OpenAI moves from a simple, retained user note to a permanent, updated and synthesized memory. This means that ChatGPT should start new conversations not from scratch but from the already collected understanding of humans, projects, preferences and constraints. So they had a massive release. We'll talk about it today. The theme of the common memory is very much I raise and raise its importance and the need for human life. And on June 4th, the OpenAI published a material called ChatGPT Memory and Dreaming. And even dreaming isn't, uh, not now a model. It described this extract they had, a more powerful, large memory system for ChatGPT. And they said they wanted to solve three problems. The first problem is memory obsolescence. The second problem is memory correctness. And the third problem is the ability to work with hundreds of millions of users and a multi-year history of interaction.
And the interesting aspect, that is, again, that memory in the artificial intelligence model and not just ChatGPT, is a critical need to be understood. I recently rented a release on the security channel, too, including the data you store and memory. But today we will talk about the usefulness of this tool and, in general, your understanding of how models are built. Because someone recently wrote a few days ago, why would you even buy a model? The models are not very different. It's a sort of outdated perception that modern artificial intelligence models are. The term of memory in the assistants often mixes into different mechanisms and, therefore, it is better to divide them to levels. Yeah, I'll split them to levels. The first level is the context of the current chat. I mean, the model sees it in open conversation right now. So you've created a new chat and what's hanging out there? And the user clarified the task, for example, five messages back, and disappears outside the window into the context of the original tip.
It's memory, right? So she has a big risk, that she, uh, does not keep it or may not be used. The second memory that exists, the second level that can be distinguished, is a clear memory. The user or system retains a separate fact, yes. And we've seen it very much in ChatGPT, and we've talked about it, talking about the accuracy and inaccuracy of using this story. For example, you might say remember that I prefer short answers. Or, for example, I'm using, remember, it's very important that I have every message to celebrate the date and the time. There's a big risk there because the fact can be outdated, incomplete and too general. Speaking of fact, about the facts and their obsolescence, I had a system that said that since you've been forty-two years old, you have to do one, two, three, four. I was asking for some medical tests, and I realized that she-- I had a little bit of a baby, not a birthday. It does not determine or understand, or recount the data accordingly. I mean, at least the data she's filled with, she doesn't even watch the date when they were filled. Of course, in the future, that will have to be rectified. I don't know why they didn't do it from the beginning, but it's what it is. Third level of memory, yes, it's a memory from chat history. It's been a very active one, I think, for months, maybe three or four, three or four years, especially in OpenAI. OpenAI is the leader in this story in the way they try and do chat. The system is looking for or synthesizing the useful facts of your past conversations that you have. And the model, therefore, remembers what you've done, what kind of work you've been doing, what restrictions you have, past decisions, and so on. Again, what is the risk and the problem is that the model can be subtracted by an unrelevant or old context. A very old context, for example, can pull. Or, again, literally from the fresh, I'm now, so I'm doing it in parallel, and I'm just creating a few systems that I'm always saying is that it's like AI native systems.
These are systems that create a purely artificial intelligence. It's not that it's not a web-based coding, it's a complete software product for itself, well, or for some, some of my own ideas and decisions. And literally the other day, the system, retracting information from different chat rooms, from different projects, mixed projects and started making very serious mistakes. And that's a big problem, especially when you have a lot of chat rooms. I have thousands of chat rooms. It's a big problem and a big problem, because a lot of data are unrelevant. The fourth level of memory to be taken out is memory from the sources involved. So you can, for example, download the context from the files, the mail, the calendar, the disk, even your browser, you can connect. And, for example, if you step away from ChatGPT and remember Gemini, Gemini is now using a lot of memory levels, and that's their very strong advantage. Given that you still use a more expanded product set at Google, yes, OpenAI has no. There are pictures, there's a search, there's YouTube and so on. Microsoft has a large infrastructure, too. In this zone, there is a high risk in this level of memory - strong personalization and increased sensitivity to data. Although it is important to understand again that whatever data you give from outside systems, it will still be impossible to process them in any way. And the fifth level of memory that exists, so that we can go through, it's totally clear, that's, uh, that's a background synthesis of memory, right? The individual process is updating, uh, the user report without a manual team. Yeah? And here, OpenAI Dreaming. There's also an Antropica, for example, that's Cloud Memory Summaries. These are examples of this, how it works. So this one has big risks, too. Yeah, it's critical to see the sight, it's editing, removal, and so on. Actually, we're gonna talk about what we released ChatGPT. Once again, Dreaming V3. Yeah, they call it that.
OpenAI describes a new background synthesis of memory that should better combat obsolescence, mistakes, scale for years. Again. An incredible, important thing in your practical application. Again, before I'm talking about Dreaming V3, I just want to show you some practical things, because otherwise you can hear so much technology in today's edition. Although I'm gonna try to do this now, we're gonna try to make very different editions on different topics. From very technical things or very applied, some cabs to the big visa. We've had a lot of visas now, and I don't know, until some time, I don't know, we'll tell you some stuff, investment, corporate business, and home use, yeah, systems. But the point, the point that is important today is that when you're using even serious models to make this serious model work on your project, she knows about your project very well. Not much, a little bit. And even if you're uploading all these data, she won't be able to analyze all these data. I gave you examples, even on this subject, which I often ask you to make, for example, fifty, thirty or 100,000 phone calls transcribating to make some conclusions. or develop a large base where thousands, thousands of pages of material are in. I always have a big story. I'm putting it hard as a chat-- ChatGPT task. We don't even talk about Codex here now because it's all different things. So he's totally on every transcribation and he's on the score, right? Or every man, in a company or corporation, took a job or took an evaluation, or every text used. I mean, I've used it all. And you always realize there's a limitation on the context window, it can't be completely conquered. How do we make it possible to keep the system from following new requests, and it's really working in this huge context window? Because when you use one chat, and you loaded a lot of calls, she made a job, you asked the second, third, fourth question, her context window, basically, it's dead, yeah, and it's starting to get started. - I'll get the data out of there. It's very visible when the systems are being developed. In systems such as Code Codex or in systems such as Codex, the context window is very rapidly being filled and it starts its automobile-- it automatically starts to purify the context window. Let's see what I have a garden, and I said, "Get me a flower." And if I don't tell you all the features of this garden around, I'm gonna get hurt by a lot of other flowers. And here's the task for the system to understand all the circumstances of your life. Knowing what you want to do tonight is not just what you need to do, what you need to do, but know the whole context of what your family, husband, wife, children, parents are doing. What you've done for years, what you're planning to do in the future, what kind of mood you've got, what's your health. A huge number of parameters.
When we go back to this Dreaming OpenAI,
This OpenAI drifter, you-- you say what it is, right? In fact, they have created a background process as a memory synthesis. The system analyses this accumulated context, updates memory, takes outdated, attempts to keep the most useful facts about the user and his tasks in mind. About all his tasks and, uh, what he's like is human. And to date, can you get that information down? Uh, practically unreal. I've tried to make aggregated reports, aggregated information. ChatGPT understands who I am. In one of the episodes, I did a description of ChatGPT, which it was thinking about me. But it's just some kind of ap-- uh, projection, abstraction. Just as there are a lot of viewers coming, who only looked me on this channel or saw me for the first time, or he doesn't know my life, he has a microcent on me. The idea, yes, who I am. That's the wrong conclusions. Okay, the wrong conclusions you're making about me, but when the system makes mistaken conclusions about you, they're about your health, it's a big problem.
What's this new drilling different from the old memory? The old model was more dependent on the obvious commands, yes, where you said, "Remember." And I remind you, ChatGPT was such a function. You said, "Remember," she kept it in a sequence of memories. The new model, it should automatically maintain the current state of memory and better take into account changes over time. And actually, uh, because OpenAI, he's going to create AGI. And this question of how to resolve their memory, of course, is fundamental because this is fundamental, including for the ta-- look, not to make you feel better. Honestly. They do, uh, AGI and make the future modern-- modern, smart, cool artificial intelligence. They want this model to be taught themselves, and if she is to study, she must work, of course, in a very correct context window, work properly. It's a question of the context window, memory. Who looks the most in memory right now, huh? Personal and long-term memory. It's, of course, OpenAI. Uh, Claude's pretty good at work, although I have questions for him personally. On a professional basis, it's considered that Microsoft Copilot works well, uh, Gemini Google. And that, yes, of course, is Anthropic.
Oh, there's a clear understanding that search-and-detection browser systems, now a lot of people have it. And one of the first such stories in searching was Perplexity. You know I'm not a super lover of this company, but I still need to mention it, yes. This is now becoming widespread in all major search engines. What's the risk of the current drifting? This erroneous or obsolete memory can affect all future responses. And therefore, the visibility of memory, the possibility of editing, removal from sources, some temporary regime without memory, some corporate policy. By the way, this is, in particular, a memoryless job. I run some sort of Internet tests on time, and I often need to run online tests when, uh, the system doesn't know anything about me. And of course, you have the opportunity to include this private session. You have a private session in every chat that's not going to last. These sessions don't use memory so you know. In ChatGPT, she doesn't use memory either. Some models used to use memory in chat rooms. There was time, Pro's model didn't use memory. What's the interesting story, how do you make the p- model remember you with a piece and the piece forget? Or, like, when can you come in and really figure out, what does she know about you? Because, uh, she could've formed a mind about you, and you're not really in agreement with that opinion. I have a lot of things like this, almost every day in different systems, which, uh, different systems, which makes one conclusion, and I think it's different. One of the last conclusions is when my team, one of my editors thought I was talking a lot about native soft right now. It's, uh, software that is exclusively built by artificial intelligence. And she thought that this AI native soft was the future, and all people would create such a soft. Uh, at least I saw it in some, uh, conclusions she made. And I said that I'm sure people won't make any software, especially software that makes a full-created intelligence.
It's gonna be the prerogative of some group of people, yeah. And it will be the prerogative of big and complex systems. But to figure out how it works, and to start using it, it's for itself now to feel the full power of artificial intelligence, it's a very important story, yes. And if you turn to the drifting, again, and look at the main strategic conclusion that is in this history, that memory turns the assistant of artificial intelligence into a long-term vision. The operating system of relationships with the user, yes. And who controls this layer, he gets a strong user-link to the product. And here we are, in our last release in the subcaste, which was now out a few days ago, talking to Elnar, raising questions with Tanya, which will be with models of what prices these models will be when, uh, um, companies like Anthropic like OpenAI will go to IPO, well, they'll be placed, public companies will be made public. And I think that, uh, the conclusion I have now reached in the view that this strategic conclusion that who controls the memory layer is that it is that it is that it is that the user is very strong in relation to the product. That's, uh, the future that's gonna own, uh, let's just say, all the money, and, uh, artificial intelligence. Of course, it's an amazing thing. I'm very happy they're making this development. What do you think about memory? Are you ready to share your memory with artificial intelligence? Do you see the security problem? Because, of course, when the system is constantly processing, indefinitely information about you, about your chat, it's important to understand that she's learning about you. And she makes a faintuning, that professional slang, specifically about you, about you personally, about your life. Because today, the fact that she has thousands of chat rooms is certainly not a faintening, because this is the building and the faintuning that it needs to be done all the time to keep the system alive. And all the time, it's just, it's just forever to communicate, analyze, and watch me react to chat, how I respond to answers, where I appreciate where I swear, where I even appreciate it. I'm telling you, where I write, that the system, uh, is very strange, bad, incorrect, and even using some bad lexico. I think they're going to the frontline of people. Sign on the canal, write. This is ToTheMoon channel. Bye, everybody.