OpenAI Dots: AI You No Longer Have to Explain Everything To? A New Way to Work with AI
What can Dots really do after two weeks of use — and how does a personal agent that sees all of OpenAI's infrastructure differ from an ordinary chat and from agents inside chats?
Dots runs on GPT-6 Astra and sees every project in Codex: it creates its tasks in Codex, works with a cloud Linux and one personal computer, and the strength of the answer depends on the model's mode — and 5.6 Sol Extra High in ChatGPT and in Codex, according to Alexander Volchek, are not the same thing. In his cases Dots finds animals on night cameras, opens cloud services without an API through a browser, checks the whole mailbox and in two seconds breaks down a company's car expenses, adding Uber and Lyft rides on its own and separating personal spending from business. At the same time it remains a half-working beta: it handles repeated tasks badly, does not create tasks in ChatGPT and once wrote people a new email instead of replying. OpenAI's main omissions, in the host's view, are that Dots cannot talk to each other outside corporate accounts and that the interface is still a chat, although it should become a space with a single entry point. For Dots to work it needs four properties of memory — storing, retrieval, updating and scope — and broad access, and the place to start is one main system.
What to watch for
Key takeaways
Alexander sums up two weeks with Dots — an agent system for working with OpenAI's infrastructure that the company keeps updating — and promises to compare it with other solutions on the market and show the difference in how the models work.
According to OpenAI's documents, Dots is handed a goal and an area of responsibility and keeps working between conversations. At launch it is available only in ChatGPT Pro; the host expects it to come to Plus, but hardly to the free versions.
Its main feature is access to all your projects in Codex. That is why the host strongly advises starting to use Codex even for those who do not do development: this is how you join a new understanding of what AI is.
Codex and Claude Code spin up multi-agent threads themselves, while Dots is a single window above them into which you can dictate different tasks in five minutes: from a question about a capital to a doctor's reminder and monitoring a temperature sensor.
The mode decides: in ChatGPT from instant to pro, in Codex from low to ultra. Dots picks the mode itself, and 5.6 Sol Extra High in ChatGPT and in Codex, the host says, are not the same thing.
Dots has never created a task in ChatGPT with Pro mode for the host. It creates tasks in Codex — mail, the network, settings, files — and works with a cloud Linux with Chrome and one personal computer, so Codex has to be set up for it.
Everything connected to Codex — ElevenLabs for speech recognition, sending SMS through a computer — is available to Dots too. But the future, in the host's view, belongs not to integrations through plugins, MCP and API but to intelligence that itself understands how to connect.
Codex sees some of the twenty-five to thirty home cameras directly, while cloud services without an API can be opened in a browser and the downloaded videos analyzed. Dots, meanwhile, is a half-working beta, and repeated tasks go badly for it.
Mail is stored not only at Google but also in iCloud, on the computer and the phone. Alexander considers OpenAI one of the most secure companies, connects all his mail to Dots, gives banks read-only access and uploads statements into Codex.
Codex broke down rentals, leases, loans, fuel, parking, garages and repairs, suggested adding Uber and Lyft rides on its own and showed the cash flow and P&L by month — faster, the host says, than any accountant could.
While closing a month, it noticed that two more people — his daughters — had been with the host on a business trip, and asked about a spa at the hotel. Hence Alexander's forecast: the bookkeeping of 99.999% of legal entities will move to Codex, Claude Code, Grok and Cursor.
Access to his wife's AI for their joint business is possible only through her Dots, and such a link exists only for corporate accounts. The host considers this an oversight by OpenAI and Anthropic or a political bet: OpenAI has been oriented toward B2C from the start.
In Alexander's view, Dots should look like Spaces — with chats, shared pages and other people's Dots — and be the single entry point. Meanwhile OpenAI sends him, a Pro user with thousands of chats, emails about images he does not use.
According to a friend of the host, Elon Musk sees the future of AI through hardware. Alexander is building his own version: microphones, a separate computer with cloud Codex and a Roborock robot vacuum that can be turned into a security device.
Storing, retrieval, updating and scope: the system should itself understand whether you went to Paris and not confuse a question about an acquaintance's health with yours — and for that you need to use it constantly and give it broad access.
Instead of Android scripts that send a call to voicemail, the Dot should write on its own that the host is in a meeting and ask how urgently he should be interrupted. If Dots could talk to each other, many family chats would disappear.
Dots began checking the whole mailbox and pulls emails out of spam, but once it wrote people a new email instead of replying. The wider the access — to the phone, to other systems — the sharper the question of where to let it in.
The host advises a casual user to settle on one main system and one additional one and set them up to work together. Dots does not work with Claude yet, and Alexander calls the voice Gemini in his car an unbelievably dumb system.
What this episode is about
A ToTheMoon special episode in which Alexander Volchek sums up two weeks of working with Dots, the personal agent OpenAI presented at DevDay. According to the company's documents, Dots is handed a goal and an area of responsibility and keeps working between conversations; at launch it is available only in ChatGPT Pro. The host explains why this is not just another chat: Dots is a single window over all your chats in Codex, ChatGPT and ChatGPT Work that sees your entire OpenAI infrastructure — and he advises starting to use Codex even for those who do not program.
Dots must not be confused with the agent systems inside chats that Codex or Claude Code spin up: it is a level above them. It runs on GPT-6 Astra, but the strength of the answer depends on the mode — from instant to pro in ChatGPT and from low to ultra in Codex — and, according to Alexander, identically named modes in ChatGPT and Codex are not equal. Dots creates its tasks in Codex, works with a cloud Linux and one personal computer, and the promised calling of Dots from chats does not work in ChatGPT yet.
The second part is about cases. Through Dots the host asks which animals were caught on his night cameras, and he opens camera cloud services without an API through a browser. He connected all of his mail: to a partner who was afraid to do the same he replied that mail is already stored at Google, in iCloud and on the phone, while OpenAI, in his view, is one of the most secure companies. From bank statements and receipts in the mailbox, Codex broke down a company's car expenses in two seconds, added Uber and Lyft rides on its own and separated personal spending from business — hence the forecast that most companies' bookkeeping will move to such systems.
Next comes what OpenAI missed. Dots cannot talk to each other outside corporate accounts, although for the business he shares with his wife the host needs access to her AI. The interface, in his view, should be a space like Spaces with a single entry point rather than a chat; and OpenAI itself sends a Pro user emails about images he does not use. Elon Musk, according to a friend of the host, does not see chat as the central system of AI at all — and Alexander is building his own: microphones, a separate computer with cloud Codex, a robot vacuum as a security guard.
The finale covers the four properties of memory (storing, retrieval, updating and scope), an assistant that should answer his wife itself while the host is in a meeting, and the limits of access: Dots checks the whole mailbox but once wrote a new email instead of replying. Alexander advises a casual user to choose one main system and one additional one, criticizes the voice Gemini in his car and promises to let Dots further into his projects and smart home right after the recording.
The episode is useful because it strips Dots of the marketing frame of "a chat for standing tasks" and shows it as a layer over all of OpenAI's infrastructure whose strength depends on what is connected to Codex and which model mode is chosen. The cases with cameras, mail and company expenses show where such an agent already saves hours, while the failures the host admits himself — repeated tasks, a new email instead of a reply — are a reminder that it is a beta. The main idea is that Dots should become the single entry point and be able to negotiate with each other; it can be tested by whether OpenAI opens links between different people's Dots.
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 94 segments: 94 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…