Hi everyone! You're on the ToTheMoon channel — tech news, insights from Silicon Valley and around the world. And today I want to do a special episode about OpenAI Dots. It's already been two weeks since DevDay, OpenAI's conference — which I was at too — where they presented and released the first Dots. It's — well, in plain language — what they call a kind of agent system through which you can interact with OpenAI's infrastructure. But I want to show you a completely different world of Dots, seen from the outside. Over these two weeks we've talked about it in various episodes.
And so that, on the one hand, you form a new understanding of this kind of system for yourselves, for the future, share some of your own experience, how it's going for you, and learn more practical things about Dots — what you can actually do with it. Because OpenAI keeps updating this system. And literally the day before yesterday Tibo — one of the people inside OpenAI who's responsible for this system — wrote about some improvements that had been made there, although in fact he wrote about improvements that had already been working for two weeks anyway.
And that's why I made this episode — so that we don't fall into any misunderstanding. By the way, today I'll try not just to tell you about Dots in general, but to compare Dots with other existing solutions on the market. To show where it really is a completely different approach as a system, and to tell you more practical things about the differences in how the models work. So, how is Dots actually positioned according to OpenAI's documents? It's positioned as a personal agent that you hand a goal and an area of responsibility, after which it keeps working between conversations. So, essentially…
By the way, at launch Dots was only available in ChatGPT Pro. I recommend you read up on it and wait for it, because it's clearly going to show up in Plus too. I don't think it'll show up in the free versions of ChatGPT. That would be strange, but in Plus it clearly should appear. So, how does it work? Basically, you set up a new window for yourself. That is, you get a single chat for all your chats in Codex. Or you could say, for all your chats in ChatGPT or all your chats in ChatGPT Work. Let's set that aside for now.
Look, OpenAI is made up of different software and different systems, and many people think they're using all of OpenAI's functionality, when they're only using some small piece of it. And that really does lead to a certain incomplete understanding of all the systems. So what is this single agent? Why do they call it a single agent? Basically, you have a single chat where you can do lots of things. And some of you would say: "How is that different from any other chat? After all, in another chat you can also do lots of different things." And I guess you could say: well, it's sort of no different at all.
I mean, I can still go into Codex, into some specific task, and set different goals in it too, or even reference other tasks within, say, that project, or pull in chats from ChatGPT. But there's still one particular point: this Dots infrastructure works a level above an ordinary chat, and it's actually built to manage, as of today, other chats and your tasks in a multifunctional way. And the most important thing — understanding when to create a chat at all and when not to. Because sometimes you create a chat just to ask what two plus two is.
Well, roughly speaking, or to ask some question — what the capital of a certain country is, right? Or in what year a certain event happened. And you kind of don't even need a chat for that. It's just a simple question you asked on its own. That's the kind of thing you can ask in Dots too. But at the same time, sometimes you can set a complex task. You can say: "I'd like to get daily news on changes in the stock price of this particular company, or news about how, I don't know, what the current status is of the war between these particular countries, or news about who'll win the election race for a certain government, or news about whether OpenAI has released a certain update."
And you can go in and set up a whole separate task for that, or you can essentially set that kind of task through Dots. Overall, if you reread all the positioning, there's this kind of dumb positioning of Dots. Once again, I think it's dumb, because I'm going to show you a more fundamental, serious understanding of this system. So that you can understand it a bit more broadly. That doesn't mean OpenAI will still have Dots a year from now. OpenAI likes to relaunch products, change them and so on. But I'll try to show you.
There's this dumb positioning that supposedly, if you want to do a linear chat — for example, research for me what the main events in my city were over the last ten days — then supposedly that's not Dots, that's ChatGPT. And if you set a task like: I want to keep track of what events are happening in my city all the time, set a reminder for that — then supposedly that's what Dots should do. Once again, the fundamental feature of Dots — it sees your entire OpenAI infrastructure. That is, it sees the projects you're working on in Codex.
I know many of you here don't even use Codex, and I strongly recommend it. It's one of the main ideas or trends I keep voicing — to use, to start using Codex, even if you don't do… Especially if you don't do development, actual programming of complex systems. In order to really plug into the new market of understanding what artificial intelligence is. Because to understand artificial intelligence, you'll have to interact with AI, give AI tasks for some kind of automation — not the dumb manual actions we're used to in search.
For example, when we open Google search, we're doing a dumb action, or when we talk to Amazon Alexa or, I don't know, Yandex Alice, we're doing one dumb, simple action. What artificial intelligence brings us is the ability to carry out — look — not just ten actions in sequence, not just ten actions in parallel. AI can work with a large number of different systems at once, in parallel, in sequence, calling up one task or another at just the right moment. That's a very important feature. By the way, Dots should in no way be confused with what are called agent systems inside chats, or inside certain projects, that the chats themselves launch.
I mean, the chat itself — in, I don't know, Codex or, say, Claude Code — spins up whole agent systems in parallel, various multi-agent threads, where different agents try to complete your task. Dots is about something a bit different, right? It's a kind of higher-level infrastructure — essentially, exactly that single window you have for interacting on different tasks. What makes it notable, interesting? The fact that you can, all at once, dictate into it what two plus two is, what the capital of wherever is, set a daily reminder to get updates on certain events, ask again about a certain email you got from your doctor two months ago — when you're actually supposed to go back and see them, have it find your test results and review them, have it create some new task, a new system for monitoring the temperature in your home, connect, say, to some sensor, and so on.
So you can do all of this in one chat in five minutes, just dictate it all to it. So essentially, it's a system where you can dictate all the different tasks you want done. Again, look, there's a comparison. People often come and say: "You could dictate this into, say, another — into another agent." Today agents are being spun up in all modern systems, right. And you can say it to Gemini. You can say it to Grok, you can say it to Claude, but the approach here is still different in terms of infrastructure. It's a little similar.
You could say GrokBot is a little bit like it, but today I still want to show you fundamental things that are different from that in principle. But first, let me tell you this. So, Dots runs on GPT-6 Astra. Astra — that's essentially the family of OpenAI's most up-to-date models that are generally available to people. And GPT-6 — that's, well, a serious level of system. Yes, there are already 6.1 systems now, but there's no Astra there yet. They could appear at any moment, including while you're already watching this video. It doesn't matter, right — if they've appeared, there's 6.1 — but clearly the six family is going to be around for a while, right?
And the seven family will probably come in maybe three, four, maybe five months, or maybe even later. We don't know what's going to happen, right? Maybe the numbers will go away altogether and it'll just be ChatGPT. We'll see what happens, what'll be in there. So, an important point: when they say GPT-6 Astra is used there, that doesn't mean you have access to a super-advanced model, because GPT-6 Astra can be really dumb when it gives you a particular answer. That's a very important point, because let me remind you that in ChatGPT, when you pick version six, you're basically also picking the mode in which the system answers.
It answers in a very simple mode, like instant, or it answers in a super-powerful mode, like pro. And by the way, I want to say that the pro mode is radically different from all the other modes, right? I've talked about this many times. This mode is only available in the Pro version. A radical difference. The models work completely differently. What does working differently mean? It means compute allocated differently to solving your task. It's as if you're riding along and a person is pushing you, pushing your cart — or there are two Ferrari engines in there. Yeah, that's a big difference.
That's roughly what this mode is. But who's being carried — which model it's carrying, whether that's a wagon, or a castle, or a whole country — that's a different question. And then which system actually runs all of this — that's a fundamental thing to understand. But in ChatGPT there's a certain nuance, because Codex and Work — because Work, it's basically chats, a system very similar to Codex. Let's stick with Codex. In Codex you can choose model six, choose within it that you're working with, say, Astra, and also choose the effort level for that Astra.
So you can choose there too, and you'll have high, or, like, a simple — a plain low, I think, yeah. Then there are different modes: medium, there's high, there's extra high, and there's ultra. Depending on the system, they're called, like, Ultra Code. For example, in Claude the strong mode is called Max. I've started mixing up these modes myself sometimes, even though I'm, well, a super-user of Codex and Claude Code. Depending on this setting — Max, or high, or extra high, or ultra — the further right your slider goes, the more powerfully the system works.
Look, in ChatGPT, when you open a chat, it says, for example, 5.6 Sol Extra High — that is not the same as Codex 5.6 Sol Extra High. Look, this is a crucial point. I rarely talk about such micro-subtleties, but I'll draw your attention to it here: to figure out Dots, you simply have to understand that 5.6 Sol Extra High in Codex is different from 5.6 Sol Extra High in ChatGPT. It'd be great if those of you who've lived through this more and understand it wrote in the comments. Share your cases too. That's really very useful. Let us know.
By the way, don't forget to support our channel and like it, share it with friends, and subscribe if by any chance you haven't yet. And even if you see our videos all the time, it all supports the channel. And plus — plus, good content, if you like it, spreads among the people close to you, too. Or when you recommend something to someone somewhere. When you move over to Pro, it turns out it's not clear at all how to choose that kind of mode, in which systems to choose it, where it even exists.
Because Codex and Work run on a different answering technology. That doesn't mean — look — that Codex in Extra High mode in a chat, in version 6.1 Sol, answers way worse. It doesn't mean that, but they work differently; they have a whole different system of analysis.
So when you work through Dots, you have to understand that this system — especially if you're used to different kinds of answers — this system is still one that, as of today, isn't yet able to understand from a particular request of yours that it needs to create a task in ChatGPT and open it in the Pro version. If you have — well, Dots is only in Pro anyway. It has never once created one for me. What does Dots do? Dots can create tasks in Codex, and it does create tasks in Codex.
That's very important. And by the way, the Dots description says that Dots can work with your cloud computer. Though there it's actually Linux with Chrome, right, a managed one. And it can work with a local computer. By the way, it says you can connect only one personal computer, and obviously it needs a network connection and the ChatGPT app open. And that's also an interesting point: you basically need to have Codex set up for Dots, and you need to work in Codex a little for this whole thing to happen for you.
Because what does Dots do? It periodically creates tasks in Codex if something needs to be checked. For example, when it needs to sort through mail or send some emails, when it needs to, I don't know, scan how your home network is doing, say, whether it works, or change some settings. That's when it starts creating tasks in Codex. Usually infrastructure ones, or ones for working with files, or for converting files. It really does start creating tasks in Codex. And of course I have a big question here: it doesn't use ChatGPT, because ChatGPT is still a slightly different infrastructure.
And so far it hasn't told me personally that it used ChatGPT for this, for example. There's some description saying that OpenAI has now supposedly announced you'll be able to call Dots from chats. So far it definitely hasn't worked from ChatGPT — it doesn't work at all. And as for calling Dots from Codex or from Work, it works more or less. It'll be interesting if you've managed to do it. I've repeatedly — I had this situation where I was working in ChatGPT, and then I said: "Set up all the reminders through Dots."
In theory — let's call it Dots for now, because it may well be called something else in other systems, OpenAI included. That is, it's a kind of single infrastructure for your own interaction with one company's artificial intelligence — and maybe later with different companies'. Because after all, if you have other models connected to Codex, or various other systems connected to Codex via API, then obviously Dots has access to that. That's a really cool thing. So if, for example, you have ElevenLabs connected in Codex for speech-to-text, you can call up a speech-to-text task through Dots.
Or if your Codex is connected to your computer's infrastructure and has access to sending SMS messages, then Dots can send an SMS message through Codex. Obviously, they'll be integrating more and more plugins, settings, details and everything else. But you and I have to understand that the future of artificial intelligence — it's not intelligence that integrates through plugins, through MCP, through API. It's an intelligence that will be able to figure out on its own how to connect to a given system.
And sometimes, to connect to some system, you'll have to plug some extra computer into an outlet, or into the network, right, or switch on some piece of equipment, or change some settings, or open the web version of some system. So everything can work in different ways. For example, I have a video surveillance system — it's more than, I don't know, 25 or 30 cameras.
And there's a system where the cameras stream directly to a computer, so Codex sees all the video footage itself, right away. I can very easily ask through my Dot: between 2 and 3 a.m., did any particular animals show up on these cameras or not? And if they did, which ones? Or ask my Dot to notify me about those animals. We'll talk in a moment about how it works in terms of ongoing notifications. We'll discuss that shortly. For example, other cameras work through a cloud service, and that cloud service doesn't have an API.
That is, it's a cloud storage system, a standard camera storage system — there are lots of cameras like that, systems like that, in the world, right? From Chinese ones — the ones that come from Xiaomi, from Alibaba and that whole sphere, from Mi House and all that, I mean Mi Home, right, to American companies like, I don't know, Ring or Blink — that's a camera system owned by Amazon. And, for example, some of them you can't connect to via API. But these systems do have, say, a web version, and you can basically give Codex access to the browser and, through that web version, download those videos and then, say, analyze them to some extent.
You can do that — well, it's not exactly a simple move, but essentially that way you can at least somehow achieve this kind of link, right? And you always have to look for links like that. For example, I have a fairly complex home network. I mean, I have lots of different devices. I have a whole series of all kinds of modem boosters, mesh networks over Wi-Fi, over Ethernet, and several satellite systems installed. And configuring all of this, for example, is just very hard to do. But if there are settings, you can give access through a web browser to configure these modems or boosters.
I mean, it's open, maybe to Pro users among others, but on the one hand it eats — it'll eat — a lot of resources, and on the other hand, it's a system that's built — well, it has lots of glitches. And obviously ChatGPT Pro users are more loyal to the company, they know where to open which systems, and they'll be testing Dots little by little on the side. They're prepared for Dots to fail at some task. I want to tell you that Dots fails at a number of tasks, and, just like ChatGPT once did, it handles repeated tasks very badly.
What makes it interesting overall? The fact that it's still a single, personal interface. You don't open individual tasks that have had some updates, right — you work with one window. I mean, that one window of yours sorts through different messages, for example. Here's an interesting thing with mine: it's connected to all my email. I recently met with my partner in America — we have a business there in re-, in real estate, with this particular person. I keep urging him — we have a lot of automation there, I've even shown part of that automation.
I'm not very involved in the day-to-day operations of that business, but I'm really pushing my partner to use Codex heavily. And he says: "Aren't you afraid to connect all your email?" I say: "Listen, where is your email stored?" He says: "At Google." I say: "And how is it protected there?" He says: "Well, probably pretty well, at Google." I say: "Do you think OpenAI protects data any worse?" And that's where the person freezes up, understandably, because email can actually be stored in more places than just Google.
If you open your email through Apple's interface, then your email also starts being stored in iCloud, for example, if you have syncing on. Your email also starts being stored on your computer. And then there's the question of whether your computer is encrypted or not, and so on. Your email is stored on your phone, right? And so OpenAI is, of course — well, probably one of the most secure companies in the world as of today, at this point, because this company has access to more than a billion people, to people's records, very important people's, right?
Even Windows operating systems — they're still offline, offline storage, right — and they don't have access to the volume of information that OpenAI has today. Although of course, in some ways they do, right, because your files are often stored on Microsoft's — on Microsoft servers, or on Google servers, or for some people on Yandex servers, and so on. So that's, well, pretty serious storage after all. The question, once again — who has access to them and when, right? And OpenAI is still a pretty serious company.
So I told him: "Of course you can connect your email and not worry." Then he asked me: "And can I connect financial systems and give Codex access?" I asked: "What exactly?" He said: "Well, you can give Codex access, for example, so that it actually manages the finances — has access, a direct integration with the bank." I say: "Listen, well, I probably don't give Codex access to all the banks, but I do give access to… Although I don't even know why — that's what I told him — but I'm sort of fine with giving read access, because I give access to a number of accounts to other systems that are less advanced than OpenAI.
It's only read access, after all, right? But I upload all my bank statements — I tell him — into Codex, and that way it can do any analysis for me. I opened it up for him there, and for one corporate environment I got an analysis in two seconds that no accountant, no finance person would give me — nobody would give it as fast as I got that data. No financial system in the world today can analyze that fast. It was an unbelievably cool result. I asked it there what the car expenses were in July at this company.
And it laid out what car rentals there were, what leases there were, what loans there were, what fill-ups there were, where those fill-ups were, what parking there was, the various garages, visits, what repairs there were. It was unreal. It laid it out in such a cool way, but really seriously — it laid it out really impressively. And then it also suggested what else could be classed — counted as car expenses. It said that, well, you could also include taxi rides, for example.
It's just that initially, since the instructions were specifically about cars, it hadn't included, for example, rides on Uber or Lyft, it hadn't listed those rides — and then it added them as well and showed everything in super detail: everything there was, in which periods, where the cash flow was, where the P&L flow was, that is, where the specific expenses were and which month they fell in. So it was an incredible thing, and it got done. Why? Because it has access. By the way, note this — my partner says something very interesting: "Listen, with just that kind of access to the statements and the bookkeeping system, it wouldn't have told you this right away."
So it turns out yours has access to the receipts, and receipts often come to email or into other systems. For example, your receipts can be stored in Uber, they can be stored in Lyft, they can be stored in Booking, in Expedia, in various, I don't know, airlines — United, whatever, Lufthansa — wherever, whatever you fly with, right? And of course a huge number of receipts — the bulk of receipts — gets aggregated in your email. And it's a huge plus here if receipts can come to your email, because then, seeing those receipts, it immediately draws conclusions quickly, and draws them incredibly well, right?
Here's another fun case I had, by the way. I was recently closing out a month, and it says to me: "Was this trip of yours to this — to this region for work?" I say: "Yes, for work." "And there were two more people with you." And my two daughters really had gone with me. And it says: "Two people traveled with you." I say: "Ah, that's personal, that Uber is personal." It says: "Okay. And were your expenses at this hotel for work?" I say: "Yes." It says: "But you had a spa here." I say: "Ah, the spa — that's personal." I mean, I don't even remember what was there.
Sure, the accountants would later sort out the various details, going through the systems — if they see everything and don't miss anything. But here the system sorts everything up front, right away — understands it, splits it, allocates it evenly. That's real mastery, of course, right? So I'm crossing off — how do you put it — absolutely all the financial systems that exist in the world today that do bookkeeping for people. Absolutely all of those systems, right? It's absolutely pointless to keep books in them.
I'm by no means talking about complex financial structures, insurance companies, credit systems, all kinds of banks, complex — even corporate — systems, investment firms. By no means. But basically, for the bulk of them — 99.999% of the world's legal entities — of course, modern bookkeeping, managing finances, is going to be done directly through systems like Codex or Claude Code or Grok, Cursor and so on. And here — well, Cursor belongs to Grok, right? SpaceX SI, right? They renamed themselves.
SpaceX Super Intelligence, right, SI. Well, because of the current new trend. Of course, it'll be done directly through these systems, because this kind of work is direct. Look, look, here's the trick. This is where you really see Dots, too. What's the advantage? Why do I say these financial systems will die? Because Codex will have far broader connections than any software. Either way, it'll have broader connections. It won't need certain types of connectors; it'll be able to integrate directly with your infrastructure. By the way, speaking of which — email.
If I have my Google mail integrated through the Gmail plugin into OpenAI, or into Claude at Anthropic, then, for example, my other email — say, I have corporate email through iCloud, or email on other mailboxes, not Gmail, then my Codex has access to that other email through my local computer, because the local computer has access to it, and it gets access through the local computer. Accordingly, Dots has access to it. This is a really cool, incredibly interconnected thing, right — of everything, everything that's happening; it's a completely different style of contact, and a whole different style of interaction and communication.
And what's remarkable about artificial intelligence? You see — imagine: by the way, I have a business, in America it's called a partnership, together with my wife, right?
Where we have some shared content and so on. And it's exactly at that point that I need, say, access to my wife's artificial intelligence — and I can get access to my wife's infrastructure, or could get it, only through her Dot, that is, through Codex instances linked to each other across different accounts. Today that kind of link only exists for corporate accounts. I think this is something of an oversight by companies like Anthropic or OpenAI — or rather, a kind of political bet, or, well, look, Anthropic positions itself for the corporate systems market, but OpenAI was positioned from the start not for the B2B market but for B2C.
And Anthropic managed to capture a new B2C market of people — the world's future automators, people who in exactly this era, this epoch, will be able to plug into thousands of different professions in an incredibly powerful way. That doesn't mean these are people who are tech-savvy like programmers. No, these people have some whole other style, a different world. Well, we talk about this a lot here on the ToTheMoon channel. It'd be cool if those of you who are that kind of person wrote it in the comments. I'd really like to hear about it.
Which country you're from, right, and how you're spread across your professions. Be sure to write. I really enjoy reading about it. It's great for other people too. Tell us about it, write right now, okay? But OpenAI has always been positioned for the B2C market. And that's exactly why OpenAI has more than a billion users — and they have more than, someone said recently, more than a billion accounts. True, a billion unique people is a big question. But still, yes, they're oriented toward the B2C market.
ChatGPT — it's the most famous artificial intelligence in the world, overall, both in its mass and in its scale. There's no more famous brand today. What's more, it's a company that became famous not because it had backing like the name Google, or backing like the name Alibaba, right? I mean, it didn't have that, or WeChat, right? I mean, with companies like those there was very serious backing. ChatGPT didn't have that, just as Anthropic didn't. Then again, Anthropic — it was never originally about the B2C market. We'll see what happens, of course, yes.
And what an oversight they made. From the start they didn't build a communication environment for AIs to talk to each other.
Maybe it's a hard problem to solve, which is why it hasn't launched. But OpenAI does have a setup where different AIs can interact with each other. For example, in Work, and inside Enterprise, and inside ChatGPT business accounts, the system sees certain — well — links. Right now ChatGPT, as we all know, has rolled out a new thing. It works very strangely so far. We're not even — I'm not even starting to shoot extra episodes about it. I think it would be strange to make videos about it for now. It's all about the feature called Spaces, and it's, well, a very strange construction.
Maybe Spaces will become a kind of foundation for different infrastructures to interact with each other. By the way, what's interesting is that I think OpenAI picked the wrong approach in terms of the UX/UI of communicating with a person. Because Dots should rather look like Spaces, not just like a chat. That is, essentially, when we go into Dots, we should land in this kind of space where we see lots of different chats, lots of different visuals, some shared pages, some bulletin boards, some communication systems where we can interact with other people's Dots, and also see our own central — say, our own central chat, or our central Dot.
By the way, OpenAI has said that for now you get one Dot, and in the future there may be many. And that, by the way — again, the people who talk about this, that's their shortsightedness. Because if you create a Dot as a kind of central mechanism, as a central control server for artificial intelligence, regardless of how it looks, then there can't be several of them. That's exactly the entry point. But if you create it as just some agent that manages different chats, then it's merely a chat that manages different chats. Right?
I hope they don't think that way. Right? That's of course a loss of a certain vision, and you'll be able to recall this video of mine when exactly these kinds of central interfaces for managing everything start to appear. Because as of today, nobody has central interfaces for managing everything. That's exactly why — look at how OpenAI as a company manages things. For example, OpenAI sends me — a Pro user with thousands of chats — updates saying four new image styles have been added. You can create pictures there, after all.
Even though they can see that I don't interact at all with those emails, specifically the ones about Images. And in general I don't use Images on its own, because they do have it as a separate thing — there's even a tab like that — Images. And when it comes to creating these various images, they have trending, they have all sorts of templates. I don't use any of it. It's a project, some half-dead project they've got hanging there. I mean, I don't use it. So it turns out they don't know me.
If they're offering me that kind of interaction, it means they don't understand at all who I really am as their user and as a person. And internally, their interaction with customers doesn't go through artificial intelligence, right? And it's important for you and me to understand that today a huge amount of communication from these companies doesn't go through artificial intelligence.
And we'll see powerful things when these systems — no matter what even Elon Musk claims — they're probably, strategically, how do I put it, visionary-wise, moving really impressively, but they don't have AI interaction — like, where AI is on the front line, right? Although at the same time Elon Musk says — at a meeting with a friend of mine, he told him personally; he didn't tell me, I haven't met him, but he told my friend that he doesn't see an AI chat as the system — as a chat — as the central system of artificial intelligence at all.
So he saw it more through hardware and through some other systems. And to some extent I agree with that. I don't see chat as the central system either, because it'll be a voice system, and the phone, some operating system, some systems, right, some robots — various systems that will be running this. Because, imagine, I'm recording here right now. I have two cameras set up that are filming me in some mode. There's a microphone, and there's a separate computer through which runs — through which a special system is launched, software that organizes the stream.
Honestly, I'd happily not turn on that computer at all — if instead some device rolled up here, some, let's call it a robot, let's call it — or not even a robot rolling up, but some system that would connect these phones of mine to its own, to its own system, all by itself, and join the stream, right? So how would that have to happen? It's not a question of a robot. Let's even imagine I'd set up the cameras myself, because there are no robots that can set them up and do it like that. It's very hard for them to do, right?
I'd tell it: "Rut, listen, I've just connected these two cameras. Please start the stream now, the recording, through them, through Riverside. We're filming. We're shooting in 4K through my Shure microphone. It's connected here to the computer, or to some system, or it's connected to you yourself. You do it." Where could this Rut be? It could be like Alexa's speaker, it could be present throughout the apartment, hung up in the form of sensors. It could be, maybe, some computer standing somewhere, it could be on a watch, it could be in a phone, it could be inside some phone that's filming.
So it can be anything at all. Look, this is very important — it's a question of interface. Today I — by the way, I'm actually building this project right now, guys. I'm about to find a microphone system that I can connect to central computers. I've just ordered myself another separate computer for the house, a modern, well, fairly powerful one. It'll have cloud Linux on it, it'll have cloud Codex on it. And to keep developing my smart home system, I'll try, by the way, to connect this microphone system, and in such a way that this microphone system would maybe listen constantly within certain ranges, when I allow it, and would connect to — connect to my computer. Funny thing, by the way.
Yesterday I… This is just for laughs, of course, right? My robot vacuum got updated. Roborock released a version — it's the S7 MaxV. It's not exactly a current-generation vacuum, but it's their most powerful model in the world as of today. This vacuum has a video camera and a system that lets it find certain objects, drive around somewhere and do things. And basically, with Codex I can turn it into practically a security device, right, one that works around the perimeter. Interestingly, a friend of mine recently suggested I buy a robot dog, but the cheapest robot dogs — the kind that can run around in the mountains and so on — cost about $4,000.
I keep thinking about it, by the way — maybe the mood will strike me and I'll buy one in the next few days, maybe not. I'll look at how available it is, how it learns, how it actually runs. All of this can now be connected directly to Codex and, sort of, linked up with Rut. And if my Rut is connected with other Ruts — well, my Dot is connected with the Dots of the rest of my family — then imagine what kind of contact that creates. You see what I'm getting at? The thing is, here's why I give examples like these — seemingly primitive, for the home, technical, geeky, supposedly uninteresting.
On the other hand, these circumstances, these different kinds of examples, really let you dive in and find new scenarios for how artificial intelligence works and how to understand it. So that you can figure out what's going on. For example, Dots has four different properties of memory.
Here's a thing it has, for example — look, this is a very important construct, right — because Dots works with a large infrastructure of the different kinds of memory you have. Let me remind you that memory comes in very different kinds, right? There's memory you've written into the system itself. For example, in ChatGPT you've written down the rules you want to work by. There's the memory of the project you're working in. There's the memory of the context window of the chat you were talking in, right?
There's memory that the model itself has learned about you — or rather, not even the model itself, but the infrastructure of the company you're working with. Like, OpenAI's infrastructure or Anthropic's infrastructure. And this is where an interesting point comes up: this isn't just some memory gimmick. Imagine: memory knows that you had a test done three months ago, but it doesn't know exactly which test you had. And, for example, today you went to the doctor, and you upload your new results to ChatGPT and ask ChatGPT: "Is my blood test normal or not?"
ChatGPT has the option of simply giving you a rough assessment of your current blood test. ChatGPT has the option of pulling up data about you. Who you are in general, what kind of person you are, where you live, in what region, how old you are, how you work. And telling you whether this is basically normal or not for you, for your makeup as a person, for your profile. ChatGPT has the option of seeing that you've already had other tests done, in another chat. Look. And, at that moment, making use of the fact that it knows you had tests done but not how they came out. Going in, additionally studying that data, and answering you — answering you based on what's actually there.
And ChatGPT may also have a memory where you've stated that you don't live by assessing your health according to, I don't know, the WHO health standards, right — that you want to approach your blood tests in a broader, different way — and for that it needs to apply this rule of yours. So it has different properties. There's the notion of storing — whether a fact is recorded at all. There's the property of retrieval — whether it's found at the right moment. There's updating the data — whether an old fact has been replaced by a new one. By the way, this one is incredibly important in memory.
Across your life, your projects, all your chats and everything everywhere, you can't constantly be on top of updating facts. Has its system updated the fact or not? You really — for example, you're supposed to go on a trip to Paris in a month. And two months have passed, and the system thinks you're supposed to go on the trip, or that you're on the trip, or that you've already been on it. It also has to know whether you actually went. And so, when you talk to it later, it has to really look into all these aspects of how things went.
And in today's world of artificial intelligence, you shouldn't have to tell the system about all this. It should somehow be in the loop. Look, for it to be in the loop somehow, you have to use these systems constantly — and use them constantly not just as a chat. You have to give this system, well, fairly extensive, broad access. That's why I gave you that example with recording, with conversations. By the way, the fourth property when it comes to a record — that's scope. Is it acceptable to use it, for example, in this task or for this audience?
I mean, this particular piece of memory that was there. Because sometimes a memory may not be relevant. You might have asked a health question that actually wasn't about you, but about someone you know. And when you ask questions, in theory the system also has to understand whether you were even asking about yourself or not. But this is a well-known case: about a year and a half ago, back then, I started telling ChatGPT clearly that I was now asking about, say, my mom's plants, not my own.
And when I ask about mine, I say — this is something I have, right. And the system clearly takes it in, looks at it, analyzes it. Very important, yes — it analyzes. And it also has to constantly work out my opinion. For example, if the system tells me: "I wouldn't do that, because there are these norms." And I say: "I don't like those norms, I want to do it differently." It should gradually remember how you live, how you — how it needs to interact with you, which norms suit you and which don't.
Because it will suggest some other norms. You'll say thank you very much, I'll use these norms. And these norms may also differ from the norms of your colleagues, your relatives, your loved ones. And, accordingly, the way systems interact with each other can vary. I've given lots of examples over the past two weeks regarding Dots — that if a Dot, if Dots could interact with each other, then between me and my wife,
well, probably at some point, in terms of interaction, any other chats would disappear — would just disappear. I mean, she and I would end up interacting through our Dots, more or less. Why? Because a lot of issues can be resolved — since we have a great many family issues, all kinds of administrative issues — they can be resolved by artificial intelligence. I mean, AI can take some of the issues on itself and remove certain complications that are there inside. It can help resolve some aspects.
It can help with what's going on at that moment, right. It can know what's going on. For example, today I had a specific situation where my wife was leaving the house and she thought there was a problem with a tire. She called me, I drove out, picked the kids up into my car and drove on. But while I was driving, my meetings started — I was already in a meeting, and she was calling me. And so, essentially, my Dot should understand: okay, my wife is calling. It sees that I have a message thread with her, that she's out there. I told her to check the tires.
I mean, my Dot… Look, this isn't about the Android system. Right now there are various systems on Android — in the Pixel, various mobile apps — that can figure things out, answer, send your wife to voicemail, and on the voicemail run some script for her and ask: "Alexander is busy right now. Is it urgent, not urgent, and so on." That's all total garbage, right? In theory, my Dot should immediately send her a message: right now Alexander is talking, he's in a meeting, he can't break away.
I can send him a message over, I don't know, Zoom. Or tell me how urgent or not urgent your matter is, whether it's worth interrupting him or not, and so on. So that's, of course, a completely different level of communication. It's essentially a completely different kind of interaction altogether. Yeah. And so Dots — once again, maybe I've run a bit far with the Spaces idea here, right, because visually these should be completely different systems. And this system — it's definitely not just a chat and not just one window. Yes, in theory I can sort of go off into Dots from any chat. Yeah.
I mean, it's a system that should be watching everything all the time, not just part of something. For example, my Dot started watching my email. And although I asked it to watch my email only for things concerning the kids' various school events, what the school sends, it started checking my email in general — and it's pretty cool. It can pull various emails out of spam and say, this is a scam, don't touch it. It can remind me that someone sent me something, that some storage space is running out, some payments somewhere else, and so on.
But it doesn't check other systems. And I'd like it to sort of check other systems too. But for that you really need to let this system in more widely.
Yes, this system needs to be let in more widely, fundamentally. And what does fundamentally wider mean? The system has to get access to the phone. And this is where we come to the question of certain conventions — conventions even on a computer. Which systems do we let Dots into? How advanced the system is, and how confident we are that this system won't do something horribly bad — because systems do bad things. Even my Dot replied to some emails, and replied to them incorrectly in terms of the email structure.
Later I saw that instead of replying to the email, it wrote these people a new email. Yeah, so it did it completely differently. Dots has an important aspect, of course, that has to do with autonomy, right. So, once again, Dots is now, first of all, available on the phone, it's available on the computer, it can work with your other computer, it can work in the cloud and work on a whole bunch of different parallel tasks. I mean, this is something incredibly important, and it's the thing that lets you work with one infrastructure.
And once again, look: if you're a casual user of artificial intelligence and you work with it through some chat or other, and occasionally have one project somewhere in Codex, then what I'm talking about now is a kind of complex future for you.
To really take in what Dots is, properly, you need to decide which main system you work with. That's one of the ideas. I talked about this — we had a whole video on the channel: ten top ideas, then ten more, then twelve more. Just recently, this year. It's worth watching all three of those videos if you haven't — that's 32 in total. Settle on one system. A second one, maybe, as some additional system somewhere, right. And, by the way, there's also the question of setting these systems up with each other. Because, for example, there are some cool things Claude can do. Can you allow Codex to, say, influence Claude, right?
Because Dots can't work together with Claude at the moment. These systems aren't developed to work with each other, and that's very bad. Yesterday I was driving, and one of my cars has Google Gemini in it. God, what a dumb system that is! What a degraded company Google is in some places. Of course, it's just — it's just an unbelievably dumb system, period. Google Gemini, the voice in the car — it's just a horror, it's an absolute horror. And Elon Musk would probably back me up here.
And yesterday I was searching the Google Store again, checking whether I could install ChatGPT. You can't install ChatGPT. So it turns out that in a car that mirrors Apple's infrastructure, the ChatGPT icon showed up. I had that in one of my cars. But the icon is there, at least, and you can use voice there and so on. And even just now, when I was on a trip in Europe, a friend of mine noticed exactly that, because the icon appeared there automatically, since I have ChatGPT on my Apple.
But here, it turns out, in the car that runs Android, I can't get ChatGPT up and running right now. I understand that maybe someone has — if someone tells me how to do it, that would be super cool, because Google Gemini — it's just a revolting thing. And of course Google has a huge minus — they're a very unfocused company, and catching up with the artificial intelligence of Anthropic, OpenAI and SpaceX will be practically impossible for them. What do you think about this, by the way? And it'll be interesting what happens — well, it'll be interesting whether it happens or not. So, Dots has proactive reactions.
That is, Dots may start doing some things you didn't ask it for. And I hope they'll develop this a lot, look at it very closely. And I, in particular — say what you like — am going to run an experiment too. And you can do it as well, those of you who are heavy users. Let's launch as many different things as possible in Dots, give it access where it can explore proactively. Let's see what happens. I'll let it in today. Right after this recording I have a piano lesson, and after the piano lesson I'll go and let it into a much wider range of everything that's going on, right, because I have lots of different infrastructure projects that I'm doing on Claude, in Codex — in marketing, huge analytics, mobile apps, ad research, well, lots of sales and so on.
I'll try to let it a bit further into my life now, and I'll beef up this smart home setup. What will you do after this overview episode of ours? We'll be filming the next episodes about Dots too. Please don't forget to support our channel here, and share it. Every comment of yours helps us here and our channel grow a lot. Bye, everyone. See you on ToTheMoon.