OpenAI Is Changing Work: Why Wait for IT When Codex Builds Your Software? A Breakdown of Codex
Who wins when small requirements turn into working software without programmers — and why should Codex be understood not as another agent but as an interface for creating your own software?
The host takes Codex apart not as a rival to Claude Code and not as another agent but as an interface for creating your own software: you describe the task, the data, the screen and the checks, and Codex builds a first version, itself proposing Vercel, Supabase or Neo4j as building blocks. The chronology of updates — on April 16 Codex went beyond code, on May 14 it appeared in ChatGPT Mobile with four million weekly users, on May 15 ChatGPT, Codex and the API were folded into one product team — runs alongside personal cases: a CRM connected by token, and a sales-training system that collects its own bugs through Telegram and redeploys itself. The main idea is framed as a new skill: not being a programmer but thinking of your work as a system. The frame stays sober: Codex will not build Windows or Zoom, not all software needs replacing, and large corporations will not rebuild themselves tomorrow.
What to watch for
Key takeaways
The winner is not the one who merely knows how to hand tasks to programmers, analysts and contractors, but the one who understands how a system like Codex works and turns requirements into working systems. It is not about everyone becoming a programmer or an entrepreneur: a person gets a new way to manage their work, time and wishes. Small requirements used to be too small for companies to assign programmers to; now they become decent software.
The market puts Codex in a row with Claude Code, GitHub Copilot and Google Jules — partly true but far too narrow: an agent is an executor, while Codex is interesting as an interface for creating your own software, and the model companies, as Elon Musk's Macrohard shows, are heading toward AI-native software rather than agents. The old world bought a CMS, BI and SaaS, Bitrix24 or Megaplan, and waited for IT while adjusting its work to somebody else's interface. In the new one you describe the task, the data, the screen, the actions and the checks, and Codex builds a first version, itself proposing Vercel, Supabase, Google Cloud, Neo4j and the OpenAI or Gemini API as building blocks.
Citing Max Grigoriev, the host reminds us that OpenAI does research and wants to create AGI or ASI, and its roughly one billion weekly users are its test subjects. Any change is released as an independent program that later gets shut down: plugins, Sora, reminders, while Deep Research, in the host's assessment, has degraded into a useless interface. Meanwhile the Mac app transcribes calls in real time, which other systems cannot — and against this background Codex arrives.
The host does research and task-setting in ChatGPT and goes to Codex to execute, because ChatGPT cannot: its chats are isolated even inside projects, and it neither creates nor manages external infrastructure — as long as OpenAI has no servers of its own for deploying systems. When ChatGPT suggested a ready-made reporting solution, the host asked why he needed Power BI or Notion if reports and lists can be built in Codex — and was told that they can. This gap troubled him especially in March and April.
OpenAI expanded Codex: browsers, images, memory, recurring tasks, more than ninety plugins, PDFs, visualizers for Excel and documents, review and several terminals. The host considers ChatGPT plugins a marketing trick — it is unclear how to use a connector with Booking or a plugin in Photoshop, whereas connecting through the API really works. Having obtained a CRM token in a real estate business, he handed it to Codex and now runs any analysis, while the head of sales has to ask the Bonza provider for a full export because the interface does not offer one.
Everyone who uses Claude told the host it was the same thing, but in his experience Claude Code is "a little terminal", while Codex lets you configure a huge infrastructure without DevOps engineers, which used to be impossible; he allows that a new version of Claude may already do this and asks for comments. He himself stopped programming in 2010, managed architecture until 2020, and now in Codex feels as if it were 2002–2004, when he dragged buttons around the screen himself.
On May 14 Codex appeared in preview in ChatGPT Mobile and in the browser, and OpenAI named four million weekly active users: from the phone you can direct long-running work and check test results. The host still recommends an ordinary computer — his Mac Studio builds three projects on Codex 24 hours a day. On May 15 it emerged that ChatGPT, Codex and the OpenAI API are being folded into one product team, and Greg Brockman, Sam Altman's partner, speaks of a shared experience for these systems.
The new skill is describing your work as a system: what data comes in, what result is needed, which actions repeat, where a person decides, where a check is needed, what screen, what gets exported and what gets remembered — though in six months, the host believes, such details will no longer need describing. The head of Anthropic says that in a year programmers will not write a single line of code — not because they will be gone but because the quality of code will be cosmic. The favorite example: there will be no CRM systems in the future, only a terminal in which a sales manager talks to the system and sees data the way that suits him.
Projects in Codex are not ChatGPT projects: a system with an enormous number of documents and files in which you build full working software. Hence Elon Musk's "I will rewrite Microsoft" and the host's conviction that OpenAI, Anthropic, xAI and Google must arrive at new operating systems, with Google, whose shares he buys, finding it harder to build something from scratch. The technical specification stops being a document for another person and becomes a conversation that turns straight into a working system, though in corporations this will not work instantly: most executives, by the host's observation on boards, do not even touch raw data in ChatGPT.
While describing a sales-training system with dynamic role-based testing, the host built in an admin panel for recording errors from the start: managers send them through Telegram, a bot clarifies the details, the panel generates automatic prompts for fixes through Codex, and then the system fixes the errors itself and redeploys itself. During preparation ChatGPT asked him not to slip into hype and reminded him about checks, security, architecture and reviews; the host agrees but stresses that nobody promises to build Windows or Zoom in Codex today.
You can build your own CMS for a channel instead of WordPress or Webflow, a client report generator, a single store of meetings from Zoom, Google Meet and Microsoft Teams, news collection, ticket tracking, recruiting tracking or a job search with WhatsApp, LinkedIn and HeadHunter, an audio processor and a shorts generator. The host himself improved the recognition of his own voice and got audio clips with an HTML page from Codex. He is not calling for all software to be replaced, but his main complaint about today's systems is endless subscriptions: even a cardio sensor demanded a new one, and he wrote the software himself, answering those sure everything is fine that 98 out of 100 people around him have the same problem.
What this episode is about
A solo ToTheMoon episode with Alexander Volchek about OpenAI's Codex. The host opens with a thesis: the winner is not the one who knows how to hand tasks to programmers, analysts and contractors, but the one who understands how a system like Codex works and turns their requirements into working systems. It is not about everyone becoming a programmer or an entrepreneur: a person gets a new way to manage their work — a dashboard without a team of analysts, a CMS without a heavy product, transcription, a CRM, reports and knowledge bases built for themselves. Small requirements used to be too small for companies to assign programmers to; now they become decent software.
The market's old mistake is to compare Codex with yet another agent — Claude Code, GitHub Copilot or Google Jules: partly true but far too narrow, and the market itself, in the host's view, is trying to do what should have been done two years ago, while Anthropic, OpenAI and xAI move in a different direction. He recalls Elon Musk's Macrohard — not agents but AI-native software and AI-native companies. The old world bought a CMS, a BI system, SaaS, Bitrix24 or Megaplan, where the host was CEO for a long time, waited for IT or a contractor and adjusted its work to somebody else's interface. The new world describes the task, the data, the screen, the actions and the checks, and Codex builds a first version, itself proposing Vercel, Supabase, Google Cloud, Neo4j, the OpenAI API or the Gemini API as building blocks.
Next comes OpenAI as a research company. Referring to the episode with Max Grigoriev, the host says that for OpenAI its roughly one billion weekly users are test subjects: the company is building AGI or ASI and releases changes as independent programs it later shuts down — plugins, Sora, reminders, a degraded Deep Research. Against this background Codex arrives, at first looking like a system for programmers only. The host describes his first task — software that takes the audio of all his talks, transcribes it through the OpenAI API and brings up an admin panel — and the question that troubled him in March and April: why ChatGPT and Codex are not merged, when research happens in one and execution in the other, while ChatGPT's chats are isolated and do not manage external infrastructure. Hence also the question of why Power BI or Notion are needed if reports and lists can be built in Codex.
The chronology of updates. On April 16 OpenAI expanded Codex beyond coding: browsers, images, memory, recurring tasks, more than ninety plugins, plus PDFs, visualizers for Excel and documents, review and several terminals. The host calls ChatGPT plugins a marketing trick — what really works is connecting through the API: he handed Codex the token of the Bonza CRM from a real estate business in which he holds a stake and got any analysis of the data, while the head of sales has to ask the provider for a full export. On May 14 Codex appeared in preview in ChatGPT Mobile and in the browser, and OpenAI named four million weekly active users, but the host advises working on a computer: his Mac Studio runs three projects around the clock. On May 15 came the news that ChatGPT, Codex and the OpenAI API are being folded into one product team, with Greg Brockman speaking of a shared experience.
The difference from Claude Code gets its own chapter: everyone who uses Claude says it is the same thing, but in the host's experience Claude Code is a terminal while Codex brings up infrastructure without DevOps engineers; he allows that a new version of Claude may already do this and asks for comments. He recalls that he stopped programming in 2010, managed architecture until 2020, and now in Codex feels as if he were back in 2002–2004, dragging buttons around the screen himself. The central idea is the new power: not being a programmer but thinking of your work as a system — describing what data comes in, what result is needed, which actions repeat, where a person decides, where a check is needed, what screen is needed, what gets exported and what gets remembered. The head of Anthropic says that in a year programmers will not write a single line of code; the host reads this as "why should they write code", and his favorite example is that there will be no CRM systems in the future, only a terminal in which a sales manager talks to the system.
Projects in Codex are not ChatGPT projects: a system with an enormous number of documents and files in which you build not agents but full working software. Hence Elon Musk's "I will rewrite Microsoft" and the host's conviction that OpenAI, Anthropic, xAI and Google must arrive at new operating systems, with Google, for all its infrastructure, finding it harder to build something from scratch. The technical specification stops being a document and becomes a living conversation that turns into a working system; in corporations this will not work instantly, and as a board member the host sees that most executives do not even touch raw data analysis in ChatGPT. He describes a sales-training system with dynamic testing, for which he built in an error admin panel, a Telegram bot, automatic fix prompts and redeployment from the start — and agrees with ChatGPT, which asked him not to slip into hype: you cannot build Windows or Zoom in Codex today.
The finale is what you can build: your own CMS for a channel instead of WordPress or Webflow, a client report generator, a single store of meetings from Zoom, Google Meet and Microsoft Teams without Zoom AI, news collection, ticket tracking, recruiting tracking or a job search with WhatsApp, LinkedIn and HeadHunter, an audio processor and a shorts generator; the host himself improved the recognition of his own voice and got audio clips with an HTML page from Codex. He is not calling for all software to be replaced and will not write an operating system for a phone. His main complaint about today's systems is endless subscriptions: even a cardio sensor demanded a new one, and the host wrote the software himself. The episode ends where it began: a person gets a new way to manage their work, and to those sure that everything is already fine with them he replies that 98 out of 100 people he deals with have the same problem.
The value of the episode is in changing the question: not "will AI replace programmers" but who will learn to describe their work as a system. The host shows it not on demos but on his own projects — from transcribing his talks to a CRM connected by token and a system that collects its own bugs — and at the same time draws the boundaries: Codex will not build Windows, corporations will not rebuild themselves tomorrow, and some subscriptions will still have to be bought. The chronology of updates shows OpenAI heading the same way: Codex, ChatGPT and the API are being assembled into one whole.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 52 segments: 0 identified, 0 mixed, 52 probable, and 0 unresolved.
Read transcript on a separate page
Loading…