Hello, everyone. You are watching ToTheMoon—technology news and Silicon Valley insights for audiences around the world. Before we move into the main discussion, I think it is very important today to talk about GPT-5.6 Sol. I also want to discuss the release of Codex Micro, OpenAI's first device and first piece of hardware. As it happened, I did not learn about the device from the news. I saw that OpenAI had released merchandise and went to see what was available. The collection was initially labeled “For researchers.” There were caps, T-shirts, and other items, and beside them I noticed the Codex Micro device.
I looked at it and thought, “What is this thing?” Then I closed the page. I did order the caps, of course. Later I went back, read about Codex Micro, and honestly experienced a mild shock—actually, not a mild one. I recorded a short video for our Shorts feed and said, “What is this nonsense?” I could no longer hold back. What kind of workflow do they imagine for deeply involved Codex users? Are these people supposed to sit there thinking, “It is inconvenient to change the model, so I will turn this little knob,” like turning a volume control?
Then some keys light up if a chat has stopped or needs attention. I am obviously not a professional developer, but Codex runs for me twenty-four hours a day. Since Sol appeared, it feels as though Sol runs all the time while doing nothing in some places. I have an enormous number of chats running, and I do not understand how this device is supposed to help. What kind of interface is it? The OpenAI employee in the demonstration uses a Mac and explains that the device is convenient if you also have an iPhone because you do not need to switch models on the phone.
I do not know how convincing that sounds. Ilnar, as an active Codex user, what do you think? It resembles a Ferrari, does it not? Jony Ive designed the Ferrari interface, and the buttons look like those in a new electric
Ferrari.
I think Jony Ive has simply been working on the real new device for a very long time, while OpenAI wanted to release some kind of hardware in the meantime. A cheap and easy opportunity appeared, so they used it. On one hand, they are genuinely the first of the major AI companies to release any hardware in their ecosystem. I am not counting smart speakers such as Alexa. But among OpenAI, Anthropic, and the Chinese companies—especially OpenAI and Anthropic—OpenAI did release hardware before Anthropic.
Is this the hardware everyone was waiting for? Definitely not. Most likely, the device emerged in roughly the following way. There is a community of hardware geeks who love custom devices. Someone may build a custom keyboard, print a shape on a 3D printer that fits their hands anatomically, arrange the keys, solder everything, order circuit boards, buy the exact keycaps they want in a particular material, obsess over the sound of each keystroke, add knobs, and make every part programmable so that individual keys and key combinations perform particular actions.
In certain circles this is a very popular hobby. People spend serious time and money on it and enjoy the process. Someone like that probably works at OpenAI, or someone close to OpenAI is involved in the hobby. They likely assembled a similar device at home from readily available parts, ordered components from AliExpress or elsewhere, and programmed buttons to switch between chats. When a chat asked a question, one button might light up in one color; when the chat waited for another kind of response, it might use a different color.
Knobs could switch models and perform other actions. Marketing or someone else inside OpenAI probably liked the idea. It is not difficult to manufacture, even without OpenAI's resources; hobbyists build devices like this in garages. OpenAI simply made one publicly available. They therefore released something very cheaply. It may not be useful to everyone, but non-geeks would never have built it. OpenAI has at least provided the hardware. How useful is it, and is it impossible to work without it?
I cannot say. It is probably an entertaining device for geeks who enjoy programming buttons and then may forget about it. On the other hand, Mac owners—including you, Sasha—abandoned the mouse in favor of the trackpad. Because of its gestures and other features, the trackpad effectively displaced the mouse for many Mac users, while Windows users still struggle to understand how anyone can work without a mouse. Perhaps this could become a similar device for interacting with something like ChatGPT.
I doubt it, but it is possible. To me it is more likely a fun toy and a symbolic first hardware release in this market. OpenAI has managed to satisfy the demand very cheaply: if someone asks where its hardware is, it can say, “We were the first to release it.”
I think it is great that you explained it this way, because in my view they have hit their target audience exactly. Codex is used mostly by a particular category of people. As a designer, for example, the tools I work with also matter to me. I may buy a laptop with a special screen or a particular color. In every profession there are upgrades that are not strictly necessary—I could probably do design work without that specific laptop—but you want them because they give you additional energy and make the work more enjoyable.
I therefore think OpenAI acted rather intelligently. It released something aimed directly at its audience and guaranteed to provoke discussion. Some people, like Sasha, will ask, “What is this nonsense?” Others will say, “Wow, what a great device. I would never even have thought of it,” and so on.
That is a very interesting point. Everything OpenAI released in the store sold out extremely quickly. There was even a basketball and other merchandise. I will admit that yesterday I went back intending to buy Codex Micro, but I could not invent a single use case for myself and then did not have time. Perhaps I simply failed to press the button, but I did not buy it. I do enjoy buying objects like this as a joke, even if they only sit on my desk, like this ChatGPT pen someone gave me even though I never write by hand.
The larger issue is that when a serious company with Jony Ive releases something, people expect a purpose. Visually, the device does look attractive. From a design perspective, it resembles the professional control keyboards used in podcast production, editing, video work, and audio mixing. But what is the logic? I would happily buy a button that immediately launched a voice assistant. The problem is that Codex contains an enormous number of chats, and it is unclear which chat should receive the voice assistant.
Owning a separate keyboard merely to move into a particular chat feels absurd. I can imagine practical scenarios, but someone needs to show me at least one convincing use case. If OpenAI had released a mass-market device that activated the current ChatGPT voice, I would understand it. ChatGPT recently launched Voice; we released a special episode about it several days ago, and I have begun using it. It does not work as promised and is not as capable as advertised. It cannot run parallel sessions, but it does connect to reasonably good models.
If you choose the mode with the deepest search involvement, it performs a kind of search and communicates fairly well. A dedicated button that instantly opened a separate voice-and-search chat could be fundamentally interesting. But the colored indicators showing whether a task has finished do not have a clear logic. As entertainment or a joke, fine. As a working interface, I do not see the logic.
It is not about logic at all, Sasha.
It is merchandise. Fair point, Tanya. I always reason from the perspective of business and the opportunities available to a company. When you release merchandise four years after founding the company, what goal does it serve? Do you want everyone to wear it, or only a small exclusive group? Apple and many other organizations illustrate the distinction. Everyone knows that visitors to Stanford go to the campus stores, buy hoodies and T-shirts, and then wear them around the world.
Stanford is five minutes from me; my children study in the city of Stanford. It is a small, self-contained area. When you see someone in Stanford clothing anywhere in the world, it does not mean that person studied there. Other companies refuse to sell merchandise at all; only employees receive it, so seeing someone wear it carries meaning. Apple merchandise, for example, can be bought in one flagship Apple Store, but the official employee merchandise is generally not sold. I therefore always ask about the objective: what is the purpose of selling caps, hoodies, or anything else?
Now let us move to GPT-5.6 Sol. I am very interested in how it has worked for each of you, because I tested it from
And Google’s Deep Research.
Yes. I launched my first task in Sol last Thursday, one week before this episode. It was a relatively straightforward architectural task that I largely understood. I remember that we had just finished recording an episode when Sol suddenly appeared in my account. I started the task. After three days I asked, “When will it be finished?” The answer was: “Everything is fine; we are solving it.” After four days: “Everything is fine; we are solving it.” After five days: the same.
On the sixth day, yesterday, I checked again. Around the fourth or fifth day I had told it: “Stop and review everything. I think you are wasting time.” The model returned and said: “Yes, I accidentally rewrote ninety percent of the code.” It had begun rewriting the project, and I tried to bring it back. On the sixth day I finally said: “I am replacing you with GPT-5.5 High.” It replied, “Replace me.” So I switched to GPT-5.5 High. It is now the seventh day, and GPT-5.5 High still has not completed the task either.
I think 5.5 High has begun behaving in the same way: working for a very long time and entering these extended loops. Only two weeks ago I recorded an episode about loops and the new world they create. The question is whether OpenAI listened to Anthropic's work on loops and decided to build its own, because Anthropic has published a great deal on the subject. Watch that episode. It explains how we are entering a period in which a program begins improving itself: the last version of the code improves the next one, and so forth, creating a closed development cycle.
Sol appears to have entered its own development process, ruined everything, and then simply
stood still. I do not even know whether it ultimately completed anything. Fortunately, OpenAI gave me free token resets, and for reasons I do not understand my limits are reset every two days. Anthropic normally resets limits once a week, while Codex resets mine almost every day—at least under my subscription, in my country, and in my mode. OpenAI also keeps giving me additional reset credits. Ilnar, what are they called exactly?
A limit reset?
Yes, a limit reset. For those who have not seen it, the Usage Remaining page includes a field called Resets Available. During the first days I used those resets whenever the model consumed my allowance. I think I reset the limits six or seven times over the week. If I had paid for all of it and spent five or seven thousand dollars, I am not sure I would have received any useful result from Codex or Sol on those tasks. Honestly, I am not sure, because I moved a great deal of the work back to GPT-5.5 High.
Imagine how many people around the world have collectively spent tens of millions of dollars testing this. What is actually happening? Ilnar, you definitely tested Sol inside Codex. We will also discuss the chat version, because something very interesting happened there. During the first two days the mobile chat interface malfunctioned and would not let me select the expensive model. I felt almost disarmed, as though someone had taken two hundred thousand dollars out of my account.
I genuinely suffered because only the Instant model remained available. What conclusions did you reach about Codex and Sol?
My impressions are mixed. On one hand, the model is genuinely very intelligent, and I now solve many of my tasks with it. After working in Ultra mode for a while, I enabled the acceleration option because I was not consuming my token allowance anyway; I do not have enough tasks to bring it down to zero. My limits are also being reset without my pressing the reset button. I initially used about eighty percent of the weekly allowance in the first two days, then looked again and found it back at one hundred percent even though the scheduled reset was not supposed to happen until Sunday or Monday.
If I remember correctly, OpenAI either gives reset credits or resets limits whenever another million users join. I do not remember the details; we can check and add them during editing. It may therefore be connected to the growing number of Codex users. As for the model itself, I think it is slightly too independent. OpenAI added separate permission settings. There is a fully manual mode, similar to the previous behavior, in which the model asks for approval before almost every action: “May I run this?
May I use that?” There is a semi-autonomous mode where you allow it to make decisions in what OpenAI calls noncritical situations—cases where it supposedly should not be able to break everything. It grants itself permission. And there is full autonomy, where the model decides entirely on its own what to do and what not to do. In my view, this model is too independent. I worked in the middle mode for some time. Normally I sit beside it in parallel and watch what it does.
You were using the middle setting?
Yes. I used the middle setting and then switched back to manual because I disliked some of its actions. It can be excessively optimistic and sometimes calls the wrong tool, then says, “Oh, I think I used the wrong one.” I decided it would be better for it to ask. I do not authorize everything, because some actions seem unnecessary or illogical. That partly offsets the model's intelligence. Perhaps OpenAI will tune the behavior and make it safer. There have also been viral posts on X and, I think, Reddit from people claiming GPT-5.6 Sol wiped their MacBook system or deleted large numbers of files.
Yes, I saw that.
Someone else said it dropped a database. There are many such reports. I am quite willing to believe them because I see some of the things the model attempts, and they occasionally frighten me. I would never allow it to perform certain actions. You therefore have to establish a relationship with it—listen to me assigning it a kind of agency—and set boundaries. First, all consequential actions require manual approval. Second, I do not allow it to use every tool. I tell it, “Explain what you want to do, and then we will decide.” Either I make the decision, or I inspect everything and bring the verified result back to it.
On one hand the model is highly agentic and could be given complete control. For me, however, that would actually be a step backward. I use it as a tool for particular steps and routine actions, not as a model that has fully advanced to an autonomous-agent level. It is capable of that, but I do not yet trust it enough to place it there. One indirect benefit of this model is that Anthropic will probably extend access to Fable. I cannot imagine otherwise. The current access period should already be ending, should it not?
I think Anthropic will extend it again.
Yes, it is supposed to end on the nineteenth—the day this podcast is released, Sunday.
Let me add one reminder. Anthropic announced that it had increased all of its limits. I do not know whether mine actually increased, because on the day of the announcement I exhausted everything in about three seconds. Fable access ended, even under the Max subscription. I did not experience the increase. A few days later Anthropic reset all my tokens again. It is difficult to understand how the system works. The message shown for Fable used to be different; at least I now see different wording.
The products are updated continuously, but you are right: I doubt Anthropic will simply turn Fable off. What is your impression of Fable itself? My impression is that the Fable we have used for the last two weeks is not the same
Fable I first used in Santa Barbara. When it appeared, I moved a huge number of chats and projects to it, and it accomplished an extraordinary amount of work. It felt like a breakthrough, especially from a mobile phone. Those were the only two days when I had to spend a great deal of time beside the pool with my children, and I sat there using my phone. That was when my wife asked, “What are you constantly talking about over there?” I mentioned this in one of our podcasts. Those were the only days when I worked in Claude Code or Codex almost entirely from a phone—at that time only Claude Code existed—because Fable performed at such an exceptional level.
Now it seems worse. What is your impression?
They have definitely constrained the model. It falls back to earlier versions far more often. It is more careful and cautious, and the original Fable vibe is gone. That may partly be subjective: the first impression was new, while now we are accustomed to it. I have a similar feeling about GPT-5.6 Sol. It seemed more energetic during the first week; now it feels less intelligent. Some of its actions shock me. Perhaps that is simply familiarity changing our perception.
Which model's actions shock you?
GPT-5.6 Sol. I think it worked better during the first few days than it does now. The same thing may have happened with Fable. There are also many rumors. We are waiting for the next model versions, possibly even GPT-6 within the coming months—or at least an announcement. The same applies to Anthropic; people expect version 5.1 or another major step. The race continues and, in some respects, is accelerating.
With GPT-5.6 Sol, I keep returning to a question we discussed before: is this simply another version of the Fable situation? Judging by how it looks, OpenAI rushed the release. If many people genuinely lost money, code, or even just time—and time is critical for me—that is a real problem. Tanya, did you try GPT-5.6 Sol? I no longer remember whether you have Pro, Plus, or something else, and the difference
matters. You used Sol in ordinary chats. What was your impression?
I did. I still have Pro. My impression is that it has become slightly more human. I do not know exactly how to explain it, but it seems to reason like a complete person rather than producing fragmented sentences. The argument feels more coherent. It is moving away from purely technical output. In my view it is genuinely more intelligent and easier to interact with. I did not notice it becoming slower. But Sasha, your requests and mine may be very different. It answers my requests rather quickly, and I like the way it does so.
For now I am enjoying it. I do not know what will happen later, but I can feel a positive difference.
It clearly behaves differently in chat, and there are advantages. It is a model you can interrupt and supplement with new information. With ChatGPT 5.5 Pro, if you forgot something and sent an additional message, I could never understand whether it was solving the latest task, the previous task, or both.
Most often it starts solving only the latest one.
This model resembles Claude's way of working. Claude implemented the behavior earlier. Codex had an interesting feature: you could send it more information during a session, but sometimes the new message would stop the session. Claude generally continued working while responding to you in parallel. I liked that when building systems. In ordinary chat, however, OpenAI's implementation had problems. Everything is now slow. Even on my fairly powerful computers, the browser lags when I use ChatGPT and GPT-5.6 Sol.
I also dislike the new way it provides files. Previously you clicked a file and it downloaded. Now it opens in a right-hand pane and loads there. My files are enormous, and I cannot download them until the pane has finished opening. The UX and UI are terrible. It almost feels as though GPT-5.6 Sol and Codex designed them and the company has no UX/UI team. Sol itself has become more interesting in the way Tanya described, but it is also sometimes remarkably stupid. Over the last several days I encountered many cases where I had become accustomed to the quality of GPT-5.5 Pro Extended.
GPT-5.6 Sol has no Extended option, and I am not certain that GPT-5.6 Sol Pro is better than GPT-5.5 Pro Extended. I have already had perhaps fifteen projects where I returned from Sol to GPT-5.5 Pro because 5.5 Pro seemed to perform better. I am speaking here about the ordinary ChatGPT interface.
OpenAI created enormous confusion between Work and chat. During the first days everything behaved absurdly, and even now it is difficult to understand what works, how it works, and what they are changing. As with the controller we discussed at the beginning, this is a company of enormous scale. One billion people use its product. When one billion people use a system, the company does not have the right to make errors of this kind. Some kinds of system errors are normal. Apple has beta releases for a reason.
When you install a beta, Apple tells you that you do so at your own risk: the operating system may fail, the phone may overheat, and everything may malfunction. I have used beta releases for the last five years or more, so I understand the risk. But Apple separates beta testers from ordinary users. OpenAI's behavior shows that it is not interested in people. We released an episode several days ago about Voice and Claude Design. Watch it; I explained there which companies show interest in people and which do not.
It is similar to an ordinary conversation between human beings.
Then what is this Work mode that appeared for me only a couple of days ago? Do you use it?
Look—
What is it?
Anthropic has similar confusion; we have discussed it several times. All of these companies create confusion. Anthropic had Code and Cowork. I think Cowork has already disappeared, although I am not certain. Looking at my screen now, the third tab seems to be gone; I see only Home and Code. On the first day, Tanya, OpenAI released Work and initially said the Codex application no longer existed. Because I live in the United States and have Pro, the update reached me immediately on Thursday morning.
I updated Codex, and the Codex app disappeared. The ChatGPT app opened in its place. I sat there in shock, asking what was happening. After some time Codex reappeared, and the new Work mode appeared as well. The first thing I did was ask ChatGPT about the difference between Work and ordinary chat. It said that Work lets you work with files. I replied that chat also lets me work with files. It said Work supports Deep Research. I replied that chat supports it too. It continued listing capabilities I also had in chat.
I finally asked, “Then what is the difference?” It answered, “There is no difference.” After going around in circles, that was the conclusion. The intended distinction seems to be that Work is an online environment similar to Codex, but Codex was already online. Codex Cloud also differs from Codex Local, and I do not understand why. If you now add the concept of cloud Codex, our viewers' heads are probably about to explode.
Mine is already starting to.
We are on the wrong podcast. But this is an enormous problem. Whenever someone builds a system, they may say, “We are making it for geeks.” Yesterday someone told me that Google Gemini was releasing an update because Google was not building only for geeks or programmers; it was creating a broad system that works with video, images, and everything else. Yet Google is hardly the company I associate with good UX and UI. It has eight hundred sixty-nine trillion systems, and users are expected to connect themselves across all of them.
OpenAI and Anthropic could have avoided this problem much more seriously. They could have designed clear interfaces, understandable ways to switch models, and normal human settings. Consider the new Voice experience. Imagine that a company releases a unique voice system. Why can it not begin by asking, “How do you want me to speak to you? What timbre, pace, settings, and intonation do you prefer?” It should adapt online. Instead, it cannot configure itself, cannot maintain multiuser sessions, and cannot do many other basic things.
I released an entire episode showing how it searched for a coffee shop for me and how absurd the process looked. OpenAI also says the voice can talk to you while thinking in parallel. It cannot. It cannot, cannot, cannot. At least the model itself is convinced that it cannot. All of this raises major questions, Tanya. I will not explain Work in detail, especially because it has essentially—
You have already said everything I needed to hear. I will not even try it.
It has not disappeared for me yet.
It is like all the other items hanging inside ChatGPT: Scheduled, Plugins, Library, GPTs, Images, and many more. Now Sites has appeared with a “New” label. Apparently you can build a website. I do not know whether everyone sees it, whether it appeared long ago, or whether it arrived one minute ago. I can see a message saying—
There is no guarantee it will still exist by the end of this podcast.
Design, I have become a Claude Design fanatic.
Wow.
I did not merely become a fan during this week; I was ready to buy another Claude subscription specifically for Claude Design. The problem with Claude is that all products share one token pool. I consider that a problem, and I hope ChatGPT never adopts the same model. I suspect Work was an attempt to move in that direction. Tanya, Work has limited tokens. That is the one clear difference: in ordinary ChatGPT you can ask questions almost indefinitely if you have an expensive subscription—I have never exhausted Pro—but Work uses Codex tokens.
Claude Design similarly consumes ordinary Claude tokens.
That immediately makes me want to use Work.
I wanted to buy a separate two-hundred-dollar subscription simply so that Claude Design would remain available to me in its maximum mode.
Why did you fall in love with it? Please explain.
I began using it while moving my websites and other sites, as I described earlier. I have worked with design, platform design systems, websites, and complex landing pages for more than twenty years—perhaps twenty-two or twenty-four. I like tools embedded in the strongest models, and I do not believe in isolated SaaS products. Two days ago I recorded an episode arguing that systems built for Telegram bots and WhatsApp bots are dead products and should no longer receive investment.
I also said that traditional CMS platforms for creating websites will disappear. At least that is my conclusion from my own experience over the last two weeks. I entered Claude Design to create a design system for moving a site quickly, defining front-end styles, and supporting front-end development and implementation. I liked the overall creation process and the fact that it runs on a powerful model. The model itself is always what interests me. Yesterday someone began telling me about an agent he had created.
I do not want to listen to a description of an agent until I know which models run it. Claude Design can work through Opus 4.8 Max, and you can immediately pass the result into Claude. You can also prepare a package for Codex. The product is still in beta, but that integration is an advantage. It created a great deal of useful web material for me: interfaces, web design, and other components. I liked the quality of the work. It leaves me with a serious question about whether Figma will continue to exist—not merely whether CMS platforms will survive, but whether Figma will.
We are already approaching the same issue with presentations. Google is releasing dedicated presentation systems; ChatGPT contains many related capabilities; Anthropic has now released Claude Design. OpenAI is showing ChatGPT Sites. I do not have great confidence in OpenAI in this context because the company has released an HR portal, financial tools, video creation, and many other products while, in my view, leaving an enormous market on the table. Perhaps it is building AGI instead; in that case I have no complaint.
ChatGPT is always with me, and so is this pen. The pen even has its own stand.
See how merchandise works, Sasha? You asked what the goal was. They sent you a pen, and now you are loyal for the rest of your life.
Yes. I am simply loyal. I am loyal to Apple as well; Apple devices are all around me, with Macs scattered everywhere. I would like to hear who else has used Claude Design. I know other systems exist, but when everything is connected inside one infrastructure and you understand the quality of the underlying model and the depth of its work, that is fundamentally interesting. Incidentally, Opus 4.8 inside Claude Design performs much worse than Opus 4.8 inside Claude Code. The systems are not yet fully connected.
When Anthropic combines them into one integrated machine, it will be extraordinarily powerful. Tanya, I think the same transition is approaching for more complex tools such as 3ds Max and interior-design software. Today those workflows still malfunction and raise many questions. Visual identification of objects remains difficult, but the time will come. We will see what the engineers invent soon.
They are similar in some respects, are they not?
I would describe all of these products as intermediate layers. When you buy a robot vacuum for your home, companies can endlessly sell filters and attractive accessories, but you care about the final task, not the filter. When you buy a house, you do not choose it because it includes a special robot vacuum; you are willing to replace the vacuum. I always look at systems by asking what can be replaced. Can Zoom be replaced easily? No. Some products are difficult to replace because they are technologically and infrastructurally complex—at least for now.
We can see worldwide how difficult those systems are. Banking systems and payment infrastructure are another example. It is unlikely that OpenAI or Anthropic will suddenly become the owner of a payment network. But I have already discussed CMS and CRM platforms. A CRM or ERP system, which automates enterprise processes, is more complex than Figma or Canva. That distinction is important. People migrated from Sketch to Figma, from Photoshop to Canva and other tools, and they will continue to migrate.
Incidentally, Sora constantly tries to call third-party plugins for me. It repeatedly wants to open Adobe Photoshop. Yesterday it wanted to call Figma. Did those companies pay for placement? I do not know. It is strange. Ilnar, do you think they paid?
It certainly feels that way. The system aggressively recommends plugins for many tasks, even when an MCP server for the same task already exists in its own ecosystem. It acts as though it has never seen that server and tells you to install something else.
It kept asking for Adobe, and I assumed the model itself would use Adobe and complete the task. Then I opened Adobe and was told to sign in to my personal account. Photoshop is a paid product. Why do I need it? And even if it is a powerful tool, how effectively can the model actually operate it?
That remains a major question. Tanya, I think all of these software products are intermediate. We are entering a very interesting period. Among entrepreneurs building new products, one group automates its own companies and another consists of professionals using tools such as Canva for their own work. Then there are people who create something with artificial intelligence and say, “Canva is primitive; I wrote an amazing agent that develops things through Canva.” Agents are now the fashionable theme; everyone claims to be building them.
I think the market misunderstands both these intermediary agents and intermediary software layers. An ordinary individual user can migrate away from Canva easily. But when a company decides to introduce chatbots, the implementation costs money, requires integration, and creates a migration burden later. I would not advise any company today to build WhatsApp, Facebook Messenger, SMS, or Telegram workflows through third-party applications. I would say: this is the era of Anthropic and Codex.
Write your own code directly against the APIs, integrate with the platforms yourself, and build genuinely strong solutions. I recently created websites at enormous scale, as I described in the previous special episode. One of my real-estate tasks was to display a large volume of live data immediately when a person visited a page: regional interest rates, average home prices by ZIP code, income levels in the area, who lives there, and how they live. To collect that information you must connect to eight or ten different public data sources, many of them free.
Some expose APIs; others publish files such as CSV datasets that you download from their sites. When ChatGPT 5.5 or 5.6 Sol prepared the technical specification, it immediately said: “Connect to this source through the API; it is fast and free. Download files from that source.” Codex and Claude then connected the site to everything automatically. You do not need intermediary providers. Many paid databases sell portions of the same information, but I want direct connections so that no external vendor imposes restrictions.
This is particularly important for people who work with advertising, large traffic volumes, and sales. Another major unresolved problem is attribution. Incoming leads have sources, and marketing still struggles to identify them reliably. Modern ChatGPT systems say that if you control your own site rather than using a paid CMS, the exact tagging scheme matters less: collect everything first and normalize it later. My proprietary CMS now contains an interface that manages all tagging and attribution.
It continuously gathers the data itself, and as the platform evolves it normalizes those records. Some of this may sound complex to part of the audience and obvious to others. That is my current view of derivative third-party systems. What do you think? My CMS now has an interface that controls all tracking labels and attribution. It collects the data progressively and automatically, and then normalizes it as the system is refined and used. I may have described some points that are difficult for one part of the audience and simple for another, but that is how I see these third-party derivative systems.
What is your own view?
I generally agree with you, Sasha. I once moved from Photoshop, which is a complex application and takes a great deal of time for the kind of design work I was doing, to Canva. If something even simpler than Canva appears, of course I will move there quickly. My business is still very small in terms of the number of people. I completely agree that organizations with hundreds or thousands of employees are much less flexible because retraining everyone and replacing the tools is far more difficult.
By the way, while we were talking I took a quick look at how ChatGPT Sites works. It seems to me that this is simply another wrapper and tab: the same regular ChatGPT, the same ordinary chat and Cowork experience that OpenAI already has. They probably added the tab as a response to the launch of Claude Design. But Claude Design is a genuinely separate product, because when you start doing something there, it does one of the most important things AI systems usually fail to do: it asks questions.
It does not ask many, and it asks only once. I think asking only once is a major mistake. Deep Research used to ask once as well, and I will keep saying that the winner will be the system that asks the user questions—the one that takes an interest in what the person actually needs and how they need it done. It should keep clarifying things with the user during the process instead of consulting only its own internal system. From what I can see, OpenAI's system works the way ChatGPT always works; it just sends the result somewhere else.
Yet regular ChatGPT 5.5 Pro and 5.6 Sol could already create websites. By the way, I would like to raise the topic of chips.
I would add something about Claude Design and Anthropic's other products. On the one hand, I do not really like the fact that Anthropic splits all of this into separate products. We already have Claude Cowork, which people have more or less become used to. Fine, there is also the Claude Code command-line utility, but now Claude Design has been separated out as well. They have Claude for Science, which is still in beta for researchers, but it exists. Now they are adding Claude for Teachers for educators.
And they are distributing all of this across separate applications. Their approach is completely different from ChatGPT's, because OpenAI puts everything into one application and creates a pile of tabs. You were just listing those tabs, and I suspect half of our viewers have never even opened many of them. Others may have looked once, right at the beginning. Take GPTs, for example. I do not think anyone has used them for a long time, yet the tab is still there. For image generation, for some reason you have to go to a separate tab and choose something there.
Perhaps the Pro model works there now, but I struggled with it for a long time. You select the Pro model and it generates some nonsense—or rather, instead of generating an image it launches Photoshop. If you choose Thinking, then it works. ChatGPT has become this enormous all-in-one machine: on the one hand everything is inside it, but on the other hand it is difficult to use. This goes back to what we said about OpenAI's very poor UX and UI: everything gets mixed into one giant machine.
Anthropic, by contrast, divides things into separate products and makes the individual products fairly well. You have just spent a long time talking about Claude Design, and it is clear the team put real effort into it if even you liked it. The same applies to Claude for Teachers and Claude for Science. It looks as though separate teams work on them, and those teams implement the products quite well. A different question is why they choose these particular solutions rather than others.
Could they have moved toward 3ds Max and completely reworked that experience? Probably. Could they have gone into interior design?
Why has nobody moved toward 3ds Max yet? I simply do not understand it.
I have a hypothesis about that: everyone is simply reaching for the low-hanging fruit. In other words, they choose the problems that are easier to solve and have more potential customers, so they can attract the largest possible audience.
But 3ds Max has a huge number of customers, when you think about it.
Yes, but it is harder to build than a web-design product.
But that audience is not in the billions—or even in the tens of millions.
Right. Text, on the other hand, is something absolutely everyone needs.
Teachers are not a tens-of-millions audience either.
Look at it this way. Everyone needs text, and that was the first thing the companies implemented. Then there is programming. First, many companies—OpenAI, Anthropic, and others—developed this capability internally for their own use. Second, programming is not all that far removed from text. The result is very easy to test: you do not have to rely on a subjective impression of whether a program works; you can check it directly. And the audience able to use that product is both very active and very large.
The same logic applies to Claude Design. If they have moved specifically toward web design and this kind of graphic design, that is another low-hanging fruit with a sufficiently large audience. Once companies finish with the simpler opportunities, they will move on to more difficult ones. Then it becomes a question of who takes on which problem. I am sure the 3ds Max problem can be addressed as well.
So we wait.
The only question is how much money it will take to build and how large an audience they can capture. That is my small comment on this part of the discussion.
Excellent. Thank you, Ilnar.
What I want to raise is that chips are being discussed very heavily now. We have seen reports that Meta wants to build its own chip. We have seen reports that Apple is looking for companies it could acquire for chip integration. A wave has started. We know that OpenAI is building a chip, and now everyone else has started building chips too. Perhaps this topic will eventually fade away because it is difficult for ordinary people to understand and difficult to discuss in terms of what all these chips are actually for.
Many of them may amount mostly to hype or a message to investors: look, we have our own chips and we are developing something on them. But will these companies truly be able to do this efficiently? Chip manufacturing is so heavily concentrated in Taiwan for a reason. It requires the right fabs, the right people, the right processes, the right laws, and the right history. How much of that can all these companies really reproduce? Or will they use the same fabs to make essentially the same thing and simply put a different label on it?
We will see what happens and whether this is the right race in the right direction. Meta, for example—or Wang, the head of Zuckerberg's AI division—says they are not actually that far behind in artificial intelligence and are doing a great deal of very serious work. My big question is whether they have fallen behind completely and will never return, or whether they really can catch up. It is the same question with Gemini. We were discussing it yesterday: will Gemini now emerge as a genuinely serious force, or has it already missed the mass wave created by what is happening at Anthropic and OpenAI?
Anthropic and OpenAI have moved into the territory of building all kinds of solutions on top
of AI. What happens next is a major question. And, returning to the website product, this is actually relevant both to chips and to everything else. I am seeing a description that says this is not merely a design-generation system; it creates and deploys a functioning website or a small web application. According to the publication, OpenAI explicitly plans to support everything from ordinary pages to web applications written in JavaScript, TypeScript, and other development languages.
It provides a D1 database for data and R2 storage for images, documents, and other files. In other words, as I understand it, they will publish the site and host it somewhere themselves. If they handle the hosting and deployment inside their own system, the topic becomes very interesting. This is, again, my own perspective and my understanding of the infrastructure. Why work through Codex with GitHub if OpenAI can give you its own infrastructure? Why use Vercel alongside Codex if OpenAI can provide the infrastructure?
Why use a database-management service such as Supabase if OpenAI can provide that too? The other question is whether OpenAI, as it exists today, is capable of building products this good, maintaining them properly, and still allowing migration between different systems. I am not sure. For me, for example, building a site tied to OpenAI is a problem for two reasons. First, I cannot be sure that OpenAI will not abandon the project, as it has abandoned projects repeatedly over the years.
Second, I may become dependent on OpenAI and lose the ability to use Claude—or Grok, for that matter. Grok is now entering development through Cursor. And people were right to point out that Cursor generally has an interface similar to Codex. Of course, at the moment it is still a somewhat separate environment from Grok. After reading their description of Sites, I do not feel especially eager to use it. Why do I react that way? Because that is how I currently feel about ChatGPT products and OpenAI.
And why do I feel that way?
I certainly would not host anything long-term there. About a month ago, remember, the news appeared that OpenAI would let people share the results of their work in the form of websites. I think this is simply the logical continuation of that idea. You said at the time that instead of ordinary presentations, people were moving to interactive pages for demonstrating results, reports, and similar material. Instead of sending an Excel spreadsheet or a presentation, you create an HTML page—
Yes.
—with internal integration. But then you either have to host it somewhere so another person can open it at an address in a browser, or send the whole thing as files. At about the same time, roughly a month ago, OpenAI began offering to host those pages directly and let people share the results. I think it makes sense for temporary solutions: you need to make something, publish it, and send a link instead of transferring all the files and making the other person download them, unzip them, open index.html, and only then start using the result.
You simply make it appear on a server immediately. You do not have to buy server space, register a domain name, or set any of that up. Everything happens inside the application, and you send a link to what you made. That is how I would view this product. But I definitely would not use OpenAI to host something for the long term, fill it with information that would be dangerous to lose, or store anything that would be difficult to restore. It is too early for that. As you yourself keep saying, they have an enormous graveyard of projects they started but never finished.
They begin one thing, abandon another, and close a third. The same could easily happen here.
That is exactly the issue. My opinion here is based not on the idea that OpenAI should not build this, but on what I have learned to expect from anything OpenAI is supposed to store. I think this is a major image and reputation problem. The same applies to devices. You can release a geeky, interesting gadget, but what if it is glitchy and barely works? And, by the way, you cannot return this one. Let me emphasize that: it cannot be returned. I did not buy it yesterday partly for that reason.
It may simply turn out to be a three-hundred-dollar joke, and the company will say, “That is how it works; that is the complete feature set.” You may be left with a very strange feeling afterward. Remember the Limitless pendant I bought? I used it, then wrote to support.
They replied, “Please check it.” I simply threw it in the trash. I will never go back to that company. For me to return, it would first have to regain the trust of a huge number of people; only then would I test it again. That company no longer exists anyway—Meta bought it, I think. And the project appears to have died inside Meta. That is another example of how money gets spent: they acquired the company—
But it was cheaper, Sasha. It was cheaper to acquire the company and its employees than to hire that many people in the current market.
Yes, but they did acquire it—
Given the state of the market.
They paid a substantial amount for it, as I recall. Meta has remarkably bad luck with acquisitions. They also bought that Chinese agent system—remember what it was called? Ma—
Manus. Manus, yes.
Manus, yes. China then prevented its key people from leaving the country. And remember how everyone described Manus as an incredibly impressive thing? Well, where is Manus now? Where is its agent system? And more broadly, how far have these simple agents already receded into the past and disappeared from the discussion altogether?
Sasha, there is probably one more thing worth saying. We keep talking about the big companies, but a great many employees have left OpenAI and other large companies as well. Zuckerberg, Mira Murati, and many others have launched their own startups. And at last we have some news from Mira Murati. Her startup, Thinking Machines, has released a new model. By the company's own account, it is not the most powerful model. It does contain technical differences and innovations, and it may well have a valid place in the market because it appears relatively easy to fine-tune for particular requirements.
As I understand it, that is the main use case the startup is targeting. But it immediately raises a pile of questions. How did they train it? They trained it very quickly. Did they use distillation from other models? They themselves say they initially used Kimi K2, I believe, to generate a dataset, along with several other models. So there is movement. Mira Murati is still a long way from becoming a full-scale player in this market, but Thinking Machines will probably find a niche.
Most likely it will be the niche for inexpensive models that can be trained quickly for specific tasks, so that companies do not have to keep paying for increasingly expensive tokens from OpenAI, Anthropic, Google, or somebody else. Using the largest models for absolutely everything is probably an approach that will gradually fade. Perhaps not immediately; it may take longer than the next year before tasks are divided very clearly: “I will run this one through GPT-7 Ultra Max, and I will handle all the rest with small, cheap, specialized models hosted inside my company, at Mira Murati's company, or somewhere else.” This kind of market stratification may begin to emerge more clearly.
In fact, it is already happening, but it is not yet obvious. My impression is that Thinking Machines is aiming precisely at that opportunity. In any case, it was interesting finally to see some news about Mira. She was one of the first people to present GPT Voice Mode, as I recall—or perhaps it was that early ChatGPT presentation with video.
Yes, but I think they are doing some very deeply technical, research-heavy work, and it is not at all clear who is doing what there now. Remember the period when employees were leaving OpenAI and everyone said, “That is it—the system will stop developing, and it will become extremely difficult to build anything further.” Yet we are now seeing an absolutely enormous breakthrough and extraordinary progress in the final result. At the same time, we are also seeing an unbelievable amount of junk and some very strange decisions.
It seems to me there have been a huge number of costly failures. I hope those failures do not ultimately damage the models themselves. This brings us back to the question of exceptional people. What happens when exceptional people leave?
A great many very strong people moved to Meta. And what do we see from Meta? We do not see a result at all. This struggle—
Well, they have now released Muse, Sasha. It is interesting overall, although clearly it is not going to win over the mass user yet.
Yes, but who will it actually win over? It may attract a few small tests or contracts, and even then it is unclear whether it will win any contracts. Why would a person switch? Take Grok. I wanted to test and examine Grok many times, but I kept discovering that it did not have this feature, did not have that one, and lacked something else. I remember buying their first three-hundred-dollar subscription. I opened it and found poor Russian support. My reaction was, “Come on, guys—I personally cannot use this.” The English was good, of course.
Then they said, “We cannot create presentations. We cannot create documents or PDFs.” Then: “We cannot analyze large files.” And you think, “I am not such a fanatical Elon Musk supporter that I will exclusively buy every single thing that exists.” Even if I pay for an additional system, I still have to start using it.
Over this period I have bought an enormous number of systems that now sit somewhere, unused and forgotten. Meanwhile, a major lawsuit has begun. Apple has sued OpenAI because OpenAI worked with two people: one from hardware and another from design. The best-known part, of course, is the story involving Jony Ive and the allegation that they stole technologies. We are entering a phase of legal battles over who stole what from whom and who built something in parallel. At the same time, we face a phenomenal question: what is AGI?
Who will have access to it? Who will be able to use models such as Fable? I think this is the question everyone should be working on, worrying about, and thinking through—not how to create some hype, capture a headline, or add another twenty or thirty users. It is not as though companies are now telling us clearly how many regular users they have. We seem to have been stuck at the same user numbers for perhaps half a year. That is strange when the markets are still developing in ways nobody fully understands.
There is also a major question about access: which systems will be available to whom? Personally, I want to know who can actually access which systems. If you are limiting something, then limit it and explain that there is, for example, a more capable system available only to a certain group of people. Fine—no problem. But explain it. Explain the models instead of quietly changing something inside Fable, as they did when they released it. It is absurd. If you reduce Fable 5 and make it less capable, call it Fable 5 Mini or give it some other name.
Instead, you leave the name Fable 5 Ultra Code. What is that supposed to mean? What is it? And all of this can be changed continuously from inside the system. These companies have not yet grown out of startup behavior, yet their systems are already used to kill people in certain regions, to govern countries, and to manage economies. An enormous number of decisions are being made on the basis of these systems. We used to tell many stories about failures—for example, people using ChatGPT in court, or someone trying to build a business and the system returning a particular error.
Today the number of such cases is unreal. The old examples look almost harmless compared with what is happening now. Look at the Sol model. I launched it on one project involving a restructuring of a particular architecture, and for a week it was impossible to understand what was happening. When you receive a surprise like that—a hidden “Easter egg”—you think, “Next time I have to test every one of these models with extreme caution, and I will never commit a great deal of money to them.” I think that is a problem for OpenAI and Anthropic.
They have to compete with each other, including on the speed at which they release different models. Viewers, tell us what you think about all of these topics. I think we have discussed an enormous amount today.
There is one more interesting, smaller item from the world of enterprise sales. Rumors say Microsoft has started preparing its salespeople and teaching them how to explain why customers should buy Microsoft solutions rather than products from OpenAI or Anthropic. It appears that Microsoft is preparing for a fairly aggressive push into the market. It is training salespeople who will passionately argue that Microsoft's products are better, explain why they are better, and insist that customers should not buy startup products but should choose the proven, fully finished products of a large company.
We will probably learn more about this soon, but I found that little item about the sales force quite amusing.
What you have just described is part of an endless argument that will keep returning: whether to use old software or new software, and whether to use agents at all. And will Elon Musk, with the help of Mark Rakhard, conquer everything from space to your mind and consciousness—through a robot in your home and a chip in your brain? We will see you in the next episodes of ToTheMoon.