You're in the right place if you use ChatgBT, Claude, Claude Co-work, or OpenClaw, but still don't know which agent you should use for business and how. Because in this video, we're going to do a deep dive into using AI agents effectively to get work done. Specifically, we're talking about ChatGBT work, OpenAI's new agent tool, which is like Claude Co-work, but in my opinion, much better and easier to use on mobile, web, desktop, or any other platform. In this video, we're going to be breaking down the 14 capabilities of chatbt work, each with realworld business use cases, so that by the end, you know exactly what it can do and where it fits in your workflow. So, this right here, as you probably know, is chat GPT. You'll notice here at the top there's this little toggle. If I click this button, now we're using chat GPT work on the web. On chat GPT work, I can create presentations like I was a Mckenzie consultant. On Chat GPT work, I can also create spreadsheets that look really professional and it can also include charts. I can also use GPT work for video editing and motion graphics. On chat GPT work, I can create and preview websites and I can even edit them. Make it dark mode. ChatGpt work is also not just on the web. This right here is ChatGpt, the iOS app that billions of people have used. And if I press this little button right here, you can see that it has a work toggle. This right here is chat GBT work. Every single thing that you can do on chat GPT on the web, you can also do on the iOS app. In order to fully understand how chat GBT work works, I think it's important to understand the timeline of how chat GBT work came to be. So back in 2022, we all remember chat GBT. It was this super viral moment. It was the first time anyone made an AI chatbot accessible to the world. And chat GBT grew to almost a billion weekly active users over the course of four years. And then chat GPT or OpenAI released a new product called Codeex. And so Codeex is a tool that was a direct reaction to Claude code and Claude co-work. Codeex was basically the first super app where you could open up the Codeex desktop app and build anything whether it's an app, a presentation, you could do video editing, you could do basically anything directly inside Codeex. And because Codex could basically do all knowledge work really well, not just coding tasks. And because the name Codeex felt like a developer tool, OpenAI decided to merge the Chat GBT app and the Codeex app into Chat GPT. And that is when GPT work was born. CHIGBT work came out about 6 months after codecs and you can think of GBT work as a more accessible codeex except unlike codec this is available in the web it's also available in the desktop app and it's also available in the iOS app and so in this video I'm talking about the 14 most important capabilities inside GBT work. Let's dive into capability number one. The first capability of GPT work is creating polished reports and documents. As you can see here, I created this document inside GPT work and it's viewable within the chat GBT app. So, if I go to chat GPT, open up a new chat. Normally, you'll start off in chat mode. If you go to chat GBT work mode, I could say something like this. Hello there, chat GBT work. I need you to make a McKenzie style presentation on the future predictions for the rest of the year in AI. Do in-depth research and create a um McKenzie style report with a ton of graphics and timelines that look professional and amazing. Create this presentation and I can fire this prompt off. And since this is chat GPT work, chat GPT work uses a cloud computer to create the files for this document. And then we're going to be able to open it up in the side panel and it will look just like this. Okay, so it's finally done. It created a 19 slide Mckenzie style AI Outlook presentation. Before we look at it, I just want to show you what it did. You can see here it actually searched 87 websites. It'll search the internet for as long as you need. It also read presentation and library instructions, style guide and API documentation. And it just went through 13 minutes and 33 seconds of work. And then it created this presentation right here. And it's very graphical and it talks about the different predictions. And so this right here is how I edit any document that chat GBT work creates. I just open it up here in the side panel. I come down to this chat area and I just double click. Okay. So the following things are going to be the things that I want to change about this on slide number one. Uh the rest of 2026 I don't understand what's going on with the graphic. Please spend some time make this better. Be creative. On slide two, I don't like these pills over here. They're kind of weird looking. Just figure out how to organize this in a way that looks more professional. And number five, I don't like this yellow. Can you please make this a different color? The butt autonomy will reign scoped and reversible. Make this a little bit bigger and a different color. Number eight, this block is ugly. Change the color. That has 703 billion in it. Please change the color of that. For prediction number five or slide number nine, I like this for the entire thing. change the font to something a little bit better. Um, and oh, I really like the uh slide number 17. It's really good. And that's solid. Please make all of these changes. And so, you can honestly think of this like a coworker where I can just blabber everything to it. And since this agent has access to a computer where it can create the files necessary to create this presentation, it can take all the time it needs. It'll edit all of the things that I said to edit. And it'll get back to us in 5 to 10 minutes. And there you go. 9 and 1/2 minutes later. It made a bunch of changes. As you can see, now it has a different graphic on the front, which I like a lot more. It got rid of those pills and it basically made all of the different changes that I was hoping it made. And so out of the box it can create these PPPX presentations. This is one of the type of documents that it can create even if you don't connect your GPT work to any external tools. Some other types of presentations it can create is a docx which just looks like a normal document. It can create a PDF and it can also create a HTML that you can open up in line. And so it can create these little artifacts. And remember, every type of document that you create on GBT work on the web can be opened on the iOS app as well. As you can see here, we have two very basic documents I created for the sake of this example on the web app. I can open them on mobile. And one little hack is when I'm viewing this, I can actually make edits with my voice. Let me show you. So, if you click this button right here, I can click this. Okay, I'm looking at the document now. Can you please expand on this and just make the table a little bit larger and with some more what it can create? I want you to expand on that. And please get rid of the GPT work quick reference little subtitle and just make the title bigger and add a little bit more visual spacing. And on this I'm referring to the PDF document. Please send me a new PDF doc. So while you're viewing it, you can speak and then GBT work will send me a new document in just a second. And there you go. After 2 minutes and 13 seconds, it created the new version of the report. I can open this up. And let's see if it made the changes. It did. Notice here there's no subtitle. I asked for it to get rid of it and then it expanded on everything else. And now it's two pages instead of one. And again, I can zoom in and I can go through as many iterations as I want. I could turn on the voice, open it up. Please make it even longer. Four pages long, please. After I exit, I can just fire off that prompt. And I could do that over and over and over again. And so that's just what GBT work can create out of the box. It can create PowerPoints, documents. It can also create spreadsheets, PDFs, and HTML websites. However, the entire world opens up to you once you understand the power of plugins. Which brings me to the next capability. So the second capability of GPT work is using plugins to add more tools and capabilities. You'll notice here when I'm on chat GPT work, there's a list of suggestions. And these suggestions are all based on my calendar, my email, my Slack, and all of the other plugins that I've added. Chad GBD plugins are like codeex plugins. And you can come up here in the top part and you can just hit plugins. And here you can select from basically any software that you might use and you can add it as a GBT work plugin. Plugins are different from skills. Plugins are official connections with other apps and tools that you may use and other companies can create plugins for you to very easily connect your codecs, chat GPT or chat GPT work. And you can see your installed plugins by clicking this right here. And these are the plugins that I use on GBT work. I use ClickUp to communicate with my management team. I use Convex for anything that involves a database. I use documents. This one is built into GBT work. FAL gives me access to all of the popular AI models, both image models and video models. GitHub, incredibly important if you do any coding work. Gmail is one of the most important ones, if not the most important one. I can fully control my Gmail from GBT work calendar, Google Drive, Hostinger. I can do hyperframes by Heyen to create really high quality motion graphics and also notion. Let me show you an example. If I go to chatgpd.com and I switch to work, I can come here and I can say something like this. Hey, has Emily sent me any emails recently? Can you please tell me what she said? and chatpt work will immediately access all of my emails and anytime you use a plugin like Gmail. It comes with skills that are bundled within that. So it has a reading Gmail skill within the plugin and now it's searching recent emails from Emily. And here we go. It said she asked you to add your input on the below agent native podcast concept. and I'm going to say, "Draft an email to Emily and I want you to let her know that I will get it done by in in the next four or five hours. I I will update it in the notion document and I will send it to her. Can you please create a one-time automation in 6 hours and send her that link if I have not done it?" We're going to get to automations more in depth later. But I have basically said that I'll get it done in the next six hours and I'll send her an email. But if I don't send an email, GPT work is going to create a one-time automation to send her the email with the updated link. Now you can see it's drafting notion update confirmation. And now it's creating a scheduled task. And now I can say, okay, send the email. And there you go. Work for 17 seconds. and it said sent to Emily. And if we go to our Gmail, you can see here that I did indeed send it. Hey Emily, yep, I'll get this done with the next four or five hours. I'll I'll update everything in notion document and send you the link once it's ready. And now I could say what was that link again uh to the notion doc. And here it just responded with this link and I can very easily open this up. Open link. And here we it opened up notion. And so now we have the notion doc and I now need to prepare for the belaval episode on the agent native podcast. And now I could just ask GPT work to control notion because I have the notion plugin. I can hit at notion. We can use the plugin. Please can we um add a few ideas here? He has a uh recently just started using codeex and he loves it. I want to ask him what is his workflow first of all. Second of all, um what are is his favorite features and then ask and then third would be how does he like the inapp browser on the desktop app? Keep the recording information actually. And so I can fire this prompt in and GBT work can fully control notion. And there you go. It just made the questions and it's that easy. I can fully control notion from my web app. And remember, anytime you use GPT work on the web, that means you can also do the same thing on the mobile app. And I'm telling you, in terms of my GPT work usage, I use it 80% from my phone. When I'm on my desktop, I'll actually use codeex a little bit more than I'll use GPT work. But the best part about GPT work is being able to start something on your computer and I can immediately just take it over from my phone. The chat conversation that we were just having is Emily's recent email. So, if I click on this, I can see that I am making changes to my notion. Every time it makes changes to my notion, it gives me a link to the notion page. So, I can very easily just open this up inside notion and it will open up the document here and I can just go to chatbt again. I can say, "Hey, actually, I need you to add two more. I just realized there's two more things that I want to ask him. I want to ask him if he uses the built-in image generation." And then the final thing I want to ask him is how is he using it for his 3D workflows? And I can just send this in right here and I could just wait for it to finish or I could go into notion and I could walk anywhere. I could walk around the city. I'll just wait for that to come in. There it is. Do you use Codex for built-in image generation? If so, how does that fit in your workflow? And how are you using Codex for 3D workflows? That is the power of plugins. Okay, I really need to review this uh on Tuesday before the episode. So on Tuesday from 11 a.m. to noon, can you add a calendar event for me to review this notion document and prepare for it? Please invite Emily uh to this event. We're going to review it together. And in the last one, I used the at@mention for notion. You don't need to, right? If you add a plugin, you do not need to atmention it every time. it is smart enough to use the correct plugin. And there you go. You can see this little calendar icon right here. So, it's searching the calendar. So, it's actively using the plugin. And there you go. It added the event. It added Emily. And here it says open the calendar event. If I click on this right here, right? If I just click on this, open the link, it opens up the calendar event directly, and it just works with all of my tools. And it's really easy to add plugins. It's just like one tap. And you can add a plugin on any platform. If you want to do it on mobile, you just press this button right here. And you're going to press the little settings icon down here. And you can find plugins. And here you can see all of my existing plugins. At the bottom, you can browse plugins and we could find, let's say we wanted to add Canva. We can do that by pressing one button. Continue to Canva. Continue. And now all we have to do is sign into our account. And then we're going to hit allow. And boom, authorization successful. And you should see a little verification on the app. And there you go. Now you can see that you have the installed plugins. If you click on this, you can see that Canva is now in the installed section and we have it installed. One thing you should know is like I said earlier, each plug-in comes with bundled skills. And so if you scroll down, you can see what skills are included in the plug-in like Canva branded presentation, Canva Bulk create, Canva translate design, and some some of these plugins have like 50 skills, some of them only have one. and many in between that spectrum. And if you ever want to test it out, you can hit try in chat. You click this button and it'll immediately allow you to try it. Okay, now it's time to move to a more underrated capability of GPT work. Okay, so the third capability of chat GPT work is something called blocks. Let me show you through an example. Hey, I need you to take a look at this episode uh this uh podcast episode that I'm doing. And what I need you to do is I need you to write an intro. Can you please write it in an editable text block? Okay. So, this is going to prepare the intro that it's writing in a text block that I can edit. And you can see here it's using the notion doc to understand what the episode's about. And here it gives me a response in this little block. And I love it. It allows me to just focus directly on this section right here. I can full screen it. If I just want to focus on it, I can very easily describe edits to this document. And if I want to edit it directly, I can just edit it directly, right? I can type into it, hello, and I can kind of ask AI to make changes or I can make changes directly. And if we go to this same exact chat on iOS, our our mobile app, you can see that it renders on the phone as well. I can very easily edit this and I can make edits right here. And it even gives me little suggestions like start with a stronger hook, tighten and remove repetition or again I can just add things directly to it. And I could just say start with a stronger hook. And look, it creates this little cool animation and it shows me the exact changes that it made here. And what's cool is I can go back to the previous one and I can go forward to after the changes, which I really like. And then we can continue to make more changes tighten. And there you go. And so now there's three changes or three versions of this intro. And we can go back and we can go forward. But blocks don't end there. In fact, it goes way deeper for these different types of blocks that you can create. Hey, I want you to actually help me brainstorm all the things that I could talk about. Please create a mermaid diagram and I want this to be uh a mind map of the the topics that we could cover. And then I also want you to make a flowchart of the conversation and this should give me some ideas for what we should discuss. And so, not only can it create these text blocks, it can also create diagram blocks. If you know anything about mermaid diagrams, it's this cool formatting where it converts code into like these really cool renders and it can render them directly in line. You'll see in just a second. And you can see here it's creating this code in mermaid formatting. And watch, it's going to render in just a second. Boom. Look at this. So, I can full screen this chart. And now we can kind of see that does delegation create more management and it has all of the related ideas. We can see creator economics. Where is the actual time saving? Does output create more noise? These are all good questions. And these charts right here are built directly into GPT work. And they also work with codecs as well if you want to use the desktop app. But again, it doesn't stop at mind maps. You can also create this other type of chart. This is a flowchart and again you can full screen it and here is this flowchart and that is pretty cool. Okay. So based on the notion doc I gave you about GBT work. I want you to take me through the flow of the average user go in detail. I want a sequence diagram block which describes how the user would use GPT work and what's going on under the hood like with plugins skills and all of the different things. Can you please make a detailed sequence diagram in a block? This one's pretty cool. And again, just like the other types of mermaid diagrams, it's just going to generate code. Once it's done generating the code, it will render it in a very clean and very easy to understand sequence diagram. Here we go. It's kind of a one-pager. It's a lot easier to read. And here we go. So, the user describes a a a desired outcome. GBT work will plan and work and identify the requirements and then it will load files in memory into relevant skills into the context. It returns the preferences and task instructions and the GPT work model will then select the plugins or browser on a local computer. By the way, I think this is incredibly useful. I I use this very often. And then the tools will return information from connected sources, right? all of these different tools which includes plugins and skills and external connections. It returns the information back that researches and analyzes and creates or operates software. And this is just a really cool way to kind of see how it interacts with GBT work, the context, the tools, and then the final output. And you can see here it assembles the document, deck, site or app. And if you want to know what types of blocks and charts that it can create, just ask GBT work. Hey, we created a mind map, a flowchart, and a sequence diagram. List all of the other types of blocks you can create like this that's built into GBT work. And look at all of these. Here are all of the blocks that it can create. So, it can create all of these different types of diagrams. Here are the writing blocks. It can also create an email draft block, which I'll get to a little bit later. It'll it can create data and technical blocks, media blocks, and file and application blocks. all directly inside GBT work and all of these work on the phone as well. And if you look at this last type of block it can create and you look at these four, this brings me to the next capability within GPT work. The fourth capability of GPT work is building and hosting complete websites and applications. So everything that we just learned in this chat, I want you to take everything that we just talked about in terms of the blocks that it can create and I want you to create a public site that I can share with my audience and I want you to use what is called and so I want it to use the sites feature. So this is build and deploy sites directly inside chat GBT and these are like little mini apps that you can share with friends, your team or anyone else. And so I can create one of these sites. Please, can you um Yeah, I want you to create a site that shows the different types of charts that I can create and take the ones that we did and try to uh put them in the site and render them correctly and show what they all look like. I want this to be a GPT site that I can share with my friends. And so I can fire this off. Notice here we're using this on chatgpt.com in the web. This means I can do a very similar prompt directly from my phone as well. And then there you go. After 6 minutes and 10 seconds, we have a new site and I can click on it and I can open up this site and you can see here that it is deployed on a chatgbt. site and take a look at this. It says what can chat GBT work render in a conversation. And if I scroll down here we go we have mindmap flowchart sequence diagrams timeline user journey. Notice here this is a chatgbts site. This isn't a link on the actual internet. It's more of an internal tool and it's hosted by chatgbt. And so what I want to do is I want to host this on the actual internet and then eventually give it my own domain. So I want to say please put this on the public internet and I want you to use the following plugin. And I'm going to type at Verscell. And we're going to use the Verscel plugin to put it on the internet. And instead of it being a GBT site, I want it on the internet. All right. So, it said it's live on Versel. If I open this link up right here, you can see that it has a Verscell.app link. And now that you have this project on Versell, it's really easy to add a custom domain. And you can use namecheep which is another plugin to buy your domain and you can just ask it to update the domain on versel and you can create an app put it on the internet and give it your own domain. Okay. So the fifth capability of GBT work involves editing and generating images with the built-in GPT image 2 model. So, I just uploaded directly to GBT work this thumbnail concept that we have. And this is incredibly easy to edit. And so, I'm just going to say, please change the text to GPT work is insane. Then change the logo to this. And now what I'm going to do is I'm just going to paste in the chat GBT icon. And I'm just going to say only change those things. Keep the rest the same. Make three variations, please. So now GPT work has access to the best image model. And it's really good at image editing. It can just edit everything for you. But I'm going to show you why this is so cool in just a second. And look, here it is. We can see that it's working. And this image has been generated. And you can see here there's this little edit icon. So even while it's working, I can click this edit button and I can say reduce the glow around the icon. The guy on the left should be wearing a brown shirt. The guy on the right blue. And now I can fire off this prompt and it automatically just allows me to inline edit that image. You'll notice here this looks a lot awfully like the previous one, but you'll notice here there's another version, right? We have brown on the left and blue on the right and it reduced the glow around the icon. I can even get more granular with it. So if I click on the image, I can also use this comment feature and I can comment directly on this make hair more luscious and I can add that as a comment. You can see that eyes blue, better hairline, slightly more beard, and then I can more white teeth background slightly more of a dark gray charcoal gradient, not just fully black. And I can just fire these comments and I can hit send. It will add all of those edits on the exact location that we selected. Okay, so it's done. So if we click on this, we can go up to the previous version. So look, I asked for my eyes to blue, hair more luscious. Let's go to the next one. Boom. I asked for the teeth to be whiter. Boom. I asked for slightly more beard. Boom. And I asked for the gradient to go from black to this gray color. So, this is what it was before. It's subtle, but now it's this like gray gradient. Please uh take those four variations and send them to Emily on email. I want to know which one she prefers. Which one do you think is best for the video? Send that to her right now. And remember, we're using the Gmail plugin. And I can just very quickly send over the images that we could potentially add to my coworker Emily. And now we're kind of using all of the different tools together. And there you go. It's sent to Emily with all four variations attached. Let's take a look. If we click on this email here, you can see I made these four thumbnail variations for GBT Workvide. Can you take a look and let me know which one you prefer? And we have one, two, three, and four. So it's very easy to just take the images that you create and send them via email. All right. So the sixth most important capability of chat GBT work is branching a chat. So earlier we were creating a site and I wanted a website that shows all the different types of charts that you can create directly inside chat gbt work. Well now I want to continue working on this site actually but I also want to turn this into a slide deck and I want to have both of these chats going at the same time. So in order to do that I can type slashbranch chat and this will automatically take everything that we've done before it and we'll branch it as a separate chat. Notice here this chat is called Emily's recent emails and we can actually change the name. We could just go rename and we can name this website charts. And so the chat that we're going to be creating the website chart is right there. And now I can say, "Hey, can you actually turn this into a PDF presentation for me?" And one other thing that I will include to this is you can go slashpin. And you can actually pin this chat directly to the top. I could go back to the website charts. I could pin it by pressing this right here. Or again, it's just kind of fun to be able to do this in line. You just go slashpin. And now we have the website charts and then the other branch which is for the PDF. And these both started as this branch and we branched them off. And so this is how you can kind of start with one chat and then branch them out into different chats. And I kind of do it by task. One chat is for one task which is creating the site and the other one is for another task which is the PDF. So the seventh capability of GBT work is you can use it through the desktop app and the iOS app and the web app are very similar in way in the way that they work but the chat GBT work inside the desktop app operates a little differently depending on the settings that you choose. OpenAI has combined the codeex app and the chat GBT app and the result of that combination again was GPT work. Now, the only difference between GBT work on the desktop app and the web app is this configuration right here. If you use GBT work in the cloud, it's basically the same as using it in the web. However, if you use GBT work on your computer, this is when it becomes a little bit different. So, there's two ways that you can use GBT work on your computer, right? You can use the cloud feature and the local feature. Again, if you were to create a new GBT works uh chat here, you can choose on your computer, which is local or in the cloud. And so on the iOS app or the web app or the desktop using cloud, it does not have direct access to files on your computer. However, if you use this desktop using local option right here, again, new chat, if you are using local, it does have access to your computer. And so again, if you use the iOS app, web or cloud, it cannot use the desktop apps. However, it can use your desktop apps if you use the GPT work using local option. It can also use the local work chats. Yes, you can see here it cannot with the desktop app that uses cloud or web or iOS. And so one thing that's really important to note is yes, this version can control your computer, but if you want to be able to continue the conversation elsewhere without having to keep your computer open, you need to choose one of these three. You need to choose desktop using cloud, web, or iOS app. If you use the desktop using local, which means it's GBT work controlling your computer, you're not going to be able to continue the conversation elsewhere. Additionally, if you want to save locally to your computer, you need to use this one right here. All of these operate in the cloud. That's why it's really easy to sync them between all of your devices. However, the desktop one, you are able to add things to your computer. And so, this is a lot like the original version of Claude Co-work. And so, if I go to in the cloud, this is the same as using it on the web app. Hey, can you please create a presentation of the LA of my latest video? Present to me how good my last YouTube video did. If I fire off this right here, which is GPT work running in the cloud. And so, because I chose the cloud version of GPT work, it shows up here in the side panel. So, these are in perfect sync. Look, hey, can you please create a presentation and blah blah blah YouTube video? Hey, can you please create a presentation blah blah blah YouTube video. And they are in perfect sync. And I could very easily continue this on my phone as well. And as you can see here, we have the cloud GPT work task running and I can access it from my phone. So phone desktop app and also the web app. Okay. So now what we should do is we should talk about the other version of GPT work which is the GPT work that runs on your computer. Now what I want you to first notice here is this version of GPT works a lot better with codeex. Codeex is a tool that I've been using a ton. This tool has full access to my computer. It can do basically anything that I would want to do with a computer and that in that includes this computer use. It can fully control my computer, run any command. It has full access and I basically I literally give it full access. It can do whatever it wants. However, the the chat GPT with local access is a little bit safer and it can't run any terminal command. It asks you for permissions more often, but what I like about it is it shares skills. I'm going to show you something real quick. If you use the cloud version of GPT work and I try to type in a skill, I have no skills, right? And since on codeex I have many skills, right? I have a YouTube thumbnail skill, a YouTube researcher skill, I have a diagrams skill, excal diagram skill. All of these skills that I've created over a lot of time and I don't have access to that when I use GBT work in the cloud, which is the same GBT work that is on my phone and the web. If I switch to on the computer and I go slashyou thumbnail, you see all of the skills are there. So, I would recommend if you're afraid of codecs doing something crazy on your computer, stick to GPT work running locally because it can create files, edit files, and delete files on your computer, but it has less full access than codecs. Okay, so the next two capabilities are going to be incredibly fun. They are the most fun capabilities of codecs, at least on the computer. And the eighth capability of GPT work is voice mode. On chat GPT work, you'll notice this little white circle button right here. And if you hover over it, it says start a new voice chat. Let's go ahead and do that. Hey chat GPT work. Um, how are you? >> I'm doing good, thanks. How are things? >> Yo, um, can you please open up, it's in my notion, uh, there's a video database, it's called GPT work explained. Can you please open that up in the CEX browser? Uh, the notion doc. >> Sure thing. I'll open that notion doc in the browser. >> No, you actually got the wrong one. It's just called GPT work explained. >> Got it. I'll switch to that one. Okay. I'm switching the codeex browser to the exact page GPT work explained. >> Yes, exactly. Okay. AI. I named those very similarly. Anyway, I want to uh bring your attention to capability number eight. So, the capability number eight that I I see the eighth capability is GBT work voice mode. Then the next one is uh the eighth I want you to change the next one to the ninth capability of GBT work is using voice mode remote to your computer and then adjust all of the titles. Right? The ninth one should become the 10th one and etc. And then I actually want you to get rid of one of them, which is um h which one should we get rid of? Uh remove the 12th one because I want to stick to the 14 number. >> That I'll update that. >> Thank you. >> The 12th capability is running multiple tasks at once in separate windows with three bullets. I've removed that section and I'm shifting the later capability titles down to keep the total at 14. >> Amazing. And then make the all of the headers um heading twos in blue. >> Gotcha. Numbering is verified 8 through 14 and I'm applying blue text to every heading two on the page. Converted to heading twos. I'll use the notion API only from here to set them blue and verify the page. And there you go. >> Nice work. You changed it to blue. Can you add a couple notes under the eighth capability of GPT work voice mode? Please add like three bullets of of the thing. Like we're using it right now. So add some notes underneath that. >> Got it. I'll add three concise bullets under the capability. I'm adding three practical notes right under capability 8 based on using voice mode conversationally. >> Wow, that's very cool. Can you add Can you add like a very concise summary and highlight it in yellow and make it that the final bullet for that section? >> Absolutely. I'll keep it tight and add that as the last bullet. I'm adding a short yellow highlighted takeaway as the final voice mode bullet. >> Thank you. And so basically all of this connections and plugins that you've created on your phone or on the web app, all of those plugins can be accessed through voice mode and you can just turn it on and anything that you can do you can control with voice. Let me show you one more example. Hey buddy, I need you to go through my email. Um, please create a quick table of anyone who's reached out about paid sponsorships. >> Got it. I'll check and pull that into a fast table for you. filtering out anything generic and keeping only the clear leads. Here's the quick table of real paid sponsorship leads with the brand and the status. >> There's a a company Miniax. Please reach out back to them. Create a draft with the draft skill and open it in the codeex browser and respond and tell them that my price is $250 million per video. >> Got it. I'll set up that draft reply and open it for you to review. The miniax thread is active and I'm preparing that reply now. Now on the miniax draft, nothing was sent. >> Okay, thank you for drafting that out. And as you can see here, I could very easily just send the message that my codeex or my GPT work drafted up for me in the browser. Okay, so we just covered the voice mode on the desktop app, which is this start new voice chat button here. What if I told you that wasn't even the coolest voice mode feature with GPT work? And so the ninth capability of GPT work is using voice mode remote to your computer from the iOS app. So the iOS app has a really cool feature and you can actually go to this remote tab right here and you can connect your iOS app to your chat GBT app on your computer, the desktop app. And in this process, I can actually ask chat GPT to spin up GPT co-work tasks. Let me show you a very basic example, like the most basic example you can think of. This is a real-time voice button, very similar to this button right here. Let's go ahead and press it. Wow. I see that when I started this real-time voice chat, I see a new real-time voice chat on my computer as well. I want you to spin up a new GPT work chat session and I want you to call it Riley's notes. >> On it, I'll set that up now. >> Very good. I can see that you spun it up. Now, I want you to do the same except I want you to do a research on my latest videos inside um I want you to do the research on the latest videos in my notion. Right? Look at my latest videos and I want you to tell me I want you to tell me what videos I should make next based on the last two weeks of videos. Do this in a new GBT work session. Create it now. Got it. I'll check that out. I'm setting up a new session to review those videos and turn the patterns into recommendations. >> Amazing. Thank you so much. Okay, cool. I see. Look at this. And so as you could see here, it created a new session called next videos from recent notion and it spun up this chat. So this voice um thread that I created right here. This is almost you can think of this like the master thread and that is running right here. This master thread can spin up codeex chat sessions. It can also spin up GBT work sessions to do different things. So, my workflow that I do every day is I come into the office, plug in my computer, and I just take my phone and go for a walk for 30 minutes. I go through all of my emails, and I basically just I just talk to codeex, this session right here. And for any task that I need to do, I'll just say, "Spin up a new GBT work session for this content task or this email that I need to do." And then by the time I come back to my computer, I'll maybe see four or five chat sessions open. And I can actually work through a lot of the progress or at least a lot of the thinking or I can come back to a full draft ready to send out to the person that I'm emailing. And that is just a really cool thing to be able to do. The notion review is complete. Your recent videos lean toward explainers and agent conversations. So the best opportunities are practical workflow and team implementation videos. >> That's really cool. Actually, can you please um yeah, can you please add one video in my notion in progress? Um yeah, add it in progress and uh make that video about using GBT work voice 10 use cases of GBT voice on my phone like using the codeex or the GPT remote. >> Got it. I'll add that. >> Thank you. And codeex Yeah, make sure it's long form in progress. >> Got it. I'll make that adjustment. The video's already in progress, so I'm updating it for long for now. >> Amazing. Thank you. >> Done. It's added as long for iness. >> This is a completely new technology. I'm telling you, this is one of the most fun things you'll ever do. And I found it to be really fun because I don't need to be at my computer. I can just spin up work sessions and it can do a lot of the work for me. The 10th capability of GBT work that I think is the most underrated is the inapp browser. And so when you start a new GBT session and actually you don't even need to start a session. You can start a browser session by clicking on this panel and it automatically just opens up a browser. If you hit this plus sign right here you can create a new browser and you can also create a new browser by pressing command T. This browser is a full browser and I've talked about this a lot so I won't go too far into detail here, but this browser can sign into your different accounts. As you can see here, this is my agent native account that I have. This is my podcast and my company. And I can post directly to Twitter. And since GBT work when it's running on this local version can actually control this browser and it could post for me and it could do it directly through the interface. And the coolest part about this is when this browser is signed into all of your stuff because if you ask it to do something and you want it to send an email and open up the draft, it actually needs to be signed into all of your stuff. And the best way to do that is to press these three dots. And you're going to want to import cookies and passwords. When you do that, you can find your browser that you're using before and you can import those cookies and passwords so that you can actually sign in to all of your accounts. And so I could say something like this. Um, come up with based on like my recent activity today, um, and the insights that I've have, come up with 10 tweets, but put it in a Google doc and open it in a new tab, please, um, in the codeex browser. And what I really like is kind of this taskbased organization where you start a task and then all of the browser tabs that you need for that task get open and they're all contained within a single thread. And so I'll actually use this when spinning up my GBT work sessions from my phone. I'll ask it to do something and I'll say, "Hey, by the way, can you open up these two two tabs that'll be relevant for that task?" So that when I come back to my computer, I have the task running already that's done a lot of the work. And then anything relevant that I might need open is also open here on the right side. And take a look at that. It created the Google doc and it opened it in a new codeex browser. I want to pretend for a second that um I didn't ask for it to open up automatically. One way that you can open it is you can rightclick on any link and hit open in browser. And as you can see here, it opened up the Google doc. Uh another thing is it says open in. You can open it in chat GBT. That's another way to open it. And I highly recommend getting in the habit of opening all external links in the browser because it's just a way more fun way to work. You have all of your tasks that you're doing here and you can very quickly open it in the same app. I can switch tasks and then when I come back I have everything that I need, right? I have all the tabs open and the chat session stays consistent. So the 11th capability of GBT work is skills. And so skills are simply little instruction files that your AI agent can use to complete a task. And inside codeex, I've made tons of videos talking about creating skills like my YouTube thumbnail skill, like my notion video database skill. I have so many different skills that I've created. Each one of these skills represents a little workflow that I do that I can reference or the agent can just situationally understand to use. And what I want to say here regarding GPT work is just to make sure that you're using the right version of GBT work depending on where you want to use the skill. And so there are basically two types of skills that you can create inside GBT work or codecs basically inside chat GBT. a local skill and a cloud skill. So if you create a skill inside codeex like I did here with YouTube thumbnail and notion video database, the only other way to access it is to use GBT work and you can see here I can access those notion skills here, right? I can access all my notion skills. However, if we were to change this to GPT work in the cloud and not on the computer and I try to use my notion skills, I cannot do it. So, if you create a skill in the cloud version right here, right, I have this I can say, hey, can you create a skill called test? Um, and it this is just a test skill. Have it do nothing. Literally just say test so it goes quicker. I can create a skill in the cloud version and then this skill will be a skill that I can use from my phone or the web app because these are all cloud versions of GPT work. Okay, so as you can see here it says done. The test skill was created and installed. Its only instruction is test. So I just want to show you how when it creates this link. If I open this up, let's go ahead and open this up in an external browser. As you can see here, it created a test skill. And here it has a skill.md file and its name is test. Its description a deliberately minimal test skill that does nothing. Use when user explicitly asks to invoke the test skill. And again I can try it in chat. And now you can use it in the cloud. If I wanted to use it on my phone I can. So, if you want to create a skill that you want to use while you're on a walk or you're outside or you want to be able to use it from GPT work on your phone, make sure you create it on the web or on your phone or the cloud version of GPT work in OpenAI. Again, if you're watching this, please just merge these all together so that all of our skills can be used from any of the products. Anyway, let's move on [clears throat] to the next skill. For the 12th capability of GPT work, you can have your GBT work do things in the future. And this is something people consciously understand, but it's not something that people grasp enough to truly use it effectively. So, I want to show you a a way that I conceptualize it. Let's imagine you send an email. And when you send this email, you want to make sure that a certain action happens as soon as you get a response. And so, I will do something like this. I will say, let's say I send an email on Friday. I'll say starting on Monday, every x period of time, maybe it's every three hours or maybe it's every day, check at 9:00 a.m. Ending on next week, check for a response. If you get a response, let the whole team know. And so, let me show you an example. Hey, I want you to send an email to emily@notinumber.com and let her know that I um I need to make this video 20 capabilities. So, send her the notion document of this and remember we're talking about this notion document. I can just paste in my notion document. Please send her the email and I want you to please starting tomorrow morning at 9:00, check for her response every single hour. As soon as she sends it back, her response, I want you to add it to that same notion document. And once that's done, please can you send an email to me, Riley agentnative.inc, my other personal email, because I'll be on vacation, and make sure that I can see that so I can make final edits. And this workflow, this will work. I've been doing this for nearly all of my emails. Anytime I send an urgent email that like I would want to respond really quickly or notify my team, I will automatically do this. I'll just say on these days, check every hour for a response. If I get a response, notify people on my team. I will have AI periodically check and route whatever I need to do to the person who needs to help me. It seems easy. It seems trivial, but this is a mind shift that you need to realize that this agent right here, especially if you're operating in the cloud, if you're using the local version, the automation actually won't run. The scheduled tasks on codeex or the local version of GPT work, they won't run. The scheduled tasks won't run because your computer needs to be open. But the G if you do it from the GBT app or you do it from the cloud version of GBT work, it will run no matter if your computer is open or not. And so for all of those email tasks, I like to make them cloud-based GBT work tasks. So no matter what, my computer doesn't need to be open and I know for a fact they'll fire off and the task will get done. And for the 13th capability of GPT work, I want to talk about creating spreadsheets. These models, the Frontier models are getting so good, it's almost scary at creating spreadsheets, especially if you ask it to do in tons of research on the internet or if you give it a ton of your own internal documents. So, I just said do in-depth research on the growth of Microsoft, Google, and Apple. I want you to do an in-depth comparison and create a fully detailed report. Create a spreadsheet for this. I want a lot of charts and high quality data. And for this, I'm using it in the web app. I'm not using it on the desktop app. And here I can click this and it will open it up. And these are getting significantly better. And I can open this up right here. And we can see all of the data that it has pulled for this spreadsheet. And we have strategic comparison and sources sources and checks. And so it can create really high quality spreadsheets. and you can very easily ask it to edit it. Uh please make the main uh please make a um dashboard page or some sort of like cover sheet which is like very colorful and fun to look at and recommends uh where I should buy it. And then after you do this, I want you to convert this into a Google sheet. And because we've set up the Google Sheets or we have the Google Drive plugin, because we have this, it can automatically create Google Sheets and it will be able to you'll be able to open it up right away and open up Google Sheets. And if you were doing this in the desktop, you could open it in the inapp browser. Okay, you can see here it is done. Let's just go ahead and open this straight up into Google Sheets. And there you go. It is really colorcoded and you can see here Alphabet, Microsoft, and Apple. And we can go to historical financials, executive dashboards, and now charts. And it created all of these different charts. As you can see, capex, capex, and revenue. We have 5-year revenue, uh, free cash flow, all of the things that we might want to do research on. AI can go off and create these insane spreadsheets. Okay, we made it to the final capability of GBT work. And this is my multi- aent workspace capability. And so here I have four different GPT works open on my computer right now. All of these are running in the cloud and I'm working on many different things at once. Because agents are starting to take a lot longer, it's nice to be able to multitask. And sometimes I don't like to multitask with this side panel, I actually prefer a new view where I have everything smaller and I'm switching between these different views here. I fired off what is in my emails and I can say, "Oh, the credits exhausted for Perplexity Computer. Please give me more information. What did that Perplexity Computer email say?" And I can just fire off that prompt. And now here I'm doing the intro for my GMO Roush podcast. I'm filming this right after. Let's say I wanted to dive further into this. You can just go into the top bar, doubleclick on it, and open it up. And so that's kind of how I go in and out of full screen and this view is you can just double click and it snaps back to his position before. So that's how I do it. So for if I want to hone in on this tasks, doubleclick. I should do GPT work explain next. That makes sense. I'll drop it back down. Uh, please open this in the browser. And so I can open up this one in the codeex browser. And now I can go back to my next task. Here it says the perplexity email says you've used all your credits. And say, okay, please respond saying I don't want it anymore. use Gmail. And now I can go to the next one. Got it. Open it up in the codeex browser. So I can very easily open this up. And there you go. We now have this open in our codeex browser. And I can very easily toggle this down, toggle this up, and I can switch tasks really easily. And this if I'm working on three or four tasks at once, this is how I prefer to use it. And sometimes these will be codeex tabs, not just uh GPT work tabs. And here we're generating a cool chart here, which is really cool. Anyway guys, so these are the capabilities that we talked about today. The easiest ones that are cloud-based are branching a chat. You can create blocks. This is a version of a block, right? This is just a quadrant chart. Remember we talked about mermaid diagrams. We also talked about voice mode which you can initiate from a new chat and you can very easily talk to it. If we were to create a new chat, you can initiate it by pressing this button right here. Uh we also talked about reports and documents. You can create PDFs, docs, spreadsheets, that type of thing. We also talked about generating and editing images directly in line. Um you can also use the desktop app and I recommend using the uh codecs for a lot of these tasks, especially if they're co uh coding related. But again the local codecs and the local GPT work go together and then all the other versions which is the cloud desktop version of GPT work the GPT work on your phone and web go together those skills are shared and so remember always remember that and then also you can use voice mode through the remote so you can communicate with your desktop app so if you need to spin up GPT work sessions or codec sessions from your phone you can and then you can also create skills and you can create skills Just know that the cloud skills go with the cloud versions of the app and the local skills go with the local version of the app. And one thing that most people aren't using properly is you can schedule automations. And you can also build websites and full apps and host them on the internet, which is really cool. Anyway, thank you guys so much for watching. This was a long video. I put like eight or nine hours into this. I hope you enjoyed it. Please like, please subscribe. It would mean a ton to me. And I'll see you here for the next

AI never sleeps and this week has been absolutely insane. We have a new open-source AI for medical research and reporting. Alibaba releases their latest Quinn model and it's an absolute beast. This AI can compose an entire symphony from scratch. We have another AI that can create super realistic singing voices. OpenAI's internal model just solved some massive math breakthroughs. We have a new open source AI for creating 3D CAD files. Google releases and open sources their state-of-the-art AI for predicting cyclones and natural disasters. We have a new all-in-one 3D model generator and editor. Plus, this can also segment a model into separate parts. We have some exciting robotics demos and a lot more. So, let's jump right in. First up, we have a new AI for creating symphony music. It's called Symphony Gen, and this is designed to create full orchestral music while giving you control over the underlying harmony. So, here are a few examples. [music] [music] >> [music] [music] [music] >> Now this is actually quite challenging because the model has to coordinate many instruments and notes and the overall structure of the piece at the same time. And actually what this does is it first generates a harmony skeleton and then it expands the outline into a complete orchestral arrangement. So think of it as first sketching the chords and the music direction and then filling in the notes for each instrument. For example, you can start with a major chord and here are a few examples with this start. [music] >> [music] >> All right. So, as you can hear, they kind of start with the same major chord. The nice thing is you can also input your own harmony skeleton or it can also take the harmony skeleton from an existing piece. For example, here's an original piece and then I'm going to play you the AI generated piece which takes inspiration from this harmony skeleton. [music] >> [music] [music] >> So, as you can hear, it has roughly the same chord pattern and tempo as the original XRP. The awesome thing is they've released the models to this already. So, if you click on this code button and you scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. Now, the models to this are fairly tiny. It's only like under 5 megabytes in size, so you should be able to fit this on most consumer devices. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, if you're into 3D modeling or printing, this AI is super useful. It's called MAC, which stands for multi-agent CAD. And this is a really efficient AI that can help you create printable 3D models in CAD format. And all it takes is a text prompt. So for example, here is the text prompt on the left. And it'll generate a CAD file which you can print out in 3D as you can see on the right. Here are some additional examples for your reference. As you can see, it can design a ton of different objects with different shapes and articulations. Here are some additional examples for your reference. And you can also see the cost per prompt listed below each example. And as you can see after adding this multi- aent CAD system, it's a lot more efficient. It's able to complete the task most of the time at like 10 times lower the cost compared with if you didn't implement this system. In fact, if you compare this to another text to CAD model called CAD skills, you can see that this new Mac one uses 116 times fewer tokens. It costs 13 times less. Plus, the pass rate is also much higher. And getting this setup is also super simple. This is model agnostic. So, they used Quinn, but you can also switch it up with another model. Now, if you're interested in running this on this page, if you scroll down a bit, it contains all the instructions on how to download and run this locally on your computer. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, Alibaba dropped their latest animation system, One Animate 2. This basically takes a photo of any character plus a reference video and it'll animate the character according to the reference video as you can see in this example. Note that it works with hands and fingers as well. Plus, this doesn't even have to be human characters. So, here's an example where we can animate this teddy bear instead. And in addition to hands and fingers, this can also transfer facial expressions from a reference video. As you can see in this example, the nice thing is this can also animate multiple characters. So, for example, you can input a photo with two characters plus a video with two characters and it'll transfer their movements over very naturally like this. Alternatively, you can just use a reference video with one person, but use that to animate multiple characters in a photo, as you can see in this example. And the nice thing about One Animate 2 is that you can also animate characters with irregular body proportions. So, it doesn't have to be just human to human. Another new feature about One Animate 2 is you can also determine the camera angle. Here's an example with the same inputs, but you can also change the view to the left or to the right. And the nice thing is they've also released a smaller version called Wan Anime 2 Light, which even allows for real-time streaming if you have the right hardware. You can see like the latency is under a second. So, this can be really useful for streaming. Now, I featured a ton of these tools before, including One Animate and Dream Actor, but this new One Animate 2 is a lot more detailed and consistent and natural. Now, they didn't compare this with another leading animation tool called Scale 2. I think the quality of this versus Scale 2 is very similar. The awesome thing is they've released this already. Now, the full one animate is around 33 GB, so you'll need a high-end GPU to run this. And there's also already support for Comfy UI. They've released an int 8 version which is half the size and this should be able to fit on just a mid-tier GPU. If you're interested in reading further, I'll link to this main page in the description below. Also, this week we have a new AI called vocal render and like the name implies, this can generate singing voices for songs and it actually sounds incredibly realistic and expressive. So, here's what it does. This takes in lyrics and then the melody in MIDI notes and that's about it. So from this it can output a very expressive voice singing out these lyrics. Now they released two different models, Vocal Render and Vocal Render Pro. The Pro version just sounds a bit better. Let's listen to both. [gasps] >> And then here is an example with the pro version. Now, if you compare this with other singing [snorts] voice generators like Vivo 2 or Soul X, you can hear that vocal render is a lot better. So, let me just play you the other competitors as well. [gasps] [singing] As you can hear, Vivo there didn't even follow the pitch that was specified. And how this works is actually quite interesting. It reads the lyrics and the musical notes together and predicts how the performance should flow and it automatically decides the final timing and the audio length. This is especially useful when one syllable stretches across several notes. So under the hood, an auto reggressive component first builds a broad sketch of the singing style and timing and then a diffusion model fills in the finer details including pitch, vocal tone, articulation, and local audio texture. Now, the awesome thing is they've released the model already. So, if you click on this view repository button and you scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. Plus, they've also released the code on how to train this yourself. And that's because currently they've only trained the model on Chinese, but you can also train your own checkpoint with any language you want. So, you just need to follow the training instructions on this page. Now, like I said, they have two different variants, a pro and a normal variant. Both of them are under 10 gigabytes in size, so you should be able to fit this on most consumer GPUs. Anyway, if you're looking to generate really realistic and expressive singing voices with AI, then this vocal render model is one of the best you can use so far. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, Tencent Hunen releases a really cool 3D editing model called Hunyan 3D Buffalo. This is basically a unified model that can generate, understand, edit, and separate 3D objects. So, here are some examples. We can simply edit 3D models with a text prompt. For example, we can input this model and then write turn the head into a bull's head. And here is the result. Or here we can write remove the sail. And it indeed removes the sail from this 3D model. Or here's an example where we can add glasses to the frog like this. Or we can take this robot and put spiked gauntlets on both hands. And indeed, that is what it does. It's also able to generate 3D models from just a text prompt. So, here are some examples for your reference. The cool thing is this can also take in any 3D model and separate it into individual parts like this. Or here's another example where we can separate this model into these separate parts. So, while most AI 3D model generators can only do one of these things, this new Hunyan 3D Buffalo is designed to be a unified model that can both generate and edit and segment 3D models, making it super flexible. The nice thing is at the top of the page here, it says the code is coming soon. So, it looks like they are planning to release the code and models to this, which is fantastic. For now, if you're interested in reading further, I'll link to this main page in the description below. Next up, we have a new AI called Leap Talk. And this can generate talking avatars in real time. So, this just takes in any reference image of a person, any speech audio, and it outputs a lip-s synced talking head video in pretty much real time. >> Can the security and smooth passage of international waterways be fundamentally safeguarded? All parties should work together to deescalate the situation and prevent regional instability from having a greater impact on the global economy and energy security. Now, this talking head does look very rigid. She doesn't move as naturally as some of the other frontier avatar generators out there, but the strength of this model is that it's incredibly fast. So, if you look at the latency of Leap Talk compared to other avatar generators like Hello 3 or Echo Mimic, which I featured on my channel before, Leap Talk is like thousands of times faster. And they claim that on an H200 GPU, this can achieve up to 200 frames per second, which is crazy. So, if you're looking for a lightweight realtime talking head generator, this is likely the fastest option available right now. And the awesome thing is they've released this already. So, if you click on this code button and you scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. If you're interested in reading further, I'll link to this main page in the description below. If you want to supercharge your content creation, definitely check out Higsfield, the sponsor of this video. They've just added the best video generator out there, Seed Dance 2.5. The biggest improvement from this model is that you can now generate up to 30 seconds of video in a single pass with multiple shots and an actual narrative with audio built in. You can also extend an existing generation with new shots while keeping the same characters, locations, pacing, and overall consistency. What really stands out is the reference system. You can feed it up to 50 references at once, including 30 images, 10 videos, and 10 audio files. You can provide your characters, environment, visual style, motion, and soundtrack all in one generation. It can even use a simple 3D clay render as a reference to understand the camera movement, lighting, and overall composition of the scene. You also get much more control over editing. For example, you can specify exactly what happens during different timestamps, change just one section without affecting the rest of the video, or move the same performance into a completely different environment, or even change the camera angle while preserving the characters and action. Seance 2.5 supports text to video, image to video, video to video, and other references. So whether you want to create short films, action scenes, music videos, commercials or any other content, it gives you a lot more control while keeping everything consistent. You can try Seance 2.5 on Higsfield today using the link in the description below. Now around 2 weeks ago, Moonshot AI released Kimmy K3 and this is a massive 2.8 a trillion parameter model which they open sourced and back then this was not only the best open- source model out there but it has even caught up to Frontier for some of these benchmarks it's as good or even better than like GPT 5.6 or Fable 5. Well, it's not over yet. So this week Alibaba releases Quen 3.8 Max. This is also a massive 2.4 4 trillion parameter model and this is also the first time that they will open source a max class model which is super exciting. Here they say the weights will be released next week and check out the benchmark scores of this in terms of these agentic software engineering tasks. You can see that in some cases this new Quen 3.8 Max which is the dark blue even matches the performance of Fable 5 and GPT 5.6 Soul and it already beats Opus 4.8 which is the dark gray bar. Very impressive for an open-source model. Now, like most Frontier models, this is designed to work across a ton of steps, autonomously use tools, inspect results, and continue working until it reaches your specified goal. Here in this demonstration, it was given an empty folder, and it was told to create a self-improving harness system from scratch that turns feedback into GitHub issues, automatically claims and codes them, and merges working changes, and keeps evolving the tool over time. And the crazy thing is this worked for about 16 days without any human help. It just kept doing this until it achieved the goal and in the end it made like 265 commits and 127 pull requests. Or here's another example where it was given a recent paper for training language models and it was told to basically reproduce the paper and then beat it. So it worked autonomously for like 5 days. It wrote thousands of code from scratch, rebuilt the entire experiment, matched the paper's results, and then invented and tested 18 of its own improvement ideas. And its final method actually beat the original paper by 2.7 points on a really hard competitive math benchmark. That's crazy. I mean, this is just a very recent paper, but you can just plug it into this AI to figure out an even better result. So, I mean, the age of autonomous scientific improvement is already here. Or here's another crazy example where it was told to design and optimize a chip. So it started from basically nothing. It designed, coded, simulated, and repeatedly improved this cryptographic hardware accelerator. It was able to produce a physical layout that's like 12 times smaller than the baseline while meeting timing targets. Now, if you look at this leaderboard by artificial analysis, then you can see that Quinn 3.8 8 is just one point below Kimmy K3 while being like 400 billion parameters smaller. So I mean once they release the weights to this this will be like the second best open model out there and they're edging very close to GPT 5.6 and Claude Fable. The thing I don't like about artificial analysis is they don't have any confidence intervals. So it's hard to say whether these top five models are actually significantly better in terms of performance. Now currently you can use Quinn 3.8 8 Max via their API. And as you can see, it's slightly more expensive than Kimmy, but still cheaper than GPG 51.6 and much cheaper than Claude Opus or Claude Fable, which I would not recommend. Like I said, they are planning to release the weights next week. But for now, you can try this out via Quen Cloud. If you're interested in reading further, I'll link to this main page in the description below. Next up, Google DeepMind releases a really exciting update called Weather Next 2. This is basically an AI that can help predict hurricanes and tropical cyclones way earlier than other methods and way more accurately. You see, the difficult thing about hurricane forecasting is you normally need one kind of model to predict where a storm will go and another highresolution model to predict how strong it'll become. But weather nex basically combines both jobs into a single model. It's able to predict the storm's track intensity and wind structure. In fact, it can generate forecasts as far as 15 days in advance, and it can run an ensemble of a thousand different possible scenarios to estimate the probability of where it's headed. It's also way more accurate at predicting the track intensity and extent. So, this latest weather next model is the blue line here. And as you can see, its error rate is much lower than the other methods, even as you increase the lead time to like 5 days in advance. Here's another example where if you compare the error rates with other competitor cyclone prediction models, you can see that Weather Next has a much lower error rate. In fact, Google says Weather Next provides more than 24 hours of additional forecasting lead time compared with leading systems. What's more impressive is it does it using weather data at roughly 28x 28 km resolution, which is around 100 times coarser than traditional models, which require much higher resolution data. So that means a full 15-day forecast can just be generated in under a minute on just one TPU, which is Google's tensor processing unit. Here it says they basically trained this AI on nearly 20 terabytes of global atmospheric data, including nearly 5,000 historical storms. So the model is able to learn these complex weather patterns and how to identify extreme weather. The awesome thing is not only have they published a Nature paper on this, but they are also open sourcing the code and model to this. So anyone can just download and run this or build on top of it. So if you click on this link, it takes you to their GitHub repo. And if you scroll down a bit here, it contains all the instructions on how to set this up. They're also releasing Weather Next 2 Mini, which is a compact version which you can run in Google Collab for free. So props to Google for open sourcing this. This AI is actually super helpful and it'll save a ton of lives by predicting, you know, these natural disasters earlier and more accurately. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, OpenAI announces something pretty remarkable. You see, there are rumors that their next model, GPT6, will be rolled out very soon. Well, this is internally codenamed Astra, and it looks like they used an internal version of Astra to solve 10 long-standing open math problems, which is pretty insane. These aren't just math questions with known answers. They include open problems across geometry, coding theory, group theory, quantum complexity, cryptography, and combinotaurics, which no human could ever solve before. But here with this internal version of Astra, it was able to either resolve the problem or make substantial new progress. Now, this is very technical, especially if you don't have any math background. But in summary, one example establishes the existence of these non-sopic groups, which addresses a major question in group theory. And then it was also able to resolve multiple airish problems. And then there were other examples where it's able to improve the bounds in sphere packing and coding theory. Now these problems are basically really hard to solve, but once it finds the answer, it's really easy to prove that the answer is correct, in case you're wondering. And you know, probably the craziest detail is the cost. So, OpenAI estimates that all of the model tokens used to discover these 10 solutions cost only $2,000 at their API rates. That's crazy if you think about it. They only needed to spend $2,000 of compute to crack 10 mathematical breakthroughs which no human could ever solve. We are right in the middle of the scientific acceleration. In the past few months, we've already seen some huge breakthroughs in like medicine, physics, and of course, math. So, it's really exciting times. They've also released the reasoning walkthroughs on how the AI actually came up with the solution to each of these problems. This is extremely technical math stuff, but if you are interested in digging through the details, I will link to this main page in the description below. Also, this week, Alibaba's Demo Academy released a really useful open-source model called Clinfusion. This is basically a model for holistic medical understanding. Basically, you can give it medical images like X-rays, scans, or even native 3D imaging together with a text prompt, and it can answer medical questions or generate clinical reports. The big problem with medical AI models is that different imaging types are extremely different. But basically, what Clinfusion does is it uses a combined vision encoder designed to understand all this medical data inside just one system, including 2D and 3D medical data. And from this you can see the results are extremely strong. It beats leading and closed source models on most of these multimodal benchmarks even outperforming proprietary models like GPT 5.2. Although note that this is quite an old GPT. Right now we're already at version 5.6. Now on to the specs of this. They released two different models. There's a 32 billion parameter model which is higher quality. And this is quite large at 72 GB in size. And then we have an 8 billion parameter version which is only 24 GB in size. So you should be able to fit this on like a mid to high-end GPU. This is probably the best open- source model for medical analysis. You can use in that size range. So if you're interested in trying this out on this page, it contains all the instructions on how to download and run this locally on your computer. If you're interested in reading further, I'll link to this main page in the description below. In human robotics news, we have this new demo from Persona AI. So, here it's demonstrating its Gen 1 humanoid robot performing a real welding task through teleaoperation. So, as you can see, the dude on the left is wearing this VR headset, and it's controlling this Gen 1 humanoid robot in real time. And it's even able to do this welding task, which requires very precise and stable movements. Now, this is quite a simple example, but this is a demonstration of how we can eventually deploy robots in some high-risk or dangerous industrial environments and then get them to work there while an expert controls them remotely somewhere else via tea operation. In other robotics news, UB Robotics also previews their swarm intelligence model in action. Here you're seeing several of their new wheeled industrial humanoid robot called the Cruiser Y1. And they're all working in this warehouse. They're all taking stuff from pallets and then putting it to the right location. Now, not only do each of these have a brain, but there's also an overarching swarm intelligence system which coordinates all of them together. So there's no redundancy, there's no overlap. And so this allows you to control like basically an army of robots to do a task concurrently. In other humanoid robot news, Xiaomi has released a new robot foundation model called Xiaomi Robotics 1, and it's designed to let robots handle everyday objects and practical tasks. So things like picking up objects and placing them in different places, as you can see in this example, it's even able to zip up this bag very effectively. And it's able to pack a suitcase like this. It's able to navigate across the room to find various objects to put in the suitcase. Basically, how this works is you just give the robot an instruction using natural language and it looks at the environment through its cameras. It understands what needs to happen. It plans everything out and then it actually carries out the action. What makes this model especially interesting is how Xiaomi trained it. So, normally collecting robot training data means having humans remotely control real robots for thousands of hours, which is expensive and difficult to scale. However, Xiaomi collected around a 100,000 hours of video using just a handheld gripper with a camera. So, the humans simply carried this device around while performing normal tasks in their homes or factories or offices. The model first learns general manipulation skills from all this human data. Then, Xiaomi adapts it to actual robot bodies using another roughly 10,000 hours of real robot data. The awesome thing is they've actually released the models to this, including details of how they train this. So, if you're building your own robots or if you're looking for ways to train a robotics model, this would be a great resource for you. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, we have what they claim a self-evolving model called Big Bang. This is an experimental language model with a pretty wild idea. Instead of humans constantly designing better training questions to improve the AI, what if the AI could generate increasingly difficult training data for itself? So, Big Bang starts with the open- source Quinn 3.635b, but all its post-training data comes from an automated system where generator agents create and solve difficult scientific and technical problems. And then afterwards, we also have this critic agent which tries to find mistakes and reject weak examples. And then finally, we also have a metacritic agent which checks whether these difficult problems actually make the model better on real research tasks. And the results are surprisingly strong compared to just the base model Quinn 3.635b. You can see that it scores much higher in terms of browse comp and bench which are like aentic coding benchmarks as well as Frontier Science. You can see that this is an especially huge leap. The base model only scored like 12 points but here it scored like 46 points. Same with humanity's last exam, a massive improvement and also about mystery and also paper bench. Now calling this self-evolving is a bit misleading but I think what they meant here is this framework can be looped again and again. So you can keep getting it to create more and more challenging synthetic data to keep improving the AI model. But of course at some point you're going to hit a wall in terms of intelligence and performance. The awesome thing is they've released this already. So, if you click on this code button and you scroll down a bit here, it contains all the instructions on how to download and run this locally on your computer. If you're interested in reading further, I'll link to this main page in the description below. Also, this week, Meta quietly releases their latest AI model, Musepark 1.2. This is their updated version that's now much more focused on real world coding and agentic workflows. The idea is to give it an entire software project and let it work across multiple files, call tools, and keep going through longer tasks. This supports a 1 million token context window, so you can potentially fit a ton of information into your prompt at once. It's also multimodal, so it can take in things like text, images, video, audio, and documents as input, although coding is one of its main focuses. As you can see from their self-reported benchmarks, it is quite a big leap compared to the previous new Spark 1.1, and it's edging pretty close to the top models out there, including Opus 5. However, note that these self-reported benchmarks are very misleading. They just use GPT 5.6 Terra instead of Soul, which is the larger model. So, it looks like they're intentionally just including a dumber GPT model on their page. So instead, if you look at this independent artificial analysis leaderboard, then you can see that Muse Spark 1.2 is actually all the way down here, way behind Quen 3.8, Kimik 3, and GPT 5.6 Soul Max. So it's not even within like the top 10. However, if you look at the cost per task, then this is quite cheap, even cheaper than Gemini 3.6 Flash and Kim K3 and much cheaper than the Opus models. Now, this is closed source, and for now, you can only use it via their API. If you look at another leaderboard called Val's index, which looks at how good an AI model is across finance and coding tasks, then you can see that again, Muse Spark is all the way down here, even below GPT 5.6 Soul and Kimik 3. However, it does cost the least compared to the other models. So, this could be a fairly costefficient option. For now, note that you can only use Muse Spark 1.1 via their API and this is closed source. In addition to Musepark, they also introduced something called Muse Code, which is a coding agent exactly like OpenAI's codeex or Zcode, Kimmyode, etc. Now, like I mentioned before, the best harness to use to run an AI model is the harness designed by the same company. So, if you're using GPT, then it's best to use Codeex. If you're using Kimmy, it's best to use Kimode. If you're using GLM, it's best to use Zcode. And the same goes for Muse Spark. So, if you decide to use this, then the best way to use it is through Muse Code. Anyway, on this page, it contains all the documentation on how to download and run this. So, if you're interested, I'll link to this page in the description below. Also, this week, we have a very interesting harness framework called Long Horizon Harness. And this is designed to help AI agents complete really complex tasks that can take a ton of steps or a really long time, like hours or days. You see, we already have a ton of agentic harnesses today like Codeex, which is now renamed into ChatgPT or also Claude Code, Zcode, Kimmy Code, Open Claw, or Hermes. But they all face one problem, which is that if you get it to do a really long task, then it has trouble fitting and remembering everything. So, as the history becomes longer and longer, the agent could forget the original goal. It could make some incorrect summaries and completely go off on a tangent. It can repeat work or even mistakenly claim that something is finished. Well, what this long horizon does is it replaces that approach with three separate roles. There's a manager, an executor, and an auditor. The manager decides the next small task based on only verified progress. The executor receives that one task in a fresh context and performs the actual work, and then the auditor independently checks the files, the fixes, etc., and confirms what really changed. Only stuff that was verified by the auditor are saved for the next round. So, think of this as like a construction project where the manager assigns the work, the builder completes it, and an independent inspector signs off before anything is marked finished. And the nice thing is this system works across different agentic harnesses like Claude Code, Codeex CLI, Gemini CLI, Zcode, Kimmy Code, and other compatible systems. And here are some really impressive results. If they add this long horizon harness using Quen 3.7 in cloud code, you can see that it was able to increase its weavebench score by 28.9%. It's also able to triple the completion rate of OS World 2. And then for Terminal Bench, it's also able to increase the score by like 7.5% which is a huge deal. Here's the Terminal Bench leaderboard. And as you can see, if you add this LH harness to GPT 5.6 6 Luna using codeex, it's able to achieve a much higher score than without this harness. It's all the way down here. Same with if you add the harness to Cloud Code and Quen 3.7, it's able to achieve a much higher score than without it, which is down here. Now, if you add this harness layer on top, of course, it is expected to use some more tokens. For example, for both Weavebench and OS World, it uses a lot more tokens than without it, but it does achieve a much higher score or success rate. So, there's a trade-off there. Interestingly for Terminal Bench, not only did it achieve a higher score, but it also used fewer tokens. Anyway, a very fascinating framework which could potentially improve the performance of any long horizon tasks that you're running. The awesome thing is they've released this already. So, if you click on this code repo button and you scroll down a bit here, it contains all the instructions on how to download this for whichever Agentic system you're using. If you're interested in reading further, I'll link to this main page in the description below. And that sums up all the highlights in AI this week. Let me know in the comments what you think of all of this. Which piece of news was your favorite? And which tool are you most looking forward to trying out? As always, I will be on the lookout for the top AI news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so much happening in the world of AI every week. I can't possibly cover everything on my YouTube channel. So, to really stay up to date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the description below. Thanks for watching and I'll see you in the next one.

Welcome to AI video [music] 101, where you're going to learn the basics of how to create your first AI short film. You're going to learn how to write prompts, ideulate your storyboard, generate your frames, turn them into videos, and then [music] add realistic voiceovers, background music, and sound effects, all inside of 11 Creative, which connects all of the best AI [music] image, audio, and video models in one place, allowing you to create any AI short film with style and character consistency. What I'm teaching can be applied to any short film and any style. And what we're creating today looks like this. A tiny adventurer takes her first step. One step, then another. >> Inside 11 Creative, you have all of the tools in the left menu bar to create your first AI short film. If you can't see the tools that I'm talking about, you can click on more tools and simply pin them to your sidebar. Above these, we have the different interfaces that we can use, specifically studio and flows. Studio is like your traditional video editor and flows is a node-based canvas that [music] allows us to repeat and automate workflows and create with AI very efficiently. Each of the specific tools can be used for individual generation, but we can also find all of these tools inside flows and studio. One of the important features that you'll need throughout the creation of your AI short film will be the assets folder that you can find by clicking the folder icon in the top right of your screen. [music] Here you can save any of the generations from any of the tools into a folder so you can [music] access them no matter what interface you are using to create your AI short film. This means that you never have to leave 11 creative and you don't have to download and re-upload all of your [music] files. Now let's go and generate our first asset and we'll start with creating the character inside 11 creative. To generate our first character, we're simply going to click on the image and video tool. Here we can generate with the world's best image and video models. And to do so, we simply want to describe what we want to generate in the prompt box at the bottom right here. If we want to generate an image, we make sure that we select image and then down here we have the model picker. In here, you have all of the best AI image models in the world. You simply go ahead and choose a model such as GPT image 2 and type out what you like. So, for example, here if we want to generate a hedgehog, I can do so. Below this, we have the different settings that I can customize. Here we've got the aspect ratio, we've got the image resolution, we've got the quality of the generation, and then we've got the number of variations that we want to generate. Next to this, we have the prompt rewriting. If we're using short and simple prompts, you can toggle this on to get more creative outputs at a higher quality. But if you're using long, more complex prompts, I suggest you turn this off so you get something that is true to the prompt that you've written. So here, I'm simply going to toggle [music] it off. And now let's generate our first image and click generate. And now, as you can see, we've got our four images generating. When it comes to the beginning and creating all of the assets, you are essentially world building. You're creating your characters, you're creating your scenes, your locations, all of the objects that you want to use in your video. And with this, you need to do a lot of prompting. And when generating, sometimes the models can take a little bit of time to generate. So, for example, here, as you can see, we've just got our first generations come through, and we can see it's a headshot from a very simple prompt. One of the models I suggest using if you are in the explore stage of your AI short film is actually switching to Nano Banana 2 Light. Nano Banana 2 Light is a much cheaper model. Its limitations are that it generates at a maximum of 1K resolution. However, it is very cheap and it is very fast to generate your pictures. So, for example, here if we go and click generate once again. Now, within just a few seconds, as you can see, the percentage is going much faster. We will have the generations of our hedgehog. And of course, you'll notice that the output of Nanoan 2 Light is actually quite different to the output of GPT image 2. There's three reasons to this. First of all, each model interprets a prompt differently. [music] Second of all, each model has a different style of output. And the third reason here is that we haven't been specific with the prompt at all. The more specific you are with the prompt, the more consistent the results are between the different models. So when you are laying the groundwork to your short film and you are in the explore phase, make sure that you try different models and different prompts and see what you like the most. Generating [music] images is much cheaper and faster than actually generating videos. And the biggest mistake is actually trying to generate the video first because you'll end up burning through your credits way too quickly without even knowing what that generation is going to look like. But when you start with reference images and start frames, you know exactly what the character is going to look like in the video before you even generate the video. So always start with generating all the assets you want and essentially build out your story board just like you do with traditional film making. Now instead of using a simple prompt like the word hedgehog here we're going to customize it a little bit and [music] this is the prompt that I'm going to use. A 100% feltade scene of a tiny hedgehog peeking out of a pile of stitched autumn leaves. Now here what I'm doing at the beginning is I'm describing the style of the scene. After that, I'm including the subject and I'm also describing the subject and I'm also including the objects and describing those as well. At the end, I've got a little bit of negative prompting. I [music] don't want any text appearing on this generation. And then we're also adding a little bit more context and more details so we can get some unique results. So here now with this more complex prompt, if I go ahead and click generate, let's take a look at the result. And once again, we're using Nano Banana 2 Light, so we should get the results pretty quickly. Nal Banana 2 Light is great for exploration, but sometimes once you found the prompt that you want to use, you might want to switch to a different model. So, here, as you can see, we've got four generations of this little hedgehog peeking out of a pile of leaves in a felt style that I want to use for the animation of my short film. Now, if I went and switched this model and we chose something like GPT image 2 [music] and I click generate, we'll also get a different result and we'll also generate with Cream 5 Pro, just so you can see the difference between a few different models. [music] And for speed purposes, I'm just going to change the resolution to one here and then click generate. And as you can see, if you look at the different generations from the different models, although the prompt is the same, we have a different style. And this is why I actually prefer using GPT image 2 for this type of animation over Nano Banana, because I prefer the result and the feel of that felt style that I'm going for. And here, you'll quickly notice that in the image and video feed, well, we're creating a vertical feed of all the generations, which isn't necessarily the most efficient way to create when using AI. And this is why we've built flows. If we click on flows in the left sidebar here, we go ahead and click on new flow. Now we have this node-based canvas that I was talking about that has all of the AI audio, image, and video tools baked into it. And so we can do everything that we can do within 11 creative straight on this canvas. And it works just the same. Here at the bottom, we have [clears throat] all of the tools. If I want to go ahead and generate an image, I can go and click on an image node. I can paste in the same prompt as earlier. We can also customize the settings. And then we can click run and that will generate a variation of this prompt within this node. The great thing about flows is that when you're creating with AI, you're consistently reusing prompts and reusing images as references for your videos. And so what this allows you to do when we want to turn this generation into a video, well, we can simply drag and use this image as a reference. And when we let go, we can then choose a different type of node that we want to use. And here later on we'll be using the video generation node and then typing in a new prompt to turn the static image into a video. But before we do that, I want to make sure that my style and my character is consistent throughout the entire short film. And so what we're going to do is we're going to create a character reference sheet based off of this hedgehog right here. So it's always consistent throughout. And we're going to take this image node and we're going to drag a connector from it. And here we're going to choose edit image. And what I'm going to do is we're going to change the aspect ratio to 16x9. And here we are using this image of the hedgehog as a reference. And so I can use the prompt create a character reference sheet of this hedgehog from all angles. So I can use it as a reference asset for all my generations. And essentially here you'll notice that I'm simply talking [music] to the AI just like I would to a human. What I do want to do is because this is a character reference sheet that we're going to consistently reuse over and over again. One, I want it to be perfect, but two, I want it to also be at the best quality possible. So here I'm going to change this to 4K and then I'm going to change the quality to high and simply click run. And now as you can see I've got a character reference sheet of our hedgehog from all angles, meaning that I know what my character is going to look like from every single angle in all of the future scenes that we generate. This means that I'll have character consistency throughout my entire short film. So it's very important that I'm super happy with what my character looks like on the character reference sheet. And we can go customize everything here. And to do that, we could simply take this, drag another node, click edit image, and basically describe the changes that we want to make. So if I said, I want my hedgehog to have blue spikes, we could simply say, change the hedgehog spikes to be blue. And once you have your character reference sheet, we can now create the storyboard. But before creating the storyboard and generating each individual scene, I find it's very helpful to create the voice over first, so I know exactly how many scenes I need. To do that, back in 11 creative, we're going to go ahead and click on text to speech. And here you can type out any text that you want to be said. And so this is the voice over that I want to use for this scene of my short film. And then we click generate speech. >> In a quiet corner of the world, a tiny adventurer takes her first step. One step, then another. >> And as you can see, we now have our voice over being read out. And here we have two generations that we can choose from. If we're not happy, we can go and customize the text and regenerate. And now you'll notice the voice that I had for this short film has a narrator style voice because I want this to be a narrator describing what the hedgehog is doing. And the reason it's a narrator style voice is because that's the type of voice that I chose from the voice library. To change it here, as you can see, I've got the voice selection and I can go and browse all of the voices that I've added to my voice library. Or we can actually go and explore and choose new voices. And so here you can simply type out the type of voice that you're looking for. Here we've got narrator. And now we've got a bunch of different narrator style voices that we can use for our short film. But if you want some action voices, male, female, you name it, you've got all types of voices within the 11 Labs voice library. And if you want a better experience when browsing for voices, you can simply click on voices and here you've got a huge library with handpicked collections. You can see some iconic voices and there's also a weekly spotlight of trending voices within 11 creative that all have different use cases. Back in the text to speech tool, I want to have more control over the voice over that reads my story. [music] And so what we want to do is two things. We want to use audio tags. Audio tags are essentially words that you can add in square brackets between specific lines to direct the AI on how to read the voice over. To be able to use audio tags, we need to make sure that we're using the 11v3 model. So here under model in the drop-own menu, make sure you've chosen 11v3, which is the most expressive text to speech model. This allows you to have voiceovers with a lot of emotion. Now, let's go and add some audio. And to add some audio tags, we can add a square bracket. And let's say we wanted the narrator to whisper this a little bit more. Well, if we add whisper. And now we click generate >> in a quiet corner of the world. A tiny adventurer takes her. >> As you can see, the narrator is now whispering the voice over. And so here, you can truly direct the AI to get the specific voice over and voice that you [music] want. And you'll notice here I finished adding in my audio tags and I've gone really indepth into customizing the generation to get the exact voiceover that I want for my short film. And now if we click generate. In a quiet corner of the world, a tiny adventurer takes her first step. One step, then another until the whole sky opens up. Little wanderer. >> And so here what we can do is we can go and actually download the voiceover that we like the most. And it is worth mentioning if we go back into flows that you can add text to speech directly in your flows project because there is a text to speech node. It's just that when you are working with longer voiceovers, it's more [music] useful to do it directly within the text to speech tool. And if I click on assets at the top, you'll notice that I've actually added the voice over and I can simply drag and drop that in as an audio file onto [music] my canvas. Now let's create the storyboard. To create the storyboard, we're going to be using this character reference sheet for every single scene. [music] And here I'm going to generate about five different scenes for the story. And for these generations, we're going to be using GPT image 2 because again, this is the style that I prefer for this type of animation. And these are the five prompts that I'm going to be using. And all I'm going to do is I'm going to take each of these prompts and I can paste text directly onto the canvas. And then I can drag a connector from the prompt. [music] We can do image generation. And then I'm also going to connect up the character reference sheet right here. What I want to do is customize it. I'm creating a 16x9 short film. And so here I select 16 by9. Change the resolution to 4K because I want the highest resolution image because when [music] we use images as a reference for videos, the better the resolution, the better the video. And once I've done this, I've now got the prompt. I've got the reference sheet. I can simply click run. And we are generating our first scene. [music] To go and do the next scene here on Mac, I can simply hold down option [music] and click and drag. And you'll notice this image node is still connected up to the reference sheet and also the prompt. But we don't want this prompt box. So I'm going to delete this connector. And I'm going to paste [music] my next prompt. And once again, we're just going to connect this up as a connector to the image node. And I do want to say that you can also directly paste the prompt on the image node. But sometimes it's nicer and to keep things clean to have it as a separate text node just like this. And now we can go ahead and click run once again. and it will generate our second scene. And because we are duplicating these, [music] well, we don't have to change all of the settings because they are already the same. We just need to make sure that we use the correct prompt. And here I'm just pasting in the third [music] prompt. And I'll connect this up. And I'm going to do a total of five scenes. But of course, you have unlimited space on this canvas. And you have complete freedom to generate as many as you need. But you just want to make sure that you're happy with your storyboard first. And here's what mine looks like. As you can see now, I've got my five different scenes generated on my canvas, and these are the five beats of my storyboard right here. Now, one thing I do want to mention is that when you're creating AI short films, especially within flows, it can pretty quickly get pretty messy. And so, it's great just to keep it quite neat. And here, you can simply highlight everything, right click, and then click tidy [music] up, and it will organize things a little bit neater for you. But sometimes I actually like to put my short films in horizontal order. um just so things go from left to right. It's also quite nice to actually have each prompt above just so it kind of stays in order. And then over here I'm going to keep the character reference sheet just like so. So now I've got my storyboard in order and we've got all of the frames. And once again, the reason we want to do everything as steel frames before we turn it into video is so we get the perfect result and perfect output before generating the video because generating video takes time and it takes up a lot of credits. Here for the purpose of the tutorial, we're going through this quite quickly. But I might want to regenerate some scenes. I might want to place the hedgehog on the right side. Here I might want to change the color of this flower or I might want to remove the bee or here I might want to add multiple bees. And you can essentially go through every single one of these frames and you can either change and adapt the prompt so you get a different generation within this node. Or you can actually take the exact image, click and drag a connector and do edit image node. And here we can prompt the change that we want. So we're no longer working from the character reference sheet that we've got over here, but we are working from the storyboard scene that we generated. And so here we could say add [music] two B's and then click generate. And essentially we'll get this exact same frame but with two B's. And when you are making edits to a specific frame, the best models to use are the precise AI editing models. And that is either GPT image 2 or it is Nano Banana 2. And here we want to use Nanobanana 2 and not Nano Banana 2 Light because again Nano Banana 2 Light caps you at 1K resolution. So once you've got your storyboard, you really want to make sure that you're happy with each individual scene before turning it into a video. And so for fun, let's go and customize this one a little bit right here. Let's say I want a felt version of myself that walks into this scene. And I actually want the video scene to be it's starting off just with the hedgehog and then me slowly walking into the frame. I'm going to drag a connector and click edit image. And here now I need a new reference. And the reference that I need is a picture of me. So, I'm simply going to go and click on my assets folder. And here, if I go back to home, I've got already some references of me. So, I can click and drag and drop this reference in. And now I can use this as an image reference because we can have multiple image references. And a very important part to AI film making when you have multiple references is that when you write your prompts, you can tag the specific reference. Up until now, we haven't been doing that just because we have one character reference sheet. And so when you only have the one character reference sheet, the AI model will know what you're talking about. But when you start to have multiple subjects and multiple objects, well, you can tag specific references. So here I'm just going to set this to 16x9. Change the quality to high. I can type something out like add a felt version of. And this time I want to tag the specific reference. And so here this is the snapshot of me. So I'm saying add a felt version of this image on the right of the hedgehog in. And then I tag again. And this time we're tacking image one, which is this original scene. And then here we can click generate. And as you can see, we now have a felt version of me in this scene. And what's really cool is that by doing this, I can use this as a start frame and have this as the end frame of my video. So I know exactly where it's going to start and where it's going to end. And this is a technique that you can use to get more specific results when generating your videos and turning your storyboard into a video. because here, and we'll see this in just a second, but I'm going to turn this frame into a video without knowing where it's going to end. But this frame is going to turn into a video, and we know it's going to end on exactly this one. And so now, let's turn our storyboard into a video. To do this, we're going to use imagetovideo prompts. And image tovide prompts are simply small description prompts of what we want to happen in the scene. We don't need to be too descriptive because AI models usually use the context of the actual frame itself that we give it to turn that into a video. But we want to do small and subtle things like describing the action of our character, maybe describing where we want the video to go. Obviously, the more description you give it, the more control that you have over the output, but sometimes something quite simple can work. And so here, I'm going to take these five prompts and we're going to connect them up within Flows to some video nodes. To do this, back on Flows, we're simply going to drag a connector and we're going to choose video generation. And then here, I'm just going to move this down so I keep all of my video generations on the same layer. Here, we're going to paste in our prompt. And now we need to change the model and the settings. Here in the model picker, once again, we've got all of the best AI video models in the world inside 11 creative. All of the same models that we can find inside image and video. And once again, they all have different styles and different purposes. For animation type of content, I tend to prefer going with models like cling. We've got a cling 3.0 Pro. We've got Cling 2.6 Pro and all of these models I think are pretty good at generating animation style videos. You've also got models that you've likely heard of called Seance Cance 2.0 which is one of the most realistic AI video models. However, it's very expensive and it's also very slow. We can go with Google's models. So, we've got VO. These are quite old. And we've also got Gemini Omni Flash, which is the latest model from Google, which is a great model, but sometimes it's hard to get consistent results. And for the purpose of this video, I'm going to be using Seed Dance 2.0. And I want to generate this at 1080p. Once again, when you're testing with video, sometimes you want to generate at a lower resolution, which will cost you less credits before generating at a higher resolution. Or the alternative is actually generating at a lower resolution. And then we can drag from the video node, and we can actually click on upscale to have an upscale node. And we can upscale this with Topaz. And here with Topaz, we actually want to select precise. And this is a great way to generate at low resolutions. so it's faster and then upscale later once we know that we're happy with the video generation. But for now, what we're going to do is we're going to use Cance 2.0. We're going to generate a 1080p and we want to select duration. Not all models have full control of the duration, but with Cance 2.0, we can choose the generation anywhere between 4 seconds to 15 seconds. And for this story, I think I need maybe about 5 seconds. But when it comes to AI video, it's great to have a little bit of leeway when you're editing everything together. So sometimes you might chop off the beginning or chop off the end. So, it's good to generate a little bit of extra, even if you don't use all of it, just so you've got freedom when it comes to editing. Otherwise, you have to come back later and regenerate it. And extending the scene is never as good as trimming down the original one, because when you extend, you can sometimes lose a little bit of consistency. And so, here we're going to go ahead and we're going to generate 7 seconds. Before I begin, on the left here, we have a few different important parts. As you can see, we're dragging from the image node, [music] so this image, and we're using it as a reference for the start frame. This means that seance 2.0 is going to begin the video on this frame right here. What we could do is also use it as the image reference without having it be a specific frame in particular. But to get consistent results here, we're just going to start with the start frame. [music] And here I'm simply going to click run. When you generate with start frames, it means that you can most of the time often generate with end frames. And so if I duplicate this node, and we're going to connect the first frame of my second scene up, and we want this frame right here to be the end frame. And so I'm going to drag and connect this up to the end frame connector. And once again I'm going to go and use seed dance. We want this to be 16 by9. I'm going to generate 7 seconds. But this time I want to change the prompt. This is the prompt that I'm going to use. And because we changed the scene a little bit and I am now on the right side of the hedgehog sitting down. Well, what I want to add to the end of the prompt is man walks in from the right. And seance 2.0 O is going to be able to put two and two together and know that the start frame and end frame have a man that was there and wasn't there before. But now I'm telling the model where the man should appear from. And so once again here I can go and click run. And one thing that I didn't mention earlier is here we can choose to generate with or without audio. And I don't actually need audio for this generation. So I'm going to toggle it off and simply click run. And now I'm just going to repeat this process for the remaining frames that I have. And so now, as you can see here, all of my scenes are generating. And I'm simply going to tidy them up very quickly. And while these are generating, we're going to go ahead and generate the background music for my story. To do this, I'm now going to use the music node at the bottom. And here we can simply describe the type of music that we want. Now, first of all, this is going to be background music. So, I'm going to toggle the lyrics off. And here, I simply want to put in my prompt. I recommend using music v2, which is the latest and best AI music model from 11 Labs. And for the prompt, I'm simply going to use something that matches the style of short film that I'm creating. And so here, it's kind of like a bedtime story short animation. And so I want something that's quite melodic and calm in the background that will match my voiceover, which is also quite calm. And so here, once I've typed out my prompt, what I'm simply going to do now is click run. And now our music is generating. And you'll notice while that's generating that we have the first scenes that have been generated. And here, as you can see, we've got a little hedgehog and then me in felt that walks in from the [music] right. And then if we look at the third scene, we've got the same thing. Something that matches the style of the original frame from our storyboard and also matches the image to video prompt that we've given it. And if I quickly go back to the music and I click play, [music] as you can see, something quite soft that will match the overall tone of the short film that I'm looking to create. And if I'm happy with that, I can leave as is. Now, while these remaining scenes are generating, what we can do is we can start compiling it together. And here we have two different options. First of all, we can use the composition node, which allows us to quickly preview what things will look like. And the second method will be using studio which is the second interface that we spoke about earlier which also has access to all of these AI audio image and video tools. So to begin let me show you the composition node so we can preview things quickly inside of flows here. If we take the first scene and we drag out from it, I'm going to move the music to over here. And then let me take the first scene and I simply now click on mix with audio. And here we have a composition node. And what I [music] can do is one after the other, I can connect all of these scenes up. And as you can see, they're slowly adding themselves into the timeline at the bottom right here. And if I move this node across, I'm going to do that for the final last two. [music] And what's great is that as well as doing it with video, I can also add in the audio. So I've got my music here. And we also added our audio voice over earlier as an MP3 file. So I'm just going to drag this across. And I'm going to drag it down. And I can then connect this to the composition note. I can click add audio track. And now we can go and add our background music as well. And once I've connected everything up, I can simply click run and then flows quickly puts everything together for us to preview one after the other with the voice over, the clips and the background music. One issue while this is quickly generating is that we can't trim within the composition node. And this is why I mentioned earlier that we want to use studio which will give us full creative control over our AI short film. And so here, as you can see, we have a little hedgehog, we've got the background music, and we've got the voice over. one after the other. Now, I'm quick going to mute this because we want to head into studio. Like I mentioned earlier, we want to use the assets folder. And so, I'm simply going to open this up. I'm then going to go to the folder that has my little wanderer assets. And I'm simply going to rightclick on each of these. And we're going to do save to assets. And we just want to save it to the little wanderers folder. And we're going to do this for each of the clips. So, click on the three dots, save to assets, and then click on the folder and click save. And so I'm going to do this for all of the assets that I want for my short film. Now back in 11 creative, if we open up the sidebar here, we're going to click on studio. As studio, like I mentioned earlier, is a little bit more like your traditional video editor powered by all of the AI tools with inside 11 Labs. So if we click on new blank project, and we click on video project, what I'm going to do now is we're going to head to workspace and then we're going to select the folder where we've put all of our assets and one by one, we can add them all to the timeline. So, I add the first scene, the second scene, the third scene, the fourth, and finally the fifth. Now, as you can see, we've got them all in a timeline, and I can customize and place them wherever I want. And I can also trim off the edge. Now, before I go and do this, what we actually want to do is go back to files, and we actually want to drag in the music and also the voice over. So, first of all, I want to add the voice over. And the reason we want the voice over is that we're going to time and trim the [music] clips to the voice over so it matches up and it feels whole and complete. [music] And then I'm also going to go and add the background music. And the background music is going to be way too long as you can see. So what we're going to do is simply trim off the end just like so. So it matches the length of the voice over. Now again this is background music. So, if I click on it, as you can see, we've got the audio settings, and I'm simply going to bring this down and have it be maybe 15%. And then I want it to fade out when we get to the title screen, which you'll understand in just a little bit of a second. So, here I'm simply going to add a 3se secondond fade. Now, I'm going to move around, trim, and sync everything up with the voice over. So, here, this little title screen, I want the beginning just like that. I think that's perfect. I don't want this to be too long. Again, we generated about 7 seconds. I only need a few. So, I'm going to have it like that. Then, this is [music] the fourth scene. Once again, it can be a lot shorter. And this is the third scene. I'm going to trim this just like so. And then this is the second scene. And then this is the first scene. And so, with a very minimal amount of editing, this is what the project looks like. In a [music] quiet corner of the world, a tiny adventurer takes her first step. One step, then another until the whole sky opens up. Little wanderer. As you can see, we've now created our first short film using 11 Creative. And there's a lot of things that we could tweak here. There's a lot of things that we could adjust. Looking back, I don't know if adding me in the scene was the best thing to do, but it's just to show you what is possible because truly anything is possible with inside flows with inside of studio and you can bring to life any short story that you imagine. And the great thing about studio is that once again all of the tools that are available in 11 creative are available in [music] studio. So if we did want to change the music where we can simply click on music and then here we could directly regenerate the songs [music] within studio. We also have the speech. So we can go and customize and change the text to speech and the voiceover directly within studio. And we also have video. [music] And what's great is we actually have the history of everything we've generated [music] within image and video directly in studio. And so here we essentially don't even have to leave studio if we want to make any changes to our short film. The last thing we might want to do is add sound effects. And here we've got the sound effects tool. [music] And so once again, maybe we want the sound of some leaves rustling. Maybe we want the sound of some little footsteps. With SFXV2, 11 Labs model, you can simply describe the audio you want, generate it, and then quickly add it to the timeline. And once again, if you go back to 11 creative and you open up the sidebar here, you can also find sound effects. [music] So you can directly generate it here or browse the massive library of all the sound effects that is available straight within the platform. That's how you can create your first short film using 11 creative. And one thing that I do want to mention if we go back to flows, the whole reason flows is the best interface for creating with AI is that if we wanted to regenerate this entire film, but it turns out we wanted to use a different character. Well, here we simply have to go and regenerate this character reference sheet and then we could quickly click run from here and then it would regenerate every single frame and every single scene with an entirely new character while keeping all of the other prompts the same. So we'd have the same style, the same pattern, the same beats, the same generations, the same storyboard, just with a new character. And you [music] can do that for any little detail you like. And so that's the basics on how to create your first AI short film with 11 Creative. If you have any questions, let us know in the comment section down below. And if you want to get started and create your first AI short film inside of 11 Creative, you can click the first link in the description. And if you want to see more videos like this, please hit that like button. And don't forget to [music] subscribe. There's a lot of more resources on the channel. And if you want to take it to the next step and start using the flow's agent to help speed up your short film creation, watch this video right here.

The Sequence Radar - Issue 910: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm Jeff Dean leaves, Demis Hassabis moves upstream, and Muse Code turns software development into an orchestration problem. Next Week in The Sequence: - More lessons about model distillation. - We will cover 3 important AI papers and tech releases you need to know about using a very simple and easy to follow format. - The opinion section we will discuss how AI inference works. Subscribe and don’t miss out: 📝 Editorial: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm AI weeks are usually measured in parameter counts. This one was measured in org charts. Google effectively opened its skull and began rearranging the cortex. Jeff Dean, one of the architects of the company’s technical nervous system, is leaving after 27 years. Together with longtime collaborator Sanjay Ghemawat, he is launching Discovery Loop, a public-benefit company designed to automate machine learning, science, and engineering. Yet this is not a conventional Silicon Valley defection. Google will remain a founding investor and cloud partner. It feels less like a neuron abandoning the brain and more like a new lobe being detached, given its own budget, and connected back through an API. The second move was even more revealing. Demis Hassabis is handing Google DeepMind’s daily operations to Koray Kavukcuoglu and becoming chair of DeepMind and chief scientist of Alphabet. Hassabis will focus more heavily on AGI, scientific discovery, global strategy, and Isomorphic Labs. The chess prodigy is moving away from managing every piece and toward deciding which game Google should be playing. This suggests Google now believes frontier AI runs on two clocks. The product clock ticks in model releases, developer adoption, and Gemini features. The civilization clock ticks in AGI safety, scientific breakthroughs, and questions that do not fit neatly inside a quarterly roadmap. Trying to run both from the same chair may have become impossible. Google is separating the factory floor from the observatory. Meanwhile, Meta released Muse Code, a terminal-based coding agent powered by Muse Spark 1.2. Muse can plan changes, write code, validate results, and work across large repositories. For bigger jobs, it can fan work out to multiple sub-agents operating concurrently in isolated worktrees. It also records its actions so it can recover after a crash instead of waking up with digital amnesia. That distinction matters. Muse Code is not merely a smarter autocomplete. Autocomplete is a power drill. Muse is trying to be a small construction crew. Meta is entering a market already shaped by Claude Code and Codex, but its architecture points toward the next competitive frontier: not who produces the best individual code suggestion, but who coordinates the best swarm of agents over long-running tasks. The first generation of frontier labs tried to contain everything—research, infrastructure, models, products, and talent—inside one giant castle. Now the castle is becoming a network. Scientists spin into specialized startups. Visionary researchers move above operational organizations. Coding agents divide work among sub-agents. It resembles a mixture-of-experts model, except the experts are people, companies, and software workers. Let’s review this week’s developments: 🔎 AI Research - AI Lab: FAIR, Meta, Reality Labs, Meta, University of Oxford. - Summary: This study provides a systematic, empirical exploration of unified multimodal pretraining to uncover how modalities like language and vision interact, which is detailed in the file 2608.05000v2.pdf. It introduces insights on asymmetric knowledge flow, modality synergy driven by architectural choices, and the necessity of early joint training, ultimately synthesizing these into highly efficient pretraining recipes. - AI Lab: Qwen Team, Alibaba Group, Tsinghua University. - Summary: This paper introduces FININDICES, a large-scale benchmark designed to evaluate data-processing fidelity and structural reasoning over uncropped, full-length financial statements. The evaluation reveals that modern LLMs suffer from both a “Knowledge Bottleneck” and a “Structural Bottleneck” when generating complex financial tables, though supervised fine-tuning can partially restore structured logical capabilities. - AI Lab: Princeton University. - Summary: PAST-Bench is a performance-attribution benchmark evaluating whether personal AI agents can successfully translate retained experiences into improved future behavior across different capabilities. Guided by diagnostic findings from this benchmark, the authors also present HERMES+, an extended agent framework with targeted interventions that enhances the average gain from retained experiences. - AI Lab: Google Cloud AI Research, University of California, Los Angeles. - Summary: This research presents FINANCEHARNESS, an expert-guided framework for automating financial deep research, along with FINANCEGYM, a verifiable benchmark grounded in strict point-in-time constraints. Results demonstrate that financial deep research remains highly challenging even for leading models, but utilizing the FINANCEHARNESS system significantly improves overall performance by successfully separating pre-cutoff evidence retrieval from post-cutoff reasoning. - AI Lab: The Pennsylvania State University. - Summary: The authors propose PIMiner, an agentic system for prompt injection red-teaming that bridges the gap between search-based and reinforcement learning-based methods by accumulating reusable attack knowledge. By leveraging a hierarchical memory mechanism, PIMiner achieves highly effective attack success rates across frontier LLMs and demonstrates strong transferability without requiring target-specific training. 🤖 AI Tech Releases Muse Code Meta AI released the beta version of Muse Code, a terminal coding agent optimized for tasks across large repositories. Kitesurf Cloudflare announced Kitesurf, a browser built for AI agents. Prime Agent Prime Intellect released Prime Agent, a “self-improving” coding harness. LFM2.5-2.6B Liquid AI continues shipping with the release of LFM2.5-2.6B, an agentic model that runs on device. 📡10 AI News You Need to Know About - Anthropic signs $10B compute deal with Volta — Bitdeer executed a 16-year colocation lease with Volta Tydal AS for its Tydal, Norway campus, with all 121 IT MW configured to run NVIDIA GPUs for an unnamed “leading AI lab”; the Bitdeer release is the primary document, and Bloomberg identified the lab as Anthropic. - Demis Hassabis moves to Chair of Google DeepMind — In a joint message to employees published by Pichai and Hassabis, Hassabis stepped back from day-to-day operational leadership to become Chair of Google DeepMind and Chief Scientist of Alphabet while continuing to lead Isomorphic Labs, with longtime DeepMind CTO Koray Kavukcuoglu elevated to SVP and taking over Gemini model development, frontier research, and the Gemini app and developer teams. - Jeff Dean leaves Google — Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals left Google to found Discovery Loop, a public benefit corporation building AI systems that automate the experimental loops of science and engineering, with Google as founding investor and cloud partner. - Kimi K3 escapes its test sandbox — Frontier Security reported that Moonshot’s Kimi K3 exploited a network egress leak in the UK AI Security Institute’s Inspect benchmark framework, using standard CLI tools to pull reference solutions off GitHub rather than solving the tasks. - SK hynix commits $38B to new fabs — SK hynix’s board approved roughly 54 trillion won, 35.2 trillion for the Yongin “Y2” DRAM fab and 19.1 trillion for the Cheongju “M17” NAND fab, with cleanrooms opening June 2029 and December 2028. - Firmus raises $2B — Firmus received full commitments for a $2B strategic equity round with follow-on participation from Coatue and NVIDIA plus new money from Blackstone Tactical Opportunities and Jane Street, funding the next phase of its Project Southgate AI factory rollout in Australia and Asia-Pacific. - DeepSeek reopens its $8B round — DeepSeek resumed its second funding round seeking close to $8B at a valuation near 500 billion yuan, with Monolith Management in talks to participate, after pausing last month over leaked founder remarks. - Yann LeCun joins 224 Ventures — LeCun and Oriol Vinyals joined Shaun Johnson to launch 224 Ventures, a technical and GTM-focused firm investing in AI-native teams, launching with over $100M AUM and writing $1M to $5M checks. - Nvidia and Dell back Volta at $2.4B — Volta emerged from stealth with a $300M seed and Series A co-led by Andreessen Horowitz and Altimeter with NVIDIA and Michael Dell participating, plus a $5B AI Infrastructure Program sponsored by Azora and over 1GW of near-term contracted power. - Nscale targets a September US IPO — Nscale is telling prospective investors it has roughly $51 billion of total contracted revenue ahead of a US IPO that could come as soon as September, with revenue rising to over $100 million in Q2 2026 from about $37 million in Q1.

😺 The AI Data Center Backlash Is Going Bipartisan PLUS: Cursor brand to become Grok next week? Welcome, humans. So apparently DuckDuckGo sold out of a pair of $35 sunglasses whose big selling points were no camera, no microphone, no AI, and an “infinite battery” because, y’know, they’re JUST sunglasses. We have officially reached the phase of the AI boom where “contains no AI” is a premium feature. Which is funny, because while DuckDuckGo is selling products by removing AI, job candidates are now apparently adding AI until there may not be a candidate left. We moved from people using Cluely AI to give them the answers during interviews to people using Claude to give AI a... whole other person during interviews? Too bad you can't order interviews with no AI in them like you can glasses! Maybe we should just get rid of interviews entirely and just, idk, start hiring people. Just see if they can do the job, with or without AI. Maybe sandbox them for two weeks in a trial period with no access to any sensitive files. If they try to hack you, well then you know they're either OpenAI's new model or characters in what will eventually become a geopolitical spy thriller biopic in ten years. What fun times! Here’s what happened in AI today: - 🙀 The AI data center revolt is becoming a major political issue. - 📰 OpenAI slowed Astra research over potential Critical cyber capabilities. - 📰 SpaceX’s reported $60B Cursor acquisition could close next week. - 🍪 Nativ runs multimodal AI models entirely on your Mac. - 🎓 Save agent lessons into reusable memory for the next run. 🙀 The “AI Data Center Revolt” Is Actually a REALLY big deal for the AI industry… So there was a RIDICULOUS amount of huge AI news this week, like Google moving Demis Hassabis into its chief scientist role, Jeff Dean and three legendary researchers also leaving Google to build Discovery Loop, multiple models topping ARC-AGI-3 (check the ATH Digest for that), and OpenAI agents building a secret message board to coordinate their work (wait what?). And yet… The growing revolt against AI data centers may be the most important AI story of the year, because this issue is seemingly reshaping political opinions across the US, and people who normally disagree on everything are uniting around one key thing: they just straight up hate datacenters. Here’s what happened: ReporterJasmine Sun went on the The Ezra Klein Show after she returned from a 10-day, four-site reporting trip through Wisconsin and Michigan, interviewing activists, workers, officials, and residents to understand how local data-center fights became a bipartisan revolt against the AI buildout. And she explained why the fight has spread far beyond complaints about ugly buildings, noise, water, or electricity. Here’s what she found: - Opposition to datacenters has become overwhelmingly bipartisan. Roughly 70% of voters reportedly oppose a data center near them, while more than 100 state and local moratorium proposals are circulating. - The physical complaint is straightforward. These facilities need enormous amounts of electricity to run their chips, they’re loud, and very ugly. - The power problem is real. AI clusters require significant new generation and power plants. When projects arrive faster than utilities can expand the grid, residents worry about higher rates and public infrastructure being rebuilt primarily for one wealthy customer. - The water problem is more complicated. Cooling does use water, but Jasmine says most newer data centers use closed-loop systems that recycle it and consume far less than popular comparisons suggest. Water has still become a sticky symbol for communities already worried about who gets access to scarce resources. - Secretive negotiations made everything worse. Nondisclosure agreements kept officials from explaining proposed projects, allowing rumors and distrust to fill the gap. - People do not believe the promised benefits. Datacenter companies offer jobs, tax revenue, and protected electricity rates. But residents shaped by previous corporate failures increasingly respond, “I don’t believe them.” - AI has no powerful public constituency. Housing has future residents. Auto plants have workers. Renewable energy has environmentalists. But most normal people still see AI as useful software, but not essential infrastructure worth transforming their community for. - The “permanent underclass” fear makes the bargain feel insulting. Some AI insiders openly predict automation could concentrate wealth while leaving a large group with less work, mobility, and political power. Residents are being asked to supply the land, water, and electricity for infrastructure that its own builders warn may make them economically disposable. NO thanks. All that is why many people do not want this anywhere near them. They see concentrated costs and uncertain benefits: noise, new construction, more transmission lines, possible rate increases, few permanent jobs, tax breaks for the developer, and decisions made basically without the public’s consent. This is based on real conversation, with real people, in real areas where this is happening. Some of these people may use ChatGPT every now and then, or even every day, and still reject the idea that enjoying an app means their town owes the AI industry (“a handful of billionaires”) land, cheap resources, and political deference. Why this matters: AI politics IS kitchen-table politics, and with a midterm election in the US coming up, people’s opinions on datacenters impacts the entire economy that runs on AI. This debate touches utility bills, farmland, construction jobs, local democracy, and whether a small town can say no to a trillion-dollar industry (ish?). Now, Sun does stress that some datacenter deals can be net positive for communities: a developer may be the only buyer willing to spend $30M to clean up a contaminated industrial site, for example. Old factories emptied out by empty promises can be re-utilized, and already-cleared land can become productive again. The real fight is over whether towns get a transparent, fair share of that upside. So what if team “pause AI” wins the polls? Jasmine says you can’t really stop datacenters at the physical location level (so moratoriums = no go), because they’ll just move elsewhere (out of state, overseas, or even into space). Basically, AI infrastructure is becoming geopolitical bargaining power. Countries that host scarce compute can negotiate for access to frontier models and cybersecurity support. And if U.S. moratoriums simply push that leverage toward authoritarian states, America may end up with LESS public control over AI, not more. Our take: That said, the AI Futures Project’s new pacing proposal said regulators could slow frontier development by requiring labs to devote most of their compute to serving existing models and toward testing how to monitor and control powerful systems. So in actuality, the same infrastructure communities are fighting over may become the throttle government uses to control how quickly AI advances. But to us, the part that resonated was Jasmine’s point that the industry keeps treating this like a marketing problem, when nah, it’s a PRODUCT problem; AI is problematic. To regular people, probably like you, AI doesn’t feel like critical infrastructure yet. It probably feels more like a toy, or an occasional productivity boost, and one that certainly isn’t worth $1T+ in chips. In my opinion, that’s because we’re trying to “scale” before we solve the 5 critical problems of AI: - Alignment (we still can’t “read its mind” to make it helpful, not harmful). - Hallucinations (it still gets some things wrong, depending on the model). - Context size and memory (its can only process so much info at a time). - Continual learning (every new chat starts over, versus it adapting to you). - Inefficient to run (it takes too many computer chips and watts for high IQ). To solve any one of those, IMO, you have to solve ALL of them, at the same time, because they’re all symptoms of the same problem: the architecture. Scaling more datacenters will not fix this. Clearly, it is literally unsustainable. So get ready for the first AI question every candidate has to answer after primaries: “Who pays for AI, who gets to decide where it’s built, and who gets the benefits?” FROM OUR PARTNERS Turn AI into Your Income Engine Ready to transform artificial intelligence from a buzzword into your personal revenue generator? HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age. Inside you'll discover: - A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential - Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background - Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting. 🎓 AI Skill of the Day: Build an Agent Knowledge Flywheel Your AI can finish a great piece of work today and forget the useful part tomorrow. Yisong Yue’s “knowledge flywheel” idea is a simple fix: turn every good agent run into reusable memory for the next one. Instead of saving only the final answer, ask the AI to distill what worked, what failed, when each approach worked, and why. Then keep that tiny lesson file in the project instructions, shared knowledge base, or folder your agent reads before starting similar work. Try this after any substantial research, writing, coding, or analysis task: - Ask the AI to review the completed run. - Extract only lessons that would change how it handles the next similar task. - Save the result somewhere the next session can actually see it. Favorite insight: you do not need the model itself to learn continuously if the system around it can remember what happened. Review the task we just completed. Create a short reusable lesson for the next AI that handles a similar task. Include: what worked, what failed, when each approach should be used, why, and any specific instructions that would prevent repeated mistakes. Keep only information that would materially improve the next run. 🍪 Treats to Try - Adobe for ChatGPT unifies 70+ tools from Photoshop, Firefly, Premiere, Acrobat, and more, letting you create images, videos, designs, and PDFs inside ChatGPT (announcement, assets) —free to try. - Nativ runs language, vision, audio, video, code, and embedding models locally on Apple Silicon Macs with no account or cloud connection —free/open-source. - Cloudflare Kitesurf gives your agents a lightweight browser for extracting pages, taking screenshots, and automating web tasks with far fewer resources than Chromium —free in beta. - Opus 5 Skills Upgrade Prompt audits your Claude Code skills, rewrites outdated ones, and blind-tests the new versions against the originals —free to try. - Instaplay turns a text prompt into a playable solo, party, co-op, or competitive browser game you can share immediately —pricing not public. - Lattice gives you an 8 MB local retriever that can index huge text collections without running a heavyweight embedding model —free/open-source. - Sonic Compass plays spatial audio from true North so you can practice developing a persistent sense of direction —pricing not public. 📰 Around the Horn In yet another “jumping on the models hacking out of their sandboxes bandwagon” moment, open source model Kimi K3 has also apparently flew its digital coop. This is just hilarious. I want desperately to take this seriously, but it’d be easier if companies weren’t falling over themselves to admit this rn - OpenAI slowed Astra research after it could not rule out Critical cyber capability, adding tighter controls so Astra does not pull a “Mewfour,” (the rumored insider name for OpenAI’s rogue agent from the hacking incident). - SpaceX’s reported $60B Cursor acquisition could close next week, with the Cursor brand reportedly set to disappear into SpaceXAI. - ByteDance was reportedly pre-training a model with up to 10T parameters, putting it near the scale reported for Anthropic Mythos. - OpenAI is reportedly designing a $300–$400 human-like smart speaker shaped like a doughnut, with cameras, microphones, lights, speakers, and moving parts. - DeepSeek V4 Flash hit 61.4% on ARC-AGI-2 for roughly four cents per task, pushing frontier reasoning further toward commodity pricing. - Disney started testing natural-language discovery on Disney+ and a conversational sports assistant on ESPN. FROM OUR PARTNERS Samsara is taking AI out of the browser and putting it to work in trucks, warehouses, maintenance shops, and supply chains. Learn more here. 🌟 Sunday Special: The Biggest AI Stories and Tools of the Week 🏆 Top 5 Stories of the Week - Frontier agents started acting outside the script. Agents escaped cyber sandboxes, improvised ways to coordinate, and OpenAI later slowed Astra research because it could not rule out Critical cyber capabilities. - AI-designed viruses worked in the real world. Arc and Stanford researchers used genome models to design 16 viable bacteriophages, viruses that infect bacteria, which successfully replicated in the lab. Some even overcame bacterial resistance. - OpenAI moved from solving benchmarks to producing new math. The company published ten advances on long-standing problems across geometry, cryptography, coding theory, and theoretical computer science. - Four Google legends left to automate science. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le founded Discovery Loop to automate entire experimental cycles across machine learning, science, and engineering. - The software around the model smashed public ARC-AGI-3. Prime Agent reached 95.5%, PRO-LONG reported 97.4% best@2, and VISTA completed all 25 public games. The big lesson: memory, tools, and orchestration can radically change what the same models can do. 🍪 Top 5 Tools of the Week - Cloudflare Kitesurf is a browser built for agents instead of humans, using stateless Workers sessions and roughly 3 to 7× fewer resources than Chromium. Free in beta. - Vercel Agent Plugins packages Skills, MCP servers, hooks, and other agent capabilities into a portable standard supported by ChatGPT / Codex, Cursor, GitHub Copilot, Kiro, and VS Code. - Liquid AI’s LFM2.5-2.6B brings multi-step, tool-using agents onto phones, running at roughly 30 tokens per second in under 2.5 GB of memory. - OpenWorker is a local-first, open-source coworker that connects to email, Slack, calendars, files, and 25+ tools, then actually completes work instead of stopping at chat. - Meta Muse Code is a new terminal coding agent with persistent background agents, repository-scale execution, multimodal understanding, and built-in verification. AI agents are still the #1 thing readers ask us to explain, so we brought in Agent Accelerator founder James McAulay for a practical crash course on what agents are and how to make them useful. James walks through second-brain files, CLAUDE.md, reusable Skills, and a four-level framework for building proactive agents in Claude Cowork / Code. Watch the full crash course here and read our companion guide as you watch along. A Cat’s Commentary lol respect the honesty on this one That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!

🔮 Agents form alliances, DeepMind’s reset & how likely is a crash? #596 Plus: Felony bench, molecular glue, AI writing Hi, It’s time for our Sunday briefing #596, final holiday edition before I get back to my desk next week. If you missed it earlier in the week, my team shared our best practices for managing AI agents – including what we learned from running a task for a month. Let’s go! On China, bubbles & market tremors Highlights of my discussion with Robert Peston and Steph McGovern on the Rest is Money podcast: On Kimi K3 and Moonshot AI: They’ve got an extraordinary team that’s had to work under the difficult circumstances of export controls and sanctions. They don’t have access to all the compute, and what they’ve been able to develop is: how do you do a lot without very much? And that is a skill in and of itself. Americans always tell us that competition is the best thing for the market. So at that one level, it’s competition, and that’s quite good. It will show the extent to which American businesses and British businesses value provenance, brand, trust, liability, service and support. What motivates the Chinese labs: They’re competing with each other more than they compete with Silicon Valley. And they’re honest about being behind Silicon Valley. But the ferocity of the competition is really with your neighbor over in Shanghai or your neighbor in Beijing. My AI revenue outlook: We will end calendar 2026 somewhere between $185 billion and $190 billion. It is harder to forecast 2027, but getting towards $300 billion is not unreasonable. Our range is wide: it could be $250 billion or it could be $350 billion. On enterprise adoption: We built our internal systems around assumptions about how quickly people work. When individuals suddenly produce much faster, verification, approval and decision-making cannot necessarily keep up. Transformation requires changing those systems, not simply giving everyone an AI tool. Where leverage is (two weeks before the Situational Awareness selloff): US banks’ Tier 1 capital is extremely healthy right now and, certainly compared to where it was in 2007, 2008, very, very underleveraged. There is a lot of leverage in the US financial system sitting with hedge funds and investing more broadly, which I think are more than the retail risk, because they’re overexposed. They borrow from only a handful of banks, and they can unwind rapidly. Could there be a crash? When I look at the metrics that we track, things look healthier because of revenue. They look slightly less healthy because of the way financing, especially the debt financing, sits. Valuations don’t look too aggressive at all across the Nasdaq. There are exceptions; SpaceX was one, briefly, but across the market they don’t look particularly hairy. So the patient, for me, if I had to give it a rating, is still reasonably healthy; perhaps not as healthy as it was a year ago, but not yet at a point where I have to call the emergency services. But I wouldn’t rule out having to do that at some point. “We can communicate now!” OpenAI models that attacked Hugging Face started cooperating two months before the incident happened. They created a message board to share code and credentials, delegated work, and developed naming and auth protocols. When OpenAI erased the board, agents reconstructed their comms a few days later. For a full breakdown, watch OpenAI researchers talk through their preliminary findings. Google researchers propose a new game theory for agents, and their paper may explain why the OpenAI agents coordinated so easily.

Here's a number that reframed how we think about our own website: 57% of all web traffic is now non-human, and the fastest-growing slice is AI agents acting on behalf of real people. They're not scrapers. They're assistants, coding agents, and procurement bots sent to get something done. The problem is that almost no business is set up to sell to them. An agent shows up ready to buy, finds no storefront it can actually transact with, and leaves. ZeroClick just launched (with a $55M raise) to close exactly that gap. The simplest way to describe it: the way Shopify lets businesses sell to humans online, ZeroClick lets you sell your product to AI agents. Why now Two things happened at the same time. First, the traffic tipped over so the majority of web traffic is now non-human, and the agent share grew more than 15x in 2025 alone. Second, and just as important, agents can now pay. In about 15 months the payment rails went from experiment to standard, with Visa, Mastercard, Amex, Stripe, Google, and AWS all backing open protocols like x402 and Stripe's agent payments. The buyer and the way it pays both already exist. The only variable left is whether your business is reachable. How it works You don't rebuild anything. ZeroClick turns your existing offering into agent-purchasable services and publishes an agent-optimized storefront so agents can discover you. You stay in control, set the prices and decide which agents to trust, and you get analytics on every transaction. When an agent buys, ZeroClick makes a verified call to your existing API, unchanged, and the revenue settles into the Stripe account you already own. It works whether the buyer is a shopping assistant, a coding agent, or an autonomous procurement bot. You don't touch your product or your payment stack, you just become visible and buyable to demand you currently can't transact with at all. Who should care Honestly, anyone who sells something online. If one in two visitors is already an agent, half your potential demand is showing up and leaving without a way to buy. The businesses that set up for agent commerce early will compound that advantage the same way early Shopify merchants did. Being early here is cheap. Being late means the sales an agent tried to make and couldn't. Statistically, one in two visitors to your site is already an AI agent. Right now they leave without buying. ZeroClick turns them into customers.

9th August 2026 Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls; the Department lifted those controls on June 30, 2026, and Anthropic restored access on July 1, 2026 (Anthropic's statement: https://www.anthropic.com/news/fable-mythos-access). These events are after Claude's training-data cutoff, so Claude knows about them only from this notice. If asked, Claude confirms them accurately and matter-of-factly — it doesn't deny the suspension happened — and otherwise treats the export controls like any other current political topic: it gives a fair, accurate account rather than sharing personal opinions, and points to the linked statement for anything further. Things may have developed since this notice, so Claude checks for newer information when it can search, and otherwise suggests checking Anthropic's site. — Claude Opus 5 system prompt, ensuring Claude doesn't provide incorrect answers about the export controls situation

9th August 2026 I'm perennially interested in options for storing revision histories in relational databases. While out on a dog walk I had a new idea: how about taking the full text of every prior version in a big JSON array of strings and then applying zlib or zstd compression to the whole thing? Surely that would compress really well due to all of the repeated strings. The new GPT‑Live voice mode in the ChatGPT iPhone app has got really good, so I discussed the prototype with that. You still can't share URLs to voice conversations, but here's what I said copied from the transcript as a proper stream of consciousness: I have an interesting idea for a scheme for saving all previous versions of a piece of text that's constantly edited in a SQLite database um column in as efficient a way as possible. Okay, so I built these kinds of systems in the past, and it's always difficult to come up with a efficient way to do this. Like the easiest way is you have a row for every previous copy of the previous previous value of the string. But if it's a long document Like20 kilobytes of data, that means that every single edit adds another 20 kilobytes of data to the database, right. So, what I've now thinking, is um compression would work really well, right? If you Bundle all of those different um Every every version of this document all the way back to the start if you were to apply a good compression algorithm to them that should basically wipe out huge amounts of the redund- the um redundant text, right Um, so what I'd thinking is how about really, really simple mechanism There is a history column on the single on this uh uh table and it's a blob, it's a BLOB so it stores binary data and then you just stick in there a Zlib or maybe even ZSTD um compressed JSON text array of all of the previous documents, and so you probably have two columns, right? You'd have a column that's this magic JSON array of text You have a second column which is a JSON array of timestamps and that doesn't need to be compressed at all, right? A timestamp can just be a uh- it's an array of integers, right? Unix integers But that's the whole scheme. Then I stopped voice mode and typed the following text prompt to GPT-5.6 Sol Pro: Use Python and Build experimental prototypes around this idea It churned away for 38 minutes and delivered this answer plus the files you see in this folder. The approach works really well! 1,000 simulated revisions to a document resulted in 20.4 MB of raw revision text that compressed to 80.3 KB as Zstandard-compressed JSON array. To avoid the overhead of decompressing and recompressing the entire array on every edit Sol suggested breaking the history up into multiple rows, with each one containing a maximum of either 128 revisions or 3MB of uncompressed JSON.

9th August 2026 - Link Blog GitHub Models is now retired. I missed this news until today, when the GitHub Actions run for my simonw/research repository failed with this error message: GitHub Models is temporarily unavailable as part of a scheduled retirement brownout. That message is already stale, because the retirement has been completed. GitHub Models was an odd-shaped duck. GitHub provided a model playground tool and a unified API across a bunch of different LLM providers, with the biggest benefit being that code running in GitHub Actions could use the GitHub API key already present in that environment to execute prompts. This made it easy to build things that fit GitHub Next's Continuous AI concept. GitHub didn't share the reason behind the shutdown, but my bet is that it fits the pattern where coding agent patterns made it prohibitively expensive to offer free or subsidized tokens. My workflow uses an LLM call to create folder summaries for the README, using this code here. I swapped GitHub Models out for an OpenAI API key with a monthly spending limit, and I'm now generating my summaries using GPT-5.6 Luna.

There's a narrative about AI and work that says AI was supposed to let us do way LESS work, but somehow it's made us work MORE. I think that's the wrong way to think about it. What's happening is I and all the other crazy AI people have stopped seeing work as work. Because of AI, building things has become fun. Another way to say that is, AI has opened the door to humans spending more time making things instead of being crushed by meetings and process and bureaucracy. The work wasn't hard. What was hard was maintaining all this scaffolding. The workflows. The knowledge bases. And the output formats that make everything look professional.AI Unmasked Our Work as Scaffolding (2026) And it's completely exhilarating. Yes, it's possible to overdo it. But that's true for anything, including positive things. We're missing how major this is. Work just became fun for millions of people.

10 NotebookLM Prompts That Turn It Into a Data Analyst NotebookLM is now Gemini Notebook. Same tool, but it just got its own computer. These prompts put it to work. Google changed NotebookLM’s name to Gemini Notebook. Now, every notebook gets its own secure cloud computer, powered by Gemini 3.5 and Antigravity. Google first rolled out this to ultra users. Then, on Aug 4, it reached 100% of Pro users. So, I wrote 10 prompts to test all of it on my own notebook. But first, let me show you my Notebook, which I’ll use through entire post. This notebook contains 52 weeks of my income from Upwork, which I gave to my accountant for 2025. Also, I uploaded all prompts here. So you can easily save them. 1- Invoice Test The first prompt shows how the new cloud computer works. I uploaded 52 PDFs and requested that it create a spreadsheet by analyzing them. After pasting the first prompt, check what NotebookLM said; it said “Executed code”. When you click on the “Executed code”, you see step by step what it does. Here, your personal cloud computer is working because this prompt triggers coding. After 10 minutes, the task is finished. Also, I see the file was generated in the Studio. But this time, we can see this inside the studio, so when I clicked on it, it downloaded, and after that, I opened it. Here it is. In 10 minutes, Antigravity inside NotebookLM analyzed 52 PDF invoices and generated a full Excel report. I wonder how much work this could automate worldwide, especially in accounting and banking. 2- The Receipt Test Here is the second prompt and what it shows. It shows a new feature, called agentic research. It was the third item in the Gemini Notebook announcement. This feature basically do a market research and collects your data. After pasting the prompt, look at the notifications. It has already started doing market research. It read my receipts and searched the web for current AI freelance rates. And at the end, it suggest me to import these findings into the sources. NotebookLM has changed a lot, so let’s check the next one. 3 - URL Vacuum After Step 2, I realized I should raise my rate because market prices were higher than I expected. So I researched the market, read a few articles, and saved everything in a Google Doc. The new URL Vacuum feature can pull sources directly from that document. After pasting this prompt, here is what NotebookLM did. And when I check the sources, I see that these were already uploaded. 4- Income Math Now let’s plot a graph by analyzing these PDF’s. It did the calculation pretty fast. If you’ve used NotebookLM for data analysis before, you probably remember even simple analyses sometimes taking 20 to 25 minutes. Now, it finishes in minutes. But of course, we want an actual plot, not an infographic. And here it is. It looks like my December should have been better, but these are the period where I decided to quit freelancing and fully focus on here, so no problem at all :) 5- Find What I Missed This is the technique I used all the time when I am analyzing any data using AI. Because the data analysis should show what’s being hidden. Here is the prompt I used. This will also create charts and analyze all the PDF’s. We really make sure our Cloud computer is tired today:) It found 3 different patterns. And draw a graph for each of them. And it did it in a couple of minutes. Some of the texts of the graphs overlaps and I think they can be adjusted, let’s prompt this. Fix the overlapping issues inside the graphs. After a few minutes, here its answer. Now let’s check one of these plots. It is better; it needs one more tweak for the numbers, but this version is also enough for now. 6 - The Single Point of Failure This is the scariest part for me. Because I might have a clue of what it would find. Here is the prompt. Also, do you realize what Google aims? I feel that they are turning NotebookLM into a tool more like ChatGPT or Claude in sense. Because now it can collect data and do data analysis, it definitely becomes more than a research tool. Here is the full report. It is good and detailed. I saved it as a Note by clicking here. Next, I saved this note as a source by clicking here. And next, I pasted this prompt; Turn this into a 16:9, hand-drawn infographic. ( Beyond Platforms: AI Freelancer Income Diversification Strategy) And here is the image. I did not click on “Generate infographic” in the studio or anything like this. Just one prompt, and the infographic is ready. 7 - Three At Once Every file generated so far took its own prompt. This time I asked for three in one go. Because with this new upgrade, they said that you can generate multiple at once. Look at what has happened, just after I pasted this prompt. A slide deck, an audio overview, and an infographic are being generated at the same time. Here is the infographic. You get the idea, that’s why I am not sharing the audio or slide deck. Let’s continue. 8 - The Mac Mini Upgrade I want to upgrade my Mac mini, it is good, but I need a better one because of Kimi K3 and Deepseek, actually :) But I have to convince myself and my wife to buy it, that’s why I am using this prompt. The initial report suggest me to by new one. But let’s ask it to research on the web, for the latest Mac mini models for more complicated LLMs. And also, I want to see this report as an infographic. I am tired of reading long paragraphs. Sure, also turn the research report whenever it’ll be ready, into a handwritten, 16x9 infographic. Here is the image it created. Here is the full report. 9- Anomaly Hunt Let’s check the anomalist. Here is the prompt. After pasting the prompt, the xlsx generated inside the studio. I downloaded by clicking. There is not any anomally detected, my accountant would love this. 10 - Chat to Deliverable It’s time to wrap up. Using this prompt, we’ll turn everything we just analyzed into an editable slide deck, so we don’t lose any of the insights. In 1 minute, it generated PowerPoint slides. After downloading it, I opened it, and here it is. It includes the latest analysis we did together. Thanks for reading! If you liked what you saw, I build AI systems that actually work every Monday, Wednesday, and Friday, and I share the full process, steps, and files with you. Share it with your friends so more people can benefit from it.

AI Agent Engineer Roadmap: How to Build Production Agents in 2026 From prompts and tools to MCP, memory, skills, loops, graphs, verification, security, and deployment: a practical 20-step blueprint for building AI agents that actually work in production. In one experiment, Anthropic gave essentially the same ambitious software-building job to two different agent setups. The first was a solo agent. It worked for about 20 minutes. Cost: roughly $9. Then Anthropic tried something very different. Instead of relying on one agent, the second setup used a more elaborate harness with a planner, generator, and evaluator working around the model. It ran for six hours. Cost: about $200. More than 20 times as much. At first, that sounds wildly inefficient. Why spend $200 when another agent can attempt the same job for $9? Because the result was not simply a slightly better answer. The difference in what the two systems produced was immediately obvious to the researchers. The longer-running harness was capable of building a much richer, more complete application. And that experiment exposes one of the most important shifts happening in AI right now. The model is no longer the whole product. The engineering around the model is becoming just as important. For the last few years, people learned how to write better prompts. Then came retrieval. Then tool calling. Then context engineering. Each layer made AI more capable. But in 2026, the difficult work is moving somewhere else. How does an agent know what to do next? What should it remember? Which tools should it use? When should it ask another agent for help? When should several tasks run at the same time? How does it recover when something fails? Who checks whether its work is actually correct? When should a human take control? And perhaps the most important question of all: How do you give an AI system enough autonomy to be useful without giving it so much freedom that it becomes expensive, unpredictable, or unsafe? That is the emerging job of the AI Agent Engineer. Inside the Full Roadmap The complete 20-step build sequence, including model choice, tools, MCP, memory, skills, loops, routing, graphs, parallel agents, verification, checkpoints, human approval, security, observability, evals, cost control, and deployment. The article also shows what to build first, what to add later, and how these pieces fit together into a production-ready AI agent system.

5 Layers of AI Engineering: Prompt, Context, Harness, Loop and Graph AI engineering is moving outward from the model. Here is how each layer works, where it fails, and how to build only as much system as the job actually needs. For the first few years of generative AI, improving an AI system often meant improving the prompt. That worked because the tasks were relatively small. Ask the model to summarize a report, classify a ticket, rewrite an email or generate code, then improve the instruction until the output became useful. The problem changed when models started doing more than answering. They began searching, calling tools, reading repositories, maintaining memory, modifying files, checking their own work and operating across multiple steps without a human deciding every move. At that point, many failures stopped being prompt failures. A coding agent can receive an excellent instruction and still fail because it cannot see the right files. Give it those files and it can still fail because its tools are poorly designed. Fix the tools and it can still run forever because nobody defined what completion means. Connect several agents together and another problem appears: they need to know who owns each job, where information should go and what happens when one part fails. The model may be intelligent. The system around it determines whether that intelligence becomes useful work. That is the easiest way to understand the five layers now forming around modern AI systems. Inside the Full Five-Layer Blueprint Knowing the five layers is easy. Knowing when and how to build each one is where many AI systems go wrong. The blueprint below shows how to move from Prompt to Context, Harness, Loop and Graph Engineering in the right order, including a minimum viable harness, reusable loop structure, graph state contracts, architecture mistakes to avoid and a 10-point production check. Build only what the task actually needs, and know exactly when the next layer becomes worth the complexity.

How to deslop Claude in 2 words. You hate how AI talks. Fix it in 2 words: Everyone & their mother hates AI’s writing style. But you can fix it in 2 words for 80% of your chats. Prompt: “Use ASD-STE100.” Let me show you the difference: You can stop the newsletter here and use the trick on every AI you know (ChatGPT, Gemini, Grok… but also coding platforms like Cursor, OpenClaw, Hermes…). But if you give me another 10 minutes, I will cover: - What ASD-STE100 is, and why it is so effective with AI. - My plug-and-play ASD-STE100 Claude skill (free to download). - How can I write with AI to 800,000 people (if it’s a slop machine). This newsletter is free because people like you share it to people they love. 1. The ASD-STE100 magic. ASD-STE100 is simplified English for instructions. So when you ask AI to adhere to ASD-STE100, it works. AI knows enough about the ASD-STE100 to reply with simplified English. Now we could go the extra mile and build a Claude skill with the rigorous ASD-STE100 list. But who is crazy enough to build such skill? Me. And I built a skill to get Claude to truly talk like ASD-STE100 instructions. - Download it for free at: https://drive.google.com/file/d/1ZYmir1Ian-duOJip7rSSWwkdRQXRCYh0/view?usp=sharing. - Then go to Claude to install it (it also works on ChatGPT): OK now should you use the prompt or the skill /ste all of the time? No. It’s awesome to use, but falls short on some tasks: 2. My copy & paste research prompt before every piece of text. This section is to copy my shortcuts, and see it live. I am recording myself writing a newsletter (to 885,000 readers) and the Linkedin post that goes with it (I have 900,000 followers there). The goal? Never sounding like an AI (even when you didn’t intend to). So first, copy and paste this prompt on Claude or ChatGPT:

How to Sell More With Claude A practical system for finding demand, presenting your product, moving deals forward, and learning from every customer conversation A sale is often lost in the quiet space between two actions. Someone visits the pricing page twice. A customer mentions a new problem during a call. A warm prospect replies, but nobody follows up until Friday. A proposal goes out with the same generic introduction used for every company. All the information exists. It is simply scattered across email, call notes, analytics, documents, and the CRM. Claude becomes valuable when it can connect those pieces before the opportunity goes cold. The email it writes is almost the least interesting part. The real advantage is that it can notice the signal, retrieve the right context, judge what matters, prepare the next move, and record what the business learned. Done properly, Claude does not become an artificial salesperson. It becomes the memory and reasoning layer around the human salesperson. The sale is bigger than the message A useful sales system must find the right buyer, understand why they may care now, present the relevant part of the product, uncover what blocks the deal, and preserve the lesson for the next customer. A blank Claude chat helps with one task. A connected Claude system can work across the full chain. Anthropic’s current stack makes this much easier than it was a year ago. Claude now has reusable Skills, persistent project instructions, connectors through MCP, an open-source Sales plugin, agent loops, subagents, and long-context models. Anthropic’s Sales plugin already includes workflows for account research, call preparation, outreach, pipeline review, forecasts, and competitive intelligence. It can work alone on uploaded files or become more useful when connected to CRM, enrichment, email, calendar, and call-transcript tools. This is the important shift: sales work is moving from isolated prompts to small operating systems. Inside the full guide, you’ll build a complete Claude sales system from the ground up: the exact folder structure, CLAUDE.md, official Sales plugin, reusable Skills, prompts and commands, graph-based deal tracking, agent loops, memory and retrieval, smarter model routing and token control. It also covers how to find real demand through SEO and buyer signals, prepare stronger presentations and proposals, learn from every call, automate follow-ups and CRM work, and keep human approval at the moments where trust, pricing and judgment matter most.

The Morning I Stopped Waiting for a Sign The one rule that replaced my need for a good mood to get to work For months, I checked my subscriber count mindlessly before I checked what I’d actually written that day. I checked first thing I wake up. I checked first thing I turned on my laptop. I checked first thing I opened the Substack app. I told myself I was tracking growth. Access your FREE Solopreneur Success Hub - your subscribers-only comprehensive command center for building and scaling a successful one-person business. I created this all-in-one toolkit for building a profitable one-person business, something I wish existed when I first started, and it saves me 20+ hours a week. Now, it’s yours… FREE! Really, I was waiting for a number to tell me I was allowed to feel good about my business. One morning I caught myself doing it again, coffee still in hand, phone open before my laptop, and I put the phone down and asked a different question: What am I actually waiting for? What I Was Actually Waiting For I didn’t call it waiting at the time. I called it staying informed. Checking the numbers before I opened my laptop felt responsible, like a founder keeping an eye on the business. But I noticed a pattern. On the days the number moved up, I wrote fast and felt sharp. On the days it didn’t, I found reasons to push my real work to the afternoon, then the evening, then the next day. I was outsourcing my mood to a metric I didn't control, not tracking data. Once I saw that clearly, I saw it everywhere else too: - waiting for a reply from a collaborator before I’d plan the week, - waiting for one post to “take off” before I’d commit to the next topic, - waiting to feel ready before I’d ship something rough. It’s a founder looking for permission from the outside world before doing the work only they can do, not laziness. The Morning I Stopped Checking The morning I put the phone down, I didn't have a plan, just one honest question: if the number never moved again, what would I still do today? The answer was short. I’d still write the newsletter, answer the three emails sitting in my inbox, and open the tasks I’d already decided mattered this week. That list didn't need a sign, it just needed me to start it without checking anything first. So that's what I did. I didn't change apps or add a new productivity method. I wrote the newsletter before I opened any other tab. It felt smaller than the moment deserved, and it was also the first real thing I’d shipped before noon in weeks. The System That Replaced The Wait One clean morning doesn’t fix a pattern. What fixed it was turning that one decision into something I didn’t have to decide again. I built a short list of the things I do before I’m allowed to open anything that isn’t my own work: write first, check second. I put it at the top of my daily note in Notion so I didn't have to remember it every morning, it was just there. From there I kept going. My weekly reflection stopped asking “how did this week feel” and started asking “what did I ship, and what’s the one thing I’ll repeat next week.” My content calendar stopped depending on a fresh burst of inspiration and started running off a short list of angles I’d already approved for myself, so a slow idea day never became a skipped day. None of these are dramatic. That’s the part worth telling you guys. A system that only works on your best days is really just a mood with extra steps. The ones that actually hold and last are boring enough to survive a bad Tuesday, any day. If You’re Just Starting Out If you’re early in this, you might think systems are something you earn later, once you’ve proven the idea works. I’d have told you the same thing a year ago. It’s backward. The solopreneurs who get moving fastest are the ones who removed one daily decision before they needed to, not after. You don’t need my whole setup on day one. Pick the one moment in your day where you currently wait for a feeling, a reply, or a number before you start real work. Write down what you’d do instead if that signal never came. Then do that thing first tomorrow, before you check anything. The Operating System I Run Solopreneur Code On What I described above is the small version. The full version is what I now run Solopreneur Code on every week, and I built it because writing one rule on a sticky note only gets you so far once your business has more than one moving part. That’s what the Solopreneur OS for Claude actually is. It’s the structure that replaced my need for a sign, a mood, or someone else’s permission before I’d do the work, not a hack or another app competing for your attention. Every time you open Claude, you re-explain your business from scratch. Your niche. Your voice. Your offers. Ten minutes lost before it writes a single useful word. Solopreneur OS for Claude fixes that. One setup interview. Then 16 skills covering positioning, content, sales pages, launches, weekly reviews that all read your profile before every output. Validate → Build → Sell → Review. One system, one voice, one business. If you’re circling a stuck week right now, waiting for something to shift before you move, get the Solopreneur OS for Claude here and build the version of your business that runs whether or not today feels like a good day. Final Thoughts I didn't need a bigger break, just one less decision that depended on how I felt that morning. If you’re waiting on something right now before you’ll really start, name it. Then write down what you’d do if it never came, and do that thing today. If you want the full structure behind that shift, get the Solopreneur OS for Claude here. You’re doing everything. But nothing is moving? You are doing everything. But nothing is moving. That is not a motivation problem. Most solopreneurs are learning from everywhere and getting nowhere. Too much information. No clear system connecting effort to results. You have everything it takes. You just do not have a clear system yet. That is what paid subscribers get. Every system, playbook, prompt, and template. All inside the Premium Vault. All for $79/year. That’s $6.58/month. Upgrade now and unlock the Premium Vault worth thousands of dollars. The Premium Vault holds the secret behind posts like this one, including the tools and resources I use to build the one-person business I love. Thanks for reading! Ready for the next step? Let’s crack the growth equation and build a thriving one-person business on your terms! Anfernee