Hi everyone! This is the ToTheMoon channel. Tech news, insights from Silicon Valley and around the world. Our Sunday podcast, right? Just a reminder, we now come out many times during the week. By the way, speaking of which, at ToTheMoon applications are now open so that you can shoot content for ToTheMoon media, and the link will probably be somewhere below, or our editorial team will put out a QR code. So, listen, Ilnar, what I actually want to start with — of course, I think it's a really cool topic that Tibo writes about on X, and a lot of people just don't know about it.
Tokens are being reset in Codex. And I think even people who have a free subscription or a twenty-dollar subscription can get a pretty solid amount done over a whole month, because they'll have to reset the tokens. We'll talk about that in a moment, but first I want to start with a topic that, that's been hotly discussed this past week. Lots of people are writing questions and comments about it — about Dots and, in general, the advantages of OpenAI Dots over, say, Hermes or GrokBot.
And it'll be really interesting to see in the comments how many people actually use Dots, and what scenarios you use it for. We'll discuss that a bit now. And the story I want to tell here is that… By the way, Ilnar, have you started using Dots, or haven't you switched to it yet?
The point of application is unclear to me. A point of convenience — a bit of a tautology, of course, but yes. But somehow there's just no need.
That's why I'll bring you up to speed a bit now, and maybe Ilnar, by the way, will get into it himself a bit more, right? So, first of all, I want to explain the difference, a sort of top-level difference, for when we start comparing every system out there and saying: but didn't DeepSeek have something like this? Didn't other models have something, something like this? And it may really seem that having some chat with a so-called, in quotes, agent that performs various tasks — that's the same thing that's been discussed, including on the ToTheMoon channel, for, like, a year and a half.
Yes, we talk about how you can create all kinds of agentic setups like that. You know I'm not super fond of the word "agent," on the one hand. On the other hand, we should immediately understand who's making this system, right, and what it actually is. If you're comparing Hermes and OpenAI's Dots, we should at least understand, together, that in this case GrokBot or Dots could be called the engine, and Hermes could be called the car that different engines get put into, right?
Because Hermes, basically — they do have their own agentic system, but Hermes is an open-source product, MIT, right, I think they are. And you can connect different systems to it — including via API you can hook up OpenAI's ChatGPT, and Claude, and Gemini, and so on. But the basic…
And Qwen as well, yeah, that kind of thing.
Yes. And, or, like, the Chinese market. Once again, Hermes themselves have their own model, although, by the way, they say: for complex agentic tasks, don't use our model, plug in something else. But look, let's break down a real case, a real, a real understanding of Dots versus everything else. And why, for now, I still want to tell you about it, right, and talk about it at a basic level. And when other systems roll this out, so that we'd, well, probably be in a different setup, in a comparison, a different comparison.
First, a very important thing — with Dots there's also, probably, a second layer when we talk about the comparison. The second layer there is: Dots runs on GPT-6 Astra. And actually, in terms of the system available, the announced agentic system, it's top-notch as of today. I mean, it's the top system for agentic work among these products today. Because even if you take GrokBot, what's been announced functionally is that GrokBot, inside Cursor, decides which Grok model to plug in.
And there's no guarantee, for example, that you'll get Grok 4.7, and if you do, in what mode you'll get it, right? And I urge everyone to be very critical when reports like this come out and say, for example: Grok won at, like, this championship. You open it — I open it right away and immediately look at which ChatGPT was in there, or which Claude was in there. And you open it, and it's ChatGPT 5.6 Luna. And you're like: okay, got it, closed. Not interesting, right? Because we all know that today there's already ChatGPT 6.1 Sol, or there's ChatGPT 6 Astra. And you're like: I'm not going to compare this at all, I'm not even going to look at it.
So, the next area where you actually need to compare systems like Dots. We'll talk about using it in a moment, Ilnar — first, a bit more broadly about the comparison. Dots works with your entire OpenAI infrastructure. I mean, besides the fact that Dots — it's Astra. Dots has ChatGPT memory, Dots has all your past conversations, Dots can connect all the plugins that exist in the system today. There are certain limitations; they keep working on that here and there, lifting them, because OpenAI's infrastructure isn't that simple.
They have a lot of different apps and systems; in some places you can turn things on, in others you can't. Dots has its own cloud computer, Dots has its own browser and dedicated capacity it can use. Dots can work with ChatGPT Work and Dots can work with Codex. And this is where a really cool thing comes in. I mean, Dots can answer you without creating chats, or it can work in parallel with your existing projects and tasks in Codex or in Work, or it can create them. And I even had a case where one of these neat tricks came up. You know I have a lot of projects in artificial intelligence, and there are ones like — Ilnar, Ilnar and I discussed pet projects last time.
Actually, it's not a pet project. And in the recognition system — I'm interested in developing a smart home system for my house. By the way, I was just texting with Max Grigoriev about a server computer for this, a particular one. By the way, he promised we'd shoot an episode together with him in the coming weeks. Who remembers Max Grigoriev? Drop a like for Max. We've had several episodes with him.
And Max has a lot of hardware in his garage, yeah, he can put something together.
Really cool! Have you been over to his garage, or not?
Yeah, yeah, yeah.
Very… Oh, you've been there, right? It was funny — he and I found a tea we'd been producing back in 2013. From China, well, for our educational company. And somehow, from somewhere, he had that brand. That was really funny. In San Francisco. Yeah, Max is from Google, from Airbnb, and from the AI crowd here in general. So we'll look at that separately. But here's the thing.
So, it turns out that for this video camera project I set the following task. I wanted the system to alert me if at night a person showed up on camera and didn't go any further, like, toward our place, and had no car. I mean, he somehow ended up on our street on foot. Ilnar, you know how things are laid out on our property — there shouldn't be people there at night. There can be animals. I have a separate system set up for animal alerts, just to watch them, because otherwise it's impossible to see cougars, lynxes, right, skunks, raccoons are hard to see, and, like, a coyote, a deer.
Those are all very easy to see. But a cougar like that is very hard to see. Yeah, recently, when two cougars came at night, serious ones — of course, well, you can only track them on the cameras, how they move around, by the way. It's really interesting, of course, how they hunt, how they split off from each other. You know, they move as a pack. Speaking of which, my Igor — friends are coming over to Igor, my son, for the weekend, and he kept wanting to sleep outside in a tent.
And my wife and I are like: "Are you sure you'll be outside? There could be cougars living out there, there'll be a cougar, a lynx." Well, in the end we agreed that on one, on one of the decks, the balconies, we'd put up a tent for him. Hopefully a cougar and a lynx won't come up to a person. And so I gave Codex the task of alerting me about people, and it decided it could alert me through Dots. And so Dots then sends me these alerts.
Dots has background tasks, besides Codex, and it has all kinds of action-checking systems, different restrictions for the different permission levels you can grant. And it can work with both your cloud environment and your local one, Codex's for example. That's really cool. And with the local computer. For example, my Dots connects to Codex, which is installed on the computer and which can send, for example, messages through the computer, well, in iMessage, or send emails, because Dots can be limited.
I mean, for example, you can connect a Gmail plugin, but to some email services, for example, you can't connect plugins, or you can't connect a proper plugin to iMessage, but it sends, for example — it aggregates, it reads all my email constantly. By the way, speaking of how advanced artificial intelligence has become, it'd be cool to hear which of you have built large-scale, extended systems like this, because this isn't an agent, right? It's already a kind of multifunctional artificial intelligence. That's what you could call the concept of artificial intelligence in today's world.
Of course, I don't really think it deserves the term artificial intelligence, but still. It's cool. I gave it the task of aggregating emails from four schools and sending me only the important information, constantly learning whether I need the information or not, and sending it to my wife. For some reason my wife still likes these messages in iMessage. Maybe she thinks I'm the one sending them, or something. And recently Dots sent a message at, like, one in the afternoon instead of in the evening, and it wrote: "I want to let you know that your daughter's school ends early today."
I mean, it kind of changed the rule that it's supposed to notify me, like, once a day in the evening, and it didn't notify me on the weekend. What was also cool is that over the weekend messages still kept coming from the school, but it didn't send them to me. I mean, it sent a delayed message a bit later. I mean, it adjusted itself, in particular in choosing which messages to send me and when to send them. By the way, I'm not saying it works super great; it has a number of glitches, a number of particular nuances.
But I want to say that OpenAI has started expanding this overall infrastructure more seriously. Before, you couldn't get into the chats from Codex at all. Now from Codex — by the way, I want to tell everyone: Codex will assure you that it can't delete or search in chats, in ChatGPT; that it can work with Word, with Codex — it can do this, and you don't need to give it browser permission. I just categorically don't give browser permission. I don't give it permission to roam all over the computer, right? And I don't like that.
I don't want Codex watching my mouse all day long and everything else. There's enough of all that as it is. But it definitely can search through all your chats. And remember, I had a project where I exported all my chats, and actually neither Claude nor Codex could, could solve it for me infrastructure-wise. Now the time has come when you can get closer to that, because Codex itself can start working with those chats. What's more, it can also delete them, and that's a very important thing.
Sometimes you use ChatGPT, you've got some big chat, for example, and you suddenly want to kick off something in Codex for some household process. By the way, those of you who use Codex outside of work, write us your cases. I think it's really important for people to see them, because I see that people don't fully understand how Codex applies to their everyday lives. Take Polina, my wife — even though I've talked about this a lot, I still haven't managed to get her hooked on using Codex.
Even though Codex would clearly be super convenient for her in terms of certain setups. Plus they made Codex Cloud, and Codex Cloud — that's exactly the ability not to depend on your computer. It works pretty well, it's a bit of a pain, because you have to create an environment there, but I think they'll be moving away from that. So the main difference with Dots, once again, is that it works with your OpenAI infrastructure. And it's not just some little agent off to the side somewhere. We get a lot of comments.
Like: compare it with this version of Moonshot. Compare it with the released versions of Alibaba's models. Compare it with DeepSeek here. Compare it with Gemini. Once again: on what specific question are you comparing? What's your regular infrastructure? What do you use all the time? What do you do all the time? Well, it'll be interesting to see whether Dots has become your regular tool. For the first forty-eight hours, maybe even three days, Ilnar, I used it a lot, really forcing myself to use it a lot.
Then I see that — since the interface is a bit clumsy right now, you have to open the mobile app, find Dots at the top, then switch over to Codex somewhere, that kind of thing. Until they've done that — since OpenAI still isn't an operating system, right, where you open it and you've got, like, everything, everything right there, right? Then it's a bit of a hassle to open it, to find it. You can't connect it yet. For example, if I could switch it on instead of Alexa, it would be pure bliss for me.
By the way, in — I think in the five-hundred-dollar Pro version you can connect it through a third-party interface, but I don't think they've done Amazon through third-party interfaces yet, and it'll use your, your volume of tokens. I mean, you'll be spending your own limits, the ones you have inside, the ones you've got in there.
So, Ilnar, I strongly recommend you install Dots and give it some name for yourself. I hope there'll be Dots that can message each other, and my Dots will be able to message your Dots. And that would be really cool, because sometimes I need to send you some message, but I don't want to distract you wherever you are. Or I'd like… but at the same time I'd like you to maybe be sure to get it, or for your system to decide with you whether messages like that should be sent to you, and how many.
Because imagine, I could send you... You and I discuss the news on the ToTheMoon channel once a week, right? We could send each other news, and each set up our own Dots for how it, in what form it presents them to us, right? I mean, I think this is a very important, important topic, really, well, insanely interesting, one we should, in theory, be moving toward, right? And I've expressed the opinion that if, if Dots could communicate with each other, then my wife's Dot and my Dot would take on a huge range of logistical communication, right?
Or, say — okay, me and my wife, but we actually have quite a big household here, and a family. Well, there's a whole lot going on in our lives. But even in terms of work projects, since I'm involved in a great many projects, if people have Dots, or various assistants have Dots — and not even assistants, but normal, ordinary people — these Dots interacting with each other would be cool. Last time Ilnar voiced the thought that he wouldn't like it if someone else's Dot talked to him and he didn't even realize it wasn't a person.
And that's exactly what I'm highlighting here, Ilnar: that it should happen in a good, normal way, among Dots rather than among... So that I'd, like, be presenting myself to you through this Dots. Take our comments — on ToTheMoon, bots don't reply to comments. And by the way, what amazes me is that YouTube is against bots, but when I — very often I reply to lots of comments myself, I think you can see that in the comments and tell where I've replied. I can reply from the ToTheMoon account, and I also have other YouTube accounts, I can reply from those. But the point is that YouTube constantly offers to let me mass-reply to other comments, or suggests replies to me.
And that bothers me. On the one hand, YouTube is against this, like, AI thing, fake AI interaction. On the other hand, it kind of shoves it at you. I mean, I don't want to reply with AI comments. I want to write my own real comment. One that says what I think, right? And it can actually be very different from — YouTube doesn't suggest I reply to people like that.
Listen, well, your story about Dots sounds inspiring. I'll have to give it a try and take a look after all. As for my slight indignation when people make their AI assistant answer instead of them, the question I have here is more like: then why do I need these people? I mean, if they're saying: instead of me solving your issue, go to ChatGPT. Essentially, you can rephrase it like that, right? Like, when he says: Dots, take this on — I understand the person is saying: well, better ask ChatGPT. And after a couple of iterations, I'll most likely start doing exactly that.
I mean, that's more the kind of dissonance I have here. If we're talking about Dots communicating with each other, yes, there may well, well be something to that. And so, apparently, I need to dig into this a little dee-, a little deeper too. What bothers me is that this thing might be short-lived. But given what you're saying about the integrations OpenAI has made and keeps making in this direction, maybe Dots is here to stay. So maybe it's already worth diving into this. So thanks for being a trailblazer.
I'm not sure, I'm not sure, Ilnar, that it's for the long haul, actually, because, look, first of all, it's very heavily loaded infrastructure. Yesterday I was discussing with my friend and partner why OpenAI doesn't have video generation. They've shut it down now, because I don't think spinning up video generation is a problem for them right now, but it's a load — a catastrophic, very big load. Why? Why didn't they release Dots to everyone? Well, let's be honest. It would obviously pay to open it up to everyone, right, the way they open GrokBot up to everyone. Can you imagine what the load is there?
There's GPT-6 Astra, damn, which is constantly working with your whole infrastructure, right?
GPT-6 was rolled out to everyone literally just the other day. It was only day two or three of this marathon we're going to talk about.
Keep in mind that they've only just, only just rolled out 6. And you, Sasha, are already talking here about Dots, which works with different structures and can think for a long time, and so on. It can think for a long time, right? That's very important. Actually, Ilnar, by the way, before we get to Tibo, and I think that's a very important topic, I want to share that GPT-6's Intelligent UI came out yesterday. First of all, GPT-6 was rolled out, and the Intelligent UI came out, and in the Intelligent UI they've announced... I don't know, was that what you wanted to say, or were you talking in general about GPT-6 having only just been released?
Well, Tibo, in his tweets, actually noted, as one of the achievements of this marathon we're about to talk about, that GPT-6 had been rolled out to all users, as the achievement of, like, day two or three. I mean, obviously we, as pro users, like other paying users, most likely saw GPT-6 ages ago. But it's only now that it has reached all accounts. My point is that the resource load on OpenAI is very heavy, and it's clear why they're being stingy. And the same goes for the video generation you're talking about, it'll eat up a huge amount of resources, and they can't afford that yet.
They don't have that kind of capacity to spend on video generation.
They clearly don't have the capacity, they clearly don't even have the capacity. I've always said that maybe they don't ask too many questions, because it all means incredible costs, including very big costs. Yes, all this research, the constant work with memory — those are huge costs. And maybe they didn't even set up Codex with ChatGPT, because, well, you can imagine what you'll be able to start doing now. Plus there's also the thing that people couldn't really clean out their ChatGPT, it was impossible to clean it out.
So it turns out it may be in their interest to keep this whole feed, because they, if you haven't switched off the checkbox so that nothing of yours can be analyzed. By the way, GrokBot offers to switch off this checkbox right away, but maybe OpenAI now also offers to switch it off when you sign up, I don't know, then… well, to remove this training, then of course they have access to, well, to the information you were in. And by the way, again, someone wrote about this at some point.
Look, the fact that ChatGPT trains on your chats, it doesn't train to make things better for you, it trains to develop the system itself, right? So, on the one hand, there's a plus in this. You're helping a company like this develop and helping it truly show everything it's got. My sister Anya, who also hosts episodes for us on ToTheMoon, talked about this. There really is a plus in this — not hiding information in ChatGPT. After all, you're helping it develop. But on the other hand, look, it's not training the way you know it, because, as you know, that's a slightly different, a different piece, right?
And you even turn it on separately in the settings. About GPT-6 Intelligence. There are two very interesting things they announced. The first thing. They tried to make it many times. And look, they tried to make both of these things many times, which is why my reaction to them is along the lines of, well, it kind of seems this already existed. And I think right now in our chat we should see: so Google has this, and Claude has that, or something else. But still, yes, they, OpenAI very often released something, that project died, and it ended up somewhere else. And then something super new appears, right?
And you sit there thinking: is this new thing really new, or is it a continuation of the old one, and it'll die the same way. That's how Ilnar is worried about Dots. And so am I, to some extent. Will they really connect Dots with everything, right? It's still not connected to voice, for example, a really good one. Look, GPT-6 Intelligence. Thing number one that they did.
There's now an ability, well, more precisely, not an ability. They have an interface. What's it about? That ChatGPT can start answering while it keeps thinking over the question and searching for information internally. So OpenAI calls this approach alternating thinking and answering, right? Interleaved thinking and answering. Now, an important thing for everyone. This doesn't work in Pro right now. By the way, Ilnar, this was also a very interesting story for me yesterday, when they suddenly posted about it, right, they only just announced it. We're recording about it.
Fresh stuff, right? And so why isn't it in Pro? I mean, you understand that the Pro model, the very concept of Pro in ChatGPT — it's some different construct. I saw this, I mentioned it, I think, in the episode, yesterday I was shooting an episode. By the way, watch it, it's about the new film they made about OpenAI and about OpenAI's approaches to developing AGI for every person. Watch it if you haven't. And in particular, what I mentioned was that all my modes except Pro started answering instantly. I mean, even Extra High answers just like that for me.
So in Extra High mode I barely wait anymore. There was a time, remember, there was a time when you turned on thinking and you really waited, say, five or seven minutes. You'd turn on Pro and it took forty minutes or an hour. Now in Pro mode I wait, and sometimes it can stretch to an hour and a half to two hours. If it's serious stuff, really big. In Extra High mode the answer is basically very fast. Obviously, if in Extra High mode you throw it an Excel spreadsheet with a hundred thousand rows, it won't be an instant answer, but still, the answers are fast.
And you can see that Pro is something completely different for them. It works differently in Codex and in Work, I'll note. Basically, there isn't even such a concept there, right? Because in the chat, the concept of ChatGPT-6 Pro isn't the Extra High modes, or the Max mode, or Ultra and so on. It's a different mode, right, inside. So, how, look, how has the process changed? Before, the process went like this: the system first analyzes, does a search and gives you an answer. And so the user waits until that answer shows up inside.
An important, important point: OpenAI has already tried this many times before, I think they first talked about it two years ago, when they'd give you a quick answer right at the start, and many of you saw it at the top. What's their thing now? That they've got the work interleaved within the answer. We'll see how it works. Those of you who've already noticed it, tell us your first impressions. What does it mean? It means first there's analysis, then you get the first part of the answer. I mean, not that micro-answer in the reasoning that used to be there, but you get, like, a full-fledged answer. Yeah.
Then you get new analysis, a search runs, and an additional answer comes out. And I just had a case where, since I'm a Pro user, I didn't even notice the basic one at first. I asked the system itself how it works. I asked in Extra High mode, and right away it very quickly gave me a whole lot of text. Like, really a decent amount. So it wasn't a reasoning answer. And then it kept reasoning, and yesterday I saw glitches, glitches started in the browser version. It was as if my chat stopped working.
That thing kept spinning the whole time, right, it was just sitting there inside. And then it started giving me a very big answer. I mean, a really detailed, huge answer. So it finished studying it, finished reviewing it, did a search and so on. I think this quality of answer will only be available to ChatGPT Pro users, because it's still Extra High mode, of the Pro plan anyway, right? And I don't know whether this Extra High mode works in Plus, as far as I understand, it doesn't. There's Extra there, right? Or High. I think there's High there, right?
But there's no mode, but there's no… Yeah, there's High, sorry, but there's no Extra High mode, and so, as far as I understand, it won't work quite that great. The second… Ilnar, I'll move on to the second thing now, by the way. The second thing is very interesting, an OpenAI update. And by the way, OpenAI, well, today someone will say I'm advertising them and so on. I wasn't planning to talk about them this much, but they released, damn, the Intelligent UI. What can you do, right?
They've been putting out releases for the third week in a row, Ilnar.
And they'll keep doing it little by little, practically every day. That setup you talked about…
You'll tell us about Tibo in a moment, right, why they'll keep doing it, they'll have to, right. What were you saying?
About that setup you described, where it first gives a small answer, starts thinking in parallel, and then gives a big answer. It's very similar to how in voice mode, about six months ago I think, they were all saying that now our model starts talking. Meanwhile, in parallel, another agent is gathering information, then pulls it in, and your conversation comes out smoother. Roughly speaking, you ask ChatGPT, say, what kind of tea this is, Chinese or some other kind. It says: "Ah, uh-huh, let me google it. Yes, wait a little. Coming right up."
Then it gives the answer. And meanwhile, in the background, another process is running that gathers information, say, googles or something, or generates reasoning, or something else. And then it starts delivering it. So this will probably give smoother communication with a shorter delay before a useful answer starts, if you can introduce such a metric, and it'll probably be more comfortable to talk. Yeah, so you're no longer watching the reasoning, instead some more or less sensible answer is coming to you.
So they keep moving toward comfortable communication with the chat. Most likely, that's a good thing. I don't think it'll have any big impact on the end result. If anything, GPT-6.1 Astra will be more useful than this, but using it will be much more useful. I'd also like to add something here. You were talking about the integration of Dots with chats, with Codex and so on. Overall, I'm really pleased with how OpenAI has leveled up memory and context sharing between chats over the last six months.
Because just recently I had some tests done, lab tests, and I threw them at it to look at the trend, and I only threw in the latest results. And the ones I'd had done six months ago or a couple of months ago, it found them by itself. I didn't ask it to. It pulled them in by itself, built a chart of how various parameters changed by itself. And so from there we were analyzing not at a single point but over time. And all of this is under the hood, with it doing this across different chats. And I did it in a new chat with no link to the old one, because the old chat was created back on GPT-5.5, probably.
Then one was on GPT-5.6, now it's on 6. And for there to be information sharing between them, right…
Notice that this isn't just memory anymore, that it isn't just memory anymore, that apparently it actually, what it did during the research…
It searches across chats, yeah.
From memory, maybe it had some micro-recollection, but then it most likely pulled up the chat and studied it in detail, because otherwise it couldn't have answered that well. And in a big, branching structure of a huge number of chats, I have thousands of chats, it really knows its way around. There are things it maybe doesn't see, but like the health example you gave, that's cool. I had exactly the same thing. I was referred for an exam, I asked it about it, and it tells me: "Are you sure?
You have a different question for the doctor, then." Where did it get that from? Well, it must have found it somewhere in the chats. I mean, it must have searched really seriously to find it. And you realize, you know, it became clear that you can no longer train the model yourself the way it trains itself. I mean, well, on its own infrastructure, right? How would you even do that? And what would you feed it, are you going to give it some files or something? I mean, you need to use the chat every day.
And let's remember the rule we constantly recommend — use the system every day, right? Use exactly one system, use it a lot. Then you'll be able to see the kind of tricks Ilnar is talking about, by the way. If you keep migrating, jumping around or saying: "Oh, it gave me the wrong thing." Well, it may not find out, not see, not look through things, right, it won't understand how you're used to working and functioning.
Ilnar, I'd like to add, by the way, about the parallel things it does. They have a whole list of components they use while thinking and forming the answer. So in GPT-6 Astra they have the concept of interleaved thinking. That is, it can keep thinking over the task after the visible answer has started. It has a separate concept called partial answers. That is, it delivers useful results step by step, without making you wait for all the work to finish. Next, it has the concept of wait-time awareness.
That is, it takes into account how long the user is waiting and tries to show useful information sooner. It has a separate thing, it makes tool calls, it turns to search or the available tools to get additional data. It has a separate thing, look again, that works inside. It's the search decision. Who's out there saying it's just some dumb mathematical model, right? It has a search decision. That is, it determines when it really needs to look for information on the internet and when its own knowledge is enough.
Look, it's a cool separate component. It's called search decision, right? And in general, it's a very important thing, right, where you need to go to the internet and where it's enough for you to look inside the system. It has a separate thing. That's the search for confirmation — evidence retrieval. This system finds the information and sources that back up the answer. And in GPT-6 this mechanism has been greatly improved. And they have a separate thing — that's answer composition, answering composition. It combines intermediate results into a coherently connected answer. I mean, then it gives it to you.
Look, I think this is, well, really very serious. And with all the caveats of comparing, for example, with Anthropic within the chat, well, I'd debate here which is cooler, ChatGPT or Anthropic, in terms of how this kind of system works, right? I mean, if we go into Claude Code and Codex in terms of building infrastructure, layout, displaying everything — okay. But if we... There, there you can compare, everyone can fight for their own company. But if we look, and there, for example, Cursor will still lag behind with Grok. Well, it will. They're lagging, really lagging.
A small comment. About what you were saying, the checkbox "train on your data". You just need to understand the scale. OpenAI has, like, how many is it? They're already claiming 1.2 billion users now. So even if, I don't know, half of them tick this checkbox, then the data you give for training the next models is, like, one user out of six hundred million. So it's a tiny, tiny drop in the ocean. It matters for training the next models, but this model won't adapt to you in any way, right?
I mean, the impact on, well, on GPT-7, for example, won't even be noticeable, if training continues on it. So it's not worth worrying about. If you're not leaking any really sensitive data, although it'll most likely be hidden before training, but one way or another. And, well, yeah, you shouldn't think of it as the chat being trained specifically for you, because at scale it all gets heavily diluted.
That's of course very interesting. But look, Ilnar and I have shown you the first block, right? This first block that came out in GPT-6.
And the second block that exists today, and it's actually been delivered in the Intelligent UI. It's Intelligent UI for a reason, after all. It's not even about the first block, it's not the GPT-6 rollout, right. Although again, GPT-6 kind of already existed. And it seems it probably wasn't like this, right? But okay, let's forgive them. The second thing there is, what is the Intelligent UI? It's a smart user interface for answers in ChatGPT. And seemingly it's the same old thing. We've gone over this a lot.
ChatGPT tried many times to make all sorts of interactive things, and they worked for them. They even once launched questionnaires in the form of these surveys. I think that was about two years ago. I, a few times — it was there, then it died. It still worked in Deep Research, right? Who even remembers Deep Research? By itself, the one that, you know how it worked.
That's the thing, Sasha, sorry, Deep Research — that's the thing that made me end up buying Pro. Because remember, I was a firm believer that you should only buy Pro when you're running into some limits. But…
Yeah, yeah, yeah.
…with the dollar I couldn't hang myself.
Me, I bought it right away. Yeah, yeah, yeah. And you waited.
And I only did when Deep Research stopped being enough for me. I mean, when, I don't know, there were twenty of them or something, I don't remember for what period. And only when they stopped being enough for me did I switch to the Pro version, and I never went back.
Right. And in general they had already started drawing various charts. You all know that if you ask ChatGPT, for example, how the stock price of, say, NVIDIA has changed, they can draw you a complete chart, the kind that's drawn these days in Google, or in Yandex, or on other portals. So they started making a lot of interactive stuff, really great interactive stuff overall, convenient, fun, building it in. Let me remind you once again that right now you can make crazy interactive stuff with HTML.
And, for example, Claude Code and Codex increasingly prefer not to answer with text inside at all, if you're doing an audit, they themselves create for you in the background, like, HTML, these MD files, which are then broken down, and so on. But here's the thing: they've made a new update to this intelligent interface here. I hope this will really keep developing progressively into something platform-like and operating-system-like, I don't know, into something else. What are they adding?
First of all, they're adding a choice of format. That is, it decides when to use text, tables, comparisons, a diagram, various interactive elements. They have the concept of various built-in components. That is, it can use buttons, forms, some charts, input fields and other information tools. And by the way, this is what's cool about Codex. Here's one of the standard, the standard — well, cool features that Codex can do now. You can ask it, for example, to create the structure of a Telegram bot for you.
I keep saying: stop using middlemen like FastTrack, MoneyChat and so on, and use it yourselves in your companies. That's for those of you who have businesses, for marketing and sales: create these flows yourselves, these tree structures in Telegram and everywhere, in WhatsApp too, for example, right? And ChatGPT lets you build the buttons right in Codex, do the whole thing completely. You can click all of them, press back. I mean, it can build you an interface inside, right away, right in the working window. And here, in ChatGPT, they're starting to develop this.
Some buttons will appear, input fields will appear. They're already there, right? They have a separate thing called "interface assembly." That is, it basically combines elements for the user's specific question and turns the interface descriptions generated by the model into various displayed elements. It does it inside. Then it gives them to you — exactly what we've been discussing: since it can give you its answer to the question in parts, it can address them to you separately or not, insert all sorts of interactive features and so on.
In short, it's a cool thing. Again, cool in what sense? Look, on the one hand, we've kind of been living with this already anyway. On the other hand, we all understand it's time to stop. Well, I at least am very vocal about that. It's time to stop using a huge amount of intermediate software, the kind that's sort of pointless, right? And I don't know about you, but I basically don't open Excel anymore. I don't know, I don't know.
Ilnar, do you open Excel?
No need.
No need. I mean, if someone hands you an Excel file as input, Codex or Claude or ChatGPT or something else processes it for you. If you need to create an Excel file and pass it on to someone, Codex or Claude or ChatGPT creates it for you. If you need to visualize or process something somewhere, there's simply no need at all. What's more, the quality with which OpenAI's systems and Claude create documents like PDF, Word and Excel today is, of course, tremendous. The systems may differ a little here and there, more or less, but the need really does disappear. And that, by the way, I think, is the fundamental story — Google.
Because Google is facing a very big problem, a dilemma. So what do you do with the Gmail interface? What do you do with Google Docs? What do you do with Google Sheets? Because, essentially, you can endlessly build agents into Google Docs and Google Sheets, but people will increasingly replace Google Docs and Google Sheets, move away from them completely, work with them through some kind of infrastructure, and then get rid of these systems altogether, right? And of course, this story is, well, for me personally, really instructive. In particular, what's going to happen next.
Excel or Google Sheets — well, you can think of them as an interface over a database for ordinary people. That is, when you're not working with, I don't know, some kind of... Hold on, how do I explain this more simply? Everything Excel does and everything Google Sheets does, programmers can do without it. But then you'll have to work with the database, or with the data in the form in which it's stored, right? Python developers have libraries like Pandas. And, well, I don't-don't-don't want to go into technical details now, but I want to say that these things simply weren't made for programmers, and in the process you lose some of the functionality that could have been built if you worked with the data directly.
When we start working with data either through ChatGPT and other systems, or with the data directly, the way programmers do, the limitations tied to the interface disappear. And you can do more things, and do them better, if you don't use this layer. It was really important back when you needed, I don't know, accountants and other people to use data the way Excel taught us to — they were very important at one time. Whereas now — I completely agree with you — this layer is becoming less necessary, because more and more often you start working with data directly.
But I have to add a disclaimer here: directly meaning you work through ChatGPT, say, but without the interface limitations that are imposed on you. And yes, they'll have to transform, because more data consumption, more work with data, will go around these interfaces they've created.
You know what other question there is here? For now it seems to me that OpenAI, by doing these, these micro-things, isn't working on the core UX/UI of its infrastructure, this tree-like system of projects, right? I mean, well, in the world of artificial intelligence you can't order data linearly, right? And with some unclear wording and names, I think a lot gets lost. So on the one hand they seem to be trying to help the ordinary user at a basic level, and on the other hand they make things that, well, functionally, a person will at some point say: "Listen, I'll still check on another site, because the way you've presented it to me in the interface is, say, uninteresting and I don't like it.
Well, I, for one, don't like how you did it." Yeah, can that happen? It can, inside. And it would be better if they built more serious components and thought about more serious components. Although let's see how this plays out, how it'll work for them, and how much understanding of this they actually have. Because on the one hand, they're doing a thing like, a thing like Sites, where they put in everything: domain storage, and hosting, and absolutely everything, right? On the other hand, they make some micro-gadget of their own and limit themselves to it. You're like: "Uh, and what am I supposed to do with it next?"
And then, at the same time, they make some connector and say: "Actually, go to Adobe Photoshop, and with Adobe Photoshop you'll be able to make it happen." Just recently, when you asked it to visualize something for you, it started connecting the Adobe Photoshop plugin, because you'd turned it on at some point. You tell it: "Why are you connecting it? You can do your own image models yourself, after all." So OpenAI is clearly at a crossroads between how to get revenue as a company, how to work with this huge number of users, and at the same time build this AGI.
So this, of course, well, I think, is a super serious question for them. And this update they've released — I hope this update won't turn into, well, just another thrown-away thing in terms of various capacities, and that we won't see some half-baked terminal interfaces, right? After all, a very big bet was made on, for example, working with the terminal. I mean, you all claim people won't be writing code, so what the hell do you need a terminal for, then? It's exactly the same story. If Tibo says — we're moving on to Tibo now, right? If Tibo says that programmers, all of them, definitely won't be writing code, but developers will be building systems.
But how can a developer build a system if, to build a system, you have to be a product person, you have to understand business analysis, you have to understand business, everything, you have to understand the end task? This topic, of course, sounds phenomenally interesting. Tell us what you think about it.
Yes. So, let's do it this way: first what happened, then, in my view, why it happened. Let's put it like this: there's an OpenAI employee, Tibo, a very senior employee, responsible first and foremost for Codex and everything connected with it. And if I understand correctly, on October 5 he launched a marathon on X, where he says: we either put out something really useful, some new release, some update, some other useful improvements for users, or we do a limit reset. And so, since October 5, he's been posting about it on X every day.
What you and I were already discussing a bit — on day two, I think, he posted that they'd rolled out ChatGPT-6 in chats to all users. As I understand it, it's finally reached free users too, the non-paying users. And you can sort of gauge how long this process took. Besides that, among the useful things they've rolled out, there's "Approve for me" in Codex. Remember, we discussed that it eats extra tokens, and so, basically, if you tell it: make the decision, whether this action can be done or can't be done, instead of me — it'll send requests via the API, and those requests will eat into your token limit in exactly the same way.
And he posted that this won't happen anymore. Well, I mean, that amount of tokens won't be taken from your limit. So for the user it'll be a free action, so to speak, for whoever is using, say, Codex. I hope I haven't confused anyone, but the point is that they've started fitting a bit better within the limits we have. And updates like this happen every day — as you say, there'll be another three and a half weeks, three weeks left after our video comes out. I recommend checking in if this part interests you. Now, as for why they did it.
The people who read Tibo are first and foremost those who use Codex very actively. For the most part it's pro users, programmers and people close to that field, IT folks who use this system very actively. And after DevDay, it's precisely this part of OpenAI's users that suffered the most, in my view, because they cut the limits and didn't give anything new. And it's precisely at this layer of ChatGPT users that the marathon is aimed. Tibo posts what changes they're making, and very generously does limit resets almost every day.
So everyone's happy that you can run all your tasks on Ultra. You can run them sped up. You burn through your limits in huge amounts, and the next day Tibo resets your limits. And the idea is that this will go on for a month. So I think it's aimed precisely at the audience that was left very unhappy with the results of DevDay. And it's for them that they did this. What does this mean for all the other users? If by any chance you haven't been using Codex actively until now, now's the perfect time, because you won't have to worry about the limits, you'll be able to run lots of different experiments and dive into this as deeply as possible.
Once again, we recommend looking at how Codex applies to your tasks, because you have plenty of time for experiments. There'll definitely be three weeks after our episode comes out. There'll be plenty of limits.
Yeah, maybe, maybe three and a half, or maybe they'll keep going, because I think they could — you'd agree — they could actually keep this thing going in general, right?
They very well could, very well could.
Listen, I really like this whole thing. To be honest, it's even an unexpected thing from OpenAI, and it shows, on the one hand, that it's a corporation, such a complex structure, on the other hand, there's a group of guys who have the right to reset tokens for a billion people. They do reset tokens for everyone, right? Not only, not only in the ultra paid versions, right? And by the way, to be honest, Ilnar, this week I worked more in Codex than in Claude. I came back again.
My counter in Codex is already at 53.3 billion tokens. I mean, there really were days this week — I had a day this week with seven hundred million tokens in a day, there was 439 million, 460 million tokens a day. And by the way, I tested Ultra Fast. In our notes about Tibo we wrote: Ultra Fast — that's a super speed that ChatGPT added only in the five-hundred-dollar ChatGPT Pro plan.
That's for hardcore geeks who decided to switch to it, like me. So I tried running it in Ultra Fast, Ilnar, and it gobbled up... In short, it just ate 100% of my entire Pro mode. Just like that: brrr-drrr.
Yeah, yeah, yeah.
It ate the tasks. But the mode is really cool, because if you need an answer fast, like I was saying, it's super useful. You really start thinking about where to turn it on and where not to. What was the downside? I switched Ultra Fast off later. I did a token reset. I've got a whole bunch of resets there, of these, well, credits. I did a token reset, and after the token reset I went in and switched off Ultra Fast mode, but without a restart — I mean, without stopping the session — it kept eating.
I look, and my tokens are at 90 percent again, 80 percent, 70 percent. I stopped all the sessions, switched it off manually in every session, because I realized this wasn't going to work. This way it'll eat a ton of my stuff, and who knows where it'll go. In general, by the way, I want to say that anyone who's already starting to work in Codex or in Work and starting to keep track of their tokens, it's important, because tokens — that's an environment nobody understands at all, and it's unclear how ChatGPT answers questions, and where.
I recommend that all of you double-check with these systems what can be done specifically today. Because it may be completely sure that it, I don't know, can only do this via the API, and not know what it can do itself. And the system may be sure it can't connect a plugin. In reality, today it already can connect a plugin. I mean, ChatGPT's versions, the versions are updated constantly. By the way, Ilnar, they haven't had any updates in the past twenty-four hours, I think, have they? And something's supposed to pop up, maybe it'll happen now, by our Valley morning time.
They like to drop things at some point. And did you read it, or not? Like, he posted today about what updates he'd made. By the way, it was fun. I liked that when he, when he did the token reset yesterday, he wrote: "Actually, we had some cool new changes. Here are four changes, but the community decided we're resetting tokens, and we're resetting tokens." That was really fun, because people just voted for a token reset.
Yeah, he runs a poll, like, I liked the changes or not, give us a reset. And everyone just hits "give us a reset."
And the cool thing here is that you sit there thinking: so, the changes — were the changes bad in the end or not? Or should ChatGPT reset the tokens after all? Well, they reset them, and that's that. You can eat 100% all over again. But by the way, I had my tokens reset yesterday, and in my Pro mode only 50 percent got eaten at normal speed, yeah. But 50 percent got eaten in a day, although I wouldn't have minded eating the whole hundred. I'm thinking, so, it won't eat it all, I'll buy more or something. Or, like, he'll reset it again.
Let me remind you that, for example, in the five-hundred-dollar Pro mode, a token reset is, I think, about twenty thousand dollars' worth of tokens. That's a lot, a whole lot of money, just a whole lot of money. And by the way, what else is cool right now, I think, about the paid plan in general, right, and the Pro plan, is that they reset it for you. By the way, Ilnar, you've got the two-hundred-dollar Pro, your limits are extended until the end of October. So for you the resets are just cosmic. Cosmic stuff.
I don't feel them as much in the five-hundred-dollar version as you do.
If anything, I'm worried that once October is over, this whole thing will become more significant for me.
Listen, but on the other hand, you and I can see how much everything keeps improving, so maybe these tokens will be enough. Although every day you and I, and all our viewers, are get-get-getting used to the fact that now, at the snap of a finger, you can spin up things that, well, used to be impossible to spin up, yeah. And I talked about this too, among other things. There was a whole series of episodes — there were three episodes about the coolest topics of this outgoing year. Be sure to watch them.
Support our Telegram channel, if you like watching us together. Don't forget to recommend our system. And let me remind you that we have an open call for new speakers on the website. Go down below, there'll be a link there somewhere, or on our website there's a link in the section — I think if you go to the projects section, there's a link there to register and become a speaker. Yes, I can see it right now. And it's on the home page too. There's a button there, "Become a speaker." I think you'll find everything and be able to sign up everywhere.
Obviously, all of this was built with Codex and Claude, and it already looks pretty decent in terms of authorization and everything else. And we'll see you on this podcast in exactly a week, and in a few days watch our special episodes. Bye, everyone.