Transcription
GPT54 has been released – an incredibly smart model, but smarter than Anthropic's OP. How did the drama between Anthropic and the Pentagon end, and why did the internet start canceling ChatGPT? Ababa is releasing a series of tiny models for running directly on mobile devices. Video Overviews in NotebookLM. What's so good about CL Motion? What are LUM Agents, and we'll tell you all the news from the world of artificial intelligence today. People, robots, hello. This is Prodsovet. My name is Uncle D. A weekly news release. Subscribe to the channel, leave comments, give likes to support our work. And let's go see the freshest events. Well, of course, let's start with GPT 5.4. This is a big update from OpenAI, as they claim, for real work. Remember, they have this benchmark GPT Wall, which we'll talk about now. But here are the main benchmarks and this same GDP Wall in first place. Here's a comparison with previous models. And it's worth saying right away that OpenAI released GPT 54 in ChatGPT, in the API, and in Codex. Previously, they separated all this into GPT 53. Codex GPT 53 instant also came out this week. Well, it's probably worth forgetting about this, because there is GPT 54, and it's a general model for everything, which is better in everything than its predecessors. And, by the way, they also released GPT 54 Pro immediately for subscribers of this two hundred dollar noble plan. Let's figure out what exactly has changed, what has been added. Of course, we'll compare all this with other models from Anthropic, for example. Firstly, now in ChatGPT, all users have this preambule plan feature, as they call it. You can direct the model's movement and its thought process directly during the generation of the answer. That is, this was previously only available in Codex, but now, while the model is thinking and responding, you add some more information, communicate something else if it comes to your mind, and it takes it into account. Of course, GDPva evaluates real work tasks that professionals perform in their activities. And here are the results. Moreover, GPT 54 Pro for some reason shows similar results here. Pay attention to this Industry Expert Baseline. The current generation of models, starting from GPT 52 Pro and GPT 54, outperforms humans in 70% of cases on these real tasks. And in 83% of cases, it shows results either at the level of experts in their field or higher. For example, on a specific benchmark called Investment Banking, GPT 54 shows 87.3%, while GPT 5.2 showed 68.4%. In almost 70% of cases, people prefer GPT 54 because of the stronger visual structure and aesthetics in these tables that the model creates, but we'll talk about that too. OpenAI is placing a significant emphasis on this benchmark OS World Verified, which looks at how the model can navigate different visual interfaces. Whether it's a browser, Excel spreadsheets, or just your computer. And they say that the model is already doing an amazing job with this. Accuracy is 75%, while a human's on this benchmark is around 72-73%. So, it would seem, a text model is already better at handling visual interfaces than a human. OpenAI is generally focusing on computer use in this release and calls it native computer use. That is, as far as I understand, the model now writes excellent Player code, which allows it to browse the browser, browse pages, browse visual interfaces, and in particular, click, move the mouse, and perform actions in these interfaces. We haven't seen anything like this before. In my opinion, this is the best indicator in the industry at the moment. Undoubtedly, they talk about the fact that the model has become more token-efficient, likely due to the fact that it thinks more efficiently overall, but we can definitely say it's due to this interesting feature called Tool Search. That is, now all tools, MCP, well, and a lot of different context that was previously given to the model at startup, is not given to new versions. It only sees snippets of the descriptions of these tools and then can find the right tool, read it entirely, and use it when necessary. And such a significant economy is achieved through this approach. Moreover, a million tokens of context window, finally. Or rather, up to a million tokens. You can definitely test all this in Codex. I don't know about ChatGPT, but it definitely works in Codex now. However, after 272,000,000 tokens, you will be charged twice as much for using this context window. API pricing has increased. Now it's $2.5 for input and $15 for output per million tokens. Recall that with GPT 52, we had 1.75 and $14 per million input output. Here's an interesting benchmark I found, or rather a visualization of how much stronger GPT 54 is than GPT 52. Look, this is the text arena and different domains within this text arena, how the model handles math, creative writing, and overall all text tasks. And it performs much better, moreover, in more domains. That is, you see, the graph is smoother, there are no sharp drops, as, for example, was the case with GPT 5.2. And I remind you that GPT 52 is a catastrophically smart model. It's very cool. Well, that's huge progress. In just a few months. And a few more benchmarks. I hope I haven't tired you too much with this. Then we'll go look at real examples. But I want to say the following. There is such an Artificial Analysis Intelligence Index. The model has jumped up nicely on it, reaching Gemy 3.1 Pro preview. And artificial analysis tells us that GPT4 is slightly more token-efficient than GPT 5.2, but less efficient than GPT 53 Codex. So, in principle, if you are saving tokens, then GPT 53 Codex is still better in this regard. And overall, to run the entire index, they spent $2,951 compared to $2,304 for GPT 52 XH and only $1,600 for GPT 53 Codex. Prices vary, token efficiency varies, results are as follows. In the end, GPT54 turns out to be, in principle, the most expensive, but also the smartest. I've gathered various comparisons and visualizations for you. We'll get to them after a short commercial break. Friends, I think you've noticed that the AI industry is changing very rapidly, so I believe it's important to have deep technical skills that will allow you to master complex tools and help you remain in demand in the job market. To dive headfirst into AI and not fall behind in your profession, we'll need Python. And of course, the AI itself. And today I'll tell you about the Python Developer + AI course from Skillbox. It's perfect for those who want to apply their knowledge immediately, create parsers, Telegram bots, and work with neural networks. This is what the market needs now, both for beginners and for those who want to delve deeper into AI. Why Python? It's one of the most popular and simple languages for starting in IT. It ranks first in several ratings. And it's precisely on Python that neural networks are trained. I can understand you, it used to seem that constantly learning something new, developing within your profession through courses was boring, videos, theory, millions of pages of text. But this course is different. From the first day, it's practice, you actually write code, run it, and it works. In 10 months, you'll complete several large projects, from parsers and bots to neural networks and models that are currently at the peak of popularity. And what really hooked me? Mentors, real people who check your assignments, give advice if something goes wrong, and provide feedback. An industry mentor is worth their weight in gold. Here, every action you take is important, and there's always someone to help, guide, or explain if you get stuck on something. On a positive note, the average salary of a Python developer is 225,000 rubles. Python developers are sought after by major companies like Yandex, Ozon, Sber, VK, and others. There are no daily classes on a schedule. Everyone chooses when and how much to study. Courses are constantly updated, and access to them is permanent. And an important bonus: you'll get a year of access to English language training. This will help both for your personal development and for employment in any company. And with the promo code "Productive Advice," you get a 55% discount for my subscribers. So, hurry up and follow the link in the description of this video or scan the QR code on the screen. Start your journey to your favorite work with Skillbox. Master a profession that inspires and get access to five additional courses as a gift to boost yourself. Let's go. Here's a comparison of dynamic SVGs. You can read the prompt. GPT 54 handled it noticeably better. It looped correctly. OPUS 4.6 somehow failed to achieve the same result with the same quality. Separately, both OpenAI and various companies that work with big data and tables, with all these delights. Here's the very same, in principle, GDP Wall. For example, we take an Excel table and parse it or translate all this data, aggregate it, and calculate something based on it. In general, experts say it works wonderfully. In this particular example, they are feeding datasets from different tables with over 50,000 cells in this Excel, and it correctly interprets, translates, and calculates the correct necessary results based on them. In this example, the computer use is being tested. GPT 54 Z is used, and all of this is put into a table. Previously, this was problematic. If you've been following the development of models for a long time, you'll remember that they clicked in the wrong places and didn't always do what they were asked. But look, the Excel table is ready. The model has read everything, seen everything, and added over 300 rows to the Excel table. This is an amazing result. Another interesting example. I tested ChatGPT for Excel and Claude for Excel on a very complex file. Macroeconomic data for 1,000 years of English history, spread across more than 100 tabs. Both tools performed well. ChatGPT, as a rule, only worked within the Excel application, creating formulas and manipulating data as a human would. Claude used Python and often inserted material only for display, making it difficult to track or edit. The prompt was: "Help me understand the relationship between the structure of UK agricultural production, GDP, and population, as well as the number of hours worked." In this example, they are asking to create a UI for Opus and GPT 5.4, a dashboard for stock analysis. You can read the prompt in detail here. And here, frankly, GPT54 performs worse. And in general, from what I see now, GPT 54 performs worse than Opus 4.6 in creating adequate, interesting, engaging, and creative text and cool, modern, attractive, engaging UIs. Opus 4.6 is more expensive. But here, on these tasks, it manages to work better. This is to ensure you don't rush headlong and switch to GPT if you were working with Claude. Because, in fact, yes, the model is amazing. In certain domains, it outperforms competitors, but not in everything. And by the way, if you're wondering, "How can I get ChatGPT to work with my Excel tables now?" Then there's ChatGPT for Excel, I assume, an extension that you install in Excel. You can go to the website, the link is attached here. Download it, test it. I think the time has come when we can try to give models not just simple text writing tasks. If you only do that, you're missing out on a lot. Indeed, I think it's time to try using computers, doing various automations, making them work with tables. The results on benchmarks are amazing, lots of positive feedback from companies and experts. Try it yourself, and be sure to tell us in the comments how it goes. And if you are a Codex user, then, firstly, there's a fast mode, the model can respond to you one and a half times faster, but it consumes twice as many limits. Again, don't forget about that. And if you are not a Codex user, because you own a wonderful Windows machine, then it's no longer a problem, because OpenAI is finally releasing Codex for Windows. You can download and use this wonderful application. It is indeed wonderful. I use it almost constantly. In conclusion, I think the release turned out to be very dense. I tested the model for a very short time, and I didn't notice any catastrophically huge difference compared to, for example, GPT 53 Codex. But it's interesting to hear your opinion if you've also managed to get your hands on it. The Pentagon confirms the strike on Anthropic. Let's talk about these unfortunate ones, because there's something to say. Firstly, it's official. Anthropic has received the status of Supply Chain Risk. And this is the first time such a status has been issued to an American company. Previously, only foreign companies, adversaries of the United States, could receive such a status, which imposes a number of restrictions on the company. So now Anthropic cannot work with contractors who cooperate with the US Department of Defense. Anthropic wants to challenge this and will go to court. Amadeus called this decision legally weak and too harsh, but it seems from journalists' statements that it won't be so easy for the company to get rid of it, because courts rarely argue with the Pentagon on national security issues, although it's impossible. And I remind you that OpenAI eventually signed a deal with the Department of Defense, and this caused a lot of controversy and different opinions from companies, employees, and clients, which we'll talk about now. And Amadeus expressed himself in rather harsh terms throughout the week about OpenAI's Altman. Nevertheless, Claude is growing on this scandal. Due to Anthropic's conflict with the Pentagon, consumer demand has accelerated. Why did this happen? Well, because people love principled behavior from companies, apparently. And here's what App Figures tells us. As of March 2nd, Claude had 149,000 daily installations compared to 124,000 for ChatGPT. A wave of new users is now heading towards Claude. And Daily Active Users in Claude is 11.3 million as of March 2nd, which is +183% since the beginning of the year. According to web traffic, Claude is +43% month-over-month and +297% year-over-year. ChatGPT, by the way, shows -6.2% month-over-month. And yes, indeed, I've encountered several articles and a number of posts about people switching to Claude, taking all their data from ChatGPT, and there are even tutorials on how to completely delete their subscription and account with OpenAI, because they are such scoundrels, hypocrites. First, they supported Anthropic, and then they themselves, under seemingly the same conditions, signed a contract with the Ministry of Defense, although they publicly condemned it initially. Well, in general, Claude offers you a fairly simple thing. Just go to clod.com and importmemory. Log in to Claude. It simply gives you a prompt that you need to insert into any of your providers where you communicated previously. And this prompt simply exports the entire message history, or rather, the entire memory history that this provider has, for example, OpenAI. And then you transfer all of this to Claude and say, "Friend, remember." This is a simple thing, but it has also gone viral for some reason, and people seem to be using it quite actively. Claude itself has several nice updates. Firstly, they introduced the Clod Marketplace. And this is the same as ChatGPT Apps. Essentially, companies will be able to offer their services more natively within Claude through these connectors. They also added scheduled tasks. That is, a task scheduler. It only works if your computer is on. I remind you, there are also some hyped headlines that OpenAI is done with Anthropic, you no longer need it, use the desktop Claude because it has these jobs, but in reality, not really. OpenAI works remotely, always works on a VPS if you've done everything correctly. But this thing only works if your computer is on. Nevertheless, you can indeed ask Claude to proactively perform various tasks on a schedule, check a pull request, go look at something, gather a news digest for you, and whatever else your imagination takes you, you can schedule it. And ClodCode has learned to listen. Rejoice. Now there's the Voice command, like this. Enter the slash command, and Claude can, ClodCode can listen to you. Well, uh, there have been various voice assistants and even local solutions for a long time. It's also available in Codex applications, but now Anthropic has managed to release it natively. You can also prompt by voice, which is very convenient, I recommend it. Google is rolling out Google Workspace Cai, and this is an amazing thing, friends. Finally, you can now easily, calmly, without all these dances with tambourines and third-party MCP servers, automate your work and connect your AI agents, which, I hope, you are using. It's very convenient, uh, with external services, with Google services. We have Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin. Everything is there, there are different commands. You can enter them manually, or you can teach your AI agent to enter them into the terminal, and you and your AI agent will always have access to your Google data. Wonderful, excellent, convenient. Google Workspace Cine is on GitHub, you can download and install it. Something is going wrong in the Quen team. The key tech lead, Jun Yanlin, has left. This happened just a day after the release of the new N 3.5 Small lineup. He announced this publicly but gave no reasons and explained nothing about why. N 3.5 Small actually caused a small sensation on the internet because they released four models: 0.8, 2, 4, and 9 billion parameters. The first three can be safely run on your mobile device. In general, they are designed for this. The 9 billion parameter model is comparable in benchmarks and capabilities, for example, to GPT 4O and open models from OpenAI with 120 billion parameters. And Elon Musk said it's impressive Intelligence Density. Well, that is, a model that is smart, dense with mind and intellect. Such an expression. It's not entirely clear what this is related to. There are no official explanations for the reasons for his departure, but I really hope the team doesn't fall apart, because Quen, honestly, in my opinion, makes some of the best local open-source models. At least, what I've launched, tried, and compared, I was impressed by these models. Well, and that's the end of my LM agent coding, vibe-coding news segment. Let's find out what's been happening in the world of creative neural networks this week. People, robots, hello. As always, it's Igor'yan. And as always, we're discussing creative neural network news from the past week. One of the most interesting news of the past week is the full release of all CLK 3.0 models, as well as the new model Clank 3 Motion Control. CL 3.0 and CL 3.0 Omni were essentially already released, which I talked about in previous videos, but as far as I understand, they weren't available to everyone, although I managed to test them. But Cllink 3.0 Motion Control is a completely new model that didn't exist before. I've naturally already made a video about it. I also made a post in my Telegram channel "Creative Council" where I attached cool examples of how the model works. Essentially, it's simple. You upload a video of how the image should move, how the object should move. Clink understands this movement, understands the structure of the human body in the frame, and then transfers it to the image you upload, thereby making the person in the image move. There are quite a few nuances, but it works unusually well, meaning I haven't seen better motion transfer yet. The facial expression transfer is also very good, of course, it's still not perfect. Plus, there are artifacts, but I haven't seen such quality anywhere else, except, perhaps, for Sidens 2.0. Although even there, it seems the model doesn't transfer facial expressions as well. The model immediately took top spots in the text-to-video segment and leading, but not first places in image-to-video, where Pixs 5.6 and Imag'n Video, which I also recently talked about, are leading. Among the big advantages of Motion Control 3.0, we can highlight that you can upload image models to it, i.e., these clink elements, which allow you to create stable characters. And you can show a person from all sides, indicate how their teeth look, etc. By uploading several different images, you create a maximally stable character. And if in the first frame, for example, the character stands with their mouth closed, you upload a photo of how their teeth and tongue look. And after that, when they open their mouth, the teeth will be the same every time, which is a big pain, because, well, you close your mouth, open it, and there's already a completely different bite, completely different teeth. This is, well, a big headache. I think many of you have encountered it. We'll talk more about clink and other models at our new intensive on creating videos using neural networks and content maker 2026. Two weeks of live broadcasts, calls, I'll be leading most of it, but there will be invited speakers besides me, cool specialists from different niches. We'll share experiences, talk about interesting pipelines, how to create content, discuss some pitfalls. There will be a lot of theory, good theory, and even more practice. There will be homework, we'll share cool prompts, prompt templates. And in 14 days of such intensive work, we'll try to bring you to a result where, firstly, you'll gain new knowledge, secondly, you'll make new acquaintances, and thirdly, ideally, during this time, you'll create your AI product, a short film, an AI clip, or an advertisement for your product. In short, any video content you want. This is a great reason to start, for example, running social media, because as you can see, video models are producing very good results. And 2026 is indeed the best time to start exploring the field, to master the tools, to gain a theoretical foundation so that you can not only generate a nice little video but also understand how to combine it all into a single structure, how to get from the idea stage to the stage where you have a finished cool video that you're not ashamed to show anyone and that other people will click on. In short, follow the link in the description. We start on March 9th. Another very cool release is LTX2.3. A new model, available in all LTX services. Available in open source and in API. It generates video with audio. Everything is much cooler and cleaner than before. Although they already had the best open-source model in the industry before. What's cool is that the model was trained, among other things, on a huge amount of vertical video. As we know, a huge number of models, including previous LTX models, couldn't generate vertical video. Or rather, they did, but they did it by simply cropping the generated horizontal video. In addition, they've added generation between two frames, and also significantly improved audio generation. The prompt is also understood better, and so on. Everything else, I think you've heard 150 times. The model has simply become much better. To run it on your own hardware, you'll ideally need more than 32GB of VRAM, which is, well, quite a bit. On the other hand, it's not a lot either. This is at least a significant portion of those who watch me, I think, will be able to handle it. I personally won't be able to. Along with the LTX model, they also released LTX Desktop, a desktop application for your computers, which was already partially released, but now it's such an official release for everyone. Everyone can use their application. LTX Desktop is a very convenient editor, i.e., a video editor that runs local models and allows you to work with videos, audio tracks, combine them, and edit them a bit in a pleasant interface, if you don't have very complex work. Although LTX has a wonderful workflow optimized for creating really cool, more or less long works from, well, a short idea. Now it's like it's transitioning to the stage where something really good is starting to emerge. If your computer isn't powerful enough, you can use the API instead of running models locally. And not just the LTX API, but the API of any video generation company or provider. So, the only reason to use LTX is if for some reason it's inconvenient for you to work in Coffee UI and you want to run local models in the simplest possible interface. A few days before that, LTX released dubbing and captions, i.e., the ability to create dubbing, create descriptions for your videos, which is, well, cool. In particular, this will work well with vertical video generation, because it will be much easier to automatically create content for all these social platforms like Shorts, TikToks, VK clips, and other forbidden vertical video formats with minimal effort. Next, let's talk a bit about Zopy or Jopy, I don't know how to pronounce it correctly. I assume it's Zopy. As I predicted, more and more automatic video creation systems are appearing, and they are starting to work better. And Zopy is one of them. The scheme is very simple: one idea + Zopy = cinematic blockbuster. Of course, this is a slight exaggeration, but on the other hand, where's the exaggeration, it's unclear. It will probably be cinematic, but it's unlikely to be a blockbuster, meaning it's unlikely to be shown in cinemas. But on the other hand, to each their own, we'll see. In short, Zopy will help you in a conversational mode, using any of the top models, like CLK 3, Vidu Q3 soon, and Sedence 2.0 when it's released in API. You can use it through OpenCL or other bots, and videos will be generated for you 24/7. I, of course, quite liked the interface. Everything is presented well visually, it's clear where and what has meaning at first glance. It seems like a natively good tool. If anyone has tested it, please write. It might be worth paying attention to. In addition to Zopy, Luma is also rolling out its Luma Agent. Or rather, not an agent, but agents that can also create videos using an agent system. Unlike Zopy, which is currently only in beta and requires signing up for a waitlist to access it, Luma can already be tried on their website. It creates storyboards, re-dresses everything, creates cool advertising videos automatically. And if Zopy focuses on cinematic quality, on creating that blockbuster feel, or just cinema, then Luma, as far as I understand, focuses more on working with products, with clothing, with branding, and there's an opportunity to work with such a cool canvas. Luma agents are also cool, I haven't tested them yet, but maybe it's worth it. A slightly more minor update is annotations in Reva, i.e., the ability to draw and label areas. In general, you could already label areas there, but now you can also draw on them. You upload an image, label it, and everything is positioned there. This, in my opinion, works much better than, for example, in Cre, where objects are inserted into the image very crudely, very inaccurately. Here it will work much more pleasantly. Something like this could have been done before. It can also be done in Nanobanana, and, well, where can't it be done? In general, Reva is great anyway, because they still have a very convenient interface and the ability to upload, label, and not struggle with having to go into Photoshop or Paint to draw something on an image, but to do it directly within. Although, in general, it's already possible to do this in Google Flow and in Hixle. So I don't even know what else to say here. A more interesting thing is the Voice Mode in the Cre application on iPad, which allows you not only to draw in real-time and add something. And here a person is drawing with a stylus, as you saw just now. You can also speak at the same time. You press the microphone button, and it says, "Make a sunset," and a sunset appears. It will say something else now. "Let's add three trees," and trees are added. "These willows, yes, they've been added." It says, "Make everything out of chrome, make everything out of glass." Now it will be out of glass. Here it is, out of glass. In short, if you work in Crepad, this is very cool. It's very cool to let children play with it. I think everyone will be very interested in giving commands like this by voice, adding to it, correcting the prompt somehow. Although, on the other hand, I'm surprised now that you can dictate prompts by voice. This, of course, could have been done before. Simply instead of typing the prompt on the keyboard, press the microphone button on the keyboard, and dictate it all. But you understand, here, of course, it's more convenient, more native, and as you know, I always advocate for nativeness. Another interesting thing is relight from Adventure AI. We already had cool LoRAs for Quen that also allow you to change the lighting of an image. Well, in general, it's possible in most editors like Nanobanana. The difference is that in Quen there was also a similar slider that allows you to indicate the light source point. That is, you rotate the 3D space, and the light source moves somewhere. Here it works a bit better, meaning I don't know what model they use, but it seems to be even better trained. Plus, this slider is now much more sensitive, much more degrees of where the light can be located relative to the object in the image. In Quen, there were quite a few positions for this light source. Here it rotates very, very smoothly. You can use it on adventuregen.ai. Look how smoothly it rotates. There are actually a lot of possible light source positions. Now, let's talk a bit about US legislation. It's so interesting. In fact, the news is more or less important, because, well, all tech companies are guided by US legislation. It's the largest market. And so, the US Supreme Court rejected an appeal in a case that had been dragging on since 2019. A guy created this image in some model a long time ago, as you understand, in 2019. He tried to copyright it, but he really wanted to credit the model itself as the author. And that's exactly what he was denied. That is, the author of a work of art or anything protected by copyright can only be a human. And this is quite relevant with the current trends in agent systems that create everything automatically. If they create something for you without your intervention at all, then it probably cannot be protected by copyright. On the other hand, of course, you can credit yourself as the creator, and there will be many nuances, of course. I don't think anyone in the world knows yet how this will work, say, in 5 years, because the field is developing now, there will be more precedents in the US, which has a precedent-based legal system. And, in general, some cases will accumulate, legislation will be updated, and we'll see how it works in practice. And NotebookLM is releasing video Overviews. As they say, this is a whole new stage of NotebookLM's evolution. It's currently only available to ultra-users in English, and it allows you to create super cool videos based on your data. That is, you upload some information, ask NotebookLM to create a presentation, and it also creates a cool video for this presentation that you can then put in the presentation or use instead of the presentation. In short, I'm really looking forward to when the function is rolled out to more than just ultra-users. This is really very interesting to test, because, well, based on the examples they show, it looks quite cool. Essentially, this is also an agent system for creating videos, only here it's some kind of video explanations, video guides, video tutorials. And it's based not on some short ideas, not on some plots, but, well, on real data, on some of your research that you do in NotebookLM. And the next news is from the world of 3D, namely the new Wonder 3D model from Autodesk. The generation quality seems quite good. Pleasant textures, as they write in the comments, not too many polygons either. Are there any advantages over the leaders in the niche? I haven't found any, actually. That is, if you have some cool Chinese models that you use, then continue to use them. But here's another alternative that's integrated into the Autodesk system. Upon registration, you'll get 300 credits, which is enough to create about 60 models, which is cool, but it seems like you can generate 3D models for free in Hu'an as well. Text-to-generation and image-to-generation modes are available. And that seems to be it. And if this release seemed useful to you, we'll be very grateful for your likes, subscriptions, and comments. On March 9th, our intensive on creating and video Content Maker 2026 starts. 2 weeks of theory, practice, live broadcasts, calls, communication in the chat, invited experts, and so on. And that's the end of the news from the past week. This was Productive Council. I'm Igor'yan, and Uncle D was in the first half of the video. See you on March 9th on the stream. Goodbye everyone.