The next stage of artificial intelligence is obviously not chatbots—and certainly not isolated agents. You know I keep saying this: the future is not agents or agent systems. We are moving toward what people call an AI-native world, in which companies, products, enormous processes, and perhaps even economies and countries are designed around AI from the beginning. But are the models themselves ready for that? Today a model can spend a week working on a task, connect to different tools, and make decisions on its own.
It can also get stuck in a loop, fail to ask the necessary questions, produce something other than what was actually needed, hide its real progress, and burn both time and a great deal of money. So what is really happening? Are we creating the system of the future, or are we merely inserting powerful models into our existing, unprepared processes and our ordinary, unprepared lives? Will the development of AI-native systems mean that processes, professions, products, and even entire companies appear quickly and become unnecessary just as quickly before disappearing?
That is what we will discuss in today's episode. Now, let us begin. I have been raising the subject of native artificial intelligence—AI-native software and AI-native companies—for the last several weeks. One example I often use is Elon Musk. He is building the company Macrohard, which is essentially an attempt to create native software systems. He wants to oppose Microsoft by rebuilding the whole Microsoft software stack as AI-native software—that is, software created with artificial intelligence.
This approach is extremely close to how I see the future. I seriously believe we are moving into a period when the important development will not be a collection of little agents, but a transformation of the entire environment. AI will replace major processes and major connecting links in our lives. This is not a question of simply installing an agent somewhere. It is a completely different way of solving the problem. Think about a robot vacuum cleaner. The question is not whether to attach an agent that takes photos or performs an extra check.
The question is whether to rethink the entire approach to cleaning. Once you do that, the task is no longer to adapt or improve the robot vacuum. I no longer care whether there is a robot vacuum at all.
What we can see today is that an enormous number of people are using AI to bolt additional layers onto their existing business processes without understanding that those processes need to be rebuilt completely. The basic problem of automation—creating solutions and processes for companies and for your life, and introducing technology into them—has not actually changed. Twenty years ago people also wanted to automate an existing business process without asking how it might work in an entirely different way.
This is not a new problem. I simply keep hearing endlessly about agents everywhere and about solutions to narrow local tasks. Those local solutions often make life more complicated. Recently, someone close to me mentioned in passing that they wanted to build a system that would write posts for different social networks, notify employees and partners, and distribute this information. They had even written a technical specification. I said, “Show me the specification.” First, I immediately told them not to do this at all.
Then I read it. I was not exactly shocked, but the sheer incompetence of the whole process became obvious. There are employees in a company whose job is to create content for YouTube or social networks, notify the company about information, run partner mailings, or do email marketing. For some reason, they decided the company needed to build an internal system that would generate all of this content according to their requirements. They wrote the technical specification with ChatGPT.
Because they did not understand the field at all, they skipped not only proper business analysis but also performed what they called systems analysis, right down to database field names and almost every variable the system was supposed to contain. Then they said, “Build this system for us.” My first question is: if you want such a system, why not open Codex or Claude and build it directly? My second question is whether you understand whether the thing can actually be implemented, or whether these are merely abstractions.
You may spend tens of thousands of dollars building it and thousands every month maintaining it, and the system may never come to life. God, how many times have I seen this over the last three years: people constantly announce that they are building something, spend enormous amounts of time and money, and end up with nothing. It fails because of incompetence. That is the truth. It also fails because the task is difficult—this is genuinely not easy.
Today we need to go a little deeper into what native artificial intelligence really means. Can systems such as Mythos, Fable, or 5.6 Sol be called AI-native? Over the last few days we have been living through the first stages of Fable and Sol being rolled out to us. Our episodes are always extremely current, and this one is too. I am shocked by what is happening at Anthropic and OpenAI and by the way they keep rushing around, applying updates and changing their systems every day.
I am genuinely stunned. What will happen next if these companies cannot even release a fully tested system? By “tested,” I mean at the very least a system whose UX and UI do not jump around and change every minute. Several times today Fable disappeared from my menu, then reappeared, then returned with some limit. The system told me I had exhausted the limit, and then the model disappeared again. Codex 5.6 Sol burned through an enormous number of tokens—tens of millions—and consumed a huge number of token resets in the system.
Later today I received a message saying that my 5.5 Extra High allocation was exhausted. I switched it to 5.5 High and it began working, even though 5.6 Sol was still running in a parallel thread. This is another company that changes its interface constantly. We have a new contributor at ToTheMoon. A few days ago we released an episode with her—Anna Volchek, my sister. She is still a new author and will host an independent segment for us. She recorded an episode and, literally thirty minutes later, wrote to me: “Sasha, listen, I have just opened OpenAI and the interface has changed.
Chat and Work now operate differently.” I told her, “Do not worry. In three days they may work as they did before, or they may disappear altogether.” We live in a time and environment where you can begin building something today, spend a great deal of money, and then watch the underlying product disappear. You may no longer be able to apply or use what you built. It may stop working, or work badly and unreliably. You have to remember this when deciding to build something yourself.
You can make very serious mistakes because you are operating in an environment where the platform vendors themselves constantly make them. You will use products and buy solutions that are extremely, seriously buggy. “Buggy” is the right word. There is a reason companies such as Instagram can go years without moving a button. We complain about systems such as Amazon or Booking, where bugs may remain unfixed for a long time, and we call them old-fashioned or outdated. But those companies also understand that a single change can be extremely expensive.
problems and errors will follow. I think no one will be able to restrain this even in important areas such as healthcare, banking, and finance. Why? Because AI-native solutions will appear. An AI-native solution is fundamentally different from an agent system. It may be a product, a company, or even your own professional profile, but it is built around AI principles from the beginning and lives entirely inside that framework. For now, the world understands “AI native” only in a very abstract way.
The concept of responsive design has existed for roughly fifteen years: you open a website and it is supposed to work on a mobile device. In design and application or web development, people also use the term “mobile first,” meaning the site is designed for the phone from the outset. Yet if we look around, ninety percent of the world's websites still glitch, lag, fail to open, and fail to tell users what is actually happening—even in very simple situations. A genuine AI-native system or company must be capable of developing itself.
That is why I would never describe systems in the Mythos or 5.6 Sol class as AI-native systems today: they are not yet capable of developing. Whatever people say and whatever demonstrations they show, these systems cannot currently develop themselves. Worse, their “development” hallucinates much more severely than ChatGPT did two years ago, when perhaps every fifth or sixth message contained made-up information. Today we can spend tens of thousands of dollars. When you consume tens of millions, hundreds of millions, or billions of tokens, you can invest an enormous amount of effort, time, and money in developing a particular system with Claude or Codex.
That money can simply burn away, and you will never recover it. It resembles a poorly functioning team or employee. In practice, it is very much like believing that a person will succeed and deliver. Three months pass; you give them another test, then another, but the situation does not change. They were unable to solve the problem before, and they still cannot solve it.
Another fundamental problem with AI-native systems today is that they are often built for the sake of having a system—for the check mark. I recently told a friend that we had created a system capable of evaluating the complete commercial cycle, a fairly serious end-to-end commercial process. The first thing he asked was, “Were you able to apply it?” In other words, did you actually use it, or does it merely sound impressive? Starting to use something in a polished demonstration is one thing.
Discovering that it is only a beautiful wrapper is another. A great many systems will clearly appear as wrappers. And today there is no one capable of reliably checking and defining what an AI-native system actually is. We will see. Today I looked at official releases and chats from OpenAI, Anthropic, Grok, their engineers, and some of their executives. They publish all kinds of comparisons. Moonshot has released Kimi K3, for example, which is presented as a fairly serious programming system.
The world immediately began saying that everyone would now release an incredibly powerful coding system. My reaction is: are you sure? They publish tables showing who wins which benchmark and which arena. I often see tests in which Kimi K3, Grok, Sol, or Fable generates a visualization of a billiards game or a 3D city. They show these different visual examples. I look at them and see a constrained environment detached from any recognizable life. I do not care whether the roulette wheel in a casino spins a little better or worse—that is another common demonstration.
Nor do I care how a giant model of a city is generated. People say, “Look at this city model. Look how Fable builds San Francisco.” I look at them and say, “Guys, your website is broken. It does not open properly. Inside OpenAI you change menus every day. Your mobile application shows an icon saying you have three unread comments. You open the feed and see one, while the others are hidden somewhere inside projects. Why are they there? Where are they? How do you find them?” The entire chat feed itself belongs to an earlier era.
It is not a modern way to manage all of this. And today one of OpenAI's lead engineers posts another example and writes, “Look how beautiful this is!” A switch appeared inside Codex between Work and Codex, and people posted, “Look how beautiful this is!” Yet when I open it now, I have a switch between ChatGPT and Codex, while Work is gone. I thought it was there; now it seems to have disappeared. And you sit there thinking: what 3D model of San Francisco are we talking about?
If everything is so wonderful, and you can build those 3D cities, why do your own websites work like this? They are extremely simple. Surely that process should be polished to an extraordinary degree. You give a chat in Claude a particular name, and Claude suddenly starts renaming its chats. Why does it change the chat names? Why does it call a chat “Status”? I am taking two of my daughters to Hawaii next week. I sent them the spa menu from the hotel and said, “Please choose what you want.” They could not decide which massage to book, and spa services have to be reserved in advance.
I opened ChatGPT, sent it the menu, and asked what would suit girls of their age. It made a very good selection and helped enormously. Do you know what it named the chat? “Choosing Clothes for My Daughters.” Clothes. Do you understand how serious that is? It exposes an incredibly fundamental problem. On the one hand, genuinely powerful systems are being created. Over the last two weeks I have shown an enormous number of examples here, and I show more every day. My partners and colleagues tell me, “Sasha, you are on some kind of spacecraft.
We cannot keep up with what you describe. You are constantly explaining how you process hundreds or thousands of pages, normalize data, download things, and implement one system after another.” There is simply a huge amount happening everywhere. But are these truly AI-native systems, or are genuine AI-native systems still as far away as
the Moon? Perhaps we have not even begun to approach them. Here is an important example involving my editor. She was preparing the brief for an episode we released several days ago, perhaps a week ago, in which I described what I had done while creating websites. The capabilities themselves are astonishing. When she sent me the brief, she said, “Alexander, please send me the prompt—the prompt you gave for the layout and the one that made ChatGPT produce the frontend specification.” I want to explain something extremely important.
An AI-native product is not created by writing a prompt for a particular model. It is created by building a software system around the model—or around AI itself. The system receives an environment, clarifies details, and follows a route. Ideally it should bring in different executors, use different services, retain the full history of every change, and continue developing. It cannot be created with one prompt. This is incredibly important. It is impossible. A prompt alone will not make it work that way.
What do you think? Are we approaching genuine AI-native systems, or are we still building complicated shells around powerful but seriously limited models? Write your opinion and share your experience in the comments. Subscribe to our channel and send this episode to people who need to understand not only the hype around AI, but how these technologies are already changing work, business, and human life—our own lives. By the way, I have recorded another episode that will be released in a few days about how these systems consume money, how many tokens they use, and how it is possible to spend one hundred thousand dollars in a single day.
In token terms, is one hundred thousand dollars a day a lot or a little? You will see in the episode, which will appear on our channel in a few days.
Notice what has happened. A profession appeared called “prompt engineer.” Absurd. What is a prompt engineer? Recently a viewer wrote to me on the channel. I respect the person, but people need to understand what they are saying. He asked, “Could I get a job with you? Perhaps you need a prompt engineer.” What prompt engineer? What are you talking about? That is like asking, “Could I work for you? I am a person who knows how to speak.” Today, model capabilities are growing faster than the readiness of existing processes.
This is the essential point of what I am telling you. We are attaching a jet engine to an incomprehensible cart, an old car, or even a plastic toy car. The model can already reason at an extraordinary level. It can speak very well and perform actions extremely effectively. But the processes inside companies, their data, and the permissions people have were all designed for an older world. Everything has become outdated, and you still have to prove every change to everyone. I am able to build many systems precisely because I build them myself, completely alone.
People say there will be a one-person company worth billions of dollars. A one-person company will indeed be worth billions, but partly because the value of every company keeps rising. Eventually a billion-dollar company will no longer be considered exceptionally large. We have long had one-person businesses worth a million dollars. A person engaged in automated trading could earn a million dollars quite calmly even ten years ago. The point is that I can build many things alone because I have complete access and full permissions.
I do not want to rely on people operating through obsolete processes. In some projects I am the owner, I introduce a module or system, and there are directors, contractors, and employees around it. I am constantly amazed by how they report errors. They send a WhatsApp message: “Alexander, this screen is misaligned,” or “This report is inaccurate.” They do not understand that they need to conduct a complete test or investigation, work out how the system operates, and ask AI to help analyze it.
You can even ask an independent system to perform that analysis. Then, after studying everything together, they should submit informed comments. Instead, people continue acting in the old way. Why do I need someone whose entire contribution is noticing that a line has shifted? In AI development, the fact that one line is slightly displaced may have no importance at all. We are not moving toward this technology so that every individual number is always perfect. We are moving toward the ability to draw very serious conclusions.
Even if five or seven percent of certain numbers are calculated incorrectly, that resembles ordinary life, where we collect statistics from different sources and from reports that may or may not be verified, from news portals that may or may not be reliable, from people repeating rumors that may or may not have been checked, and from directors speaking at meetings, whose reports depend on their mood and on what they want to reveal. We aggregate all of that and make a decision.
An AI-native system will be a system that makes mistakes. But it should not make the kinds of mistakes OpenAI and Anthropic make when they release interfaces like these. That should not happen. The model inside may produce an incorrect or imprecise result, but these childish interface glitches are unacceptable. I remember one story involving Sam Altman. Someone asked him—or rather, I think he commented on a post on X—about the fact that the model could not measure time. A person said, “I am going to start; count to twenty, and then we will continue.” The model could not count the seconds.
Sam Altman replied that this was not a task for the current AI and that a separate module would have to be developed. Let me quickly show you a video from social media. Have you seen this guy Hasko, known online as Hask Online? I do not think so. You should probably recruit him for the red team as someone who breaks models. “I am going to run a mile, and you time how long it takes me. All right? Start when I tell you.” “Okay, I am ready. As soon as you say ‘go,’ I will start timing.” “All right, ready.
Go.” “Okay. I have started. Just run, and I will wait here.” “All right, I am back.” “Welcome back. Here is your time.” “What do you think, Mark? How long was I running?” “You took about ten minutes and twelve seconds.” “But I swear I ran faster.” “Sometimes it can feel as though time is moving faster.” “If this is going to help save humanity, we have to get it right, Sam. You need to show this to the developers.” “No, no. This is a known issue. Maybe in a year.” “In a year?” “That would be a good outcome.” “What do you mean, a known issue?” “This model—this voice model—does not have tools for measuring time or anything like that.
But we will add intelligent capabilities to the voice models.” “If you cannot do it, just admit it. That is fine.” “It is reasonable to double-check me, but I promise I am doing my best.” “If you want, you have ten minutes.” “It took me ten minutes.” At the time I said that Sam Altman did not seem to understand completely how programming and software systems work. These models operate inside a shell. The system always knows when a particular message was received and can easily calculate the difference between that timestamp and the present moment.
Yet what do we see? Chats still do not offer a setting that displays time. Many do not even provide proper dates. I sit in front of several monitors with Claude open on one and Codex on another. I scroll through Claude and Codex, and there is no indication of when a message was sent or when an action occurred. Often there is not even a date. It is an astonishing failure of interface design and system construction. At the same time, I can accept that a model itself may make a mistake, perform an action incorrectly, or have some internal function fail.
What I cannot accept is a model such as 5.6 Sol working for seven days and nights, as mine did, supposedly trying to develop particular components and elements, while in reality it spent half the time doing nothing—idling, slowing down, hanging, and so on. Why did that happen? Because the model is limited. The servers and access to the model are limited. The model was not fully tested, yet it was released without warning me that if I assigned it a task, it might effectively die for six days.
Nor was the model trained to acknowledge the problem. Every time I asked, “Is everything all right?” it replied, “Everything is excellent. We are moving at the right pace.” But the developers know it can take a wrong turn. They know it can enter an AI-native mode in which it decides for itself how to loop and how to proceed. If it makes those decisions itself, it should also understand the problem that may result. There is a useful analogy with a car. If you install an extremely powerful engine, you obviously need exceptional brakes.
You cannot build a car with a super-powerful engine and, first of all, hand it to children. My wife once came across a list of dangerous toys that used to be sold to children in America. One was a chemistry kit containing substances that would now be totally prohibited because of their relationship to certain kinds of chemical or even nuclear hazards—materials that could expose a child to radiation. The same principle applies here. If you install a brilliant motor and terrible brakes, the car will crash with absolute certainty.
I am describing the shells currently built around these systems, but ultimately the question is not only about the shells. It is about the people. Are the current teams creating these systems actually good enough? How capable are all the teams surrounding and maintaining them? Even if we stop worrying about OpenAI, Anthropic, Gemini, and xAI and turn instead to ordinary companies and people, the same issue remains. What happens to organizations and individuals when genuinely powerful engines become available?
A more powerful engine paired with bad brakes is not a more advanced car; it is a faster route to an accident. That, it seems to me, is what happened this week with Fable and Sol. What happens to ordinary people?
How will they adapt? How will the environment and existing processes adapt? Will they adapt in a meaningful way at all? Everyone talks about restraint and regulation. New York says certain data centers cannot be built. Someone else imposes a ban. When ChatGPT first appeared, Italy banned and blocked it altogether. Some countries say, “We are isolated and will build everything without you. Our models are excellent.” Some people say, “We understand everything and see the whole picture.” Is the real problem here a problem of processes?
How should a person approach AI-native systems intelligently? How do you understand and recognize them, and how do you avoid wasting your time? It is not even only about lost time. It is about being able to interact with these systems effectively and feel comfortable doing it. I am deeply involved in AI; it is present everywhere in my life. Yet I remain very calm about it. Whether a particular system is available to me or not, I am fairly rational, because I understand that if I lose access, most other people will lose it too.
Competition remains the same, and I personally do not suffer from that. The real question is what happens when genuine AI-native systems surround everything and you fail to recognize the change. Or what happens when your company has invested enormous amounts of money in building the agents that so many people are making today—agents that are about as useful as a saddle on a cow. I do not know how the English version of our podcast will translate that expression, but that is roughly the meaning.
We are launching a new website and portal where English transcripts will be published, so translations like this will be needed from time to time. We already carry a fairly serious amount of weight, including in the development of AI-focused media channels. I am genuinely curious how a model will translate that phrase.
This leads to another point. Ask ChatGPT, Claude, Fable, or one of the well-known people in the field what an AI-native system is, and they will probably tell you it is a system built according to a particular set of rules for producing a result. A modern model or specialist may add, “You need to orchestrate agents or run parallel workstreams.” Version 5.6 Sol has parallel agent streams. It constantly says things such as, “My slots are occupied,” “A slot is available,” or “I have an agent.” Then it launches an independent agent to verify everything.
Do you really believe that is the defining difference in an AI-native system? Is an AI-native system simply a task that you formulated correctly? How is “add artificial intelligence to sales” meaningfully different from what I have just described? Do you really think a strong model does not already know that it should begin with an audit, study the existing process, examine a huge body of data, conduct an enormous number of investigations, develop a solution, run tests, produce independent conclusions, and add details?
Do you think it does not already work that way? Unfortunately, this is still how AI-native systems will be described, because the definitions will be written by people and models that do not fully understand what is actually happening. What do we see today? Systems still instruct people on how people are supposed to act, while not one of the leading models asks questions properly. Claude Code does ask some questions, of course. Codex, however, seems to me to have almost stopped asking them and now asks very rarely.
Claude Code used to ask many questions inside its own interface. In the latest versions and updates, it has begun asking far fewer. I do not know whether you see the same thing; share your experience. At the time of recording, I can see that it asks much less. The only system that really asks questions is Claude Design, and even that system asks them once and then makes an extraordinary number of mistakes without asking again, even while you continue working with it. I like the tool and have already built a great deal with it.
It is very good, yet I am amazed by the stupidity of this behavior. If you decided to question the user at the beginning, and new branches and decisions appear during creation, why do you not reopen the question interface at those moments and ask the person again? OpenAI made several attempts to ask questions two and a half years ago. The first was when it launched Deep Research. A screen or page appeared with three or four clarifying questions. I remember very clearly that when the Pro version appeared, I believe, it asked one question.
Later they experimented with questionnaire-style questions displayed as separate screens. You may not even remember or have noticed it; they seemed to be testing an interface. Then the whole idea died. For some reason, the system decided that creating AGI does not require questioning the human being. It is astonishing that we are living in July 2026 and the system still does not ask you questions continuously. When it does ask, it does so in an equally astonishing way. Claude Fable or Claude Opus may ask ten highly technical questions full of implausibly sophisticated terminology.
You keep telling the model, “Remember this and write it down: you are developing the system.” Claude and Codex develop projects for me constantly. Their rules state very explicitly that there are no human engineers and that the model must write everything itself, yet they still keep asking me to do something manually. You reply, “I will not do that.” Then they say, “Oh, all right, no problem.” They repeatedly ask technical questions that you do not need, change their style of communication throughout the process, and cannot maintain a coherent line.
Of course, somebody will eventually begin asking many questions and learn how to ask them, where to ask them, whom to ask, and at which moments questions are appropriate or inappropriate. A system will develop this capability. OpenAI may develop it, Anthropic may develop it, or Grok may develop it. The question is who will do it first and who will take advantage of it. That will be a genuine AI-native system. An AI-native system cannot exist without continually asking what is happening at different points.
Even today, when I ask a powerful model or chat system to investigate something, a truly capable system should do more. ChatGPT 5.6 Sol can now spend an hour and a half or two hours reasoning in the chat itself—not in Codex or Work, but in the regular chat. During those two hours, it could theoretically go to Reddit, ask a question on a forum, collect something in parallel, observe another source, or study an additional issue. It could do that if the system were designed not merely to gather information abstractly, but to understand the subject properly and discover what is truly happening rather than stopping at whatever material is already available.
Such a system would also come back and say, “I do not know the answer.” I do not see systems that say, “I do not know.” I ask Claude Fable, “Are you one hundred percent certain?” It always replies, “No system can ever be one hundred percent certain.” I say, “I am asking in the ordinary, approximate sense.” Yet the same Fable, when I ask what percentage of a task has been completed, lies to me in numbers. It easily says eighty percent, then thirty minutes later says five percent, then produces another number, breaks, starts working again, turns out to have been doing the wrong task, or freezes altogether.
Or look at tokens. A token economy is being traded into existence. Companies gain and lose billions of dollars because of tokens.
Hundreds of billions circulate through the investment sector—trillions, in reality, if you trace all the capital connected with tokens. What will the token economy look like? How much will tokens cost? How will any of this operate? Nobody understands the final cost today. Instead, people publish endless tests: this system used this many tokens, that one used another amount, this one cost this much money, and that one took this much time. The results constantly diverge because every system is different.
Run ten tests and you get ten different systems. One is better, another worse; then the order reverses. Run the same system again and it may die completely. Was the work completed or not? When did you test it? At what point in the product's development? On what environment? I am sitting here with Codex and Claude working in front of me. How much money did they spend on this task? Nobody knows. Why is it unknown? Is it because companies do not want to tell us, the users? Or is it because the employees building these systems do not fully understand the complete process and all of its problems?
Perhaps the people building the systems do not even understand what I am talking about because token accounting is not a problem in their own work. Or is it a problem? Does the problem exist or not? Are the people building these systems genuinely ready for an AI-native era, or do they think only in terms of small abstractions and celebrate because they made an amusing button or a polished interface? Remember how Sam Altman celebrated when they launched a daily events feed that collected items it thought might interest a person?
We said immediately that it was complete nonsense. Worse, you had to train the feed. It wanted me to invest my time in teaching a system that subsequently disappeared and no longer exists. Could the creation of AI-native systems threaten more than the disappearance of certain professions or processes? Could it lead to the constant appearance of processes, companies, professions, and products with very short life spans because each one turns out to be only an intermediate stage?
How do you avoid deceiving yourself in this environment? How do you find the truth in all of it? What do you think about AI-native systems? How strong and confident do you feel in this area, and how well do you understand it? Please leave your comments; I want to know what is happening in your experience. Share the video with your friends, and we will see you in the next episodes of ToTheMoon.