Why do you need Codex or Claude Code if you already have ChatGPT? Especially if you’re not a programmer. This, I think, is one of the main misunderstandings about artificial intelligence today. And today I want to compare four systems: Codex from OpenAI, Claude Code from Anthropic, Cursor as part of SpaceX and xAI, and Google Antigravity. But not in terms of programming and some technical tests, but in terms of what they can already do today in your ordinary life, your profession, your business.
Using my own real examples, I’ll show you which tasks you can hand over to these systems entirely and which of them is worth choosing for yourself. Let’s watch the episode. You know that on the ToTheMoon channel I’ve been paying a great deal of attention over the last three months to the idea that every person should start using apps from OpenAI such as Codex. And an app from Anthropic such as Claude Code. And this can cause some confusion, because most people actually perceive modern artificial intelligence as ChatGPT.
If you even take a careful look at, well, the new films made over the last two years or so, or watch lots of different shows, the people there who aren’t in the, kind of, artificial intelligence and tech market keep saying: “ChatGPT, ChatGPT, ChatGPT.” And essentially ChatGPT has become a generic name, the way Pampers or Xerox once did. This has happened many times before, and most likely it will stay that way. For now, at least, right? And even I use the expression GPT-like systems.
And it seems to me that this split that once happened at Anthropic, this emergence of Claude Code. They also had Cowork, then they killed it, and what they’re left with now is, there’s just Claude and there’s Code, but actually it’s Claude Code. And by the way, they also got Claude Design. They’re killing it now too, and it will live inside the Claude and Claude Code infrastructure. And at OpenAI there’s the story of what happened: they had ChatGPT, and they released Codex as, essentially, the same kind of thing, I guess, as Anthropic’s Claude Code, if you can put it that way.
With Claude, I think, there’s also a big question. Given the naming of Claude, people often call it, like, Cloud and so on. I think that can cause some confusion. And today, together with you, we’ll expand this lineup, we won’t stop at these two systems, we’ll add two more systems. Why is it that I tell ordinary people to use them?
And it’s well known, I have an example, it hasn’t come true yet. I keep telling my wife that it’s time for her to start using Codex in addition to ChatGPT. And to start using Codex from the standpoint of an ordinary person. And that’s exactly the problem that happened with Claude Code, and with this whole terminal programming thing in general, and with Codex: a very big bet was placed on them as systems for programmers. Essentially, these systems will increasingly become the foundational systems of an ordinary person, and at some point in time they’ll become a person’s main system.
And they shouldn’t differ in any way. Even if you look at the new interface of the ChatGPT app on the Mac, we’ll see two tabs there. By the way, this app is called ChatGPT. And there’s a switch there between ChatGPT and Codex. I mean, a switch has appeared inside, but the app is called ChatGPT. They’ve also had a few cases where they separately called it ChatGPT Codex, and the old ChatGPT app is still around, right, on the Mac. The mobile, mobile version of the app is different from this. And the web version of ChatGPT is different too, because it has Codex inside it as a separate, as a separate, independent kind of section.
Precisely as a section inside ChatGPT, right? Here, though, there’s switching between ChatGPT and Codex. And by the way, I’ll note that ChatGPT on the Mac today contains both the chat and the chat from ChatGPT. Someone who has come here for the first time today may get confused by this whole story. And I by no means want to drift into technical-podcast territory today. I mean, on the ToTheMoon channel I personally don’t talk about that, we have other speakers, but I don’t really pay attention to some super-advanced settings where I show, you know, how to install an app inside.
I want to get across to you the kind of knowledge that will let you use artificial intelligence in your life, in your profession, at work as effectively as possible. That’s my task, my personal task, right? Although we will have episodes coming out, and soon a new author will appear on ToTheMoon too, with more hands-on things. So, once again, this story that happened with programming. Why is it fundamentally important?
Because it’s essentially, and in general with terminal work there’s also this notion, like, well, and the basic story overall is that programmers worked, you know, a lot, like, through the terminal, including in Cursor. And originally, by the way, in Claude Code. Originally Claude Code and Codex put a very strong emphasis on the person seeing the code and programmers seeing the code. Of course, we’re going to see revolutionary changes in the programming market. And among programmers, over the next ten years, there will be serious shifts in the understanding of who, who considers themselves a programmer, right?
Or who can consider themselves a programmer? For example, can I be considered a programmer today, when essentially I studied programming more than twenty years ago, right? I graduated from university in 2005, in really serious professional programming, and I started working as a programmer even before that, right? But obviously I wasn’t doing development as a programmer. But I could build such systems with the help of people, with the help of teams, while running various businesses.
And now I’ve obviously moved into a mode where I can build such a system myself. And not because I somehow particularly like it, or not because I can’t hire programmers, but because what has changed, what has changed is the notion of efficiency and of making the right decision. I mean, situations have come up in which it’s become more advantageous for me to do something directly myself. So, today I want to hold a meeting for, well, for you to really get into these systems.
Along with these systems there will also be Google Antigravity and Cursor. And the time has come, in this case I’d call it Cursor xAI, right? I’ll add it like this, right, or Cursor Grok, Grok Cursor. That’s important, right? Because if we say ChatGPT Codex, and we say Google Antigravity, and we say Claude Code, then let’s add something to Cursor, that’s very important. Either xAI or Grok. Let’s add it. And there will be changes there too in terms of names. I think the name Cursor should, in theory, disappear and dissolve, and what should remain is something together with either the word Grok or the word xAI, right?
For people who consider themselves super-professionals in Codex or in Claude Code, who make, like, a trillion programs, I’d certainly recommend getting very deeply involved in today’s discussion. At the same time, I understand that someone will say: “Listen, I’m already, you know, programming here as it is. I could, you know, do without your stories here today.” So, look, there are four systems that could be compared somehow. Before, I only compared Codex and Claude Code. Now I do want to add Google Antigravity for you, given the Google model released literally just now, the other day. Argon, right? Why do I say the other day?
Because we haven’t even been living with Argon for a few months. I mean, it’s very little. And I’d like some time to pass so that we can really test Argon together and really check how it works, right?
But now the time has come, because, I think, the process of this, this process of transformation in terms of understanding, in people’s lives, that you can automate a huge number of tasks and change the whole approach to technological access for the ordinary person, it will change enormously in the coming year, it will simply change colossally in the coming year, when the ability to apply, in your own life, some kind of automation becomes very simple and dumb, exactly the same way that registering an email account became a simple task for people.
That, by the way, wasn’t so easy before, because you had to enter the settings in an email client to get the mailbox running, right? I mean, it has become very easy. Or opening a space for storing files. I mean, that thing has become very easy. It used to be hard, and it got very simple, when Google Drive appeared, it became very easy. And Google Mail, by the way, brought revolutionary changes, in a certain sense, overall in terms of people’s lives. Well, really serious revolutionary changes, because people got access to the ability to store lots of emails for free.
So we’re going to get the ability to run a huge number of processes essentially for free. Completely for free, right? Even if some tens of dollars get added, it’s pointless to discuss that now. It will be an automatic thing in every person’s life. Without it, simply nothing will happen. Because today people do pay for electricity, right, well, or they pay for internet at home, or pay for mobile service. It’s an absolutely simple, ordinary action that will be in every person’s life.
And right now, precisely in this window of time, over the course of the year, I want every person to get going a bit earlier, to get going just a bit earlier, try a bit earlier, right, and really try to do something serious. I remember two and a half years ago, when many bloggers talking about ChatGPT said that you could, I don’t know, write a joke or make a poem, or translate some text. Those examples were very banal, very simple, uninteresting. If you listen to those very same people, they’re essentially still stuck at examples like that in their lives.
To launch something here for yourself… Of course, there are people who got started and did some serious, broad projects, and many projects won’t bring these people any results. As we know, today the use of artificial intelligence in the world hasn’t exactly increased GDP in countries, or given people the opportunity to earn more, or made their lives more efficient. Today more than a billion users have ChatGPT installed. Well, and Google Gemini about the same number. I mean, there are some five hundred million or so more or less active users in the world, which is pretty serious, right?
That’s plus or minus ten percent of the world’s population, of that kind of active audience that’s learning something and has internet access, they have it. It has appeared, artificial intelligence has appeared. How many of these people have really made money? And by the way, I hear an incredible amount about people buying various programs or asking for advice on whether they should buy a program, because where… Someone on our channel even asked where artificial intelligence will help make money.
If someone is already telling you how to make money with artificial intelligence, right, then in these tasks that’s usually already a missed opportunity, right, and lost time. Usually, as one of my partners says. It’s usually already lost time. On the one hand, I want to walk you through my reasoning, well, give you a slightly broader understanding of such systems in general, show the need for such systems in general. Maybe somewhere, for you, so that you hear some new thought and do start using something differently. On the other hand, to still look at these four systems today independently, a little independently.
Although, of course, once again, we’re unlikely to get some astronomical number of examples here. But! The basic things and my own opinions I will give you today. I’ll give them very seriously.
I am, of course, a very serious professional user of Codex and Claude Code. Possibly a super-professional one, because I use these systems, unlike, for example, programmers or architects, or just project managers, I use them in fundamentally, seriously different ways. I have a great many different projects, and because of what I do I have to, on the one hand, build some high-load systems in some places, analyze a very large amount of data in others, and in others use such systems for ordinary family tasks, right, along the lines of analyzing emails from the schools about how my kids are doing.
So, one of the simplest examples and useful things today that you can try out in all of these systems is connecting your email to these systems. I mean, take all your mailboxes and connect them to either Codex or Claude Code. I want to say right away that it’s easiest to connect to Codex, in my view. But you can connect it to something, right, I’ll be calling it a Codex-like system, because for me, after all, even Codex today is a more serious infrastructure, because there’s ChatGPT.
For me, ChatGPT still surpasses Anthropic’s system in quality and on various parameters. At the same time, I want to say that for me Claude Code, as a system for developing various solutions, is more professional, it’s more professional, it’s cooler than Codex. Right? I mean, for me Opus 5.5 and Fable, if there are more professional people here who understand what that is, it’s the choice of models, right? In, for example, Max mode, not even in Ultra mode, like, or Ultra Code, they’re better than the 6 Astra ChatGPT Codex or 6.1 Sol systems in Max mode.
Right? I mean, for me personally these systems, for me personally Claude Code matters in developing big things. Look, there’s one remark here. In life you’ll come across a lot of people who use Codex, Claude Code for various systems. I mean, someone creates some bots, someone creates websites, someone analyzes sales phone calls, and so on. Someone, some cosmic systems. People send a lot to ToTheMoon, someone is building some cosmic systems. I don’t know why people even build these systems.
Someone builds systems for aggregating different models with each other, someone builds systems for analyzing, like, token consumption, someone builds systems for managing businesses and business processes in companies. I mean, they make some million, some billion different systems. In ninety-nine point nine nine nine percent of cases it’s unclear why. But what’s the point? That a lot of people will be telling you things and sharing their own opinions, their own approaches. My task is for you personally to start using the system yourself, right?
And for that you need to start using at least ChatGPT in your life every day, in various processes, right? Without that, it will be very hard to really feel Codex, Claude Code. I mean, if you use Codex and Claude Code only for work purposes and a little bit for everyday life, it’s very hard, really very hard, to feel its variety. And for people whose life is limited in terms of the flow of events, I mean very narrow, very narrow, for example, work, a specialization, I don’t know, a person does one repetitive action, I don’t know, at a factory, for example, and at home he doesn’t have, say, a family. The person lives, for example, alone.
And he has a heavy work schedule, he works, I don’t know, twelve hours a day or fourteen hours, then it’s even harder, because there aren’t, in his life, situations and cases for which you can create various automations. And here it’s worth making things up for yourself too. Although, of course, making up scenarios that aren’t real is very hard. It’s, as you know, there’s nothing worse than developing various automated systems and trying to make money on them or attract investment.
A system that, you’ve imagined it’s big and cool, but in reality nobody needs it. Only you need it. Right? Because nowadays, with the help of artificial intelligence, you can improve systems endlessly, just endlessly, and nobody will use them, and nobody will look at these systems. Fortunately, today artificial intelligence will help you a lot to track and understand the statistics, whether your systems are being used or not. And who is actually using them and how they’re using them. But in the world we often don’t understand how to evaluate it.
Like, my wife often comes to me and says: “Look, this girl has a really cool, like, blog, she’s got, you know, a reel, like, I don’t know, five thousand people watched it.” I say: “What’s five thousand? That’s nothing at all. In reels. They watched the reel for, like, thirty seconds or twenty seconds each.” Or ten thousand people watched someone’s YouTube video. The person evidently used, say, advertising. So, of course, so with average ad traffic, the average watch time will be, I don’t know, one minute or thirty seconds.
Well, okay, it might get up to a minute and a half, somewhere around there. So. It’s just people watching for a minute and a half each. That’s nothing. I mean, we, as of today, often don’t know how to evaluate what, in reality, the effectiveness of this or that system is. Just as people couldn’t evaluate it five, ten, or fifteen, or twenty years ago, people now also evaluate everything very relatively, by some external aspects that aren’t entirely clear. I have a lot of acquaintances here in Silicon Valley.
I know a great many people, both young and older, speaking different languages, of different nationalities, who build systems that bring in a lot, many of them have a lot of money, just an unreal amount of money, and with incredibly big names. There are lots of funds in there, somewhere. But you look at their companies and you see that it’s an empty shell, that there’s nothing behind this company, that there’s no substance in this company. It’s a useless story. These people have thirty, forty, fifty, a hundred people working for them, and you understand that it’s an empty shell.
Your task, though, is to effectively join the field of artificial intelligence development. And so the very first simple task, which today, maybe, we’ll end up with a fairly long video. Once again, this is very important. I always want to tell you everything through various real-life examples. Very simply. You can always get an answer by going into ChatGPT and writing: “How do I use Codex properly? Give me ten cases.” But I want to tell you that ChatGPT never once described it properly for me, no matter how much I tried. By the way, I spent a long time preparing for this episode.
And over this week, for this episode, it, it never managed to compare the four systems with each other for me. Because the way it compares Codex and Claude Code, for example. Well, so, comparing them, or, like, Google Antigravity or Cursor — it’s all a comparison for techies, or programmers, or geeks, or a very narrow audience, an incredibly narrow audience, and not for the bulk of ordinary, regular, regular people. And some of you say: “You over there in your California have gotten fat and spoiled, you know, you think life only exists somewhere over there.” Or someone says: “We don’t have a hundred dollars here to pay for a system.” Believe me, I try like no one else to get everything across for regular people, for the effectiveness of ordinary, regular people.
This regular person could be someone who earns five hundred dollars a month, or it could be a regular person who earns five million dollars a month. Well, he remains the very same person, and he can hear for himself, at the same time hear for himself, different ideas. By the way, we have very different people on the channel: awfully serious ones, with titles, owners of serious companies or managers of serious companies, or from government bodies, with ministerial slogans, not slogans, positions.
And at the same time very ordinary people: pensioners, and young people who maybe don’t even work anywhere and are still studying. The simplest task you can do is to connect your email, right. And to connect your email not so that it deletes unneeded emails for you and cleans out your inbox — that kind of pointless thing — but so that, if you live, you have email, if you do have it, look, this is very important, so that this system starts going through that email for you every day, and you start interacting through this interface without opening your mailbox.
Look, many people, when they make a case, we even had one of the people who… We’ve had this open thing going, you can — it’s still open — you can become a speaker for ToTheMoon. Well, the links are somewhere on our channel, on YouTube, or they’re on the website. And on the ToTheMoon website, it’s all within, within the Besolid brand, right. And one person sent in that he had connected his email through OpenAI Dot. By the way, how do you all like Dot? I, I like this approach. I shot a separate episode on this topic.
And, by the way, write in again, now, which of you uses what: Claude Code, Codex, Cursor, Google Antigravity? Who uses this for serious aspects of using these systems in general, right, and not just for programming, roughly speaking, through the terminal. So, one person sent in that he had connected Grok Dot, and it helped him put his email in order. I want to say that this case isn’t interesting to me. Why? Because it’s, this case is only relatively useful when it comes to putting email in order.
Someone will need it, someone who has a huge amount of email and who, by putting their email in order — and maybe that person did exactly that — will get some interesting thoughts for themselves. For example, they’ll get interesting thoughts for themselves and draw great conclusions about their business. And that really is a plus. Maybe this person did that for himself. That’s a huge plus. By the way, I think he did exactly that, from a business point of view. But many people free up, like, some space, or set up a case where an agent corresponds by email and schedules some meetings.
But what I’m talking about is you creating a new interface of interaction for yourself. I mean, first of all, all four of these systems, but most of all, of course, Codex, are systems that give you the ability to build an interface of interaction between you and a huge number of other systems. It doesn’t matter which systems: it’s an email service, in some cases it’s even sending messages by email. Look, by the way, that same person connected Telegram. That’s a good case. You have to be a risk-taker to open up your Telegram, but actually, for a lot of everyday living, if you don’t have anything super secret to hide there, it’s also a cool thing to give Codex access to Telegram.
ah, give Codex access so that it keeps track of some of your reporting, for example, internally, which relates, for example, so that it keeps track of your bills internally. If you need, for example, to sort out certain utilities, give it access to that. And what’s cool today is that in Codex, within one task, you can log in inside the Codex browser. Look, this is very interesting. Inside the Codex browser you can log in to five, to six systems through which you make payments.
For example, mobile phones, electricity, gas, I don’t know, some kind of mail, something else, the internet. Log in through the different systems and tell Codex to collect the data from there that concerns your expenses, the ones you’re incurring. You can integrate Codex with your banks. And if you don’t want to log into a system, for example, if it happens to be logged in, maybe it could spend your money. You shouldn’t do that. But if you can log in read-only, or so that spending money means entering a code. Many countries around the world have that. Then do it that way.
Or, to analyze your financial systems, upload your statements there at least once a week, upload the statements there in order to do the analysis. I want to say that ChatGPT has a very cool new module called Finance. At first I was rather skeptical about it, and suspicious. Now I like it, because it’s integrated with a lot of banks, and once I loaded a few dozen accounts into it, it gave me a great thing: I can ask at any point in time. I even had a case just now. I was buying a mini computer on Amazon for a micro server, and I had to place the order four times because the delivery wasn’t going through, and I couldn’t tell how much money had been charged to me.
I told Codex: “Analyze it.” It was all very recent. I needed to look at practically yesterday’s money. I told it: “Analyze my statements.” It, since it has, as it were, live access to the statements, it did the analysis very quickly. By the way, one of the cool cases combining email and the bank was this story.
This, by the way, is already the level of serious analysis. I recently went on a trip to Europe. I visited: Poland, Holland, Spain, France. And I had a whole lot of different locations, payments, hotels, bookings, flights and everything else. And I didn’t check how much money was being charged to my cards, how much was supposed to be charged, what the bills were, what bonuses had been credited where. At many hotels I was supposed to get credits refunded, free, you know, breakfasts here and there, and so on, right?
And I asked Codex to analyze this. So what did it do? It took all my email, looking at what bookings I’d had. Because how could it know where I had what, how much in total was supposed to be charged. And in the email there were already statements from Expedia, from Booking, from Chase Travel, from Amex Travel, right, from the Marriott apps. I mean, well, different, you make bookings in different systems. And I want to say that it found things retroactively, going back several months, even more. It even found things from before France.
Before that, at the beginning of August, I was in New York, it found that. In July I was in Hawaii, it found that too. It found various details where hotels hadn’t applied my breakfast credits. I mean, they were supposed to apply them and didn’t. Since I travel a lot, I can’t keep track of all that. Then it, look, after that it drafted emails to the hotels for me on its own, long emails, and sent those emails to the hotels. And when the hotels replied, it analyzed those emails itself and wrote back to the hotels until it had sorted everything out.
And I want to say, by the way, that there really were mistakes, and some of the hotels refunded certain amounts of money. In some places there were no mistakes. It says: “I’ve looked into it, there are no mistakes.” Look, this case is remarkable, it may sound very simple, and if we all look at how Codex is being used, then with Codex or Google Antigravity everyone will be telling you about, I don’t know, building complex high-load systems, solving some math problem. But this case, it’s very fundamental in terms of forming, in your way of thinking, a new process of interacting with different systems, a new life with different systems.
And this, this whole thing about scheduling a meeting with someone through an assistant, I remember, was a real pain point for the Valley. Everyone said: if someone builds a startup where a system schedules meetings, they’ll win. Well, they won’t win at all, because that’s child’s play compared to what the system did here. I want to tell you that you have to be a very serious assistant, or a secretary, or a finance department to analyze documents like that, make sense of the whole concept of bonuses, discounts and promotions everywhere, and write such high-quality emails to all the hotels with an understanding of those hotels’ systems, the way Codex did in this case.
In general, when we talk about systems like these, then once again, for me the fundamental thing in all these systems is when the system does something for you, fully. And the reason I told you that today you can log in through Codex, tell Codex to connect to different systems, is that just two months or so ago this was hard to do. Codex didn’t handle it very well, and the browser inside Codex didn’t work very well in places either. Now everything works much better.
And even systems you can’t integrate with via API or via MCP are becoming less and less of a problem, right? Here’s one of my latest stories. I have a very complicated story, let’s revisit it, I’ve talked about it before. One of my camera monitoring systems, more than twenty cameras, sits on a system in the cloud storage of a company, of Amazon, and there’s no API of any kind in that cloud storage. And in essence you can’t manage anything directly yourself when working with those videos. And you had to use third-party, very complicated middlemen, which I never liked, and give access to third-party systems.
And now, as of today, literally today, Codex has finally integrated with everything completely for me, connected to those systems by itself, downloads the videos by itself, and from there I’m able to manage this infrastructure independently of this infrastructure, that is, not depend on the problems this infrastructure creates. Because, you see, imagine: you book a hotel on Expedia, they give you a certain discount, or on Booking. But I’m not even talking about, for those who live in America, the fact that credit cards give you certain benefits, right?
Because many people in the world don’t have that, but in America it’s a whole huge system, right? Many of my cards give me money back on hotels, breakfasts, guests… well, lots of different bonuses. And the hotel and this system are, in essence, relatively integrated with each other. I mean, when you get a bill from a hotel, you have to check everything there in great detail, thoroughly, in order to understand whether it’s even genuine or not, whether everything there is right or wrong.
And this case, it’s really non-trivial. I don’t want to check these bills, I don’t want to deal with the hotel, and separately with the bank, because the bank also takes money its own way and by its own rules, often, as we know. And separately there’s also a third system that I booked through. I mean, I don’t want to deal with this, I want to have some other system that will check all of it, that will study all of it completely, find everything out, look, figure everything out and find it.
And when the ticket came, it wasn’t me driving the car at that moment, it was a friend of mine. And with the help of Codex I transferred the ticket from me to that other person. And in the end I didn’t know whether it had been transferred or not, because it took time. And Codex, which was keeping track of it, at some point, yesterday I think, sent me a message: I checked through the French system, your request has been reviewed, the ticket is no longer on you. Now your friend has to check further. Let him check his email spam for a notification from this company, let him look into these links or send them a request, and so on.
I mean, the system figured this out by itself, or did it by itself. Or, for example, I recently sent, I needed to get a power of attorney done, and I sent a letter by mail to an agency, along with an envelope that was supposed to come back to me, a prepaid envelope. And I immediately opened Dots from OpenAI and told it: I’ve sent a letter, here’s the number you need to track for how the mail comes back to me. Sometime the day before yesterday Dots came back to me on its own and says: your letter is now in transit, and you should receive it tomorrow. And today I received that letter.
I mean, the system tracked everything that was happening by itself. That, of course, is a completely different level. And what I want to tell you is that we all need to understand that we’re moving toward a time when you personally, yourself, will be creating things inside these systems without extra middlemen.
This is very important. I mean, you won’t be making the many settings that you have to make today in Google Antigravity, or in Codex, or in Claude Code, or in Cursor. And I recommend that you use these systems to the maximum. Those of you who aren’t super-developers. Super-developers — okay, no problem. Those of you who are ordinary people, under no circumstances slip into the mode where you start using super-features. There are a lot of additional features there. For example, these features may be recommended to you. For example, in Google Antigravity, which is a relatively new system from Google.
Google Antigravity came out in the spring, at the end of spring this year, and now Argon has appeared there. It’s a Codex-like system. I think it’s lower-tier, really, seriously lower-tier, as an infrastructure in itself. And in this system you need, for example, for the system to carry your goal through to the end, you need to put, to write a certain command there, for example, to write a command like slash goal. Or you need, for example, for Cursor to launch a certain mode, you need to select long-running agents there, right?
It’s a separate mode for long-running work, with the plan agreed in advance and the agents checking each other. And in essence you’re forced, or led, to a point where there are a billion settings, like there once were in Windows back in the nineties. There are a billion settings that you, well, first have to install and configure for these systems to work, right? It’s like they say: write all the settings into files, agents.md in Codex, for example, make these settings. If we even just open the settings in Claude Code or in Codex today, there’s already a mess in there.
And when they say: “Set up some environment, set up some stuff,” you start wondering what that even is. I mean, it’s very hard for ordinary people to understand. You don’t need to get into it. The task in front of us today is to build in you an understanding of how the new future cases will work. Because all these things, all these additional settings you can specify by hand, they’re pointless.
You will never in your life do the setup better than artificial intelligence itself, actually, will do the setup. I mean, today the task at hand, today the task is not prompt engineering. Today the task at hand is that you come to Codex and say: “Launch a website for me, I want to see this on it. Or analyze my emails to see whether I’ve been charged extra money.” Before, you would have had to say which emails, from which hotels, and which details you want to look at inside, and which bank statements, and what you want it to show.
And note that there are transactions that have gone through, that haven’t gone through, that were temporarily put on hold, that weren’t put on hold. And look at whether I have bonuses or don’t have bonuses. Today, the task I set in Docs, I simply asked it — very important, this is very important — I simply asked it to figure out whether the hotel had charged me extra money. This, this is incredibly important. I mean, its task, this artificial intelligence’s task, is itself, on its own, to understand in advance that there are credit card bonuses, that transactions at a bank can be modified, can change, right.
They could have put one amount on, on hold for you and actually charged a second amount. That hotel bills can be sent a little later, and so on. I mean, its task is to figure this out by itself, not for you to think about it. Exactly the same goes for launching different systems. I mean, our task today is that if you come and you want to do some case along the lines of sorting out your finances, for example, how this year you need to file your personal tax return for this year in your country, the one you’re in, or figure out whether you need to file at all.
You give it access to certain documents or to tax returns from previous years. You give it access to some of your statements and say: “Do it!” If it has questions, it should ask them. Look, if the system doesn’t ask those questions, or the system produces complete gibberish or nonsense, that system doesn’t work for this task. Look, this is very important to understand. It means this system doesn’t work. It means this system, in terms of decision-making, in terms of various conclusions, in terms of various details, this system is no good.
At one point one of the crazy advantages of Codex was that Codex, over Claude Code, by the way, at some point in Codex you could spin up a pretty decent, serious project. And at the same time it integrated by itself with systems that stored your keys, like Google Cloud, or with server infrastructure like Vercel, or with a database management system like Supabase. I mean, Codex deployed everything together by itself. Of course, you might have to go in somewhere, sign up, get API keys, but basically from there it did all the setup entirely by itself.
We’re moving toward that very seriously. I think, by the way, that even projects like GitHub are bound to run into problems at some point. It would seem to be a very promising project, but either someone has to buy it in a really serious way, the way SpaceX once bought Cursor. By the way, I want to say that for Cursor it was an incredible deal, right. Tens of billions of dollars. I don’t remember what the final price was, whether it was fifty billion or not in the end. For them it was an incredible story, because at some point, if SpaceX AI hadn’t come along, they would have died. And notice what happened in Cursor.
For example, in Cursor you can no longer turn on an OpenAI product. And in Cursor by default, when you launch Cursor, they have, for example, a mobile app for Mac. And in Cursor by default, even if you open, I don’t know, the free version, you’ll have Cursor Grok 4.6 Medium set. I mean, by default it’s a thing there that works that way. Yes, there’s also a certain tie-in there in terms of offering integration with GitHub. But even now Cursor already asks: “And who are you? Are you a developer, or do you just want to manage projects, or are you a designer?” I mean, it already asks what kind of person you are in general, which, which cases you’re going to create next.
Of course, a non-linear thing like that won’t lead to anything good, but still. I mean, I mean, all these systems, like Codex, like Cursor, like Google Antigravity, they’re of course needed so that you can fully carry out any task you want in them. You want to analyze bills, you want to launch your own mini website and put your jewelry on it. You want to analyze five hundred photos from a recent trip and pick the best ones or delete some duplicates. You don’t have enough space on your computer, you want to analyze which files aren’t being used or which files in there are unnecessary.
You want to create a notification system for yourself that will send you reminders in Telegram about some of your trips, or a notification system for your calendar. By the way, notice that calendar apps themselves are becoming pointless. There’s… In development there’s this concept when… It’s a very trendy thing in the world.
There are also systems, by the way, systems that will die out completely, right? I periodically talk about systems that will die out. Here’s one more system that will die completely. There are systems that let you schedule meetings between, with people, right? And they, like, give you a calendar module. I mean, you can put the calendar module on your website, choose in that module which, what times suit you or don’t suit you, in terms of, I don’t know, say you, I don’t know, work as a massage therapist, and when it’s convenient for you to see people.
And you tell this system, add this module to your website or integrate with it, and people book with you through this module, and you keep track of your appointments. But all these systems are always limited. Today you have the ability to make your own appointment-booking module, roughly speaking, with a snap of your fingers. It’s already here now, the ability to do it with a snap of your fingers. For some people it’ll be a bit harder to do, because you still need to propose certain functionality for the system.
But I want to say that even if in ChatGPT, in a good version, you ask for the functionality you need, it will immediately suggest everything to you: that there shouldn’t be duplicate slots, that you need to send certain notifications in advance, cancellations and so on. Today you’ll be able to build a system like that with practically any messenger of yours, whichever you have, or launch it on a website. I mean, launching a system like that today won’t be any trouble. And not in order to develop systems like that yourself, but so that you can make it for yourself, tailor it to yourself.
And these systems also offer ads. Look, there’s this concept of payment systems in the world. In the world the leading, some of the leading systems out there are, for example, Stripe or PayPal. These systems, very often after you’ve paid, they offer ads. Not all of them, but many. Or they offer some infrastructure of their own, keep inviting you to sign into their infrastructure. After all, your, or rather my goal, if I’m selling my product: I don’t want anyone to be offered other ads.
I want the person to stay on my website, not even go anywhere, not have a button hanging there at the end saying “return to the site where you made the purchase,” but keep being on my site, keep looking at other things. And often you go in, pay for something, and there it says “return to store,” you don’t even go back to the store. I mean, I want the person to stay where they were, with me. And in this case, how do you do that if you use these particular middlemen that show their own ads? Most of these systems are going to die.
And you, as you bring these various cases into your life, you start looking at all of it differently in terms of how you can apply this, this, in your own life, or in your profession, or at your job. Here’s an important thing that exists in artificial intelligence. If you use artificial intelligence in your profession, right now everyone talks about how they automate everything, do everything the trendy way, and so on. Today someone told me that an acquaintance of his had automated everything so thoroughly, in some prehistoric offline company, that without him everyone will go crazy.
But with artificial intelligence today nobody will go crazy, because if you had some person working for you who automated everything with artificial intelligence, you’ll copy all of it just the same and study all of it just the same. If you’re a competent, smart boss sitting nearby, you’ll just as easily learn everything this person is doing. And on the one hand, smart, competent people who understand artificial intelligence are leading today and can earn ten times more. On the other hand, these very same people, having done a certain job, are kind of not needed, because if at some point they set up some business process, then from then on they’re no longer needed in that business process.
As someone, a person on X, said recently, in twelve months there will be systems that will be able to record a week of a person’s life and then fully repeat everything, will be able to work without them, that is, will be able to keep developing things and doing things without them. Why do I consider systems like Google Antigravity or Cursor to be lagging behind Codex, or even behind Claude Code, today?
Because of the infrastructure that exists, the full-fledged infrastructure that Anthropic has today, and the full-fledged infrastructure that OpenAI has in terms of the models themselves inside. It’s like some people come along here and say: “Try Grok Bot, it’s better than Dots, for example.” But it can’t be better than Dots for me, because I have ChatGPT linked with Codex, linked with Dot, and Dot is linked with everything, and they’re integrated and connected to a huge number of systems.
I’ve already spent time in thousands of chats in them, and they have a memory of what I’ve been doing. And I don’t have to tell this system every time where I live, what I do, what interests I have. This system knows about me better and better, and better, and better, and better. It gathers everything better and better. Some of you may say: “Then you can export all your chats from ChatGPT and upload them to other systems.” Indeed, there’s now such a thing as exporting and importing chats, and in essence it can be done; even if that doesn’t exist on the import side, you can export almost everywhere, I think.
At the very least you can then add it to Codex, or add it to some other system, and it can learn who you are. But learn which qualities? I mean, what will the quality of the end result be? Here, of course, we come to the story of having to pay money for artificial intelligence.
And you have to pay no small amount of money. And do I use all the systems I pay for, the expensive plans, to the max? I don’t. Not everywhere. In some places I do, I squeeze every last drop out of the plan, like out of the Pro plan at OpenAI. Well, for example, I could, probably, out of the $200 Pro plan, I squeezed every last drop, out of the $500 Pro plan I don’t know whether I’m squeezing every last drop. But I know that I don’t want to miss the chance for the system to give me access to some cooler model, and for that cooler model to get something done.
And people, by the way, who keep developing all these various middleman tools and send them in, I’m telling you, people send me loads of middlemen, systems that do some third-party thing. Let me remind you that these systems, they don’t work with the most expensive, coolest models in the world, such as Fable-5.5, right, such as… in maximum modes, such as Opus 5.5, or such as Astra GPT-6, or such as Astra 6.1 Sol. They don’t work with those systems. Or with the maximum modes of Google Antigravity. I’ll note, by the way, that Google Antigravity is a fairly expensive model.
Every so often someone comes along and says: “Look, just, literally last week a model appeared, that beats GPT-6 Astra and basically all models like it, and beats Fable, and costs, like, a thousand times less in 3D modeling.” In some individual cases — yes. But within general scenarios across all areas of life, where I need to analyze at the same time my health, and my lab tests, and my dealings with lawyers, and my dealings with government agencies, and with the tax office, and with business partners, and with clients, that system doesn’t suit me.
And I want everything to be together in one system, so that this one system knows more and more about me. And so, of course, for geeks, or if some version doesn’t work in your country, or if under some condition you have a special plan. A lot of people have a Pro plan with Google for various reasons: some because they’re studying at university, some got it at work, some had a Google Drive plan, and they just don’t want to buy another system. Okay, you can use that system instead of ChatGPT Plus, for example, right, or instead of the Pro version of Claude.
But still, of course, I recommend having two systems: one for $100-200, that’s usually roughly the average price, give or take, and the second one for $20 or so, to have another system. Well, if you can’t have two at $100-200. So that something is constantly being done inside these systems in parallel, right? There’s a huge difference depending on which model you’ve chosen to answer you in Codex, or chosen to answer you in Cursor. An incredibly big difference.
There’s a big difference between the countries where these systems are installed. I mean, if you ask Dots a question in Europe and ask Dots a question in America, you may get different answers. The quality of how that answer is carried out can vary enormously. In another country it may be restricted because of local laws, or restricted because the company, for example, Anthropic or Google, decided in that country to, I don’t know, follow the political will of that state more.
Or, the other way round, to follow the specific written laws more, or, the other way round, to follow established global practice more. I mean, each system decides in its own way what to do, when to do it, where to do it. In its own way. And you and I can’t find out how exactly these systems make decisions, why at some point in time they made this decision or didn’t. Look, an important aspect when you work in such systems is the number of tests. You have to do hundreds of different tasks, even the strangest tasks, in order to figure out what works better or what works worse.
Create thousands of different reports, in Excel, and in PDF, and as images, and in PowerPoint, wherever. Well, through Codex, once again, create them through these systems in order to learn to tell the difference, or to pick some system for yourself. Because you can’t judge the quality of a system’s work by a single task. As of today, the most expensive models, such as, for example, ChatGPT-6-Astra, can easily fail to finish and fail to understand the most elementary task and just go off track.
And Fable, by the way, can also go off track in Claude Code. Although for me, in terms of building complex systems, Fable-5.5 and Opus-5.5 in Max mode, that’s probably the most optimal mode, they win. Design, for example, I do with Opus-5.5 in High mode. I think it’s more efficient, it’s more than enough. But everyone has their own take here in terms of what’s going on. By the way, I want to tell you that still, for example, creating designs, I’ve kept doing it in Claude Design, because the quality of the various platform solutions, presentations, it’s of course a higher class in terms of fonts and in terms of all sorts of micro-layout details and so on.
But again, basically, look, we should be striving for a system. And you will, you will end up with a system that decides by itself where to do that kind of design. Or it will ask you: are you ready to spend zero on this design, or do you want to spend two hundred dollars on this design? And you’ll spend, you’ll just make the decision, and having spent two hundred dollars, you’ll get something of higher quality. We’ll get to that time when the system starts asking whether you’re ready to spend a bit more on this task to get it done better, yes.
Or if you don’t want to spend, then sorry, I’ll quickly give you a certain kind of design. Of course, now in modern systems you can work with audio files, with all kinds of recordings. Here’s one of the interesting tasks I did recently: I took all my voice memos, tens of gigabytes of them, which I had made and saved on my iPhone over fifteen or sixteen years.
I analyzed them with Codex and sorted them into different folders. And I want to say that, of course, I absolutely loved how it did that on the Mac. Because, first of all, it automatically saw all the text on the Mac, and, well, plus Codex, look, can today spin up a cheap local audio-to-text recognition model for you. And it’ll spin up a cheap Whisper, and that’s more or less enough for the micro-conclusions you want to make. And, for example, if you recorded your sessions with a psychotherapist, or recorded various personal analysis sessions of yours, sometimes people talk things into a voice recorder, you have the opportunity to analyze those sessions, and analyze them literally in great detail, put them in order, break them down into the form you need, a useful form, lay them out really well, beautifully, or save them for some future analysis, for when in life you need to go back to some session, and you’ll easily, easily go back to it.
I mean, you write when it happened, and it finds it for you. I mean, it’s not just a story where you have to open, say, some separate app that analyzes these voice memos, and it does it. Rather, it’s a story where you give a request in Dot, for example, well, as of today that’s the central system, a kind of easy entry point both to Codex and to OpenAI’s ChatGPT. You give it a request to find you something that was four years ago, and the system will itself analyze your email, and look at your photos, and look at whatever files were created at that time, and listen to your audio recordings from the inside, and study your notes in there. And it’ll come back to you and give you some conclusion.
You don’t even know what exactly it will, what solution exactly it will offer you. By the way, that’s one of the non-trivial tasks being solved by these companies today, and we talk about it in our episodes with Max Grigoriev. One episode came out on Wednesday. By the way, how did you like it? Let me know, watch it. And a reminder that I’ll be very interested to see from you who has, who uses which systems more. And please share cases, great cases. Let’s pour in the cases today. And, well, let’s not forget to support our channel in general, like it, right, especially if you’ve watched up to this point.
Turn on the bells and all that, and subscribe, and send all these episodes to your friends. By the way, yesterday Max Grigoriev and I recorded the second episode. So on Wednesday there was the first episode, and we’ve also recorded the next one, about artificial intelligence safety. It’ll come out next week. And Max said a cool thing. He said: “For the first time I’ve had a great, full-fledged conversation on what safe artificial intelligence is.” He lives in Silicon Valley, by the way, and he’s an IT guy, a really hardcore techie when it comes to artificial intelligence, yes.
And he, well, he’s plugged into the artificial intelligence life of Silicon Valley. And he said that from now on he’ll be sending everyone, if they want to hear what artificial intelligence safety is, here to us, to this video, so that they watch about it, learn, right, watch and understand in general what it is, right, and how these systems work, what they do, what happens with them. So once again, write how you liked it and which systems you actually use yourselves in your own lives.
In terms of Codex, Cursor, Google Antigravity, of course, I’ll be very interested to see cases. I, I hope Google will develop this system a lot. Google has a huge infrastructure, over the last six months they’ve really lost their way in the market, yes. A year ago they were really, well, I even remember Sasha Mashrabov saying: “That’s it, we’re betting on them. So they’re going to be the leaders.” I still stayed with ChatGPT and with OpenAI, as a kind of ChatGPT devotee, I guess, but Google really did surge ahead.
But I never did start using Gemini, because of their quality. They were really behind, well, I don’t know, for me Gemini and ChatGPT are just not comparable in the paid versions and in the paid solutions. I mean, for me OpenAI has stayed the leader at the top, yes. And I want to tell you that, many companies on the market that, it would seem, should be using, for example, Google or Microsoft solutions, and they have Copilot inside. Of course, Copilot can use various systems under the hood, but I think that’s an absolutely losing story for Microsoft, yes. They use Claude Code.
Many companies that are supposedly meant to use Google, Google and Microsoft solutions. And by the way, the story with Microsoft.
Microsoft is simply a dying infrastructure, and they… A few episodes back I made the remark that even a company like Apple could disappear, yes. That’s by no means my forecast, but the probability is fairly decent. I mean, it’s probably tens of percent as of today, given how artificial intelligence is developing. So, Microsoft as a company, its probability of disappearing is much higher. And here I support Elon Musk’s position, when he launched the creation of the company Macrohard, or more precisely, the Macrohard project.
It’s like a kind of anti-abbreviation, an opposite abbreviation, right, a reverse abbreviation of Microsoft, about developing this huge amount of software with Grok and with xAI. And what will Elon Musk definitely be heading toward? Toward them, at some point in time, rewriting all the software in the world. And fundamentally, ChatGPT, OpenAI recently announced at its, by the way, latest conference too, which I attended here, in San Francisco, you know, they said that they won’t, you know, be building software, and will rather make a platform on which people, on which companies will be able to develop solutions. Well, that means they’ll lose.
But I think they won’t lose. And they will, of course, build an infrastructure in which they’ll rewrite absolutely all the solutions. The time will simply come when they can really rewrite these solutions very quickly. And what wasn’t possible… Even, look, Codex, along with Claude Code, started working properly in April. Some may say March, some May. Well, that’s when it became clear that it’s a living system. And half a year has passed. And what can be done now is, of course, simply incredible.
And what can be done now with a snap of the fingers was, just a year ago, simply unfeasible and out of reach. And now it’s already possible to rewrite ninety percent, probably ninety percent of solutions.
Just now a very big problem has happened with the help of, with the PHP language and with websites, and we’ll talk about it in the Sunday episode, which comes out in a few days. And a fairly serious bug was found there, and among other things it was used inside OpenAI, and even, a huge number of reports appeared that lots of companies had started taking their websites offline, yes. Because attacks hit these sites once such a bug had been found. So problems like that… Watch on Sunday, or if you’re already watching later, go back to that episode.
But the point is that there will be more and more problems like that, and such problems will be more and more serious, incredibly serious, very serious. What awaits you and me is a time of software decompilation and of creating a huge number of copies of software. And not just software. Someone recently, there was one person who got offended at the company Photoshop and made clones of it because that company wouldn’t let people cancel their subscriptions properly. That’s a well-known story, but that company will definitely have problems. Clones of big infrastructure projects will start appearing, like Booking, like Airbnb, even like Stripe.
Although there are other things there, they do compliance after all, they remove all sorts of risks, remove all sorts of fraudulent clients. But all of that will get solved. At some point in time it will be possible to make a system that has no fraud. See what the point is? Because right now there’s actually just an incredible amount of scams and fraud in the world. There will come a point in time when you’ll be able to work in isolation and very cleanly, without these additional systems. It’s like without antivirus systems.
Antivirus companies, on the one hand, are necessary now, because they find these, find these viruses. I mean, they don’t create systems that, that protect against existing viruses. What makes the antivirus companies that exist now useful? They find viruses at the start, and then they provide this whole thing where, these antiviruses get blocked there and recorded and so on. But at some point in time, in terms of the development of artificial intelligence, obviously companies like xAI, companies like Anthropic, OpenAI, will know internally and work better than any antivirus.
I mean, these companies clearly know already. Well, let’s put it honestly. Today, internally, they obviously know how, and are able, to build a defense system better than any antivirus. I mean, we’re moving into a completely different, rather paradoxical time, where at the same time… Max and I, in our new, next episode, which you’ll be watching, about AI safety, Max said a good word, that it’s chaos. And he said that there’s chaos that brings benefit, more benefit, and there’s chaos that brings more harm.
He gave war as an example, right, that if two sides are at war, then there’s chaos there and more harm there. And there’s chaos where there’s benefit. And he put forward the idea that, from his side, it seems to him that artificial intelligence is chaos that will bring more benefit. But we’re moving toward the kind of chaos where it’s hard to even imagine what will happen with it. Because artificial intelligence will enter offline life more and more. And for that I recommend you use systems like these. In a moment I’ll list some more features for you.
Systems like Codex or Cursor, or Antigravity. I specifically recommend using them very intensively. So that you see how it will actually enter offline life. Because if right now, it would seem, a person will keep going to a coffee shop just as they always did. But is artificial intelligence capable of influencing my choice?
Look, today an artificial intelligence system can still influence my choice. And if I get into the car today and I’ve got Google Gemini running there, right? I say: “Hey, Google! Drive me to Cafe Venetia, Palo Alto.” There’s this well-known cafe in Palo Alto. Then even if I hadn’t told it I was going there, Google at that moment can say to me: “Are you sure you want to go to Cafe Venetia? I checked, there seem to be a lot of people there today, or, on the contrary, few. I know you like it when there are people around. Or the atmosphere in town is kind of strange today. Wouldn’t you like to go?
A new cafe has opened in this neighborhood. Or wouldn’t you like to go to your favorite cafe in Los Altos, Satura, and sit there in peace? I’m looking at the cameras there, there are tables available, and there are, for example, a lot today, there are some interesting people there today. Or I see that you’re driving alone, and on your own you’d probably enjoy a walk in this location more.” I mean, shifting my focus away from a certain place. And obviously, today ChatGPT, for example, has, has access to my financial system.
It can look at receipts, it can find out from the receipts whether over the last six months I bought more at Cafe Venetia or at Cafe Satura, Los Altos. Let me ask Dots right now. I’ll ask it, I’ll record it right here together with you. I’ll ask it where I had more purchases. “Where did I have more purchases over the last six months? At Cafe Venetia, Palo Alto, or at Cafe Satura in Los Altos? What matters here is purchases, the amount plus the number of unique occasions. Plus, overall, was I alone or with someone. Usually when I’m alone, it’s just one americano, coffee, tea.” So we’ve asked it, you and I, now we’ll see what happens in terms of the answer, right, inside. We’ll see. I hope our editors have put this up on screen for you.
And while it’s searching inside, right, it’ll compare. This is just an example for you, it’s not super important to me what it was there… I even, I have no idea how much it’ll say. We’ll see in a moment. These are only two coffee shops. Obviously I don’t go to just two coffee shops. I also go to, I don’t know, Blue Bottle at Stanford Shopping Center or Blue Bottle at Santana Row in San Jose. I mean, I do go to all sorts of places, and, well, and a huge number of other locations, right, and there could be dozens of different additional locations there. Some more, some less.
Or in Menlo Park, for example, right, cafes that I even go to fairly often. With this thought I want to get across to you that today we underestimate the influence of artificial intelligence on the life of each of us, and even the influence of the current models, of the current state of development of these systems. I mean, on the one hand, it seems to us that, well, everything is ticking along and going a certain way, because we don’t see any super-collapses happening in the world.
On the other hand, the changes, they’re colossal. I mean, the example I’m showing isn’t, some gag, and it isn’t making up a joke, and it isn’t writing a poem, and it isn’t even the flying car that Elon Musk showed last week. And my wife, it, it impressed her very much. It’s not even self-driving cars. This example I’m showing now, that the system can find things like this and interact with me. I want to point out that systems like OpenAI Dots, or systems like GrokBot, are systems that you have, in essence, allowed to interact with you.
That’s a totally big difference. I mean, these are systems that are in contact with you a priori. And here’s what’s interesting about Dots: when you launch it, there are lots of different nuances. It doesn’t always give you a super-cool answer right off the top, it doesn’t always open the most expensive models, for example. You don’t fully understand how it works, what it does well or poorly and so on. When will it give you an answer, like in this example I’ve launched, right?
How much time will it take? Either I’ll have to pause the recording, or it’ll all happen while I’m telling you something. But this system has gained the ability to interact with you, I mean, to tell you things. If I install this system, like GrokBot, into some interface like Amazon Alexa, then that’s it. By the way, Apple has released new home devices, HomePod. I actually think that Apple has started falling behind very seriously in the world. And yesterday Max and I were just heading out to drink more coffee. There’s this cafe of ours, we really like it there.
It was opened by the creator of Twitter. And that’s the current X, right, in San Francisco. And, true, on the way we wandered into a restaurant, but then we went back to the coffee shop after all. And as we were leaving, Max said: “Listen, but devices like Apple HomePod, what should they be doing? They should be calling up my agent. That’s exactly the right super-solution.” Look, here’s an interesting thing, that… It just found some payments. I mean, it found that at Cafe Venetia there were noticeably more, that there were payments, eighteen payments on fifteen different days, and at Satura there were five payments on four different days.
And here I want to say that it got it wrong. I’ll write to it: you clearly didn’t check all, the accounts. The accounts. Let’s tell it straight: you, you need to go through all the dozens of accounts you have access to and check them. Most likely you have incomplete information. I mean, you can see that it made a mistake. But that doesn’t matter here, I’ll look into it later. And I just realized why it made a mistake. Because often in this infrastructure I might have used accounts that, for example, it doesn’t have access to right now. But it did find something.
I just showed that it found some data. Well, because, to be honest, I saw the sum of two hundred dollars, and I realized that it’s complete nonsense, because, probably I go for coffee in town about fifteen times a month for sure, maybe even twenty-five. By the way, I’ll do a full analysis of different systems. And if it’s twenty places, then even twenty places, even if I drank one americano twenty times at five dollars, that would already be, that alone would be, a hundred dollars a month just on americanos.
Since I often go to these places together with, with my kids, or, like, I go, and there may be three or four of us, and they also often might… they don’t buy americanos, they might buy and eat some fruit or a croissant, or have some drink, so of course it gets way more. So. Today I’ve given you, a bit more broadly, information in terms of the system here.
I want to show you in detail what is basically considered a serious parameter in evaluating such systems. Although still, from my, like, you know, from my fundamental approach, it matters. But now I’ll show you what is usually considered important in evaluating such systems. One of the serious parameters is what makes it keep going, a big task, work for a long time and not stop. And I really didn’t mention this to you today. And this is actually the foundational story of systems like GPT Astra and GPT Sol.
These are systems that were originally built, even Fable, by the way, yes. These are systems that were built with the proviso that these systems are capable of working for a long time on their own. For example, like, two, three, four days. I remember the first time my ChatGPT ate, in a day, in one session, more than a billion tokens. This surprises some people, by the way, and they say I’ve got the number wrong. I haven’t got the number wrong, and as of today I’ve spent sixty-three point three billion tokens in ChatGPT.
Yes, you all remember, literally just recently I said I’d spent fifty-something billion, and now it’s sixty-three. And there was a peak day when this system ate more than a billion tokens. I’ll even find you when it was. Here it is, here — here, I found it. There was a day, so, look, here are the first times the system started eating many billions. So that’s July thirteenth, two billion tokens, July fourteenth, four and a half billion, July fifteenth, five and a half billion, then on the sixteenth, three point seven billion.
I mean, there were days like that. And before that it was eating, like, twenty million tokens a day. Some days it was a hundred million tokens a day, right, some days it was zero tokens a day, some days it was a hundred and sixty-two million tokens a day. So. Now, this week I really did have moments, because Tibo started resetting the limits almost every day. I have a whole lot of resets, a whole lot of credits, a lot of paid tokens. So I really did have days — this week, a couple of days ago, I had a day of one point three billion.
Last week it was one point one billion tokens. Yes, last week there was one and a half billion tokens. Here, on October thirteenth I had two billion tokens. Well, I mean, the system gobbled up a whole bunch of tokens. That’s the beginning of the time when these systems started working for long periods. That’s what I just said, in particular: what makes it keep going on a big task and not stop after each answer. And to this day, as of October nineteenth, when you use the expensive ChatGPT Pro model to compare Antigravity, Claude, Codex and, like, Cursor, I want to tell you that to this day this question hasn’t been studied at all.
Nobody understands how these systems work and why they sometimes keep doing a task and complete it fully, all the way to the end. Sometimes they keep doing it, complete it partially, but spend a great deal and make junk on top of that, and sometimes they don’t complete it at all and stop. I mean, this is a fundamentally serious, big parameter that exists, right, in general. And one, by the way, of the parameters, one that’s very important today when comparing these systems, is what happens when a long history stops fitting in the context, right.
And in all these systems, they have their own rules. I mean, for example, in Claude Code, like, the history starts getting compressed. In Google Antigravity they have a separate mode in teamwork. To be honest, it’s complicated there in Google Antigravity, right, there, in its team mode, the lead hands the next stage over to, like, a fresh agent, right. In Codex there’s a goal that defines… There’s even a notion there that a saved goal does not equal saving all the details. And, accordingly, various problems can happen there in terms of context compression. It’s believed that Codex doesn’t compress the context right away.
And it seems to me there’s probably some truth to that, because Claude Code works pretty well with the context window. By the way, at one of the closed sessions at the OpenAI conference a couple of weeks ago in San Francisco, one person said that his sessions in Codex, he’s one of the strong Codex developers, run, run for several months, and so, so the session has to get compressed. But I really do have sessions in Codex that have been running for several months already, probably for about three months now I’ve had sessions running, and I like how it works there.
At the same time they’re very uniform, probably, they’re very uniform. In Cursor, after compression the agent gets access to the full history, as to a file it can search. And here’s a significant nuance — it has to figure out on its own that something is missing from the short summary, and go back to the original. In general, this story, this story of how the new models work, and OpenAI emphasized this very strongly, emphasized it, and Cursor is emphasizing it in this case, that the system didn’t just, like, compress your context, but it also still has your context separately, and there are also other chats, and there are also project chats, and there’s also your memory.
And it has to figure out on its own: do I need, at some point, to step away somewhere and take another look? ChatGPT, for example, is still very strange to me. I have various businesses, and there are businesses where I just have, like, stakes, or where I, I don’t know, invested something. There are businesses I’m just interested in, and there are businesses where I help and act in the role of, like, advisers, or, like, a strategic sort of, a strategic board of directors. And ChatGPT, to this day, can, when I ask it: “So a new feature has come out, say, in Cursor, tell me about it.” It can tell me: “For your ToTheMoon channel, it’s this and this.” And you just sit there thinking, why did it suggest my ToTheMoon channel to me and not suggest this particular business.
I mean, it still has a certain understanding of its own of how these systems of its, these kinds of systems, should work. One of the important things in Codex, in Claude, in Antigravity, in Cursor is whether what you’ve taught it will be kept for the next tasks. This is a very important topic, one that often really gets on my nerves in various projects and tasks. Both in Claude Code and in Codex, every so often the system glitches and starts asking me about something we’ve already explained and gone over five hundred thousand times. I don’t want to explain it all over again.
And that, by the way, is the problem with working with other systems. Every time you have to customize the system for yourself. And more and more, I don’t want to customize the system for myself. I want the system to learn on its own and figure out how it’s done. By the way, many people consider that a minus — the fact that the system learns on its own. But if you listen to the new serious OpenAI developers, they talk about how you’ll be training the system yourself less and less.
I mean, the system itself has to learn what to do and how. One of the strongest aspects in the work, in comparing these systems, is the parameter of how they prevent the situation where each one did its own part, but together nothing works. I mean, this is one of the big problems. And we had a special episode, I made it on the topic of, where I went through a report, a very big Anthropic study. Watch it. On the topic… Those of you who watched it, write in and say how you liked it, this study, right? And also, by the way, those of you who have the agent thing working. What was the topic there?
The topic there was how agents behave in parallel. Imagine you’ve handed out tasks to ten agents: how controllable they are, and how well they understand who the boss is, how ready they are to accept, in their own view, that some other agent is better than them, how ready they are to pick the right, really good agent. It was a really cool study. We saw there that it’s not all clear-cut. And modern systems, like Codex or Antigravity or Cursor. Why do I trust Codex? Because in Codex they’re working on this very hard, and in Claude Code they’re working on this very hard.
Cursor, though, was only just bought by xAI. I really like what Elon Musk has been doing in many of his systems. I don’t agree everywhere or with everything. I speak out a lot about where I don’t agree, right. But overall they know how to do things very fast. For me it’s probably the number one company for speed. For all that, though, OpenAI and Anthropic are ahead of the whole planet on these things. Why, by the way, don’t I talk about Chinese models? I liked Max Grigoriev’s remark on this. You’ll see it next week if you’re watching this video while it’s fresh, but I sent him a bit in advance.
Max said he disagrees with the idea that China lags a little behind the American market. And while discussing it, we saw that China lags very seriously behind the American market, and it’s unclear whether it will ever be able to catch up with it. And I’m not even talking about other countries. That’s a completely unsolvable story. And if you live in some other countries, or in some way influence the development of artificial intelligence technologies there, well, you should fundamentally acknowledge this, right.
One of the serious parameters when comparing such systems, the ones that exist, and I want to tell you about it today as well, is who decides that a task is really finished. So imagine you set a task, for example, to analyze your coffee bills — the way I just did with the coffee shop. I mean, Dot got back to me, but it has to decide whether it has finished the task or not. Or it might, for example, say: “I’ll also look through your email, and I’ll also look through the files you had.
And I’ll also, for example, get access to the map of your movements, how you got around, to your location history.” I don’t know, maybe, for example, it will get access to vehicle tracking, right, or some other data. I mean, I’ll study even more and take a look, and draw a conclusion. Or I’ll additionally study, for example, texts and messages. I mean, how capable the system is of truly, fundamentally, truly making an assessment. That’s a very important component, right, today. And to what extent the task is finished. Because often we may get some answer, and we think it’s finished.
And today you and I can see that in Dots. I think it’s a really cool example. What do I like about Dots? The fact that it often gives me an answer, then says: “I’ll check some more.” Then it comes back an hour later and says: “So I did finish working through it after all, like this.” Then it comes back a day later and says: “Actually, the task is like this.” And Codex today, right, after all, Dots works in tandem with Codex and works in tandem with ChatGPT. So Codex, Claude Code to some extent, well, and all the other similar systems, Cursor in tandem with Grok Bot, right, maybe Google Antigravity, they will come to the point where these systems, to some extent, won’t finish the task at all. I believe a task should never be finished.
So what does it mean for a task to be finished? And it seems to me, by the way, that this is a fundamental problem of the research work of developing artificial intelligence, right? Because if you want to truly interact with a person, you can’t stop interacting with them. Remember, I said I’d easily spend, like, two, three, I don’t know, four thousand dollars a month, if the system constantly studies my chats or my projects that I have going on now, and constantly tries to interact with me about those projects. So why do I want that?
I want there to be an artificial intelligence that constantly studies certain aspects of my life and suggests extra things to me. That’s not easy to do right now, right? Because sometimes it starts offering you junk, sometimes telling you uninteresting things. And, by the way, it sometimes reports something to me, it says: “I found this problem here in your email.” I say: “That’s not a problem, throw it out, I’m not interested in topics like that.” Or, like: “I studied the data in one of your projects here, you have an issue here.” I say: “That’s not a problem, I couldn’t care less about that.” You kind of have to train it and set it up to do things, but, look, it mustn’t stop working with you.
I mean, real artificial intelligence is a system that constantly interacts with you. We’re guaranteed to get there; we won’t be able to do without it. That’s that good film “Her,” right? I think it’s called “Her” — She. And it seems to me that “She” is probably hard to translate into Russian here, but still. This film is about an artificial intelligence that talks with this man constantly, I mean everywhere. You can see that this artificial intelligence, it kind of functions constantly together with the person, I mean it’s part of his life.
Look, in terms of human development, if you go into the spiritual, the spiritual realm, go into, I don’t know, philosophy or a person’s personal development — that’s a big separate question. By the way, over the last few weeks I’ve been putting out a lot of videos on my channel about the human being and artificial intelligence, the human being and a person’s personal development. Watch them, right, and whoever has watched them, give it some heat and fire here, so that other people click through to those videos. And it’s precisely there that I develop this area more, and we’ll see it together.
So this, once again — who decides that a task is finished — is a fundamentally important question. A fundamentally important aspect overall in how you go on to analyze the work of these systems. I mean, how thoroughly they, each of these systems, look. Obviously, all these systems have standalone apps today, and all these systems should be accessible with all the infrastructure, from all the apps, right? Take Codex, it’s already available, more or less, in many systems. It’s available through Dots too, it’s available from the mobile app, like, from the desktop, it’s available remotely to a server, on a server.
For example, at home I have Codex installed on a server, like, computer, and I easily control it from my phone, and I like that very much. I hope the time is coming now when cool speakers will start appearing from OpenAI or from some other companies, which will have installed inside them, say, for me, Codex, or they’ll have Grok Bot, and we’ll be able to, like, use it inside and interact with it, right, and be able to look inside. One of the important things in these systems is how capable the system is of carrying out many different independent tasks at the same time, where, where those tasks aren’t connected to each other.
Look, there are tasks that are slightly connected to each other, and there are tasks that, that go on living their own life independently of one another. And this problem, of course, hasn’t been solved yet, and it’s probably very strongly tied to the story that the system has to be listening to you all the time, because your tasks in life can’t come to an end. I mean, your day can’t come to an end; for me, there can’t be an end to sorting out, I don’t know, my kids’ interactions, going to school, like, talking with my wife.
For me, the question of analyzing my financial accounts can’t get settled, and so on. I mean, it’s a story that will keep moving forward, right. And so this topic, probably the most important parameter I’ve named for you today, at the end of the video, is this ability to solve the system, the ability to solve parallel tasks, sometimes dependent on each other, and sometimes having completely different goals, from different projects. Look, from different projects. This, of course… This, of course, is the future. I’m waiting for your support here, your comments.
I’m waiting for you to forward this video to your friends. I’m waiting for likes and bells. Write which episodes on our channel you find interesting in general. Give us some support, and we’ll be waiting for you in our new episodes, of which there will be more and more.