OpenAI Codex launches macOS app: a companion to IDEs with adaptive thinking and multimodal input

[laughter]

He saw a bunch of the kick streamers and he's like, I need to start doing this.

Okay. Well, we we will dig into that more, but until then, we have the uh we have some massive news from Codeex. Let's bring in our guest from the Reream [music] waiting room. How are you doing?

It's happening.

Hello. [laughter] Nice. Nice to be here. Excited to be here.

Thanks for hopping on the show. Uh since it's the first time on your show, please introduce yourself and uh and your role and then the announcement today.

Yeah, sure. I'm Tibo. I I work here at Open and Codeex together with an awesome team. I I lead the the entire Codex team.

Okay.

And yeah, today like we launched the the Codex app on Mac OS.

Uh and it's a very capable thing. I think it's a surprising cable.

Yeah. Well, uh I mean I I downloaded it before the show. I had like five minutes and I was able to create a version of the TVPN website in the style of uh Berkshire Hathway. It feels like sort of a Studio Giblly moment for Vibe Coding in many ways, like you can just uh spin stuff up. uh what what what what are you excited about the desktop app specifically because we've seen like the whole Moltbot news it interacts with your file system but then there's also just there's just a whole host of people that are loosely familiar with programming and maybe they're comfortable with the idea of programming but at the same time when you tell them like hey set up an environment get a CLI get you know get your you know IDE set up they're just like ah I don't want to spend the hour that it's going to take me even to set up something as simple as the tools have become It's still too much. So, what were the goals? What are you excited about for uh codeex on desktop?

There are two things I'm like really excited about. It's like one, it does make it much more accessible.

Yeah.

Uh you know, you just install it, you get going, and it's like after you log in.

Uh it's delightfully simple. You know, you're just greeted by this little composer. You start chatting with it and then sort of like invites you to get things done. But also it very much leans in into the way of working that we've seen from you know technical staff at OpenAI and our users when you look at

Peter for example like open cloud just like multitasking a lot and it's easy to lose uh context when you switch like between your different terminals you don't have notifications

uh and then you don't have like the whole multimodality things if you're working with you know front ends images if you just want to use voice all of those things are like packed into the app. Oh wow.

Very much leaning in into like making people even more productive.

Yeah. I didn't even notice that the dictate buttons right there. I was uploading screenshots and that was really effective. I Peter was talking a lot about how he he texts screenshots in and and just gives more visual information to the system. Um talk to me about the models that come out of the box, how I should be thinking about model selection. It sort of defaults to 5.2 codeex medium. When would I want to switch and what are some of the different benefits? Yeah. So we just get you started with the model that we feel is the best for most people and that's done like through extensive tests, internal evaluations and just feedback that we've got and like five 52 codecs that medium has this adaptive thinking. So it's going to work fast on easy problems. It's going to like you know work harder on things that you know actually deserve taking the time. It just feels like you know a good default thing to get started with. You know we also introduced personality. So now you have like two choices. is you can go with a more friendly one and then you can you know use the pragmatic one which was like the one that we had like so far. Um a lot of people surprisingly even on the codex team has switched to the friendly one you know it's just like it's it's nice to be a little bit more uh supported and validated but of course it also we support the other uh important models 52 which was a much bigger jump I think than people expected and it's really also what we crafted the the app and the experience around. uh 52 is like really excellent at long running you know task like independently and getting it done to like a level of uh completion you know that you expect from like you know senior technical staff and that allows you to do a lot of multitasking and you know that's really what the app is built around you so you can handle like different projects different threads all running together without losing context and like sort of like efficiently you know managing a lot of things and this is sort of like a glimpse into like what the future is going to be like with agents. Yeah, I feel like a lot of people are going to come to the Codeex app on desktop with like a vague idea. You're going more proumer it feels like with this in many ways. Um, and I'm wondering how you think about planning versus implementation, walking through actually setting codecs up for success. Uh obviously like the great programmers can can just prompt better because they understand the hierarchy of files and h and what tools might be used. But for someone who's effectively non-technical, how do you think about the planning stage giving the model enough to make good decisions and actually execute properly?

Yeah, I think planning is effective in in two ways. It's uh it is effective for for the agent because it allows to be very specific about you know what you actually want the agent to go and do like you know perhaps for many hours but it also gets your own thoughts straight uh and it sort of like invites to have like a conversation around you know what is it actually that you're trying to do you know when you're sitting behind like sometimes you think like hey you know this is like a very simple uh instruction right you know just go and refactor it and you know remove this part of the back end like speed this part up and then you know maybe there actually you don't realize there like you know five different ways like with different trade-offs uh you know on how the agent could go and tackle this and like planning just sort of like helps resolve that ambiguity. So I find it like a very useful thing just as a human, you know, to go through as a process. It's like a useful reminder of like, hey, let's be crisp about, you know, what you actually want to do. But one thing that's also very important is like with the app and in general like we don't chop off like the top end of the capabilities, right? So it's like very important that you know people who are like extremely sophisticated can get you know the most incredible things done. And we're seeing a lot of adoption actually of the app like within research and you know it's fascinating walking through the office like seeing people like do the wildest things with this thing. Uh it's like very much a very it's a very capable tool.

If I look at the release schedule I see uh the chat GPT app which has a lot of functionality. Then the atlas browser now codecs on desktop. Are cl are are idees dead in your opinion? Are people just going to bring their own or do you think this grows into an ID or is that like a separate separate environment based on like what you're experiencing and like the workflows that you're seeing?

Right now it's a it's a companion it's a companion to the ID. It you know very much

uh you know can stand on its own but is enhanced by you know using an ID like occasionally. Some people will prefer like you know just being in your ID. But like it is quite evident to me that as agents just become extremely capable you just want to talk to them and they will you know get things done and then what you want to do is like you want to be able to steer them and supervise the result.

Y

and that requires something that is a very rich interaction surface and that's what we're building with the app.

How do you think about um deployment? I I I can imagine people uh vibe coding little little apps that do things for them on their computers or I I had it built an HTML page that I just opened in Chrome and was able to just see. Uh the next step is actually deploying it. Obviously, there's a whole bunch of tools that you could plug in with. How are you thinking about open ecosystem versus potentially uh offering some hosting functionality?

Yeah, we we love investing in the open ecosystem. That's something that I've been very proud of us doing as like all our code is like the majority of our code is open source. Uh you know we also support a lot of like other open source initiatives

and one of the things that we're leaning in is like the open standard around skills and we ship with a bunch of uh skills that are useful like for example you know one is like for uh versel deployments.

Sure.

And that allows you to just take what you've built and very quickly deploy it and have it you know accessible there and you know share it with others. Yeah,

it is not a native integration. We might consider that in the future, but it's already very easy to do things that are way beyond just writing code.

Yeah. Yeah. I mean, for a lot of people, even deploying to Versel is going to be ah there's going to be it's going to ask me for an API key or something, but if you're with in a chat interface, it can go back and forth and walk and you can just ask, okay, what does this mean? Where do I go? What's the web page? Where do I click? What do I copy? And all of a sudden,

that's right. You don't need to read the docs anymore. like you can just you know ask codeex to do it for you and it will do it uh to a high degree of um a high degree of quality especially given you know all the information that's packed in the scale it's like sort of like puts like the guardrails on

yeah how are you thinking about funneling uh chat GPT users to codeex on desktop it feels like

you're going to buy some ads [laughter]

yeah no but I mean there are there are worlds where you fire off a GPT 5.2 two prompt and it just writes code in the lo in the thinking step and you don't actually see the Python that it wrote and executed and the next step is okay maybe I want to run this regularly maybe I want to wrap this in you know CLI or some desktop app uh and I could imagine a flow where you know you're you're in the chat GPT app and you get to the end of the capabilities there and says hey you should you should go over here we have a we have a more robust experience for you uh have you thought about that I mean obviously on day one there's just going to be immense amount of just excitement and trial and virality but long term what are you thinking of integration between the different ecosystems

yeah I hope that with this launch we do inspire a set of people who haven't tried Codex yet to try it and realize hey you know I can really do really cool creative things with this and you know maybe like coding agents is actually for me

uh what you explained around you know the first experience in chatbt and then you have uh you know the code interpreter workflow there like where you know you don't actually need to understand the code or you use canvas and you know you have this little bit of delight there and you know you realize hey chat you can create things for me and like you know with image generation like how do we combine that and sort of like bring the magic of like coding agents to everyone but it's definitely something that we're thinking about there are considerations there around you know how you do this safely and securely given um you know the folks wouldn't actually understand the code is running under the hood

so codex for now is very much meant for, you know, technical technical adjacent audience. But then we're also thinking like, hey, this thing is like incredibly powerful and like coding can help achieve all sorts of economically viable tasks. Um, you don't need to understand that it's code under the hood, but you know, how do we bring this to more people?

Are you going to hang with uh are you hanging with Peter while he's in SF?

Uh, we're like close on Twitter and you know, Peter is just everywhere. He's hosting this uh uh con uh cloud con like for the first time which is like I heard it's like over like you know 500 participants already and stuff so I might just like drop by and say hi.

That's amazing. That's so fast. Everything he does is fast. Uh I mean what one of my big uh takeaways from the clawbot, moltbot, open claw was just that uh mobile the importance of mobile and and people going on telegram or signal or WhatsApp and and and texting an interface that could write code for them execute things and interface with some hardware. How are you thinking about uh the mobile experience in two codecs like where how how the integration loops hap happens now that there's a desktop app there's obviously the chatbt mobile app that has massive installation uh what are the what are the pros cons what are you excited about there

yeah we're very excited about the mobile experience it's it'll always be my dream of you know I can start a task I can talk to codeex from anywhere and then I can just like

walk away follow my phone and steer it and this This is definitely something that's going to come.

Cool.

Uh we are very much optimizing you know for you know professional software engineers and people are like you know really excited about getting super productive. So it felt like bringing first an amazing experience on Mac OS you know felt like the right thing priorities wise.

Yeah

it is something that will come and I think will delight people uh when it's there. And this is also why you know we we keep calling everything Codex. Codex is our agent. Uh it is how you you get things done. Uh, and then you know whether you interact through it like via the CLI or in web or in in the OS like under the hood is the same agent and like we're going to connect it all at some point.

Yeah. Is it time to buy an extra monitor? How many monitors do you run?

Right now on my laptop I only have one.

Oh, but the app is like delightful. Like you can actually like follow things like much more. So don't feel like the need to have like

five monitors. Uh at home I I do have three and like you know very as well not going to lie here.

Yeah. May May maybe that Dell the the 65 in 6K monitors in the future. We're we're very excited about that. Uh anyway, congratulations on the launch. Thank you so much for stopping by the show. Good to meet you and we'll talk to you soon.

Cheers.

Thanks for having me. Thanks.

Goodbye. [applause]

Let me tell you about vibe.co where DTOC brands, B2B startups, and AI companies advertise on streaming TV, pick channels, target audiences, and measure sales just like on Meta. Um

where to go next? Dylan Abberriscato said that Succession is a documentary because apparently uh the next CEO of Disney could be the uh the theme park division chairman. I didn't get the reference. [laughter] Have you seen Do you get the reference from Succession?

It's uh Tom is it's Tom who's running the parks.

Tom runs the parks department.

Let me know. Let me know if I

I I got to finally get through all of Succession. Um but uh oh in other news there's uh I see this post that you're on. Uh bro spent $50 billion buying Bitcoin over the past five years and is now under

talking micro strategy.

He didn't even mention who it is

and people and people just know that it's micro.

I mean not a lot of people that are non maybe state entities have have bought this much.

712,000 bitcoins. Wow.

I think I think his his um his average entry is in the 80s.

Yeah. And where's

Bitcoin is up today? It's 78,000, but obviously that's nowhere near where it was when it was up at 120. So a significant uh a significant selloff. Um but they're there.

I have some extra context on the

OpenAI news. So um yeah, apparently uh Opening with Crisis and Grock to do like custom chips and then when uh when Grock got bought by Nvidia like that stopped the the talks.

Um I think it's just because they're too slow, right? You've seen run talk about this where codeex is is you know somewhat slower than cloud code or whatever. Uh even though if the model's actually better the experience can be you know worse sometimes.

Yeah. No no I mean like there like the models can get better but they're already really really good. Like most of the deep research reports that I get back are amazing. They're they they job mission accomplished. Like you did it. You you did the research. You compiled it and um and I got my information but I had to wait 20 minutes. cut that down to two minutes and I'll probably use it 10 times as much and I'll probably pay time 10 times as much for that. And same thing with coding. You know, there's so many memes about watching brain rot videos while uh while you're waiting for codeex or cloud code or whatever you're using to come back and answer you. Speed it up and you're going to have a much faster adoption path.

Chat has confirmed that it is Tom Wgsans in the amusement park and cruise division. Then he got moved up to ATN news.

Okay. Um the uh one more thing on the uh debt side of things. Bit mine, this is Tom Lee.

Yes, former guest has unrealized losses of 6.6 billion. Now on track to become the fifth largest documented principal trading loss in history if sold. Of course, he's not selling. He remains bullish. He's he's really [laughter] testing his diamond hands. Uh it's at 66% of the size of AR Archos. It was