand this firmware. It did that. And then did you just like plug the device? Yes. So Codex, I plugged the device to Codex. And you say like, read the USB. Okay. I was like, read the USB, what is it? He was like, this device is connected. It's this device, this is the serial number, whatever. He was like, I can actually do something on this. I was like, no way. It built this doodle jump, which is really funny. What's even interesting about this, I read their API and doc, there is no mentioning of being able to do this vertical work or vertical scrolling of the screen, but Codex just figured that out. which is really crazy. And so now I have this basically endless game, very similar to Doodle Jump, like early iOS game, right? Which is, honestly, I think it's crazy, right? And then I was like... The screen is pretty good to play, actually. No, it's fun. Yeah, you should try. It's pretty fun to play. Yeah, you just move around and the controllers move. And the other interesting thing that I, that I, after I read what Codex was doing, this little analog button actually has like different sensibility if you long press and if you like short press and Codex just like, Oh, wow. Figure it out to make sure that, you know, the experience was this smooth. Oh yeah. It even nailed the double jump. Yeah. Yeah. Yeah. If you want to like double jump, if you're like too far, you can double jump. That's good. I mean, and that was true, truly one prompt. Incredible. And when I saw it, I was like, there's no way. I have never seen this divide in the world. It's crazy. That's awesome. Then I was like, well, this is all fun, whatever, but what if I could use this for work? What is the interesting thing I could do for work? I think a team of today's conversation has been this connection mind AI. What can we do with those models? What are the things we can do? You can see here basically, Codex build an app so that it could connect. itself and so like i was like okay what can i do that's like useful for work so codex build this control where if i send the text here it basically vibrates when it's done so of course you can you can even hear it but it's i feel it right and he showed me the message right title yeah it's like okay now it's no message because we we talk more i was like i keep saying like hello and so i can see the text and it's vibrating right that's cool The reason why I did this, this is going to sound so nerdy, but I have long, long session with Codex. Sometimes I feel like Codex can be very unsupervised now. I trust it. I know it's going to run for 20, 30 minutes or 10 minutes, whatever the task is, and then when I come back, it's done. But there's always this element of like... oh, what if it goes stuck or whatever, right? And so now what I do, I'm like, you know, I go grab a coffee in my office and I hear like, da-da-da, a little vibration, I know the codex is stuck. On your desk, you can actually step away from your desk and you know when codex is done. I know it's done. And this is like a really interesting way to like use a device that, you know, I never used, you know, it's almost like a remote control, basically. Fascinating. We talked a lot about like your, you know, skill set as an engineer. a creator, a designer. I'd love to talk a little bit about you as a founder. So you're building Magic Path. Yes. Can you tell us a little bit about what that is? MagicPath is basically an infinite canvas for agent and human to design and build together. We built this common workspace where you can access from everywhere. You can have codecs running in any device on your phone or whatever, and you can just design in MagicPath. What I really think is happening right now, everybody's going to have their own agent in some extent. You have your agent that has all the information that you like, all the context that you like. I have the one that I like in the context. And so the idea is like, how can we become the reference tool? for the product teams, right? Because right now there's like a big issue where you have, you know, the PM, you're trying to be the designer, the designer should be the engineer, the engineer should be the, you know, and it's also because AI is basically collapsing all those roles because AI is becoming this sort of like translation layer. It's also an equalizer. It's an equalizer, exactly. But because of that, people tend to stay in all different kinds of tools. Like, you know, some people might access 5.5 in codex, some people might access in chatGBT, some people might access somewhere else. And so we're trying to be the place. And so what's actually really cool, Codex can build in MagicPath. So it's just a canvas for Codex to explore and design and stuff like that. And so, we launched like a year ago actually. So this is almost like our year anniversary. And it's awesome. We have like hundreds of thousands of customers and Codex in particular 5.5 has been like a huge help for us because we're about to launch the 2.0 version of the application. And truthfully, we wouldn't be able to get where we are right now. if we didn't have 5.5. I can tell you that, and I'm not even trying to like, you know, over say things, it's just the reality. Because we're a small team, we're like, you know, our product team, we're like six people. And so, you know, we move really fast because you have to move really fast in this age of AI. But having Codex access to all my employees becomes sort of like a forcing function where now I have access to 100 developers or 200 developers and stuff like that, right? Or maybe I have access to five new developers that are like really, really good developers. Right. And so that has been a huge output. I would say a funny anecdote is that as a founder, there's so many things that I need to do. I need to raise money. I need to talk to my clients. I need to make sure my employees are happy, all this stuff. And now that the AI is so good and 5.5 is so good, there's almost an element of like, oh my God, I want to push so many things in the app. My PR request has been going up, I don't know, like a tenfold because of 5.5. And so there's an element of like, As a founder, it's also important to say, okay, I know you can build anything you want right now, but you also have these incredible people that you hire, so you can trust. There's this inner battle between me as a founder and as a person that just wants to... build stuff. You just want to build. But I have to do all those other things which are equally important or more important in some sense. And so, yeah, it's been an amazing experience. Yeah, thanks for sharing. That's awesome. How has Codex changed the way you work personally? Are you just spending way more time talking to Codex now? I talk to Codex so much that my girlfriend was like, what are you doing? You're always with this app open. Who are you talking to? She's always asking me, why are you talking? Because I'm always talking to Codex, right? And actually, the reason why I do that, this might be surprising, but I don't use Codex just for coding. I use it to organize the files on my computer. Or I have this music library from years ago. It fixes all the covers, the metadata, all of those things. Or I write the notes. So actually, I write the notes for today, things I want to talk to you. I ask Codex, hey, this is the kind of stuff I want to talk about, write into Apple Notes. And it just did that, right? Because it has access to all my local system, all my local file. And so I think an interesting thing that now I'm telling my friends is to actually use Codex not just like this. coding assistant but like as an everyday everything assistant right um and especially it's super helpful for me because you know because codex can now generate designs in magic path i have all my mcp connected to codex like you know my mix panel for like how my company's doing uh stripe like all those things i can just pull up all this data in codec and then i can make just like beautiful dashboard design for my investors like that's sort of like one-to-one uh you know from idea information context to general generalization it's like It's so powerful and just Codex is like... And then I use past hours, so I don't know how much money I'm spending. You guys probably need to give me some credits. But yeah, it's incredible. That's great to hear. I mean, we see that at OpenAI too. Like virtually everyone at OpenAI is now using Codex, no matter which team they're on, from like legal to finance to HR, not just the engineers. And that's why we also... made the onboarding easier because we are noticing that outside of OpenAI too, like people like yourself who are just leveraging codecs or anything. And now that we have like all of these plugins, but also the ability, as you said, to kind of drive the browser, drive the computer, any task can be completed by codecs. Exactly. Anything you want. Yeah. Well, with all that speed that you have now with your team, like how are you thinking about like the customers and what does that unlock for you guys? People talk a lot about vibe coding, but it feels like we're far beyond that point now. Yeah, I mean, I feel like because of the velocity, we're able to get our product into customer hands way quicker. I feel like the power of having your customer just giving you feedback right away, it's just you can compare that to how product used to be built before. I worked on... lots of big tech companies in the past and I feel like that process will take months. You get the user feedback acquisition, you get all this information, you have to spend months building the thing and then you might not even know if the thing that you were building is the right thing to do. As you know, as you're building, there's always some other direction you can go. Now, because we can literally spawn a feature in a day, we have a data tester group at Magipath and so I can basically just say, hey, what do you guys think about this? They get back to me, this is great, this is now, boom, done. That is something that you couldn't achieve before. Having that feedback loop. That's much faster. It's not only just the model acting like potential customer, but it's also the model enabling you to go to your customer much quicker, which is really nice. That's great. One thing that we talked about the last time we spoke together was how we're moving to a world of supervisions. I'd love to hear you speak more about this. Yes. Actually, I believe in this so much that I built my company around it. So if I'm wrong, I'm screwed. Conviction. Conviction. But the idea is, I believe that as we go into the future of work, it's going to be less about doing the thing, but more about directing the thing. I really believe in this world of decentralized agents that you can access from everywhere. So for example, as I was saying, in Magipuff I can be on my phone at OpenClaw and I say, I'm in my bed, I have an idea at 2 a.m. in the morning. I was like, can we just change that using this content from this PRD? Because I had this idea, the agent goes, makes the thing, comes back, sends me a link, I see it. And I'm like, great, no, feedback. I don't even have to be at my desk anymore. And to me, there is something very, quote unquote, romantic and like very sort of like hopeful about that, which is like, I think as a human, right, the only thing that we have that it's in common, it's time, right? And so I feel like AI is really, to me, it's like this revenge of human against time. Like, it's like, oh, we knew that it would take us six months to... build this product now it takes us you know two weeks and that's a huge saving in time that you can spend with your family with your friends with your loved ones right and so there is an element of like really tremendous optimism that comes to me when i think about this tool because it is enabled me to i mean i worked a ton because i'm a founder but you know like i can also just like go and see my friend because i know that you know i have an agent running is doing the work that i should do well that's awesome if there's like one thing that you'd like to leave people with maybe like one feeling you'd like them to have, what would that be? For me it would be this incredible feeling of euphoria that comes from knowing that I can just do whatever I want now and just build anything I want. That's awesome. And very much like... This idea for us, going back to OpenAI's mission, that we want to build tools for all builders and for everyone, really. And now everyone is a builder. Yeah, I think the one amazing thing you guys have been able to do, especially with this new model, is really democratizing something that was so unknown and foreigner to a lot of people to now something that everybody can use. The cool thing is I grew up in Italy. I was a kid in Italy, and I feel like growing up there, I didn't have the same sort of like... possibility that somebody could be in the US, but now I feel like because everybody gets access to this incredible intelligence, we all equalize in a way, which I think is really beautiful. And thank you guys for doing that, to provide all that compute to the world. Well, thank you so, so much, Pietro, for all your work. We love having you in the community. We love having your eyes and your taste and your creativity on the models right before we are able to ship them. My pleasure. And yeah, we can't wait to see what you're going to build next with Magic Path and all of the... crazy ideas and old devices you're hacking with. And I hope this inspires people to build great things. Cool. Same. Thank you so much. Grazie mille. Grazie a te.

Pietro Schirano, Founder & CEO of MagicPath sits down with Romain Huet to talk about pushing the creative edges of GPT-5.5 and using Codex to turn ideas into software.

03:45 Images into sound
07:57 Multi-agent Codex workflows
14:34 Reviving hardware with Codex
25:27 From doing to directing