Have you seen the joke—almost a spam joke—about a billboard that appeared in San Francisco advertising AI, except that AI stands for Average Intelligence? A real person simply answers questions that arrive in the chat instead of a chatbot and charges money for the service. It uses the letters AI, but not Artificial Intelligence. It began as a meme, yet it became popular enough that, according to the reports, the creator had to close registration because too many people started contacting him.
Hello, everyone. This is ToTheMoon with news—
Creative.
—from the United States, Silicon Valley, and around the world. And especially from Silicon Valley. I have just returned from a trip—
It feels as though San Francisco has no billboards left except advertisements for AI.
Exactly. We were in New York, and I landed during the night. New York has AI billboards, perhaps one out of every hundred or two hundred, so you do encounter them. But when you approach the city, there is nothing like the volume you see in San Francisco. In San Francisco the billboards are everywhere, and they have become remarkably uniform. Most of the time you cannot even tell what is being advertised: just one interchangeable phrase after another. At first it looked interesting because an occasional campaign was genuinely distinctive.
Now almost all of it looks the same, and I often cannot understand what is actually behind the message. I want to begin today with Meta. Mark Zuckerberg announced a major new push by Meta into artificial intelligence yesterday—specifically into AI for programming. Meta introduced a programming agent called Muse Code. The company already has the Muse language model, and Muse Code is now available in beta. It also introduced Muse Spark, including version 1.2 of the language model.
The version number itself is not important to most people; what matters is that the product now exists. Muse Code is being described directly as a competitor to Codex, Claude, Gemini Code Assist, and Gemini CLI. Gemini's programming tools are not discussed as widely as Codex or Claude. Interestingly, the comparisons often do not include Grok, even though I suspect Grok Code has many more users than Muse Code. Comparing these systems has become extraordinarily difficult. Even the most expensive models will give you a comparison, and you immediately have to ask: “But does the other version not have exactly the same feature?” Muse Spark is positioned around an idea I like: it should not merely answer questions but independently complete large engineering tasks.
Yet Claude Code and OpenAI Codex have been actively pursuing that same idea for roughly half a year. Meta emphasizes a multi-agent approach as the most interesting part of the update. It is worth explaining that feature and asking whether a new serious operator can still emerge. We recently discussed whether Apple might recover in AI; Tanya and I touched on it in one of the previous lunch discussions.
Yes, we did.
Meta says that instead of using one AI, the system runs several agents simultaneously. One studies the architecture, another writes the backend, a third handles the frontend, runs tests, and searches for bugs. As an active user of Claude Code and Codex, I read this and imagine an agent writing the backend, frontend, and architecture in parallel. But why do those tasks need to be performed in parallel at all? Then I encounter a lecture by one of Claude Code's principal engineers.
Its message is that engineers should no longer manage individual agents; every engineer now has a thousand agents. My reaction is: a thousand agents? If those thousand agents were genuinely working at full capacity, especially on Anthropic's or OpenAI's infrastructure, imagine how much code they would already have produced. Meta has arrived with essentially the same positioning, and I am not convinced it is the right one. People are unlikely to use every system simultaneously.
Ilnar, what do you think specifically about Meta's move into programming? Can it become a serious operator in terms of capturing meaningful market share? Anthropic has not surpassed OpenAI in total users—it does not have a billion of them—but it has become an extremely serious OpenAI competitor in AI. Today AI includes everything from a simple text field to systems that build and program other systems. Can a company such as Meta enter this market in a genuinely substantial way, or is that unlikely?
My main impression is simply that the market is becoming more crowded. There are still not that many companies: several in the United States and their Chinese counterparts, so there is room for additional players. Meta will probably become a niche operator. I strongly doubt that it will capture a major share without a serious advantage. One possible advantage would be substantially more compute allocated to the product, which would mean higher usage limits. When users hit the limits at other companies, they might switch to Meta's operator.
Another advantage could be fewer restrictions. Among people I know, for example, Grok is mainly used by those who value its lack of restraint. It applies far fewer restrictions to content generation and can produce material that neither OpenAI nor Anthropic will generate. That difference gives it a particular user base. Something similar may happen here. Meta will take some share, but I do not expect it to become an OpenAI-, Anthropic-, or even Google-level player—at least not before it releases a much stronger next Gemini-equivalent product.
There is also an image component. Major Silicon Valley companies without their own AI models now look somewhat behind the market, and that perception probably affects their stock. This announcement may therefore be aimed as much at showing “we are still in the game; our shares are still worth buying” as at capturing the market itself. A niche position and a modest share seem realistic.
Meta's stock has shown no meaningful reaction to this very recent announcement. Over the last five days it has remained almost flat, and over the month it is slightly down. Apple is different: it briefly crossed a five-trillion-dollar valuation and once again became the world's most valuable company. Whatever people say, Apple has been doing remarkably well as an investment story. It is important to add that Apple crossed the level and then fell. The stock peaked at roughly $344 and is now around $310, a decline of about ten percent—roughly five hundred billion dollars in market value.
Yet the fact that such a rise and pullback are possible reinforces the broader direction: companies can eventually be worth more than ten trillion dollars, but reaching that scale takes time. Technology stocks often rise, correct, and continue upward over a longer horizon. I would like viewers to share what they think about the market. Opinion is sharply divided: some believe in the growth, others think the market is extremely overheated, some say investors were deceived, and others imagine investors control the process.
There may be no single perfectly correct view. I always look at the long term. With Apple's infrastructure and capital, and Meta's infrastructure, capital, and Mark Zuckerberg's persistence, both have a chance. Meta has already spent an enormous amount of money, much of it perhaps wastefully. It missed part of the race and lost a great deal. Yet it may still establish itself. For now, independent operators such as Anthropic and OpenAI—rather than divisions inside enormous legacy corporations—have created the strongest systems used by large numbers of people to build the next generation of products.
The competition between Anthropic and OpenAI is now extremely intense.
One recent report said Microsoft Copilot was losing many users, in part because people were moving to Codex. A few weeks ago I asked someone, “Why are you still in Copilot? Move to Codex.” They are different environments. Copilot was originally based on OpenAI technology, was it not, Ilnar?
It was at one point, yes.
It was, yes.
I think there may have been some changes later.
They may now offer a choice of models. But a truly strong system is not one that gives me programming as a separate side tool. It combines chat, collaboration, and programming in one environment. The companies still do not know how to build those interfaces. I do not understand why. By now they could have made much more progress toward a high-quality unified product. OpenAI took on that task several months ago, with Greg Brockman leading the direction.
So far, I consider the result a failure. The integration between Codex and ChatGPT looks strange. In the main Codex application, the top navigation includes New Chat, Pull Requests, a newer Sites feature for building websites, and older sections associated with scheduled tasks and previous plugin use. Project management and even project naming remain extremely awkward. ChatGPT itself seems lost. I often take a screenshot, share it through the application, and write “Give me the details.” I can end up with forty chats all called “Give me the details.” A year and a half or two years ago, OpenAI began assigning sensible human-readable titles instead of using only the first words of a conversation.
This is AI; it can create an accurate chat title in a second. OpenAI expanded search beyond chats to many other objects, but in my experience it became worse. To give them credit, they finally launched broader search and made it more fundamental, yet the entire interface still feels childish.
That brings us to Apple and a conflict that is unusually important even by Silicon Valley standards: Apple has filed a lawsuit involving OpenAI. The case concerns an employee named Chang Liu, who spent eight years as a senior engineer on iPhone electrical systems and left for OpenAI at the beginning of this year. Apple says he failed to return his work laptop. That alone sounds extraordinary. Apple is famously secretive; I live in Silicon Valley and know people who work there, and they almost never discuss the company.
Its security culture is radical. Yet this former employee allegedly kept the laptop, then used a former colleague's laptop. While working at OpenAI, he allegedly exploited an Apple authentication flaw, gained access to shared network folders, downloaded dozens of files about unreleased products over several weeks, and advised an Apple employee which confidential materials to study before an OpenAI interview. It reminds me of situations in different companies I have been involved with.
You leave a business, sell it, end a partnership, or step away from management, and— —someone leaves access credentials behind. Companies often treat information security only relatively seriously. Then you look at Apple and ask how any of this was possible. Apple is one of the most secretive companies in the world. How could a former employee connect remotely? There is also a second former employee, Tang Tan. He had worked at Apple for approximately twenty-four years and held a very senior role as vice president of product design for the iPhone and Apple Watch.
He later became the head of hardware at OpenAI. Apple alleges that before leaving he forwarded confidential information, including supplier information; asked candidates to bring real Apple components to interviews for demonstrations; advised future employees not to tell Apple that they were moving specifically to OpenAI; and used internal codes and other internal practices. Apple has therefore made extremely serious accusations against OpenAI concerning the theft of information.
The court filings describe potential forensic oversight of OpenAI. Elon Musk reposted the report that OpenAI might be placed under forensic supervision and wrote, “Cannot trust OpenAI.” That amused me, because when you know some of the local stories about corporate culture around Musk and certain employees, you want to ask: do you not do similar things yourselves? Apple is now asking for restrictions against OpenAI and its hardware division. This could become a major shift. If OpenAI eventually releases a device, even hypothetically, the conflict may matter directly.
Do you remember the device I bought that listened to my conversations? What was it called, Ilnar?
Limitless.
Yes, Limitless. Mark Zuckerberg's company acquired it. My device eventually died somewhere in one of the rooms at home.
I bought one too, Sasha, on your recommendation.
Mine first stopped working at home, although I later washed it perhaps five times. I think it had already died before that, and somewhere inside Mark Zuckerberg's organization the product itself also effectively disappeared. Zuckerberg has been willing to spend two or three hundred billion dollars on AI. The larger point is that if OpenAI creates a device and solves the problem of continuous communication and perhaps ambient observation, it could take over part of the interaction that now happens through the phone: WhatsApp, Telegram, browsers, and other applications.
The browser market is already weakening. OpenAI and Elon Musk have finally begun saying publicly that browsers may disappear. That was obvious for a long time. Safari could lose a substantial part of its position as well. It is enormously widespread among Apple users, and Apple is the leading device seller in the United States. What is your sense of this conflict? Based on the history of U.S. lawsuits, perhaps very little will happen in the end. Yet the confrontation itself is serious and shows how deeply the companies are afraid of losing control.
OpenAI has already published an excellent response on its website. The tone is openly mocking: Apple is described as a magnificent company famous for its attention to detail, followed by an explanation that Apple made so many errors in the complaint that OpenAI was surprised. The response then lists facts that, according to OpenAI, occurred differently, including a different chronology—
Yes. In simplified terms—
Explain it for the viewers; it is worth hearing.
I strongly recommend reading the primary response on OpenAI's website. It is entertaining, includes links, and even contains screenshots of correspondence, some of it anonymized. One example is the point you just mentioned. An OpenAI lawyer has an Asian surname, and when Apple's lawyers sent certain documents, they sent them to the wrong person. The intended lawyer never received them and therefore did not respond. Apple later tried to rely on the lack of a response. It acknowledged that the message had not actually been delivered only after OpenAI directly exposed the error.
We probably will never know whether that was deliberate or simply a confusion between similar names. It may have been an ordinary mistake, or it may have been a tactic used to set the litigation in motion. OpenAI also argues that it is common at Apple for former employees to remain in contact with current employees and for some access to remain in place. According to the response, this was not a unique exceptional event but a recurring practice. There are many other interesting details.
The most memorable part for me was the opening: Apple is supposedly the company we all love for its attention to detail, yet it distorted facts in this way. It was sharp and effective trolling.
The complaint does look strange: documents apparently went to the wrong lawyers, referred to the wrong employees, and used incorrect dates. That is why I doubt the dispute will produce consequences on the enormous scale implied by the accusations. There may be fines, temporary restrictions, or narrow injunctions, but probably not the dramatic outcome suggested by the headlines. We have seen similar situations before. OpenAI was once ordered to preserve data rather than delete them.
I sometimes wonder whether an arrangement like that can also serve both OpenAI and the government. OpenAI can say that it tries to protect user data, but a court order—perhaps originating in New York—requires the company to keep the information. Why not retain it indefinitely? A government order says the evidence must be preserved, so data that would otherwise be deleted after thirty days remain forever.
We still have not seen a clear discussion of whether that material has begun to be deleted or whether the order remains in force. Another recent case involved Perplexity scanning Amazon; the lawsuit was dismissed. The underlying question was whether Perplexity could scan the service and make purchases on behalf of a user. This leads to a much broader boundary: when do we say that an action was performed by AI rather than by the person who instructed it? We discussed an example involving X, which removed tens of thousands of accounts because they automatically posted content.
Where is the boundary between paying a human to post content automatically and assigning an AI to do it? The same applies to purchases. If AI buys something on your instructions, how is that fundamentally different from hiring a person to buy it on your instructions? May AI click a CAPTCHA? Why not? What if the AI is assisting a person who cannot use their hands, or a person who simply wants to operate the system by voice? How does such a person click a CAPTCHA by voice? A CAPTCHA is supposed to test whether a visitor is human.
But if technology is meant to improve human life and expand what people can do, we will eventually have to abandon some of these tests. Even with something as apparently simple as a CAPTCHA, the future is unclear.
With AI autoposting, Sasha, the issue is closer to spam. A person can manually send a message every ten seconds—copy it, paste it, open the next window, and continue. An API or bot can do the same thing far faster. In principle, someone could hire several people to work in shifts and send messages continuously, but very few would spend the money and effort. With bots, the barrier to entry is extremely low. If many people connect them and the platform does nothing, the system will probably be overwhelmed.
The platform reacts to prevent that outcome. It may damage the user experience in some cases, but it is their platform. We may dislike the decision, yet they still have to protect the service.
I am relatively calm about the platform's right to react.
YouTube recently blocked roughly 128,000 channels that generated content with AI. Yet YouTube itself, like much of the industry, had previously presented AI-generated content as a trend people should pursue. The question is where the permitted boundary lies. What exactly am I not allowed to do when publishing AI-assisted content? I should not violate another person's rights, but we still do not have a clear understanding of which rights apply, in which countries, and to which material.
Which video or image may be copied and which may not? May I use AI to modify the background in my own video, or is even that prohibited? The boundary is extraordinarily broad. Companies that knowingly break the law, spam, manipulate, or steal should face consequences. The more difficult case is the person who did not know that a particular filter or generated image was forbidden. Tomorrow a platform may declare that a filter cannot be published even though the same platform created it.
When you hear that hundreds of thousands of channels were removed for AI content, remember that AI-generated material is not going away. Yesterday I saw a video in a feed on X. Because Grok is integrated into X, users immediately asked whether it was real. Under almost every second video, Grok answered that many of the images had been generated with AI. My first question was: so is this particular video real or not? The answer did not resolve that at all. Today AI can produce so much that you stop knowing whether waves of a certain size really occur in the ocean, whether an incident actually happened, or whether a giraffe really bit someone's head off.
I am exaggerating, but the uncertainty is real. A year ago we discussed images presented as photographs from a war—if I remember correctly, from the Russia-Ukraine conflict. One showed a city in smoke, and a large discussion followed about whether the image was authentic. Once generation quality becomes this high, you no longer know whether the pictures, people, cars, or scenes are real. Today's image and video generation is extraordinary. The consequences can be catastrophic for our ability to recognize reality.
A recent demonstration showed two people arguing with intense emotion. It had been assumed that AI could not yet generate a convincing scene in which two people speak and argue that emotionally. But the scene was AI-generated. There was effectively no way to tell. On one hand, I do not care whether a movie was made with AI if it is genuinely interesting to watch. Why should the production method matter when the result works? On the other hand, should I be warned? How does my mind react once someone says, “This film is AI-generated”?
At the moment, I may think: “Then perhaps I do not want to watch it; perhaps something is wrong with it.” Is that a reasonable reaction, or an irrational one that should disappear?
Most importantly, will everything be labeled? It will not. There are too many models and too many production details; comprehensive tracking will be impossible. This will become an enormous question, and we should not avoid it.
Labels will always influence perception. If you learn that a film was directed by a distinctive director, you already watch it differently. If Denis Villeneuve made it, you expect carefully composed scenes, long shots, and strong visual work. If Christopher Nolan made it, you expect a tightly constructed plot. He has now made The Odyssey, where the underlying story is already known, so the expectation may change somewhat, but the director's name still frames the experience. AI will work in a similar way.
Knowing that it was used will alter how you approach the film. I do not think that effect is fundamentally different from knowing that the work was made by Tarantino, Villeneuve, or another recognizable creator. Your impression is already shaped before the film begins—
Do you remember the reaction when the first Avatar was released? I think the feeling may be similar.
Yes. And what follows from that? Logically, Avatar is not “real” cinema either, is it?
Sasha, when you buy a shirt, you want to know what the fabric is: a natural material or recycled plastic, for example. I think everyone has a right to know things of that kind. Even psychologically, the absence of that information can feel like losing contact with reality. You no longer know, as in your giraffe example, whether a giraffe can really bite someone's head off. I believe AI-generated content should therefore be labeled. It is less tangible than a shirt, but the principle still matters.
The central problem is how to label it. Humanity has not even learned to label food honestly. Look at how much deception exists in food labels and presentation.
And look at the consequences.
Yes, the consequences are substantial.
For people's health.
Exactly—from food. AI adds not only factual but psychological questions. I would like viewers to write what they think and how they feel about it. A difficult situation involving a friend arose late last night. I needed to find information quickly and help him understand what to do, so I searched with ChatGPT. At one point I noticed an interesting psychological effect. I sent the information to my friend, and he forwarded it to his family and a wider circle. What did those people think when they saw extracts that said, “I recommend doing this; I recommend doing that” in ChatGPT's voice?
On one hand, I was helping. I know how to formulate the query, filter the response, and decide what is safe and useful to send. On the other hand, the visible words came from ChatGPT. The recipients might think: “Volchek, did you really just forward us ChatGPT output?” I have heard the same response countless times: “Could we not have done this ourselves in ChatGPT?” or “We already do that; it is obvious.” People say it in nine cases out of ten. Yet two uses that sound identical can differ in quality across an enormous range—from nearly zero to something worth hundreds of millions in practical value.
That is part of the labeling problem: how the use of AI will be communicated, who will exploit the label, and how the story will be presented. I remain convinced that if a film is visually excellent and I enjoy it while watching, I do not care whether AI was used. Imagine enjoying a film and only afterward being told that it was generated with AI, then suddenly deciding it was worthless. That reaction would itself be a form of deception by our own psychology. The important thing is a person's ability to make an informed choice.
As Tanya said, some people care deeply what a product is made from. Many do not care how food was produced, and many live almost entirely inside marketing narratives without understanding what is actually happening. If origin matters to you, you need the skill to investigate and see the wider picture. Billions of people will suffer from the loss of that skill. Only a limited number of millions currently understand even basic practical AI use—perhaps at the level of spending a million tokens a month.
The overwhelming majority still do not know what these systems actually are.
While we were discussing films, I remembered an excellent AI-generated trailer about Yuri Gagarin's flight into space and the adventures he might have encountered if Hollywood had turned the story into science fiction. It was beautifully made, with very smooth editing, entirely through AI. I think we are very close to seeing AI-made short films that attract significant attention. They may not “break the internet,” but many people will watch them. A convincing feature-length film is probably still farther away, but notable short films could easily appear this year.
Social networks already create difficult psychological and physical states: envy, fear, and other forms of pressure. It is one thing to envy a real person whom you know is real. It is another to discover that the person was generated by AI. But even an apparently real person on Instagram, TikTok, or Facebook may not be showing a real life. The image may represent only a narrow fragment of reality. Behind it is an enormous life that nobody sees, investigates, or wants to spend time understanding.
AI will intensify those problems for ordinary people. There is no realistic chance of stopping this, and fighting the existence of the technology is pointless. The useful response is to begin understanding it. The President of the United States recently made a blunt point about data centers. There is now a major fight over the enormous number of facilities being built in the country and the damage they may cause to surrounding communities. Let us be honest: they do cause harm.
A kilometer-long data center built beside people's homes creates noise, changes the local environment, affects the ecology, and reshapes the space. His message was nevertheless: you will have to accept it and adapt.
The argument was that communities must be shown how beneficial the projects can be. You cannot simply reject them; you have to accept them, and the government will provide the necessary protections. The community will earn substantial money, taxes may fall, and it will receive something every other region wants. The smartest people understand that. Most people in the room agreed, although perhaps a few did not. The community and the state would become wealthier. And if one state refuses the money, another will take it, so it is better to do the project yourself.
One can condemn that statement as cynical, but honestly there may be no alternative because nobody will be able to stop the buildout. This morning I read another argument about the destruction of nature and ecology. Even something as basic as food production causes enormous environmental damage. AI infrastructure also causes harm—serious harm—but the discussion is increasingly becoming detached from reality. The larger conflicts are becoming more visible. Elon Musk and Sam Altman fight like children.
Now Apple and OpenAI have begun their own battle. There will be many more. Some may even be created for marketing purposes. We should always look more broadly. Who said every conflict is spontaneous? Who said some of the people involved were not recruited or encouraged deliberately? Who said the person who left Apple and allegedly stole information did not do exactly what someone expected him to do? Similar cases have already occurred: information was stolen from Musk's companies, OpenAI, Anthropic, and others.
Where are all those engineers now? Were they punished? What actually happened afterward? Everyone forgot because public attention is addicted to the newest story. I would rather know what viewers are discovering in their own use of even the basic chat systems. The models have become genuinely powerful.
Are you noticing changes in your thinking and mental habits? I recently traveled to New York with my wife and discovered that she was using the cheapest, simplest model for complex searches. I pressed the stronger-model button for her and said, “How many times do I have to explain this?” She answered, “You never showed me.” That is funny because I have probably discussed it forty times on ToTheMoon and am sure I showed her. I am not so detached from reality that I invented the memory entirely.
People simply do not retain these instructions. Most people want an extremely simple system: they enter, ask, and receive the right answer. OpenAI has a problem here because the application does not let you choose a reliable default model. I have to check constantly. On the iPhone the indicator is no longer clearly visible at the bottom; you have to verify whether the stronger Intelligence mode is active.
If the answer arrives too quickly, you have to go back and change the model again.
Exactly. I have stopped trusting an answer that arrives immediately because the fast model clearly cannot adapt as deeply to the request. The application does not seem to preserve the user's last model choice, presumably because OpenAI is saving compute. I think that choice harms users. It is direct evidence that OpenAI is focused on building the system, not on creating the most convenient experience for people. I am not saying that is morally wrong, but we need to understand the situation realistically.
dollars, or, conversely, that it is free for everyone. Yesterday OpenAI sent me an email: “We noticed that your Codex API activity has decreased. Tell us why.” I have been traveling extensively, and my usage really did fall. Even during the trips I used Claude Code more than Codex, and I did not consume OpenAI's tokens at my usual rate. One thing I like about OpenAI is that it often resets or expands my limits. Claude Code does that much less frequently. But why did OpenAI send the survey?
Given how weak its support and service usually are, such an email strongly suggests that Codex is losing a substantial number of users. OpenAI is not known for proactively inventing excellent user-friendly research programs just to understand how customers feel. Someone once joked after OpenAI surveyed me that I had probably only agreed to another data-sharing checkbox. Perhaps that is all this survey was as well. I deleted it without opening it. I saw no reason to answer because my earlier responses to OpenAI surveys had never produced useful value for me.
There is another possible interpretation of such monitoring. The iPhone can detect a sudden crash while someone is driving and send an emergency signal. OpenAI could similarly begin diagnosing changes in a user's condition from changes in behavior: perhaps you are ill or need assistance because you suddenly use the system less.
Returning to programming, you asked at the beginning about agents. We moved from chatbots to agents and then to agent orchestration remarkably seamlessly. For many users, the transition from ChatGPT to Codex and to cowork-style environments was almost invisible. A conversation with one chatbot simply became work with agents, often with several agents at once. It is not the thousand agents mentioned by the Anthropic researcher, but the shift still happened. Enthusiasts had experimented earlier with tools such as OpenClaude, which Tanya even installed to try orchestration—
Where did it go?
The entire process simply moved on. Users only had to wait a little, and the same idea arrived on the computers of a billion people without any additional installation. OpenClaude had features involving WhatsApp and other messengers, but its central idea was agent orchestration and work across agents. What began with enthusiasts has now reached a billion or more users across the companies that implemented it. The transition was extremely smooth. The trend began developing actively around October or November of last year, and Codex usage has grown sharply since then.
Viewers, please write and tell us whether you have moved from ChatGPT to Codex. We repeat this question often, and I apologize if it sounds the same each time, but the answer matters because the transition is still developing. Ilnar, I think you initiated this discussion about a month ago. I still have not moved completely. Tanya certainly has not; she is clearly a ChatGPT user rather than a Codex user. Let us run the question again. Sorry for interrupting, Ilnar—I wanted to make sure people answer it.
chats. Suppose you are working in a main thread and two or three parts of the Codex answer are unclear. Previously, I created branches and investigated each question inside a separate branch. That produced many chats even when I needed only one or two clarifications. Now you can highlight any fragment and ask about it in a side chat. The side chat is temporary: it disappears after a while, or you can close and delete it yourself, while continuing the main conversation. It is extremely convenient for small clarifications.
If the side discussion produces something worth preserving, the easiest solution is to ask Codex to write the result into a file. Most repositories or projects already contain a docs folder or something similar.
That is a very useful idea. ChatGPT introduced branching as well—I am not sure what the feature is called in the Russian interface.
“Branching,” yes.
My interface is in English, so I did not know the Russian label. ChatGPT lets you create a branch in an ordinary conversation, but the implementation is terribly inconvenient. The Codex side chat sounds better because it appears beside the main work. Yet even there the experience is not user-friendly. The history is not preserved, the window later disappears, you do not understand exactly what was added to the project, when you should ask the side question, or how the chat consumes tokens.
Almost nothing is explained.
I use it only for small clarifications. If a longer discussion develops and I want to preserve the result, I ask it to summarize the conversation into a file and then close the tab. That keeps the information I need; otherwise it disappears and no longer enters the context. It is an interesting feature, and I have started using it frequently. Viewers can tell us when it actually appeared, because it probably existed before I noticed it.
I think it appeared more than a month ago.
These areas are developing quickly: side chats, richer browsers, and other interface layers. I had not used the browser in Codex much, but recently I needed the system to perform several tasks through authenticated services. I have described how I migrated completely from QuickBooks into a Codex project. That project was completed clearly and quickly. By contrast, I still have not solved the problem of analyzing all of my chats. For the financial project, the data were checked through July 1 because I completed the initial work at the end of July, and the rest of July is now being added.
The system had to log into external services. The process remains awkward: it is unclear where authentication happens and what the agent downloads. Sometimes it opens an internal browser you cannot see. Sometimes it opens a visible browser and you can watch the pointer move; sometimes even that is hidden. To its credit, the browser now has multiple tabs and some ability to change the direction of the work. That is the right general movement. The problem is that the product never explains the difference between an internal and external browser, never teaches the relevant techniques, and does not show these cases clearly.
OpenAI's Instagram and website still focus on extremely simple demonstrations. Sam Altman recently shared a suggested use case for parents: let ChatGPT collect the things you tell your children and turn them into a podcast or conversation for the children to play later. My reaction is: if I am already with my children, may I simply talk to them myself instead of sending ChatGPT to do it? They will have enough conversations with AI later in life. I would rather avoid it for as long as possible.
Let them look at a tree or out of the window, throw something at one another, or listen to an old fairy tale in a good human recording. At the same time, we see endless examples about “proper” agent use: put agents into loops, create an absurd number of them, and let them interact to build an entire environment. Yet ordinary daily capabilities—parallel side chats, usable browsers, and practical project work—are not demonstrated nearly enough. Even the more progressive businesspeople I know, if they are not programmers, do not
understand how to use Codex. They do not see that it is becoming a super-tool that can replace part of the operating system. Codex is turning into a new working environment: you open a browser inside it, open applications, and create files— —with Excel, Word, and other tools inside the same space. It is not Windows or macOS, but it resembles a new category of operating environment, just as the iPhone became a new object people held and worked through. Most people do not yet understand the tool.
It is interesting that you migrated away from chats. I cannot do that completely because, for my use, the quality of answers in these agent environments still differs from the quality of the main chat. Even the strongest models often do not respond in the same way. I also have usage constraints. If I ran every serious question through Codex, I could consume all of my tokens in a day, half a day, or even a third of a day. I would like viewers to describe their own situation. When I analyze elements of my life—restaurants, meetings, people's psychology, health, interactions, or books—the requests become large and demanding.
At the same time, ChatGPT has improved dramatically. During our recent trip, my wife said, “You have apparently pushed ChatGPT so hard that it now recommends everything with extraordinary accuracy.” Nine out of ten recommendations began matching what we actually wanted. We have discussed this issue many times: the system increasingly understands what interests me, what I need, and which preferences are mine rather than the average user's. The New York trip was probably the first time AI was consistently this precise in selecting places for me to visit.
It still made mistakes. I asked it to draw a map of particular streets in a neighborhood, and the map was nonsense. Much remains unsolved, but I am confident that these problems will be solved. The open questions are who solves them first and whether companies such as Meta can join OpenAI and the current leaders. Share your view. We have also released a special episode about the singularity and another strong case-focused episode about McDonald's, robots, and a police robot—about what today's systems can actually do and what they still cannot.
The McDonald's case shows a process that appears simple but remains unsolved in practice today. You can find that recent special episode immediately before this one on the channel. Send us the subjects you want covered. ToTheMoon can grow only with the audience's support, so please share the channel and like the work. In the mass of material now flooding social networks, that is the only way strong, useful content can reach people. We will see you again in exactly one week. Goodbye.