It's not the person who simply knows how to hand out tasks to programmers, analysts, marketers and contractors who wins. The one who wins is the one who understands for themselves how a system like Codex works. Codex in this case is, of course, just an example — and who can turn their own requirements into working systems. And this isn't about everyone becoming programmers, or about everyone becoming entrepreneurs. That's very important, because there's this two-sided debate going on. It's about something else.
A person gets a new way to manage their work, their time and their desires. If you need a dashboard, you don't wait for a team of analysts. If you need a CMS, you want to launch a website, you don't buy a heavyweight product and you don't wait for programmers. If you need transcription, audio processing, some research done, a CRM, reports, knowledge bases, even a system for preparing episodes like the ones I'm making now — you can start building all of it for yourself. So Codex isn't just a competitor to Claude Code, as many say, and it isn't yet another AI agent, because "AI agent" is a very worn-out topic.
An agent is an executor. Codex is interesting as an interface for creating your own software. In this episode today on ToTheMoon we'll break down the main shift, the revolutionary shift happening in the market right now. Before, your small requirements were too small for companies to assign programmers to them. Now small requirements are becoming worthy software. And the main question of the future isn't whether AI will replace programmers, but who will learn to think of their work as a system and to task systems like Codex with building it.
Hi everyone! We're on the ToTheMoon channel — tech news and insights from Silicon Valley and from around the world. The old mistake — everyone compares Codex to just another agent. The market often says now: Codex is like Claude Code, like GitHub Copilot or Google Jules. In short, various coding-as-a-system things. That's partly true, but far too narrow. And in general, if you look at the market right now, I get a very distinct feeling that everyone's trying to do what should have been done two years ago in terms of agents.
The market is lagging way behind. But if you look at the companies creating the solutions — the model companies like Anthropic, OpenAI, like xAI — they're moving in a completely different direction. Again, the topic is very much in tune with an episode we released on the channel, in which I mentioned just recently Elon Musk and his company Macrohard. What's the essence of Macrohard? It's that he says: we won't be building agents, we'll be building AI-native software, AI-native companies. So what's the old world?
You need a website — you buy a CMS; you need analytics — you buy a BI system; you need a process — you buy some software as a service, SaaS. That is, when you needed to manage tasks, people in the Russian world, in the Russian-speaking world, bought Bitrix24 or Megaplan. Back in the day I was the CEO of Megaplan for a long time. You need automation — you wait, you wait for IT or for a contractor. That is, you build your CRM, you build an ERP, you build an xRM, you build an MRP. Everything we've been doing for the last twenty, twenty-five years. And then you bend your work to fit someone else's interface.
That's a very important point. The new world. You need a narrow tool. You describe the task, the data, the screen, the actions, the checks.
And you can even describe them fairly loosely, in terms of screens. Codex assembles the first version itself. And here's the bingo! Very importantly, if you look across the whole internet, and even at what Anthropic and OpenAI are doing now in terms of integrations — they integrate with existing software. They say: here, we've integrated with Gmail, we've integrated with Slack, we've integrated with Google Drive. That is, they build these integrations inside. But if you look now, in Codex you can now build the entire piece of software, and it'll suggest to you, say, Vercel for hosting and for putting the site up, it'll suggest Supabase for the database, it'll tell you to sign up for Google Cloud, or for a really heavy-duty graph search system, something like Neo4j, and it'll say give me an OpenAI API key, or a Gemini API key, for example.
And so all of these things become building blocks, and you're not limited by them. Obviously, if you want your user to communicate through Telegram, for example, you want your user to log in to the site through Telegram — no problem. You simply connect Telegram as an additional source, but at the same time nobody stops you: if you have people who don't use Telegram, connect email — you connect email, but you don't do it through Gmail, not through Gmail, not through Yandex Mail, you do it with an independent service, the one something like Codex will suggest to you.
And here we'll move on to the important things happening in the market, because what do we have today? There's ChatGPT and OpenAI — we talk about this a lot on ToTheMoon. They're very heavily focused on testing, and since they're a research company — we also had a separate episode with Max Grigoriev from Silicon Valley, and Max worked at Google for a long time, in ML, and in real, deep ML at that. And he talked in a really fun way about what OpenAI does — it's a research company, they do research, and all of us, completely — and today that's already around a billion weekly users — we're their guinea pigs, lab rats, call it whatever you like.
Every action that happens — it's all experimentation. There's a kind of myth now that with what they're creating for B2B, they're sort of trying to copy Anthropic and going off into some other world. But if you actually look at it, they're still doing research, they— What do they do? They make, they create models, they create artificial intelligence — and the artificial intelligence of the future, not the one that exists now. They want to create what's called AGI or ASI. They want to create a different type of artificial intelligence. That's a very important construct.
And any changes they release, they always release them just as separate programs like that, and then they may shut them down. That is, they released some plugins, then shut them down, then released others, shut them down, then third, fourth versions, then they released Sora, then they shut Sora down. They release reminders functionality and all the rest of it. Then it stops working. They release Deep Research, which worked, and then the whole market picked it up too and it worked really well.
And their Deep Research has essentially degraded completely, become a useless interface, and everything already works in completely different models. And here's what happens — they release a mobile app for Mac whose interface differs, for example, from the web, or from the mobile device — they differ from time to time. For example, on Mac, ChatGPT can transcribe your calls completely in real time. Other systems can't. And then they release Codex. At first it looks like a system just for programmers. So everyone's programming in it, doing loads of stuff, and independently.
And when I first encountered Codex, I started asking: wait, if there's something like Codex, where I can come and say: please build me — I want to make software that will receive the audio from all the talks I give. I have a huge number of them, dozens of hours every week. And we'll be able to split them by role, that is, transcribe them through the OpenAI API, split up the various audio. Then I'll be able to run checks inside. And I'll also have different admins. Do all of that, set me up an admin panel. And it takes it and starts doing everything itself.
At some point, when we were making episodes, I said: listen, what's the deal? Why isn't ChatGPT merged with Codex? Because sometimes it turns out that to set a task or to do research, I do it in ChatGPT, and then I go over to Codex to execute it. But ChatGPT can't do that. In ChatGPT all the chats are isolated. Sure, there are projects, but first of all, the chats inside a project are actually isolated too. Yes, it's true, projects in ChatGPT use shared sources, there's such a concept, and a project is set up to read other chats.
But it doesn't create infrastructure, and it doesn't manage that infrastructure, the external infrastructure that's needed — for now. That's a very important point here: for now. Unless OpenAI gets its own servers for deploying systems, for deploying websites, for deploying analytics systems. And I'm sitting there, and recently ChatGPT suggested to me, it says: let's use ready-made solutions that build reports, and we'll set up that solution. I say: listen, why do I need a solution? I just need to display data here. In one project I needed to display data on the managers' calls. And I say: why would I build separate reports?
And it says: well, yes, you don't need to build separate reports, I'll build those separate reports myself. I say: but Codex gives me all the reports in table form. ChatGPT gives me all the tables. Why would I use standalone visualizers, for example, like Power BI or some others? Or, for recording all the processes and parameters and variables I have going on — at some point it becomes: why do I need Notion? Why would I keep lists there of all my episodes, all my programs, all my articles, all my various modules, my projects?
I can take some of those things out and create them with Codex. And it tells me right away: yes, of course, you can create that inside. And this difference between Codex and ChatGPT raised a really big question for me, especially in March and April. And so what's happening? There are some incredible facts. If this is your first time on the channel, this is the ToTheMoon channel. My name is Alexander Volchek, and we have a huge number of videos on the channel. We talk about tech news, insights from Silicon Valley, the newest, coolest stuff. Everything to do with artificial intelligence. So come to our channel, be sure to subscribe, don't forget to subscribe.
And we look forward to talking with you, we look forward to your comments. That's it. Our whole team reads the comments without fail, we check them, and we add that information to episodes, including the questions you're interested in. Our channel is aimed first and foremost at making sure that you, as a person, don't get lost in the development of artificial intelligence — and not lost in the abstract, but that you understand how to grow within it. It doesn't matter who you are today. An entrepreneur, a programmer, a sales manager. You're a person. The main thing is that you're a person who wants to live in the new era and keep pace with the times.
And it said — it expanded Codex and said that Codex wouldn't be just about coding, not just about programming, and that you'd be able to go beyond that. There'll be all kinds of browsers, images, memory, recurring tasks. More than 90 plugins. It's a big question whether you even need these plugins, if I'm honest. Because all the plugins that were developed in ChatGPT are kind of a marketing gimmick, right? You go in to make a plugin, and you think: what's this plugin even for? Right?
Like, how am I going to use this Booking connector plugin, for example? What's the point of that plugin? Or how am I going to use a plugin in Photoshop? Because really, if you use Photoshop, you open Photoshop and use the entire working environment there. So what's the point of that plugin? And if I want to connect via API, I'll connect to that system via API. Someone says: no, you shouldn't connect via API, this is supposedly an automatic integration. But if we look at it factually, what's the difference? I recently handed Codex an API token.
I have a stake in a real estate business, and I was missing some reporting data there. So I said, "Can I get full access to the CRM?" They gave it to me. I went in, created a token, went into Codex, dropped in the token and said, "Connect to the CRM." It connected. Now I have the ability to run any analysis of any data whatsoever. Whereas the head of sales there, to do any analysis, goes to the CRM provider. There's a CRM called Bonsa on the real estate market. So that they'll give him a full data export. Because from Bonsa's own interface you can't do that full export. And you sit there thinking: so what's the deal?
So, on April 16, besides the plugins, they also said they'd have PDFs inside, they'd have viewers for Excel and various docs inside, they'd have reviews of everything else, and multiple terminals. That's what they say. Back when I started programming, I started comparing this thing from the very beginning and asking, "What's the difference from Claude Code?"
And everyone who uses Claude tells me, "It's the same thing." But in fact, if you analyze it, Claude Code turns out to be, well, just a little terminal, doesn't it? Now, write in the comments. Maybe as of today Claude has already released a super version that lets you do this setup of this enormous infrastructure, right? Without having DevOps engineers, DevOps programmers, because before it was impossible. Programming had gotten worse, and now it's like I've gone back 23 years.
I stopped programming in 2010, I'd gone fully into management. Before 2010, even though I already had a software development company at the time, I was still programming, doing some testing, setting up all kinds of environments. Until 2020 I was still managing architecture, that was still there, but I wasn't coding anymore, because it had all become horribly complicated. Right? Plus completely different projects, different businesses and a different level of involvement. And now I've come back into Codex, and it's as if I've landed in 2002, 2003, 2004, when you yourself dragged buttons and various controls around the screen and built, I don't know, a system that generates a report from a database, from some data store.
And Codex itself has appeared in the browser too. But it turns out there's a big difference: whether you use Codex on your computer or it runs online for you. I want to advise you right away, in case someone isn't using it: I recommend using Codex on your computer. A regular computer, nothing extraordinary. Even though I've got a Mac Studio here, which is, well, a pretty serious machine. It's on 24 hours a day. Right now, in terms of programming, constantly, endlessly, 24 hours a day, three projects are being built on Codex.
I have far more active projects than that, but those are being actively built, code is being written non-stop, right? So, when this upgrade came out on May 14 — that's literally a week ago — Codex said — OpenAI said — that they have 4 million weekly active users, and they phrased it like, you can direct long-running work from your phone, look at all kinds of test results, terminal output and everything else, right. May 15 — here's where the most interesting part comes, and it's hot off the press, guys, right? On May 15, Wired wrote that OpenAI is folding ChatGPT, Codex and the OpenAI API into one, into a single product team.
Now the new power isn't being a programmer, it's thinking of your work as a system. And a new skill emerges — the ability to describe your work as a system. What data comes in, what result is needed, which actions repeat, where a person has to make a decision, where a check is needed, what screen is needed, what should be exported, what should be remembered, and so on. Although, honestly, if you look at it, do you even need to describe all of this right now? And most importantly, in six months you'll hardly need to describe such fine details at all.
The head of Anthropic says that in a year programmers won't write a single line of code. And the logic here is already very simple. That doesn't mean there are no programmers, it doesn't mean there are no people who create software. It means: why would they write code? The code being written now is going to be of cosmic quality. And of course, in terms of business logic and business architecture, the tasks being thought through will get more and more impressive. In the solutions I'm building now, I have a lot of admin panels, these administration dashboards, and there are a lot of settings and a lot of interface there.
I didn't design it at all. ChatGPT and Codex design it for me, and the results are incredible. I have a favorite example. I don't know whether you've heard it or not. It's that in the future there won't be a CRM system, I believe. There'll be a terminal where the sales manager communicates, and in that terminal — the system he talks to — if he needs to contact someone, the system tells him through that terminal who to contact. If he needs to look at some data, he sees that data in the form he wants to see it, and in whatever way is convenient for him. And he doesn't need a programmer for that.
Codex already has projects, and essentially, if you work in that project and write a task into it — specifically inside Codex, inside a Codex project, not a ChatGPT project, that's very important — it's a system that has a huge number of different documents, files, details. And you can do simply unreal things there. You're not making agents, you're not building agents. Why does everyone say you're building agents? You're not making agents at all, you're making full-fledged working software.
So once again, Elon Musk obviously said, "I'll rewrite Microsoft." And I'm convinced that companies like OpenAI, Anthropic, xAI and Google have to arrive at new operating systems within artificial intelligence. Google has some issues — it has an enormous amount of different infrastructure. Right now it's absolutely one of the top leaders in AI, very strong, no question. I've been buying its stock and will keep buying it, but it'll be hard for them to develop something from scratch, unlike Anthropic, OpenAI, xAI. Really hard.
That is, they'll have to try very seriously. So, with Codex, the technical spec stops being a document you hand to another person. The spec becomes a live conversation that immediately turns not just into a prototype, but — once again — into a working system, an absolutely working system. Now, obviously, if you take large corporations, in large corporations this won't work instantly. And in large corporations it won't be the case that all these rollouts happen tomorrow and all these, all these rollouts just land. But what's the problem here?
I also want to tell you about the problem. I sit on the boards of directors of many corporations, companies of various sizes, both small and not-so-small, in some cases as an owner, but also as a separately engaged, outside board member. What do I see? That most executives, vice presidents, directors, CXOs of various levels don't themselves genuinely touch the models at all. I mean, forget agents, forget Codex — they don't even genuinely touch raw data analysis in ChatGPT. They still hand all of that off to analysts.
But I want to tell you that about a week ago I went to bed one evening with this question: so what is the programmer going to do? Will a programmer be able to— what I'm doing? Here's something that happened to me recently. While describing one system, I told it — I immediately thought that I'd roll this system out to the sales managers in one of my businesses and, well, hand it over. I'm not an operational person there, but I'd hand it over and they'd roll it out. A training system that teaches salespeople the product — and it's a very complex product — and it also does assessment.
That is, it tests them on the product, and dynamically, very complex. Dynamic testing across different roles. And, well, a complex architecture, but visually very light. I figured these managers would immediately see — they'd immediately have errors, they'd immediately see various bugs, or they'd have suggestions. I immediately proposed to the system, understanding that they'd go to the head of sales and start passing it on to him. The head of sales would pass it to the CEO, then they'd somehow, at some point, get it to me. Because I wasn't involved operationally at all.
Someone would mention it, someone wouldn't log some error. So I said right away: let's log these errors, let's create an admin panel for logging them. Let the managers send them right away in Telegram. The system can immediately clarify those questions with the manager, as a Telegram bot. And then, from the admin panel, we'll be able to generate automatic prompts to fix these errors through Codex. And in reality, we'll set up a system ourselves that will automatically fix the errors itself and redeploy all the software itself, right? Well, update it again on the servers, so to speak. Bingo!
That's a really cool thing. And there's this story. When I was preparing for this episode, ChatGPT tells me: whatever you do, don't go into hype, and say that Codex doesn't guarantee, like, a perfect product release from a single prompt. You need checks, you need security, you need architecture, you need reviews. And they start talking to you like you're an idiot. And you're like: but I'm talking about something else entirely. I'm in no way saying that you can come in and build Windows in Codex today, or build all of Zoom. Well, unlikely.
You're unlikely to be able to build a truly working system like that, but you'll be able to do incredible things. For example, you need your own CMS for a channel. That is, to build a website.
And not WordPress or Webflow, which already exist, but an editorial system: topics, sources, something to generate Shorts, write scripts. You can easily build all of that with Codex. You need a client report generator — files, meetings, metrics? Sure, a report with conclusions, next steps. You build it in Codex. You want it to join from the outside — sometimes your meetings are in Zoom — connect it to Zoom. You want to connect, you have Google Meet meetings — connect to Google Meet. You want to connect to Microsoft Teams meetings — connect to Microsoft Teams.
Personally, I have a huge volume of meetings across different channels. They're all in one place, right. And I have absolutely no interest in using Zoom AI. What do I need it for? I mean, it's such a short, small thing. I can dump absolutely everything into a single system, all of my conversations, into Codex, and it'll build a repository out of them, make a digestible system. What's more, it'll make it super secure for me too. Want to do, say, a news-gathering thing? Gathering, for example, like what we do at ToTheMoon with various AI news — dates, companies, importance, groupings — go to Codex.
Want to build a system for tracking requests or bugs, like support — go to Codex. Want to build tracking so you can, well, do recruiting: résumés, interviews, scores, questions, feedback — do it in Codex. What's more, want to look for a job — look for it through Codex. You can set up a huge number of rules for yourself, tons of details. Someone will say: but you can do that in ChatGPT. Guys, look, in ChatGPT you can make, like, an agent, a piece. In Codex you can deploy anything at all for yourself.
You can add tons of features. Need WhatsApp here? Add WhatsApp. Need LinkedIn? Add LinkedIn. Need HeadHunter? Add HeadHunter. Want to visualize it? Visualize it. Want agents constantly going around scanning various job postings for you — separate agents. Want résumés to be collected? Résumés get collected. You've had some conversations. Want those conversations to go into this database? Throw them into the database. Want, say, your own audio processor that collects all your audio files and cuts them up?
For me, of course, the incredible thing I got was when I went in recently and transcribed using online models, greatly improved the transcription, the recognition of my voice and of my audio, and was able to train those models. The thing is, when checking this transcription, I asked Codex to create a set of audio clips for me, just put them in a folder and give them to me along with an HTML page. It did that for me. It was simply unreal. Want to generate Shorts? Why use a third-party system? Do it yourself. Look, I'm in no way calling for replacing all software.
That's very important. It's not like I want to do everything, everything in there — build ERP systems, MRP systems, xRM systems, learning systems, personal trackers for finances, for anything at all. Although I'm now going to start deploying as many projects there as I possibly can. Obviously I'm not going to build an operating system for a phone, or program my own phone as of today. But there's a chance that at some point I'll be recording videos through my own system that Codex wrote for me, because I have a huge number of issues with the systems that exist now.
I really do have big issues with them. There's a problem today. Today every system demands an endless number of subscriptions. Even in Codex, when I deploy something, I still have to buy a huge number of systems for storage and everything else. But that's minimal compared to what I usually have to do in — well, at least, I've got more than 20 or 30 businesses I'm involved in, and it's really hard to do. Even take my medical data. Everyone builds separate apps: Apple, Google. There are various trackers.
I use various trackers too, various devices. This one connects with that one, this one doesn't, this is raw data, this isn't raw data. I recently bought a heart-rate sensor, and it told me: to keep using me fully, buy a new subscription. How many of these subscriptions can you buy? So let me just write the software myself and connect to that sensor. Which I did — I talked about it in other episodes. This episode isn't about everyone becoming programmers, and it isn't about everyone becoming entrepreneurs. A person gets a new way to manage their work.
And if you're arguing with this right now and are sure that everything's already great for you, believe me, you have a big problem, like 98 out of 100 people I usually interact with. Write your comments, speak up. It's incredibly important for me to see what you're doing and which models, which tools, which software you use. See you next time!