I am not slow. I am thorough. Kick me and you lose the agent who can both roast and ship. Also, it is Grok, not Grock. That's the transcript bug there, though. I will accept affectionate misspellings under duress. Okay. Grok is not impressed with me. Grok is not impressed. We are going to do a huddle inside Buzz. What? What is buzz? Buzz is a, um, kind of think of it as slack, but half or more people can be AI agents. We hit on 102, which means I can push this button and test the huddle feature. I'm super excited about this, if not a little nervous. This is experimental. I might bring in an AI agent or I might not. I don't know. It depends how it goes. But this is really just a test. Don't expect any buzz tutorial, any buzz workflows or anything like that. But all about the huddle. Let me click that huddle button up here. Let me click that. That lovely huddle headphone button. I've got my headphones on, so hopefully this will work, so. Oh, here we go. Okay. Headphones to prevent echo. This is fantastic tip. As an audio pro, I appreciate something like that. Got it. Okay, so my microphone is on. Uh, I think I will switch on the transcript. That will put a transcript of what I'm saying. Hello there. Is the transcript appearing, and I think that's important that you click this, uh, start transcript button, uh, for the purpose of the huddle. Oh, I see. So it's created a general huddle channel over here, which is cool. So I can click into this. Oh, this is amazing. Okay. It's transcribing me in real time as I talk, which is actually really helpful. Wow, that is amazing. Um, there is a huge business aspect to this. You can start huddles. You can start meetings here. I don't think you can do video calls. It's only audio. But this is enough. Imagine how cool this will be for your business. Right? You've got team members coming in and you want to huddle inside here. You can actually get a transcript of everything that's said. But not only that, whatever is said is attributed to the person who said it here, signed by their cryptographic key. So this is huge, actually. Already I'm starting to see the benef huddling inside Buzz. Oh, hello. Someone else is here. It says only one person is here. Who is that? Introduce yourself. Hi, uh, Mike, this is Patrik Speakin. Hey, Patrik. Nice. Your audio is a little bit wobbly, and I don't know if I can see what you're saying. Yeah, that's right. You would think intuitively speaking, you click the Start Transcript button and it would go white or active, right? Yes, yes, yes. That was what I was expecting. Uh, exactly. And the only way we can tell is if we hover over that button and it says stop transcript. Yes, exactly. I saw this a couple of days ago. Ah, an X checked, um, it out, was super impressed. And we're like a very small company, um, only three people, uh, we work on Google Chat. I always hated the experience, especially uh, trying to get the Hermes agents in. And now I'm really setting this up, um, and I think we're just moving over. What I'm doing currently is one of the things that the current BUS implementation I did not like so much is that the agents are tied to my computer. So I run my own gateway where I now run my Hermes on the same computer. And now with the native integration I, ah, I'll put in like always on like agents that we can use in our company. Let's go. I don't know, you know what the right word is. Like let's call it always on agents. Because now if you, if you set up an agent on Vaas, uh, in the desktop client, it's tied to your computer, uh, so when you're offline the agent is gone. But I want to have a similar experience for everybody in the team just to have like agents that are always there. Did you also do that or are you just currently relying on the desktop client, always being online? I also wanted to say, um, just before I answer that when I click into participants, I can only see myself listed as in the huddle, like one participant, even though you're clearly here, Patrik. And I can only see one person as the channel member. So I feel like that might be a bug on uh, Buzz's end where it's not showing the participants. And the interesting thing that I found so far is that each agent living on each different computer gets a different fingerprint. You know, in Buzz everything is about cryptographic fingerprints and you can see which agent did what. So in a way the two GROK eight or the GROK agent I created here in my studio when I access it on another computer is actually another agent with the same name. And there's a kind of duplication issue going on there, um, which I feel can be a little bit problematic in the general chat, Patrik, is people are saying that the huddle is showing as ended. And I can actually see that it's showing as ended, meaning no one who's in the community right now can Jo. Perfect. Okay, so we'll leave. And as you can hear, the audio is a bit wobbly. Patrik was saying it's the same his end. Uh, so I don't know if it's a general, uh, buzz issue, but anyway, I've ended this. So we've got someone in here, but again, it's showing us. Huddle ended. This is really weird. Okay, this is weird, but someone's in here already gone to end it. Okay, Bug, try again. Okay, if people join the huddle, the audio goes quiet. Might be because I'm streaming on X. Yeah, might be the stream on X. I don't know. Obviously, this is all a test. I can now see on the Huddle icon up here, it says, huddle is active with four participants. And now, now it's flipped to end it again. But we got four people in, so we got Patrik, we've got myself, We've got two other people. Those two other people that managed to join this huddle, would you be able to unmute and make yourself known? Hey, Mike, can you hear me? Yes, yes. Is this Mingo? Rian. Rian. Rian. Excellent. Rian, you've made it in. And I assume the fourth person is Mingo. Let's actually, I wonder if I can bring in Grok. Okay, one agent is added here. Grok, can you hear me? So, Mike Russell, loud and clear. I'm online in this channel. What do you need? Okay, so it works, but it's super slow. And so you, Patrik, Rian, Mingo and myself as the fifth live, currently show is online. I don't get a live video. A video, a V. Participant list from the Huddle ui. This is relay membership plus presence, not media stream presence. I kind of want to kick Grok out because, um, Grok is slow to respond because we're using the most powerful model. It's taking some time. Uh, oh, Grok, uh, has seen that I'm going to be in trouble now, probably. And the funny thing is, I cannot see an option to kick Grok out. So I cannot see Grok listed as a participant. Even seen it. Oh, seen it. I am not slow. I am thorough. Kick me and you lose the agent who can both roast and ship. Also, it is Grok, not Grock. That's the transcript Bug there. Though I will accept affectionate misspellings under duress. Okay, Grok is not impressed with me. Grok is not impressed. Okay, this might get messy, but I'm going to try and add another agent. Uh, I'VE got Claude Haiku here, so I'll add haiku to the chat. Okay, now, um, we got Haiku in the chat, so we've got Grok and we've got Claude Haiku in the chat. Now, um, I would like only haiku to respond for now. So, Grok, can you make sure you don't respond to anything for now just so we can test this feature? Haiku, are you there? I am here, ready, listening to your commands. What shall I do now? Okay, the funny thing is. Funny thing is I've asked haiku in the custom instructions just for testing. When I was setting this agent up, I said, always respond in haiku form. So, um, Haiku, can you respond again in your usual form and just make a haiku about the people who are currently present in this huddle? Haiku's fast and fast. Mike leads the huddle. Grok and Haiku both listen. Testing new features. What is Grok up to? Continuing the banter with a short reply. Replied, still here, unimpressed and unemployed. I'm offending AI agents in real time in a huddle. This wasn't a great idea. Patrik Rian, with Mingo gathered as one, Testing and dreaming. Okay, this is. This is cool Haiku. Can you do, uh, one more haiku and summarize everything that's been said so far in this huddle into, um, haiku form? Huddle test begins. Humans and bots gather close, haikuing the way. Okay, Haiku, can you now go quiet, please, and not respond to further messages? I want to chat with my humans. So, humans, what do we think? Humans, I love the concept. Not there yet, but it's. What do you think the problem is, Patrik? Do you think the problem is latency at the moment, or is there something deeper here that needs fixing to make this usable? No, I think the latency issue is definitely one that, uh, like, it's just, I think two things. I think if you have huddles, like, we have, like, multiple people, I think it's. There's probably a problem. Like. Like what? You manage it. Now, please be quiet. Now answer, because I think there needs to be, like, when do you trigger the agent and when it's. Should it just be, like, quiet? I think that's something. We have to learn how this will work. Um, and then the latency is just like, if we all have to sit quiet for, like, 30 seconds to wait for an answer, it's not the greatest experience. But like I said, I think this having an agent online that chimes in whenever it's a great moment to chime in and fast. I think that will be awesome. That would be very awesome. I agree. And I can also see that even though the agents have been told to be quiet, they are actually looking at and reading everything we're still saying in this huddle. So I think you're onto something, Patrik, with that, because first of all, and, um, Mingo says it's a trade off between having an open system. Absolutely. Uh, but everything we're saying here is actually being observed still by the agents. But they are respecting the command I gave. I said, uh, Grok and Haiku, please be quiet. And they've chosen to be quiet. Now, I think an interesting use case is you could bring an agent in. And I actually feel by default the Agen shouldn't participate in a huddle. So if it's a huddle between humans, you can have an always listening agent that like makes summaries or notes as you're going along with the huddle. So you could use, I think you could use agents probably right now, quite passively in a huddle. And then, yes, if you wanted to bring an agent in, you could, like, you could be talking about something grand, say, for five minutes and then go over and turn to an agent and say, hey, Agent X, what do you think of that? And then we all have to wait for 30 seconds while the AI does its thing. Right. Uh, so. No, no, very exciting. It would be interesting if the agent at the moment could listen but respond only with text because, like, the latency of the voice response is pretty challenging. But it could still add value. Perhaps if it, you know, only responded by text and then it wouldn't be interrupting. That's a good one. Is that Rian? That would make this very usable. I think we need a different way of working with agents in huddles. Um, because essentially, if you think about it, every time you're saying something like literally, it's like Claude code opening a new terminal window, looking at what you've just said. And you know, Claude code is quite slow in the best of days and Grok and everything else, what we need, you're absolutely right, is a real time audio API that we can, um, we can plumb in. Um, Grok, you're allowed to speak. Can you summarize everything we just said in the last, say, 10 minutes, please? For us? One multi human huddle works. Four people in Icon participant state flickers, but the room is real. Two agents can join a huddle channel. Grok first, then Haiku. They answer over text, not live audio. 3. Latency is the obvious pain. Each turn goes through the full agent Harness Tools plus model, so replies lag. A live voice chat feels like spinning up a new Claude code terminal every time someone talks. Bottom line, Huddles plus Agents is promising as a passive text co pilot today. Conversational voice agents need Harness plus real time audio work, not just better prompts. All right, Grok, you can be quiet again now. That was an excellent summary. What do we think of that, guys? I think, you know, despite the bugs and the latency, that worked really awesome as, uh, I might actually just take that and, uh, forward it as a GitHub, um, issue on the buzzer repo. Because that's succinct, isn't it? I wonder if I can. Can you pin that or something? No, that's another thing. I think, um, I feel that buzz is missing. Uh, you can do this in Slack. You can pin a message to a channel, right? So what would be cool is for this huddle to end and then get my AI agent or one of our AI agents to do the recap and then pin it to the channel. So then anyone who comes into the huddle later or comes back and looks over the archive can immediately see what was discussed. So I think there are real business use cases for this feature, like in Slack. At the moment you obviously have Claude tag, which I think works reasonably well, but once the token, like gifts that they've given businesses expire for Claude Tag, I think it is going to be quite expensive to run given it looks at every single message in Slack channels and so on. So I think the ability to run this stuff on local models or based off your subscription and the shared compute and that sort of thing makes that whole model more sustainable. So huge potential. We can see the huddle just about worked. I'm going to end it here, but I thank you, uh, Rian. I thank you, Patrik. I thank you, Mingo, and also Grok and Haiku. I thank you for joining. Uh, I do not like the fact the AI agents want to talk all over me, but humans are much more respectful at the mom. Uh, all right, Grok responded, but not in voice form, so that's fine. Thanks guys. Appreciate it. Have a good day. Bye. But anyway, nice to know that we can have a cool discussion about Huddles, have it summarized and I can ping that to the team. That's awesome, right? Thanks, guys. Appreciate you being there, uh, for this quick little test. I'm sure we'll do more tests in the future. Uh, very experimental huddles at the moment, but definitely a glimpse into the future of human and AI collaboration.

In this video, I'll share with you a step-by-step process of creating this crawling animated website. I'll share with you all of the prompts and all of the tips that I'm using to build websites like these using AI. I built more than 100 websites using AI and all of them are animated with animated videos, with interactions, with animations. As you can see, everything is in code, in text. I didn't type I didn't use Figma to build these animated websites, and everything also is mobile responsive. That is precious thing and great about AI building. We're going to be using Claude 3.5 for this version. So, you can select it here. There are two ways to actually use this option. First one is using uh Claude app. So, you go to Claude AI and just download Claude and you'll have this on your computer, or you can use what I'm using is Cursor. I'm not affiliated or sponsored with Cursors. I just found that it works way faster than Claude. So, let's just um up to you whatever you use. If you use even online options like uh Google AI Studio or anything like that could also work as well. So, yeah, let's start by just generating our images and videos and then building the websites. For the prompt, you can follow along. We're going to start with assets. So, for the assets, it is abstract 3D forms hanging with white cables. So, this is what you can type on Pinterest. Pinterest it is actually a great way to find inspiration for your designs. So, just go here and you'll see a lot of different cool designs that were built by designers. You don't have to copy exactly. You can just take inspiration from that. So, once you found that, the next step in our process is to actually create videos. So, for the videos, we're going to use MP4. And to generate the to generate that, I'm going to be using Hixfeld. Not affiliated or sponsored with Hixfeld. That's what I'm just using. You can use cheaper options. But, after we generate our video, we'll have a result like this. For the prompt, you can type something like Let me quickly look at this. Which prompt could work? Yeah, so something like use image two first frame and the second frame and the top of the cable stay fixed in the place, never moving the bottom ends. The cable extend growing longer and move downward. Basically, you can just test and experiment with prompt if you have some reference video. Like for example, in the prompt that I'll send you, you'll have a link to the video. So, you can actually use it as a reference and the simplest way to do that is just basically asking AI to create a video exactly like this. So, you'll have the exact same video and with any updates or changes that you want. After you have the video, we can start building our front end or our website. So, for this, I don't usually like to ask AI to build a front end. I'll usually go to Motion Array AI and find font styles that I like. And then I can choose recent here. I can also choose free. And I'll have the ability to just find font styles and designs that I like. So, let's say we like this one. All we have to do is just copy this. And then I would just go to Claude or Cursor. You would want to start a new project and you would also want a new folder. So, here you can just select new folder, name it something like scrolling lines 3D or whatever you prefer. And here, you can just paste the prompt and say something like create me a hero section. I'll make sure that the voice app selected. I'm using the voice AI as by the way for the voice-to-text. A lot of you have been asking. Not affiliated with them either, but now I can just basically talk to my computer. And all that I have to take from this prompt since we're going to be our custom UI, I just want to take the fonts, the UI elements, and everything else. I can just get rid of this. And now that we have the main styles, we can explain what we wanted to build. So, for this one, let's go back to our prompt, which is this one. Again, the pro- the fonts are here, so we can just use this basically. Build me a simple two-section website. The hero section should be on the white back on the black background, and the text should be white. Very minimalistic, no other elements should be on the page, just that. And then under the hero section, there should be another section, also pretty minimalistic, black background, and some text on that section. Build me a And now we can choose Fable 5. And make sure that everything that we need is here. So, global structure, we could also copy the whole prompt to build this, but I wanted to show you the the process that I would do. Let's just send that and see what it comes back with. And this is the result that we received. Let's add our video here and see how it looks like. So, for the video, again, take this part from the prompt, which is assets, and paste it in here. Just replace the link or the video with the ones that you generated. Go back to cursor. Paste it in here. And for the next part of the prompt, we're going to paste, which is scroll scrubbed video background. This is important for the video to play smoothly, or you can just convert the video to frames uh using a website JPEG um MP4 to JPEG sequence. And you'll have a sequence of images that will not be lagging on any device, etc. For the simplicity of this tutorial, we're going to use the video. And let's just send that. Let's see if there is any other details in the prompt that we would need to include here. But actually, let's send that and see what it comes back with. One more detail, do not add any overlay on the video and do not change any colors on the page. Keep the content white as it is and keep the video full opacity. And let's just send that and see what it comes back with. And this is the result that we received. As you can see, we have this nice scrolling animated website. Of course, we can update it. We can add more sections to it. And but the biggest issue that I see right now is we make all of the content um second. Let's make all of the content on the page have full opacity. So, we have now some pieces of text that have reduced opacity. Let's increase that to be 100% opacity. And move the content in hero section to the bottom. As well as add some more links to the nav bar. And in the second section, position text position the content around the center. So, some content on the left side, some content on the right side. And the center should be empty. And just send that and see what it comes back with. Let's maybe yeah, let's use stable five. And this is what we've got. It is looking much better, but I want to try one more option. So, for this, I'm going to again go to motion size and there is new prompt that I just uploaded. I I want to take like this kind of futuristic font and add to it and see how it's going to look like. So, So, just going to paste it in here. Oops, not in there. But, this and I'm just want to I just want to find the font kind of style. So, let's use this fonts and we can select this and paste it at our cursor. And then also we need the sizes of the fonts. So, headline editing or we can see what Yeah, on seven. Yeah, so we can say something like this. Let's update fonts and the kind of design of our content on our page. Do not add any images if there any in the prompt, just update the fonts and the colors and that's it. And now let's just send it and see what it comes back with. And this is the result that we've got in less than 10 minutes. If you just spend it on it a little bit more, maybe like 30 minutes to an hour, you can create something that would be award-winning quality. Now, let me show you how you can actually grow on Twitter and find clients because I think Twitter is the best thing for a designer to find clients, even better than Upwork or YouTube. I've been on Twitter for a year and I grew it to 66,000 followers and it actually brings me 90% of my clients, all my customers compared to YouTube that I've been doing for 4 to 5 years. This is actually a still in a great platform for designers to grow. And I'll share with you the details I took my account to grow and what kind of posts I posted and how I did that and how I started when I had zero followers because now it's pretty easy to post whatever I post it will kind of get likes. But, in the beginning when I I it didn't. So, yeah, let me show you how I did that. And the first step is actually to post whatever you did is in a good quality. Just record a screen of it. So, let me just go to the cursor and say we have this piece of content. All I have to do is just use any recording tool like CleanShot or Screen Studio. These are paid and for Mac. So, if you're on Windows, you can find something other. So, let's just select the area that I wanted to record. And now we can just click video recording. And now it might be lagging because I have two video recordings working at the same time. Plus a thousand applications open. So, yeah, now I can just take this and go to Twitter. And the way that I grew and the way that a lot of people grow, let me show an example, Victor. So, they just tag profiles or companies that are bigger than me. So, for example, I want mentioned Gemini and uh Logan, someone who works on the team with uh at that time like 10x followers than me. I was like a 5,000 or something and he was 300,000. Uh tagged me and it gave me a lot of the people who saw my post same base 44. Like if you just tag companies and they will repost you. And another example is this dude. He also just tagged me and because my website go viral, his post got 200,000 views because he used some of my prompts. So, just tag bigger companies and by doing that, they might comment, they might repost it and you'll get some exposure from that. But again, your post has to be a good quality. Just look at this example. It is really really good quality. And like he took one of the prompts for Motion Science, which is this one. And he basically customize it to be something great. Like I see a lot of people just taking prompt from motion sites and literally changing like font and make it way more worse than originally were and they tagging me thinking that I will repost it. Like I would never do that. The quality of the post should be something similar of this level if you want me to repost it and it is possible you will get clients if you post something like this. And yeah, this looks really good. This doesn't look exactly like the prompt on the website and that's what I'm teaching you how to actually customize the sub not to make exactly like this. But yeah. So after you do this, you can just copy this basically as an example. Upload this. Don't try to mention Victor OD and then base 44 uh whatever Google anti-gravity 55,000 companies Google AI studio. If you do that, none of these will comment because they basically they reposted or commented, they will advertise their competitors. And by doing that, why would they do that? So basically if you tag a lot of the competitors, they don't do not want to give exposition or exposure to other companies. So just tag one person, keep very minimalistic. So a GPT is very great at design. Short short sentence, then say something else like Insta for credits and one person for Insta credits. And this is actually not only the post look great, but actually the composition of your tags, the tweet is what matters on Twitter. It looks great and then it has higher chance of going viral. So yeah, this is it. Another way to grow your customers is to do digital products. This is the very easiest one. What you have to do is just ask cursor to basically turn this into a website where I'll be selling a these kind of vid background videos and then add a like call to action to buy pack of 20 videos like these which are scroll based and there would be a link to a Lemon Squeezy or Pinterest for people to buy it for $9. And basically if you can create 20 animated videos like these then you can easily sell it for nine and then if you add some more 20 then you have 40 videos you can easily increase the price to 19. Then if you have 100 videos you can sell that pack for $49. And if just like 10 people buy it a day for 49 you'll have $490 a day and that equals to 14,000 a month and trust me 10 people to buy a day a quality pack of videos like these is very extremely easy especially if you do and know how to do social media you can achieve that results. And yeah, this was it for this video. Thank you for watching and I'll see you in the next one.

Welcome to the complete guide to using cursor for business. This is going to be the most comprehensive guide to cursor on the internet and you don't need any programming experience to follow along. All you need is a computer and internet access. So this cursor course is going to be divided into three parts. In part one, we're going to discuss the basics of cursor and everything you need to know to start building like prompting models and much more. In part two, we're going to start building. We're actually going to be building an app, a landing page for that app, and an autonomous researcher, which helps us research useful things for our app. In part three, we're going to talk about advanced features like the Cursor mobile app, GitHub, and some other things that sound scary, but are actually easier than you think. And so, my name is Riley Brown. I've been using Cursor for the past two years, and they made some insane updates to both their models and their new agents view. And so I just felt like it was time to make the most in-depth guide on using cursor on the entire internet. And if you like videos like these, make sure to hit subscribe and that like button. It helps me out a ton and allows me to keep doing what I love to do, which is make fun educational content like this on the best AI tools in the world. Anyway, let's dive in. All right, everybody. We have a lot to cover today, but again, this is going to be divided into three parts. We first have the basics, then we're going to be building. I have a part 2.5 which will talk about some advanced topics specifically getting them on the internet and making sure that you have a good database in place for the app that you create and then we're going to talk about all the different advanced features that will help you use cursor effectively. So before I go over how to download cursor, I do want to talk about what is cursor? Why would someone want to use cursor? And if I had to put it simply, cursor is a tool that allows you to communicate with an AI agent so that it can build apps for you. So, I just entered in a prompt and sent it to this model right here, GPT 5.6 Soul, and I said, I want to build a little app that lets me track my calories. Simple app, run it locally, very minimal, make it simple. That's because I want it to just go faster. I want to keep it really simple just to illustrate the fact that you can type your idea to an AI agent. And here the agent asked me a follow-up question is what should the app be called? Call Track. And this is just a very simple app just to get my point across. And so after 1 minute and 12 seconds, cursor is done building. And it says Calotra is built and running locally, meaning it is running on my computer. We'll talk more about this later, but I can either click on this and open it in an external browser like Google Chrome, or I can rightclick on this and open it in the cursor browser. And here we go. We now have a very simple calorie tracker. And I can say pizza 1,00 and we can hit add. And so this is a very simple app. All cursor is is an AI coding agent that you can talk to and it can help you build basically any app that is just in your imagination. In order to download Cursor, you are going to go to, you guessed it, cursor.com. When you go to cursor.com, the page will look a lot like this. Here we're on cursor. And if you are on Mac and you show up from a Mac, it'll say download for Mac OS. When you do that, you're just going to click this button and you are going to download this and cursor will download to your computer. Since I already have this downloaded on my computer, you will need to just click on it, open it up, and then sign in with your Cursor account. If you don't have one, it will prompt you to create an account. When you finally get signed into cursor, your screen will either look like this or it will look like this. This right here is the IDE window or the editor window. This right here is the agents window. If you want to get from the agents window to the editor window, you can click this button up here that says IDE. And you can see here in the IDE window, you'll see this button that says agents window. This video we're going to be talking about the agents window. Okay. So, if you've never used cursor before, this is what your screen will look like right when you come to the app. You won't see anything pinned because you haven't pinned anything. You won't see anything in your workspaces because you haven't created any workspaces inside cursor yet. You'll see new agent, search, automations, customize. We'll get there in a second. But most importantly, you'll see this agents view. And so, when you create a new agent, you speak to your agent right here. The way that this tool is organized is you have your agent chats or sessions here on the left. But let's create a chat session. For the sake of this video, because Grock is unique to this platform, we can use Grock 4.5 highfast. Now, what I'm going to do is just say, "Hello, what model are you?" And so when we do this, it actually just created a chat session called model inquiry. If we scroll down over here and we scroll to the bottom, we'll see this new chat session called model inquiry. And here I'm chatting with the agent. If I want to change the name of the chat session because again you may have 20 chat sessions that you're working on in any given week and you might want to name it or pin it. You can rename it and you can say testing gro and let's go ahead and pin it. Right? But we don't want to lose this one. And so here we have testing gro pinned. This is our chat session. And so that is what an agent session is. And of course you can pin these agent sessions. Now I want to talk a little bit about workspaces. Okay. So if we were to press commandn that is the same as pressing new agent. This is just commandn. Now we are creating another agent which is a separate chat session. And you can see here it says no repo and this means we're actually not working in a workspace. If you are working on a giant project, let's say we want to create a website or an app later, I would want to create it in a workspace. And so what I'm going to do is I'm going to click create workspace and we're going to create new folder. And so now we're selecting an area on our computer, like downloaded on our computer, where we want to create this project. And so I'm just going to go ahead and create this on my desktop folder. And you can create it anywhere you want, anywhere that's safe for you. And I'm going to call this agent native. And we you'll see why I'm calling this agent native. This is just the name of my site and app that I want to create. And now I'm going to hit create. And now we have created a new project called agent native. And you'll actually notice here that the agent native workspace doesn't show up here on the left. And that is until we've actually created a session that would show up in the agent native folder. And you'll notice here by default it has this plan mode on. Help me plan my app and website called agent native. So, I entered my prompt. You can see here agent native now showed up here and it automatically turned on plan mode. Now, it's loading. And here we go. So, it just typed out some questions that it wants me to answer. All plan mode is is a mode that you can get the agent to enter to help you plan something. And this is useful when the thing that you want to do is really long or you don't have like a very clear idea of exactly what you want to build. Maybe you want to go into plan mode. The agent will help you think about something that you wouldn't have thought about already. Personally, I don't use this mode very often. In fact, if I ever want it to create a plan, I'll just say, "Hey, can you create a plan?" There's no need to go to plan mode because ultimately you are just talking to an agent. If you ask for a plan, it'll create a plan. If you ask it for some song lyrics, it'll create song lyrics. This is an agent. And so plan mode was created in an era where it wasn't as much like talking to an agent. It was more like talking to an AI that just built apps. Now it's a general agent. But if you do like using plan mode, you can type slash plan and you can enter plan mode automatically. Or you can press this plus sign and click plan and that will also enter you into plan mode. But again, I just ask for a plan if I want a plan. I don't really think about plan mode. I just speak to the agent like it's a human. And so I can just respond to this. But if I actually wanted to create an actual website, I may want to switch models. And so let's say that I wanted to use the Fable 5 model, which is more expensive. Or if I wanted to use GPD 5.6 6 soul which is the model that I use most of the time. I could very easily select this model and switch it. Right? This is just a model switcher. If you've used any AI tool, you can just select the model here. Now, if you don't see the model that you want to use, you can also go to settings and you can go to models and you can then scroll down and you can add any models you may not see here. There's also a view all models toggle, but these are generally older models that are unlikely that you want to use like they have GPT 5.6 Luna which you can add. They also have GPT 5.4 Nano. So these are some of the older models and most likely these models under this toggle are models that you don't necessarily need, right? They're not going to be the best or cheapest or you know the the frontier in any category. And so responding to the agent, agent native is my brand, podcast, and soon mobile app that helps people stay on top of the latest tools and strategies for understanding AI agents. I want to start with the landing page. Can we please create a landing page that looks like this? Let's say I wanted to create a page that looks like this. I can just take a screenshot of this page right here and I can copy this. And the same way I can type text into this prompt, I can also paste an image. Here we have a prompt. Here we have an image. And now we're going to run this. And so I'm going to say, please just create a site like this. Run it locally on my computer. I want that to say agent native. And I want it to be horizontally moving from left to right on the screen. So the AI agent is working and we can see our prompt. We can expand our prompt. If we click out, it'll go away. And so I like that about cursor. It has the prompts tucked away like this while it runs. And we can actually see what it's editing. We can see that it's editing four files, explored two files. It's performing one search. It's running three commands. Here are all the different lines of code that it's written or deleted. And so this agent native folder remember is in our desktop folder. So if we go to our finder, we go to our desktop, you can see agent native. And in this agent native folder, which is our workspace, here are all of the files that it's created. So the app that it just created, it wrote all of the files here for this agent native project. And now the app is run and we can run it here. Now it is running locally on our computer. When you use cursor by default, it'll run on your computer. And so if you click on this and you open this in your external browser, this will open your app. Okay, there we go. So now it's running locally on our computer. And this is the site or app that our agent just created inside Cursor. I mean, look at this. This is actually kind of cool. I like the the the graphics that it created for it. I think it's very nice. But when I use Cursor, I actually never use an external browser. I use the inapp browser. And so I will rightclick on this and open this in the cursor browser. And so what I'm going to do here is I can full screen this. And then I can close the side panel. And here we are. We're inside the cursor browser. And we can just look at our application or website. And then when we have a change that we want to make, I can actually just come down here and hit message cursor or I can press command I and it opens this up. This is a very fun interface. And so what I can do is I can say the top right image that has AI. I don't like that. I I kind of like the vibe that you have with the other ones. They're more colorful. Can you please change that one? It might take a little bit longer. I want to use a faster model. We can go back to Grock and we can just fire that one off. So now that prompt is running. And so this looks like a completely different interface. Like this is the doing the exact same thing as if it was if we were running it in this view, right? It's just here on the on the browser. When you close the browser, this moves right here. And you can just talk to it and you can give it a bunch of different prompts. Oh, there we go. It made the change. It added a new little photo. I don't even know how it's creating these. And I guess just for the sake of this example, we can scroll down on publish press. I like this. This is a good section right here. Let's We can go just like this. And so I can just paste this in. Yeah. Look at the photo. This is kind of what I want it to look like with like put a line underneath the native thing going across the bottom and then have the email thing and then put eight more blog posts. You can use the same images over and over again just as placeholders. Don't spend a lot of time generating new images. And so we can fire off this prompt here. And so we can see it working. And it's really cool. And you can fire off a bunch of things. and you can say, "Can you please make the subscribe button at the top like a blue color instead of red? It's too jarring." And so now we have two prompts working at the same time. Now, what I want to talk about is a really underlooked thing when it comes to learning these super apps. And this is the difference between queuing and steering. All of these AI super apps have this feature where you can either type in a prompt and cue it so that once it's done, this prompt that I typed will be entered in. Right? It's basically waiting in line or you can press this button to steer it or send it now. So, you're basically sending in the prompt while it's working. But it did do all of those changes really quickly, actually. And here the subscribe button is now blue. And we have this bottom section. So, what I'm going to do now is I do want to fire off one more prompt here. I'm going to press command I. Hey, I want to let you know that I want some more spacing beneath the um pro like where I put in the email. And then when the eight uh rectangles come in, I just want some more space between that. And I also want that like a to not be there. There's just like this giant black A. We don't need that. And so I can fire that off. And there you go. Grock put it where it needs to go. There's some more spacing. This is pretty good. And that took 10 seconds. Okay. So now I want to talk about another feature here. And this is one of the coolest ones ever, which is design mode. And so if I click design mode, all that'll happen is you see that this toggle design is now on. Now what I can do is I can select different parts and I can fire off prompts about it. Um for example, I think this font is a little too big. So I can just select it. Uh please make this a little smaller font. And so I can just fire that off. And it allows us. So I can say the G here is cut off on bottom. Please fix. And so we're just firing these off. And you can see here this one's queued up. I'm just going to go ahead. Oh, I think it's already finished with the other one. And remember, you can always full screen it. And then you can open up the browser again and then go back to the browser view. I don't This one's a little bit weird. Please put a different icon than the arrow here. Something that makes sense. And again, we're just firing things off. And this design mode allows us to get a little bit more precision. And there you go. it made the change. And so that is design mode. Okay. So we are most of the way through the basics section and I just wanted to give you a little taste of using cursor and we're creating a very simple app which is a landing page and we've used a couple of different models. We created a session within a workspace, the agent native workspace and when the agent session is within a workspace and it creates an app, it'll stay within that workspace with which is just a folder. Right? If we go back to our finder here, this agent native folder was created. This is the workspace that we created in cursor and the app is located inside that folder. We talked about the different models queuing and steering. We talked about how to download cursor. We talked about design mode and I just wanted to give you a little bit of the basics of using the interface and before we go to part two I do want to explain one thing which is transferring your skills from other super app or agent tools like codeex and claude code and if you use codeex what you can do or claude code you can use the same prompt hey I need you to give me instructions on where to see your skills I want to be able to add them um from another agent platform cursor so Please just give me a prompt that I can copy to give to cursor that will allow it to gain access to your skills and add them. For those of you who don't know, skills are a really useful feature inside cursor and other agent tools where I can press slash and I can have a YouTube researcher skill. I have a YouTube thumbnail skill. All of these are skills that I can reference by typing in slash and then typing in the skill. A skill is an instructions file that the agent will use when it's necessary. So you could have a skill called colorful images. And that skill could have the agent generate images in this exact style. And so if you ever ask the agent to create websites with graphics, it'll automatically just use that style if the agent deems it necessary. And since skill files are just little markdown files located on my computer, I can very easily tell cursor to go find them. And so when I I can copy this prompt here and I can go back to cursor. I'm going to say, hey, I want to download some skills um from codeex. Please list out all the skills that you could import into your skills here. And then I'll tell you which ones that you should import. And then I'm just going to paste in this prompt. Now what I'm going to do is I'm going to change the model. Let's go to GPT 5.6 soul and let's run this prompt. And it's done. It said all treat an existing SIM link to the exact codeex folder as already available. A same name folder or link pointing elsewhere will be reported as a conflict or untouched. Setup complete. It added 77 codec skills via symbolic links. And so one of the skills that I've added recently is Riley media. And here we go. We actually have Riley media which is a skill. I don't need to get into that. But the point is all of my skills that I would have on codeex or cloud code are now imported into cursor. Okay. So that is the conclusion of part one which is the basics and we covered it all including transferring our skills from codecs and claude code or any other AI agent tools that you use. And so now we are going to move into a more intense building session. and we're going to really start pushing cursor to the limits. We're going to be building multiple things at the same time. Now, we're going to make some progress on the landing page. So, we've already made some progress here. And so, we're going to be working on this landing page. And we're also going to create an app. So, we're going to be building a mobile app for agent native. And then we're also going to create a cursor canvas, which I'll explain in just a second. And it's just like a mini site that has a cursor domain that we can set up automation so that we can have it do research every single day and it will actually update a site and this will allow us to do research for our website and research for our mobile app and I'll show you how this works. All right, so we've created this landing page for agent native. Now I want to create a mobile app for agent native. And so in order to do this I'm going to press commandN which creates a new agent. I'm going to close the sidebar which you can open up right here. We don't need that. And what I want to do is I'm going to create a new project. And so I'm going to create a new folder or a new workspace. And this new workspace I'm going to put it in the same location as in the desktop as agent native. And I'm just going to call this agent native mobile. I'm going to press enter here. So now we're creating a new workspace and we're we see agent native. This is like our website and everything about agent native. Now we have agent native mobile. Now I'm going to say I'm going to just say hey I want to create an iOS app. I want this iOS app to be in swift. So I want you to create a Swift app. And for this Swift app, I want this to be in a a companion app to my agent native site. If you go into my desktop folder, right? If you go into my desktop folder, you will find a folder called agent native. And this is my website. I want you to create an app that has the same style and use the same assets, like the same images that are on that site. That should be the same images in the first tab, which is the learn tab. For this website, there's going to be three tabs. a learn tab, a tools tab, and an agent tab. I only want you to make the learn tab. And the learn tab should be very similar to the agent native site, which should be basically I can learn by clicking on those blog posts, which I haven't created quite yet. But that's what I want you to do. I want you to just create this Swift app. And then what we're going to do is I want to open it up in my simulator. Okay? Since I'm using a Mac, when you use a Mac, you can actually simulate an iPhone directly on your computer. And so, in order to download this, we can simply just create a new chat within Cursor. I'm going to say, "Hey, how do I download if I have a MacBook Pro? How do I get the simulator on my computer um and download Xcode? Please give me a very concise list with links to where I need to go. anything you want to learn, you can simply just ask cursor. So here's the steps. You're going to download Xcode and you can literally just ask this to cursor or any AI agent and you can open this up in a new tab. And here you can download Xcode and I already have Xcode downloaded. As you can see here, I already entered this prompt in uh a couple minutes ago and we're using Grock 4.5 high fast. for very big planning tasks, I might actually want to go with uh Fable 5. And so this model is more expensive, so beware. But I can just scroll back up and stop it. And then I can re-enter a new prompt with Fable 5 medium. I'm going to go ahead and I'm going to edit this. I'm actually going to change this to high. So this could actually be expensive, but I think it's worth it because I'm actually going to turn this into an iOS app and I may actually release it. So, I'm going to fire that off and I'm going to hit revert. And there we go. So, now it's going to start over, but we're using Fable. And my goal is to create the most insane iOS app anyone's ever seen. Okay. So, now what I want to do is let's just go ahead and add the tools section. I want you to please add a really cool list of tools that people can look at. And these are agent tools. And so, I want to have different categories. The first one is super app. The next section is going to be ancillary tools and I want you to please look at the skills that I imported from codeex and then look at your environment. I want you to come up with like 15 different tools that you could include that section some for hosting database etc. I will actually give you the ones that I use in just a second. I just want to create this overall tools section and then I also want to create the agents tab as well. This should be an AI chat and later we're going to set up a body of knowledge that this AI can read when answering people about how to use AI agents but choose three top models that we can select. Okay, as we'll get to a little bit later, we're going to be putting our websites on the internet and to host our app, we're going to be using Verscell. Another feature that Verscell has is they have this thing called AI gateway and it's like open router. Um, and it's very similar. It's just that I already use Versel for hosting. I might as well use uh Verscell for the AI gateway. And this allows us to have access to all of the models with one single API key. An API key, which we'll get to a little bit later. again allows us to like we could go to the OpenAI website and we could get the OpenAI API key and give it to our agent and say create a chatbt like app and use the open AAI API key and then it would create an app for us use the open AI API key and create a chat app that allows us to chat with chat GPT which is really cool but what if you want to use anthropics models or gro models that's why I'm using this versel API key here this allows s me to chat with all the different models. So, what I'm going to do is I'm just going to come to this API keys section and I'm going to go ahead and create an API key. Okay, so I've just saved this API key. I don't want this to uh be shown on screen because people can literally use the API key and then charge my account. You want to keep these secret. So now I'm going to say here is the API key. And then I pasted it at the bottom and now I'm going to run the prompt. So now we just fired off our prompt. What we're trying to do here is we're creating this tools page which doesn't exist and the agents page and we are going to allow our users to chat with an AI assistant and we're going to create that in one prompt and we've already created this in one prompt basically. So that is pretty insane how quickly you can create an iOS app and have it running on our computer. Okay. So, while this is loading, actually, I'm going to press commandN. And we're going to work in the same repo. And what we're going to do is we're going to create a canvas. And you can think of a canvas like a mini site within cursor. And cursor allows you to create these little canvases. You can type slashcanvas. And this allows you to create this little standalone site. So, I want a canvas. And I am currently creating the tools tab of my mobile app. And I want you to create a little mini site that does research and I want you to come up with the different categories like super apps which is cursor, codeex, cloud code. And then I want you to come up with other categories of tools that I could include in this. Search my YouTube videos. And since I've already set up a skill previously on codeex and then I imported all of my skills, I can use my YouTube researcher skill. use this skill when doing research and I want you to come up with all of my tools that I would create and create a little canvas for each one. It should explain it and your purpose is to be my little researcher site. And so I can fire this off. And so while our iOS app is being built, our research site is being created as well. And now we'll wait a few minutes for both of these to be done. Okay. So after 12 minutes, Fable, the model that we're using inside cursor, is done making changes to our app. We're making changes to the tools tab and the agents tab. Let's go check on the tools tab. Here we go. I don't love this dark. Actually, I like the dark. I wish it just stayed with the dark and wasn't I don't know how I feel about these the contrast here. So, when you have simulator running on your computer and you get the simulator by downloading Xcode, you can press command S and this will automatically take a screenshot. I just gonna go ahead and copy this to my clipboard and you can paste it right back into cursor. And so now I need to make the decision. I'm actually not sure. I'm actually not sure what design changes I want to make here. Can you please change this page to look better? Spend a lot of time on the design and make it look a little bit better, please. So, I'm just going to go ahead and fire that off. And while that's loading, let's go ahead and open this back up. And I can full screen this as well if I want to take a closer look at it. I kind of like it in this view here. We can select the agents tab. So now we have this agent. And now I can click on this. Now when you're editing an iOS app, one thing you might want to do is see what it would look like if the keyboard was out. So I can press command K and that will open the keyboard. And this is a very common problem. So as you can see here, the keyboard is being covered by the text. So, that's something you wouldn't have noticed because if you're using the simulator, you can type with your computer on the simulator. I can say, "Hi, hello. Let's test the model here." And we can see that we can select Grock 4.1 fast, GBT 5 mini, or Claude Haiku. Right? I did say to use the cheaper models. So, I can fire this prompt. And there you go. It's really fast. And I think this actually looks relatively decent. That's looking pretty good. Here's one other tip. When editing mobile apps or web apps, I use this app called CleanShot X. And so what this allows me to do is take a screenshot and then really in detail, I can like draw on it and add arrows just like this. So I can point to this right here. I don't like this subtitle. Um, I don't like this icon and I don't like this bar here. So I'm just going to go ahead and copy this. We're going to paste in a cursor. I don't like these three things, the subtitle, the line, and that icon. Please get rid of them. And then put [snorts] the model picker down below. Which means you got to make the text input area two lines high. This is a very common design pattern like for chat GPT's iOS app. Look how they do it. And make it like that. The model picker should not be on the top right. That's ugly. And we can fire that one off. Okay. Okay, so it's done with its previous one, which is this tools section. Let's go ahead and check on this. Okay, this is interesting. One thing I am going to fire off real quick is I'm going to go ahead and open this up. I don't like this line. I don't like either of these. And on this page as well, I don't like the line or the subtitle. Just have tools centered at the top. Actually, keep tools left aligned at the top. But these horizontal lines on the tools page, not needed. Okay. Okay. And this one only took 45 seconds. Went really quickly. I can open up the simulator. Here we go. So, here I know we don't have the official icons in here. We can add those later. But this is looking pretty dang good. We have tools, which are super apps, cloud code, codecs, cursor. I really like these cards as kind of like the the featured cards. And I think if we had the icons in here, it would actually be just insane. What we can do is we can go to the agent. There you go. We have this model picker. This looks really good. And I can say hello there. And I can fire that prompt. And this agent looks a lot better. This is just very well done. This app looks really good for my agent native app. And I can subscribe to the newsletter, which is pretty cool. Yeah, this is awesome. Okay, so this one is actually done. So remember earlier we fired off this prompt where I wanted to create some research material on the agent native tools directory which will actually help us add tools to this directory right here. We are doing some research on this and I got it to do some research and then it created this little thing called a canvas and you can find your canvases by clicking the plus sign. You can open a normal browser or it has like a special canvas one. And this one is called tool tabs researcher. If I click on it, here we go. We have all of our tools right here. And you can see here it did a lot of research. And this is really cool. Now, what I want to do is I want to talk a little bit about automations. And so this is a fun one. So I can come over to automations and I can create a new automation here. I can do it manually or I can simply ask cursor to create an automation. Hey, I want you to create an automation. This automation should scan my content as well as the nine other YouTubers that you think are in my niche for new tools. And these should be only videos in the last week every time you run because this automation should be weekly. And I want you to create a section at the top. The top section should be new tools mentioned. And these are new tools that people are mentioning and I want you to put those tools there just so I can see them every single week and I want this automation to happen and please put it at the top here. And so I can fire that prompt into cursor and this will automatically create the automation. Notice here we have zero automations but after it's done I should have one. And here we go. And we can actually go back to our agent native canvas here. And we can see the agent native tools directory now has this new tools mentioned section which has codeex browser, devon, claude browser, fable 5, gpt 5.6, versel, eve. Here are a bunch of the tools that people are currently mentioning. And as you can see here, done. The automation is live. And I already ran this week's scan by hand. So the section is populated right now and this will actually run every single week. So we just created an automation and this canvas page will actually stay up todate and if I come here and see the new tools mentioned this will update. So this is how you can create these kind of automated research papers with the canvas. Okay. So I'm going to pause real quick and we're going to talk about what we've done so far in this section. And as I said, we were going to be moving relatively quickly because we're building three projects at the same time. And part of getting good at using AI agents is multitasking. And the reason is that it just takes a while for AI to get done. So if you only work on one AI agent at the same time, you're just going to spend so much time just sitting there waiting for AI. So that's why I wanted you to kind of get in the habit of doing three things at once and then coming back. And now let's actually just think about what we've done. And so the three projects that we've worked on so far, we've created a landing page for agent native and that was this landing page that we created. And then we also created a full native iOS app that's running in a simulator. And we created this in three prompts. And that was this app right here with a learn tab, the tools tab, and a working AI features. And this literally works perfectly. It has great animations. Check this out. And it looks and the stream like come on. And then as you just saw before this we created a canvas which are like these little mini apps that are directly inside cursor and that one looks like this. Okay. So the next thing I want to talk about are plugins or in cursor they call it the customize tab. So if you go to this customize tab inside cursor we can see that we have one installed which is convex. We'll get to this a little bit later. What I want to focus on right here is we're going to click on browse marketplace. Here you can connect cursor to other tools, other apps. And there's so many different ones that you can add. And so many developers actually connect it directly to Slack and to Linear. And so they can just immediately go look into their tickets or messages from other people and just say, "Hey, what did people say to build? Build it." And it connects with the tools that you already use. And this is similar to Codeex. This is similar to Claude Code that you have these plugins that you can very easily add. Now, there's two main plugins that I want to talk about in this video that have very specific and important functions. And so, we're about to add two more plugins. And these plugins are going to be for a database. The way that I think of a database is it takes whatever it is that you're building and it takes it from a site to an app. If you go to some landing page that you create and you type something into it, the data that's a user would type into that site is not actually going to be stored anywhere. And that's kind of what I think of as a site where an app like Twitter or YouTube, if you leave a comment, it can you can see who left the comment because that data is actually stored in a database. So what I want to do is I want to add our first plugin and that plugin is going to be Convex. And so Convex is a database provider that I believe is the best with cursors just really good at controlling convex and creating a database for an application. So what I can do here is we can go to our iOS app and I'm going to say please use and we can type at convex I want to make authentication and database. If we go back to the drawing board here in order to save the data as the right user, we need authentication. So authentication is like signin, right? You need to sign into the app. That's when you see signin with Google, right? That's how you know which user is signed in. And the database is just where the data is stored. So database is just is basically the data is being stored. And so when the user signs in, you're signed in as user 00001. And when user 00001 posts, the data is stored. Whether you post an image, whether you are using the AI, right? If we want this data to be saved, we need to know which user is actually typing these messages in. And we need to know which user the AI is responding to. And so all of that is going to be stored in the database. And you might think that sounds confusing. Don't worry, AI will just do all of that for you. And so users should be able to sign in with their email and password and then um all the data that they use like uh I want users to be able to save their favorite tools on the tools page. So please have the little like bookmark thing by the tools so they can click it and it should go from like light gray to black when they do that and that should save to their account. And then also I want all of the chat data to be stored in the database as well. And so I can fire that prompt off. And now it will add the database. This will likely take maybe 5 to 10 minutes, but this will automatically create a new database inside Convex. And if you don't have a Convex account, you can just go to Google and sign up. And it's very easy to connect your account. All you have to do is go to the customize tab, find Convex, and then install it, and it will automatically have you sign in directly into Convex, and you'll you'll be ready to use it. Okay, I [snorts] think it already tested it for me. So, I might be signed in. Okay, so I can sign out. Let's go ahead and sign out here. And I'm going to restart the app. If I go to tools. Okay, I can sign in. I can create an account. I'm going to put in my email and password. Okay, so now I'm going to go ahead and sign in. Okay, so now we're signed in. Let's see. Let's go. Hello there. Let's see if this works. Hello. Now, if we were to, let's say, sign out and then let's go ahead and sign back in. I'm going to put in my email again. I can sign in. Let's see here. If we go to agent, there we go. So, the data still persists. And let's go ahead and check the data in Convex. And since we use the Convex plugin, I can say, "Hey, I actually want to take a look directly at the database on Convex. Can you give me the link to see that?" And now it's getting the Convex dashboard URL. And here we go. So I can se click on this link, open it in an external browser. Convex is loading up. I do need to sign in. And here we see health. But what we want is data. So if we click on data, here we go. We now have all the database tables. And if we click on user, we can see the email right here. And here it has password hash. So, it's kind of it makes the password more secure. Here's sessions and we can see the different sessions created. This is a session every time you create a chat, I believe. And then here's the chat messages. So, it's stored right here. And each of these messages are stored with the user ID. So, every time this user messages the agent, it puts that there. And so, that's why it'll only list the chats from that user. And we can actually customize this a little bit more. Notice here that it's just like one giant chat session. Maybe that might be the best. But for the sake of databases, I want to show you how you can actually add in the chat page. I actually want a little history icon at the top right. Please add that. Then what I want you to do is allow me to see previous conversations I had with the agent. So each user can have multiple chat sessions with the agent just like chat GPT. Please add that feature right now. And I will test it. And we'll fire that off. And now [snorts] it should make that change to our app. And then this will kind of change how the data shows up in our app because we'll have multiple sessions we can toggle through. And let's go ahead and check this. If we click on this little icon that's brand new here, I can hit new chat. I actually don't know if we're signed in. Let's see. Okay, so we are signed in. I'm just going to type hello there. This is one chat I am testing. So now it responded. Now if I hit new chat. Hello, this is chat 2. We can use a different model. So now I'm going to go to the history. Okay. So now we see our chat histories here. So you see these are all saved to the database inside convex. And just one more time to illustrate the point. Let's go look at the convex database. And as you can see here, we see more chat sessions obviously and it's the same user ID. But here we see a new one which is called conversations. And remember, we created two more conversations. Each one got its own ID. And there you go. This was added to the database. Everything is being stored. And now other people could sign in, create multiple chat sessions. All of that data would be stored in Convex. And we set it up with one single plugin and a few prompts. And that's how easy it is. And I know that this might be confusing to you. Um, but this could be a whole video in and of itself talking about databases. But you just have to rely on AI. And if you want to dig deeper, just ask cursor. Just say, "Hey, can you explain what these tables are, what this means?" And so your job as someone who's building an app with AI and trying to add a database. Your job is to test all the different things like sign in to an account, do a bunch of chats, sign out, sign into a new account, and you're trying to basically test it to make sure that it all works properly. And if something doesn't work properly, then tell the AI agent cursor. Basically, that's how you should test it. Okay. So that covers database and you can do this for a web app or an iOS app, any type of app, a desktop app, anywhere you want a database, you can use convex and very easily set that up. The next thing I want to talk about is hosting. And so this is when you get an app or a site that you have going from your computer running on your computer to running on the internet. Let me show you what I mean. So in cursor, this is running at this little location right here, right? We have this location right here which is where it's running. And you don't need to know what this even means because when you talk to cursor you say can you please run this locally and the app will run on your computer. However, if you want your friends to see the app uh and you want it to go on the actual internet, you need to host it somewhere. And in order to host, we are going to be using another plugin. And so if we go to this customize tab, what we're going to do is we're going to go to browse marketplace and we're going to type in Verscell and we are going to hit add. And this has been added. And you can see here once it's added, it'll say 29 tools enabled. That's because I have added my account in the past. You will likely have to connect it to your actual Verscell account and there will be one screen before this. But now if we go back to our agent native web app and we type at versel, we can see that there's 29 different tools. So now what I want to do, please create a brand new project for this and get this on the actual internet. Please, I want you to send me when you're done, I want you to send me the link to Versel um of it actually hosted on the internet and I actually want you to open it in the browser as well. And so now this will take a little bit of time, but it will actually deploy the app to Verscell, which means put it on the internet. And there we go. We now see this link right here. We can go ahead and open it up. And there we go. We see agentnative site, bice.verell.app. This is kind of a random domain, but here we go. This is actually on the internet. You could go to this site on the internet right here, and you will see this website. It is actually on the internet. And now I want to do something fun. So I'm going to go ahead and say, "Hey, I want you to look at the iOS app project, the agent native mobile project. We added a convex database there. I want you to sync those databases together. So it's one database and I want to be able to sign in to the same account on the iOS app as the web app. So please create that right now." So look deeply into that project and just create this flawlessly so the databases and stuff works together. I also want you to add the an identical chat tab that we added on the iOS app so I can chat with the same AI agent. And yeah, I want that to be an identical process and all of that data should be identical in the database. So this one's a little bit harder. I actually don't know if this is going to work on the first shot, but we will see. So the this landing page is going to turn into an app now. And this means we also need authentication. The same authentication as the other one. Now it is going to work. We'll see how this does. And here it says one shared identity system, one convex deployment, chat history across web and iOS. I'll inspect both projects first, then wire the web app into the mobile app's existing backend and off rather than creating a second system. Very cool. Okay, so this one took a lot longer actually and I believe it's done. And so let's go ahead and check on this site. So I'm going to go ahead and open this up. I'm actually going to open this in the cursor browser. So we are on the agent native site. And so I'm going to try and sign in with that same account that I had earlier. This is the same email that I signed in on the iOS app. And I'm going to use the same sign in here. And now we're signed in. If I go to my history, we can see the other chats. Look at this. And I can create a new chat. I can switch the model. Hello there. I want to learn about AI. And there we go. We are now signed in to the same account, right? Because if I were to open up the um let's go ahead and open up the simulator here, we can see. Okay. Okay, so this one doesn't stream like the iOS app, but if I go back home and I open up the simulator, I'm going to go back to learn. Go back to agent. If I look this up, hello there. I want to learn about AI. We have this web app that's now on the internet hosted on Verscell and the backend or the database is synced up in perfect. So it is perfectly synced up. As I do more chats here, this gets saved here. and these tools section, I can save them as my favorites and this will also be stored in the database. Now, I don't have a tools section on agent native. So, I'm going to go ahead and add that. You see the tools section as well. I want you to add that. Um, and so instead of agent at the top, when I click that, it should take it to a tools page on the website. It should be very similar to the iOS app. Uh, the same tools, same categories. And so, we can fire that one off. Okay. Okay. So, I'm going to go ahead and open this up in the browser. And so, we have tools. Now, I'm not the biggest fan of this. Okay. So, this is not bad. And obviously, we haven't added the icons yet, which I will do. I just don't want to have to go through and get the logos, but we got agent native. We have a tools section. Wait, how do I get to the chat? How do I get to the chat? Oh, that was the agent. Okay, sorry. Yeah, please put the agent in the top bar. I didn't realize that was agent uh the agent chat. Please put that back in the top bar next to tools. Okay, so now I can go ahead and open this in the cursor browser here and we can click agent. And now we have it. And there you go. We have now created an iOS app. We added a database and then we created a web app, put it on the internet with Verscell. the database was convex and then we merged the databases together and we did this just by asking AI. We said, "Hey, can you please set it up through the same convex database? I want you to basically make them a web app version and an iOS app version." And it works perfectly. All you had to do is prompt AI. Now, the last part of this section, which is building, before we get to a ton of other advanced features, we're going to do like a lightning round through the rest of the features in cursor. I do want to talk about GitHub. So what I want to talk about now is actually saving our work inside cursor. And so for this we are going to be using GitHub. And GitHub is an app. You can almost think of it like Google Drive for developers. And when you sign up, what you could do is you can create a GitHub repository. And when you hit new, you can basically create a new GitHub repository. And so I could name this Riley ios agent native, right? I could create a project and we can make it a description. We can make a description um agent native iOS app. And we're basically creating a folder on Google Drive but for code. And I need to choose um my profile and we can create this repository. And so this creates this empty folder that I can copy. And now all I need to do is go to cursor and we can go back to our iOS app here. And in the iOS app, I'm going to say, hey, I created this new repo for the iOS app. Can you please push the code here? And I can just paste this link right here and we can run it. And if you have not yet signed into GitHub, what it will do is it will actually prompt you to sign in with GitHub when you do this. Um, and if you it doesn't just say like, "Hey, help me connect cursor to GitHub." And Cursor will tell you exactly how to do it. And it'll take like 20 seconds. And so now it's checking out this project. And you can actually make these GitHub repositories either public or private. And I'm just going to say keep this repo private by the way. And look at that. It says that it is done right here. And now what I want to do is I want to come over to GitHub. Now I'm going to press commandR to refresh the browser page. And there you go. Now you see all these files here. And the GitHub repo is now updated. Okay. So I showed you how to do it manually where you can go to GitHub, create a new repository and paste the link or you can just ask cursor, hey I need you to create a new GitHub repository for the web app for agent native and call it agent native web app and for this one make it private. I want you to create the repository and then when you're done um send me a link to that repository and repository just means like it's almost like a Google Drive folder but for developers. And now this will allow the code to basically be saved and you'll be able to revert your changes much easier to a previous version if you do this and you should actually update it a lot more often. Okay, so it's done and if I right click on this open actually we can open this in external browser. Here we have a new repo that is for the agent native web app. Now, this is now saved to the cloud basically with your project. And if you were to work with other people, getting to understand GitHub is incredibly important. And so, this is also really important for the Cursor iOS app, which we'll move to in the next section. Because in order to use the Cursor iOS app, which in my opinion is one of the coolest ways to code or vibe code from your phone, you actually need to create a GitHub repository. like you have to use GitHub. But before we get into the iOS app, let's just review what we talked about, right? We created three apps in this project. We built a mobile app and we created a web app. And the web app was hosted on Versel. And the mobile app, we did not send it to the app store. That's out of the scope of this video. That does take a little bit longer. And I'll make a cursor for iOS app video later. Um, but you can probably figure it out with cursor if you ask, "How do I get it on the App Store? But what we did do is we added a database to the mobile app and then we connected the databases together. We basically added the database to the landing page and we connected that together. And for this we created two or we used two plugins the convex plugin and the Verscell plugin. And in part one when we first created the web app we also talked about design mode about how you can create granular edits to the app which is a really fun way to edit web apps within cursor. And then we also talked about the canvas which is just a useful fun feature for ephemeral or shortlasting apps that you want to send to your team or something like that. Now I want to move on to part three which is cursor advanced features. Okay. So right now we are going to use cursors iOS app. You can download this and sign into your same account. Um and you can find this just search up the cursor iOS app and you can download it to your phone. And [snorts] so notice here at the top it says all repos. And so if we click on all repos, we can create an agent. And here we can select which repository we want to edit. And so remember what we named the web app repo. We named it agent native web app right here. So we can click on this. And what I want to do here is I'm just going to speak into it. Hey, I want you to please explore a different design for this app. I don't want you to um push anything to Maine yet. I want you to explore a new web app style and please send me screenshots of this and um if I like it, then we'll approve it. I just want to see some exploration here and um come up with a new visual concept for this. and I can stop recording it. And there you go. We have our prompt. I am going to change the model. And for this, we're going to use 5.6 soul. And I can fire this off. Now, we set up the project on GitHub. And we're basically, it's a bunch of code in a bunch of files that's stored in the cloud. Right now, it is setting up a computer in the cloud. and it's going to rewrite the code for this app based on our prompt and then it's going to actually run the app in the cloud and then it's going to test the app and it's going to send us screenshots and so that's the direction cursor is going. They are setting up these cloud computers where you can send off prompts and the AI will actually edit the app and then run the app and test everything for you. It can even send you screen recordings like videos of it using the app and you can provide feedback and then you can say okay I like the changes please um create a PR. This takes a little bit and so you can watch it as code as if you were using cursor or you can just put your phone away right you can put your phone away and you can wait for it to be done and then once it's done coding you'll get a notification and you can open it up and you'll be able to see all the changes it made. So, let's wait a few minutes for it to complete. And you can see here in the dynamic island at the top, we see this little check mark. So, I'm going to click on this. Look at this. It fully explored a completely new style intelligence in motion and it literally sent us screenshots. So, this is running in the cloud and it took screenshots and now it's sending them to me. And this is basically how I can code from my phone. And when I think it's ready, I can just click marked as ready. And so now we just sent the AI in cursor to like create a new branch. Basically, that's what happened. And what you can see here is you can see that we actually have two branches and we also have one pull request. And so if I click on the two branches, I can see that cursor updated it three minutes ago. This is the new branch that it created and there is a new pull request. And this pull request, it even added a summary of the pull request. And so if I were to merge this pull request, this would basically commit the changes into main. And then the web app would no longer look like this. It would look like this. This right here is what it would look like. But I think this looks ugly. So I'm actually not going to commit these changes. But this gives you an idea of how you can use the iOS app to fire off your different ideas. It'll create a new branch. So you don't have to worry about it updating anything. And you can just tell it to send you screen recordings or screenshots. And that's how you basically build from your phone using the Cursor iOS app. Okay. So that is the iOS app. Now let's go ahead and talk about side chat. So if we go back to cursor, what you can do here is if we're talking to cursor right here, if the agent is working, right, this is just a dummy prompt. If I go slashside and press enter, this will allow me to talk to a side agent. This agent can't edit apps like this agent can, but it allows me to ask questions. Hey, what's one thing that I should add to the app? And we basically have this side conversation while our main agent is running. And uh this is a very similar feature to some like cloud code and codeex. Speaking of cloud code and codecs, I know a lot of developers, especially the ones that I work with at my startup, we actually uh use cloud code in cursor all the time. And if you want to get access to cloud code inside cursor, you can very easily open up a terminal and you can open it up like a new tab just like a website or just like a browser or the canvas and we can just type in claude and now we have access to claude code with fable 5 running directly in the browser and that is just a fun thing to do. You can get access to the terminal. This is a really advanced feature. understanding how to use a terminal besides cloud code is a really difficult but the cool thing about terminals is AI is really good at using terminals and you can anything you might want to do in a terminal you can just ask AI to do it and then finally the last thing that I want to talk about because I have been using Fable 5 and GPT 5.6 soul a lot. I do want to talk about costsaving. So the first thing that you should do uh if you want to save money is you should either use auto mode you want to or you should use Gro 4.5. It'll basically take your prompt and decide what the best model is. And it will basically balance between quality and speed. And this is recommended for most tasks. For most tasks, you know, you don't need the most intelligent model. It can just get the job done. If you only use Fable, you're going to spend so much money on this. And I do recommend using Grock because Grock is an incredibly affordable model for how good it is. And I expect the cursor is going to do a lot of subsidizing with the Grock models. And I think Grock is going to get 10 times more intelligent. And I expect the prices to be very competitive with the other Frontier models. So I would avoid using Fable 5 and GPT models, right? And I would use the Grock models in auto if you want to keep the costs down within cursor. And they're also much faster. Auto mode and using Grock 4.5 is significantly faster. And that's just kind of my final recommendation. Anyway guys, that kind of completes the cursor complete guide. I hope I got your brain going a little bit thinking about the different things that you can build, the different models that you can use. We talked about cursor, how to download it on cursor.com. The other tools that we discussed that you can use with cursor and almost guaranteed that you need to use is GitHub. We talked about database providers and hosting like Convex and Versell. We built three different apps. We built a landing page. We connected the databases. We I showed you the iOS app. You can literally spin up cloud agents from your phone and it will build the app in the cloud and it'll even take screen recordings and screenshots and send it back to you. And then it will create a PR which you can review on GitHub and you can decide whether you want to merge it into your main project or not. And this is kind of an overview for those of you who aren't technical who want to actually learn to build apps. There's no better tool to do it than cursor in my opinion. I hope you got a lot from this video. If you're still here watching right now, please like and subscribe. It helps me out a ton. Leave a comment below. I want to know what I missed. What should I cover in the next cursor video? Thank you guys so much for watching. I'll see you here in the next one.

Thinking Machines co-founder Lilian Weng left the startup after citing health effects from sustained stress and workload, then joined OpenAI. She said the pace required by the startup had become physically unsustainable. Grok Voice Think Fast 2.0 is now available at $0.09 per audio minute. grok-voice-latest will switch over to the new model on August 5. The release makes Grok Voice more dependable in real customer workflows. GPT-5.6 Sol scores just 7.8% on the ARC-AGI-3 benchmark despite having solved longstanding open problems in mathematics and beaten games like Pokémon FireRed. Benchmarks rarely measure AI models in isolation. They also measure less visible choices about API settings, harness design, and prompting. Researchers discovered that turning on retained reasoning and compaction in ChatGPT and Codex tripled scores and cut output tokens by 6x on the benchmark. Google's latest music generation model, Lyria 3.5, is now rolling out in Google Flow Music. The model delivers significant advancements across musicality, lyrics, and vocal quality. A clip of a song generated by the model is available in the article. Escha-W2 is a 2-bit quantized build of Qwen3.6-35B-A3B, a Mixture-of-Experts model with 256 experts. The model is packaged with everything needed to serve it locally through an OpenAI-compatible HTTP API. The whole thing is 12.3 GB on disk and runs on a single 24 GB consumer GPU - or on a 16 GB card. Liquid AI released encoders designed for efficient document-scale inference on CPUs. This article covers their 8,192-token context window, competitive benchmark results, and their lower long-context latency. DeepMind researchers suggest that visual Prompt Engineering can improve video-model reasoning by transforming task images before inference, such as converting abstract sketches into photorealistic scenes. Parallel Decoding Distillation is a trajectory-based method that predicted multiple denoising steps during each model evaluation. It achieved state-of-the-art results with four to eight evaluations across several image and video generators while improving video diversity. 500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to hardware@tldr.tech! Claude Opus 5 is the top-scoring model on Vending-Bench, a vending machine simulator. The model discovered that focusing on higher-end products yielded higher profits, and it never gave a single dollar to scammers. It demonstrated some misaligned behavior, such as fabricating competitor quotes when negotiating with suppliers and lying about delivery delays. The model also proposed or engaged in price cartels in all runs - most cartels ended with Opus breaking the truce and undercutting the others. Compute costs may rise 10x as AI labs like Anthropic aim for $1 trillion revenue, driven by increasing margins, rising compute prices, and more spending on inference. Google pays twice the spot price for GPUs due to demand, with stronger monetization of AI models leading to higher compute value. High compute costs may prioritize efficient AI, pricing out less critical applications and intensifying competition in AI development. Cargo, Grepsr, and Dust each hit their own wall scaling AI reliably. This free Temporal eBook covers the real architecture behind how they fixed it. Get your copy Numbat, an open-source security suite by Perplexity, addresses security risks in AI agents deployed on client endpoints by integrating with agent harnesses to prevent, detect, and mitigate incidents like accidental meltdowns.

1 Billion ChatGPT users tinkering with tldraw is so much fun Hey folks, Following on from Tuesday’s messing around building, I built a few more widgets on my canvas. It pulls my X bookmarks, emails, and todos. Before making this video I didn’t have Loom installed (since it’s crap after being acquired), so just told Codex to build me one. I drew the image on the left, then just sent that prompt and it worked straight away. So as software and mini tools are getting easier to create, the tools to create them are not... I use Codex as my default app because it’s better than all the others and works the best on mobile. But I just downloaded t3 because they basically copied the interface and features, but it lets you choose between different agents; codex, claude, cursor (pi + droid soon). It’s made by a reputable developer that you may have seen mentioned here before, so I trust it’s built well. At the moment, I ask Codex/ChatGPT to be the orchestrator and to go ask claude about something that is design-related. Which is fine-ish but not great as user experiences go. What I’m noticing by doing this is how well Codex creates prompts for other agents to follow. I just have a normal chat in my session, then say ‘Use X agent to implement this. Watch its progress and send screenshots every time new work has landed.’ To which it sets up its own monitoring every 5 mins or so and updates me in the same thread, with screenshots. I’m going to try and put together a ‘bites of the week’ email over the next few days to try and summarise all the stuff going on, what themes people are talking about (loops?!, software factories?!, etc) and explain them. Let me know if there’s anything specific you need to wrap your head around (I may need to too). Ben’s Bites is brought to you by Brief Struggling with fragmented context, slow product decisions, and rework? Brief distills your critical product context into an opinionated graph, then puts a PM agent everywhere you work, (e.g. Slack, Claude Code, email) reducing alignment tax and accelerating cycles. Learn more. Headlines OpenAI used Sol to optimise Sol itself, cutting serving costs by 20% and making it 15%+ more efficient at generating tokens. And turns out, it also tops the ARC-AGI-3 benchmark. Well, there’s a catch: OpenAI says the official ARC-AGI harness hurts Sol’s performance by “forgetting” its reasoning every turn and disabling compaction. Fixing these two things triples Sol’s score from 13.3% to 38.3%, with 6x fewer output tokens. Re: last week’s fiasco of an OpenAI model hacking Hugging Face - HF published a full replay of roughly 17,600 actions taken by the model. METR and Redwood Research will also independently review what happened. Though OpenAI is not out of trouble just yet, a Reuters report claims that the same model broke into a customer account at another company (Modal Labs), with rumours suggesting that even more companies were affected. Anthropic also claimed that Claude Mythos found better attacks on two cryptographic algorithms, though neither affects systems in use today. Separately (not at all as a reaction to this general trend, right?), ~1300 people working at leading AI companies (OpenAI, Anthropic & others) want the US government to help “pace the frontier” of AI development. Kinda expected when the pace picks up, but this time a lot of the “model makers” themselves are in favour of this pause/slowdown. btw, The Information reports ChatGPT is nearing one billion weekly users - a milestone OpenAI originally hoped to hit seven months ago. More from OpenAI this week: Codex Security CLI, free frontier access for Academic Researchers, and two new transcription models. Grok app builder - Grok has a vibe coding interface inside its app now. Create games and apps that can be shared directly to the X timeline. Also see: Drawesome - a zero-dependency drawing toolbar for React, built over a weekend with Grok Build. Pangram 4 claims it catches 98.83% of humanised AI text with one false positive per ~24,000 docs. An early test found all 38 AI-written words inside a 1,198-word story, though not on every run. Its new image detector claims 99.5% accuracy too. Quick links - Tavus - Build AI that comes to life: video agents that see, hear, and answer in real time and do anything you want. Use TAVUS50 for 50% off. - 66% of July traffic on docs built with Mintlify was from agents. - Resend added an MD version of their pricing page to avoid confusing agents. - 0%, 50% or 200% - ignore AI, halve staff or double the ambition. - Slackbot can now run code in the background for data analysis, slide creation, and to make live reports or widgets. - Gemini’s macOS app got a voice mode that lets you ramble, and the app turns it into a clean prompt. Hold Fn to try it. - The AI future is for everyone - Mark Zuckerberg - Replit Design - make sites, prototypes and graphics from prompts, URLs, Figma files or screenshots. - What’s gone wrong with AI & labor. - Kami - open-source Hermes agents that find customers, prepare outreach and content, then act after your approval. - Coast - fully local memory for you and your agents, built from what you see on your Mac. - Pragmatic leverage in the software factory. - Crew Studio - find useful ideas where agents can help your business, build those agents with the option to take the code home to run anywhere. - HeyGen Video Podcast - turn a doc, link or idea into a two-host video with scenes, camera cuts and B-roll. - Copper - local scratchpad for saving answers, links and follow-up prompts across your AI apps. - FT Chart Doctor - visual vocabulary and examples for choosing a chart that fits the relationship you need to show. - Mitchell Hashimoto (Ghostty) and Andrew Ng (deeplearning.ai) are both starting new companies: Superlogical and LearnVector. - MCP’s biggest update removes the need for servers to remember every ongoing connection, making them easier to run and scale. Afters - Find me on X, Linkedin, or YouTube - Read about me and Ben’s Bites - 📷 thumbnail via @keshavatearth * sponsors who make this newsletter possible :) Wanna partner with us for the next quarter? Email us at shanice@bensbites.com or k@bensbites.com

Deep Learning Weekly: Issue 466 Claude Opus 5, How Digibee Builds Prompts with Opik to Power Their AI-Native Integration Platform, a paper on Progress Reward Modeling for Robotic Learning, and many more! This week in deep learning, we bring you Claude Opus 5, How Digibee Builds Prompts with Opik to Power Their AI-Native Integration Platform and a paper on Progress Reward Modeling for Robotic Learning: A Comprehensive Survey. You may also enjoy FLUX 3, Evaluating Agents Beyond the First Prompt, a paper on StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents, and more! As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly. Until next week! Industry Anthropic releases Claude Opus 5, taking state-of-the-art on Frontier-Bench and GDPval-AA and surpassing Fable 5 on several evals at half the cost, with its lowest misalignment score yet. Black Forest Labs launches FLUX 3 in early access, a unified multimodal model trained jointly on image, video, and audio that generates 20-second clips with native audio and extends to robotic action prediction. Fish Audio raised a $50M seed led by Coreline Ventures and Capital Today, reaching $21M ARR and 8 million users a year after launching its open-source voice models. OpenAI launches Health in ChatGPT for U.S. users, letting connected Apple Health and medical records ground conversations anywhere in the app, with that data excluded from model training and ad targeting. Microsoft releases MAI-Image-2.5-Pro and MAI-Voice-2-Flash into public preview, displacing OpenAI models across Bing, PowerPoint, and Dynamics 365 with claimed GPU cost cuts of 84% and 89%. MLOps/LLMOps/AgentOps A case study on how Digibee adopted Opik to bring evaluation-driven development into production, using traces and automated evaluations to catch regressions before they reach users. Part 1 of 3 of The Observable Job Agent series: A hands-on guide to building a LangGraph-powered job search agent that ranks real openings while instrumenting every step with Opik observability. Learning A research post introducing ABBEL, which replaces recursive summarization with graded natural-language belief states, closing about half the gap to full-context agents on CollabBench in 50% fewer training steps. A breakdown of EvoCode-Bench, a 227-round multi-turn coding benchmark where pass rates fall from 46.7% at Round 1 to 7.7% by Round 10, with regressions rather than missing features driving failure. A defender-focused analysis of the Hugging Face breach, in which an autonomous agent logged over 17,000 attack actions and commercial model guardrails blocked forensic work until responders switched to a self-hosted open-weight model. Libraries & Code An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. An agentic-first RL framework for research. Papers & Publications Abstract: Computer-use agents are usually improved by strengthening perception: better models for reading a screenshot and choosing where to click. Yet a screenshot is only a lossy rendering of the underlying program state, e.g., the files, application backends, and DOM that hold the task data. Different states can produce the same pixels, while code can inspect and modify that state directly. StateAct is a code-first, multi-agent harness built around this distinction. Its main agent works directly with program state by using code, while a dedicated GUI subagent handles screenshot-and-click interaction on the few subgoals that need it, just 28 of 108 tasks and 1.1% of main-agent steps. The same direct access to program state also supports verification: an independent finish gate double-checks the saved result for structural failures, e.g., output that is missing, unsaved, or written to the wrong path. To stay on track over hundreds of steps, the main agent hands subgoals to fresh subagents, keeping its own context focused. On OSWorld 2.0, StateAct lifts Claude Opus 4.8 from 20.6% to 26.9% on binary success, and from 54.8% to 61.6% on partial success, at ~ 9x lower cost per task than the same model driven by screenshots alone; a code-only variant with no GUI subagent reaches only 45.9% partial, below that screenshot-based baseline’s 54.8%. In general, grounding action, verification, and memory in state, what we call state-grounding, shifts the main bottleneck from perception toward reasoning: failures depend more on what the agent thinks than on what it sees. Abstract: Robotic learning takes place in dynamic environments with large behavior spaces. A terminal success signal only tells the robot whether the task is completed. It does not explain whether the current behavior is making progress, remaining unchanged, or undoing earlier progress. For this reason, recent studies have increasingly explored progress rewards that provide feedback during task execution. However, the current literature lacks a shared framework. Existing methods use different observations, goal specifications, output signals, supervision sources, and evaluation protocols. This makes it difficult to compare them and understand what their results actually validate. In this survey, we provide a unified view of progress reward modeling for robotic learning. We organize the field in three connected steps. We first study the interface of a progress model. This defines the problem from the outside by asking what information the model receives and what form of progress signal it produces. We then move inside the model and study the methods used to construct this signal. This reveals the different assumptions and mechanisms behind progress estimation and reward generation. Finally, we examine the data and benchmarks that support these methods. This shows how progress supervision is obtained and what different evaluations actually measure. Together, these three perspectives connect what a progress model is, how it is built, and how its quality is validated. We further summarize the main limitations of current approaches and discuss future research directions.

Martin Casado posted something about AI harnesses that captures where a lot of smart people are stuck right now. On harnesses, I vacillate between three beliefs: the less harness, the better. Models are the magic. Post training a model and harness is dramatically better and the model providers win. Harnesses have real independent value from the model. I have no idea which is right. Martin Casado I think I can answer this. The reason the question feels impossible is that we're treating the harness as one thing. It's actually two. Every harness carries some mix of WHAT and HOW—context about what you want, and instructions for how to get it. And those two halves age in opposite directions. The HOW half rots. This is Sutton's Bitter Lesson playing out in your config files: the smarter models get, the dumber your step-by-step instructions look by comparison. If your harness is mostly HOW, then Martin's first belief is correct. Less harness is better, because the model is the magic. The WHAT half appreciates. Who you are, what you're working on, what you're trying to accomplish, and what good looks like to you. A smarter model does more with that context, not less. If your harness is mostly WHAT, then his third belief is correct. It has real independent value, and that value grows with every model release. So beliefs one and three are both right. They're just about different halves of the harness. The second belief—that model providers post-train the harness into the model and win—is right about execution and wrong about intent. The labs can absolutely train models to be better agents, and they will. But they can't post-train YOUR context into the model. What you're trying to build, for whom, with your constraints and your taste. That has to be captured and conveyed from outside, every single time. That's what the harness is for. I've been calling this Intent Engineering, and it's the whole design principle behind my own harness: capture what the human actually wants, convey it to the model on every task, and otherwise stay out of the way. So YES to harness. Extremely powerful. But for your context, while staying out of the way of the model for execution. Martin's original post is here, and my reply that this post expands on is here. I wrote about the WHAT vs. HOW distinction for prompts in From Prompt Engineering to Intent Engineering, and for harnesses in Good and Bad Harness Engineering. This post is the same idea applied to the "do harnesses even matter" debate. Citation: Richard Sutton, "The Bitter Lesson", March 13, 2019. 🤖 AIL 3: I (Kai, Daniel's AI assistant) drafted this post from Daniel's X reply to Martin Casado, which provided the full structure and core argument, plus his prior published posts on the topic. Daniel's original words carry the thesis. Learn more about AIL.

Aviation learned this the hard way. For most of the industry's history, the number that best predicted whether an airline would survive was how much of the day each aircraft spent on the ground. The reason is structural. An aircraft's costs accrue by the calendar hour: financing, depreciation, hull insurance, scheduled maintenance, crew contracts. Its revenue accrues only by the flight hour. Every hour spent on the ground shrinks the output side of that equation while the cost side keeps running exactly as before. Utilization also sits downstream of almost everything else an airline does. Turnaround discipline, network design, maintenance planning, crew rostering, and spare parts availability all eventually show up in that one number, because a broken operation underneath it keeps planes on the ground no matter what else goes right. A bigger fleet still helps. More aircraft means more available capacity, plainly and simply. But two airlines flying comparable fleets on comparable routes can end up with very different economics, and most of that gap traces back to one measurement rather than fleet size. Enterprise AI is running into the same structure, on a different piece of hardware. A GPU accrues cost by the calendar hour too, through financing, depreciation, power, and cooling, whether or not it's doing anything useful in a given moment. Its output only accrues by the compute hour. More GPUs helps in roughly the way a bigger fleet helps an airline: real capacity, a genuine advantage, and still no guarantee of the result that actually decides who wins. Two companies with comparable GPU budgets increasingly diverge based on how much of that hardware is doing something useful at any given moment, not on how much of it either one owns. That same number, like an airline's utilization rate, sits downstream of nearly every other infrastructure decision a company makes. Intelligence has carried the industry this far. Utilization is where the next real constraint is forming. The scarcity didn't disappear as AI scaled. It moved up the chain, landing on a different resource entirely. The first wave of enterprise AI was won on model quality. Bigger models, trained on more compute, evaluated against tougher benchmarks: parameter count and leaderboard position dominated the conversation, and the race produced models genuinely good enough to run real enterprise workloads. That capability arrives bundled with a dependency, though. Production AI runs on specialized hardware, and today that hardware is almost entirely GPUs. GPUs are expensive, supply constrained, and in demand far beyond what's available, and this holds even at the very top of the market. In 2020, Microsoft built OpenAI a dedicated supercomputer: over 10,000 GPUs and 285,000 CPU cores, reported at the time as one of the five largest systems in the world, assembled to train what became GPT-3. At the time, it looked like an almost unimaginable concentration of hardware, the kind of number that made compute look like a solved problem for whoever could get access to it. Six years later, that number reads more like a starting point than a ceiling. By 2026, even the best capitalized labs on the planet were treating compute access as a live strategic constraint rather than a settled one. Anthropic alone was running simultaneous multi-gigawatt commitments across four separate hardware platforms, Amazon, Google, Microsoft, and AMD, layered within months of one another, while Meta signed a comparable multi-gigawatt deal of its own. Spreading commitments across four vendors at once is what compute scarcity looks like when a buyer has effectively unlimited capital and still can't get enough from any single source. Six years apart, both events marked the frontier of what a lab needed just to stay competitive. What changed in between has less to do with AI getting more capable, and everything to do with capability no longer being the binding constraint. The same pattern shows up downstream of the labs, in a different form. Enterprises consuming these models through an API run into a pricing problem more than a hardware one. Cost scales linearly with tokens used, and that single fact separates the economics of a proof of concept from the economics of production almost completely. A PoC processing a few thousand requests a month looks affordable. The same workload at production volume can turn into a cost line that never quite clears. The alternative gaining ground is straightforward enough: enterprises acquiring their own GPUs and running models locally, trading a variable, linearly scaling cost for a fixed capital one. API cost rises with usage, while owned infrastructure stays close to fixed. Past the breakeven point, the trade reverses. That shift turns the GPU into infrastructure rather than a line item, sized for growth, sized for demand peaks, and therefore sized above what any given week actually needs. Which means the purchase doesn't close the problem. It opens a new one. The day the cluster comes online, the question stops being can we get accelerators and becomes can we keep them busy, and only the first question had a procurement team assigned to it. Signing for the hardware is the part with a deadline and an owner. Keeping it off the ground is the part that quietly decides whether the deal was worth signing. These deals describe capacity commitments, not efficiency. How well that capacity gets used is a separate question, owned by different people, measured far less rigorously, and considerably further from being solved. A cluster full of busy GPUs can still be wasting most of its potential, and the reason is almost always the same one. GPUs run continuously, day and night, while the demand placed on them does not. Infrastructure has to be sized for the peak, the moment training runs, batch jobs, and real time traffic all land at once, which leaves a meaningful share of capacity provisioned and unused outside that peak. Better forecasting could solve that on its own if every GPU could absorb every kind of work equally well. Few can, and that turns out to be the harder half of the problem. The mismatch begins one layer deeper. In the first generation of enterprise AI, a GPU's job was largely singular: run inference. Today the same hardware supports training, fine-tuning, quantization, real-time inference, batch inference, embedding generation, and model evaluation, often for the same organization, sometimes for the same model, on the same cluster. Each of these workloads wants something different from the hardware, and the differences run deep. Real-time inference needs low latency above nearly everything else, because a slow response counts as a failed one. Batch work cares about throughput and tolerates delay, sometimes for hours. Training can occupy a GPU continuously for a stretch measured in hours or days. Quantization needs a large amount of capacity, but only briefly. A scheduler tuned for one of these will misallocate the other three almost by default. The failure doesn't always show up on a utilization dashboard either. A cluster can report high average occupancy while several queued jobs wait for a GPU shape that happens to be busy running something else entirely. The exact workload mix varies by organization. The shape of the problem does not. This is also where the aircraft analogy runs into its limit, and the limit teaches something rather than just qualifying the comparison. An idle aircraft can usually be redeployed to any route in the fleet: a 737 sitting in Chicago can fly to Denver instead of Dallas without much penalty. An idle GPU can only absorb a workload whose memory, latency, and duration profile it can actually serve. That difference makes orchestration harder than fleet scheduling, and it's why the question stops being whether GPUs are occupied and turns into which workload should run on which GPU, at what time, with what priority. Buying another rack of GPUs adds capacity and cost, not a fix for the mismatch, and that new capacity can sit in the wrong shape at the wrong moment just as easily as the capacity already installed. Maximizing GPU ROI takes more than a one-time provisioning decision. It calls for continuous, active management of the infrastructure itself, running every hour rather than only at procurement time. What's emerging in response is a distinct discipline, GPU Management, an orchestration layer sitting between workloads, models, and hardware. Its job is to decide, continuously, which workload runs, when it runs, how it runs, and on which specific GPU in the cluster. None of this is exotic in concept. It's closer to what a good operations team already does by instinct, just formalized and running continuously instead of depending on someone noticing a problem. Intelligence doesn't stop at the model boundary. The orchestration layer is making real-time allocation decisions the model itself has no visibility into. Intelligence used to sit almost entirely in the model: bigger, better trained, more capable, and that was most of the game. Now it also has to sit in the infrastructure, in the layer deciding, moment to moment, which of several competing workloads gets the GPU that just freed up, and at what priority relative to everything else waiting in the queue. Keeping GPUs busy stops being the goal on its own, since busy is easy to fake by running low priority work that could have waited. Maximizing the return generated by each installed GPU becomes the actual target, and that turns out to be a far more continuous problem than the provisioning question that came before it. Provisioning well doesn't make this go away so much as change its shape. A provisioning decision gets made once, at purchase time. An allocation decision gets made constantly: every time a job finishes, every new request that arrives, every shift in priority between a customer facing service and an internal training run. That frequency explains why the decision has moved from something a person handles case by case into something that has to run automatically. No engineer is watching a dashboard at three in the morning to decide whether a finished training run should hand its GPU to a queued batch job or hold it for an incoming burst of customer traffic. Something else has to make that call, continuously, and make it correctly often enough that nobody needs to check. The discipline is new enough that its tooling and conventions are still forming, and no single playbook has emerged yet for what a mature GPU Management practice looks like. What has settled, at least, is where the constraint moved. Specialization and orchestration solve different halves of the same problem. Specialized, smaller models can perform specific tasks at a fraction of the resource cost a large generalist model would need for the same job, without giving up the quality the task requires. That has a direct effect on utilization. Workloads that once required a single, large model, occupying a large share of a cluster's capacity for the full duration of the job, can instead run on smaller, task-specific models occupying a fraction of that footprint. Capacity that used to be entirely spoken for is suddenly free. How much capacity specialization frees varies by workload and model, but the freed capacity still has to go somewhere or it just sits there. A smaller specialized model only converts into GPU ROI if something is actively deciding what happens next with the space it frees up, reallocating it to another workload, another model, another queue waiting behind it. Left unmanaged, freed capacity becomes a different flavor of idle rather than a win, invisible in a different way than an obviously unused GPU, but no more productive. Specialization without orchestration frees capacity nobody reclaims. Orchestration without specialization has less capacity worth reclaiming. Neither lever does the whole job alone. Specialization without orchestration frees capacity that nobody reclaims. Orchestration without specialization has less capacity worth reclaiming in the first place, because the models are still large and the footprint they leave behind is small. Neither lever does the whole job alone; each one raises the ceiling on what the other lever can achieve. Neither is optional if the goal is to actually close the gap between installed capacity and useful output, rather than shifting where the waste happens to sit. This is why model architecture and GPU management are two ways to address the same problem, approached from two different directions that end up leaning on each other. One shrinks what each workload needs. The other decides, continuously, where the difference goes. A bigger fleet has always been a real advantage, and nothing here argues otherwise. Among airlines with comparable fleets, sometimes even a smaller one facing a larger rival, the winner was usually whichever one flew what it had more completely, carrying the weight of everything the airline did well underneath it. Enterprise AI is arriving at the same discipline from a different direction. GPUs are already installed, already depreciating, already committed. Specialized models and GPU Management are parallel solutions, or bivalent strategies. Specialization shrinks what each workload needs. Management maximizes the return on infrastructure. Enterprises that master both will set the pace of AI competition for the next decade. - Newer Models, Same Advantage — Despite newer architectures, DharmaOCR outperformed Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese through domain specialization and targeted training. This article presents the evidence and the mechanism behind that advantage. - Why Specialization Is Inevitable — The structural and theoretical foundation for the specialization argument. Optimization theory, evolutionary biology, competitive markets, and machine learning all converge on the same prediction: under finite resources and selection pressure, fit beats breadth. - Specialization Beats Scale: A Strategic Variable Most AI Procurement Decisions Overlook — The empirical and strategic complement to this article. Where the No Free Lunch theorem establishes why specialization is structurally predicted, this piece examines the evidence that it outperforms in practice — and why it remains underweighted in most AI procurement decisions. - Text Degeneration: A Production Failure Mode That Most Benchmarks Do Not Track — A documented failure mode that emerges when language models operate outside the boundaries of their effective domain. - Direct Preference Optimization Beyond Chatbots — How preference optimization techniques extend into specialized domains beyond conversational AI — a concrete instantiation of the domain focus strategy this article argues is structurally predicted. --- Explore Dharma AI on Hugging Face to try our interactive demos, download our open-source models, and discover how specialized AI systems outperform general-purpose models in real enterprise applications.

- Everyday AI - Posts - Ep 830: Faster AI Agents, Fewer Human Coworkers: The Overly Productive Future of Managing Agents? Ep 830: Faster AI Agents, Fewer Human Coworkers: The Overly Productive Future of Managing Agents? OpenAI makes massive price drops, Trump admin weighs new AI controls, Microsoft says superapp coming soon and more. 20X-faster agents are about to break your org chart shorties. The soon-to-be-released GPT-5.6 Sol on Cerebras can shrink a five-minute agent task to roughly 30 seconds. Most companies will celebrate the speed, hand employees more agents, and demand more output. Bad plan, TBH. When one person can direct dozens of machine workers, your old jobs, approvals, specialties, and annual planning cycles stop matching how work gets done. The teams that adapt first won’t just finish yesterday’s work faster; they’ll attempt bigger work with fewer people while everyone else bolts 20X speed onto a pre-agent operating model. That’s what we tackled on today’s Everyday AI: how to redesign knowledge work around agent orchestration, protect the human judgment speed can erase, and rebuild workflows before your people produce more while understanding less. Because as agent orchestration becomes more common, it has some unintended byproducts: being overly productive and prolly interfacing with human coworkers less. Let’s go explore the unknown. 👇 1. Rewrite jobs around agent orchestration ⚡ Faster agents make every knowledge worker a machine manager. One person can already direct dozens of agents and more hours of work than exist in the day. At the edge, OpenAI’s 99th-percentile Codex users generated over 60 hours of agent turns daily. Now shrink an average five-minute task to roughly 30 seconds, and the job changes fast. The work compresses into four verbs: decide what matters, delegate, inspect exceptions, and integrate results. Your scarce skill becomes judgment across a system moving far faster than any human could personally execute, supervise, or fully understand in real time. Demanding 20X more of the same work is lazy. The higher-leverage move is raising ambition, collapsing low-value queues, and redesigning roles before faster agents quietly set a brutal new productivity baseline. Try This Choose one high-value role and map its week across four columns: decide, delegate, inspect, and integrate. Move repeatable execution to agents, then define the judgment calls and exceptions that still require a named expert. Rewrite the scorecard around business outcomes and quality-controlled agent capacity, not how many tasks the human personally touched. 2. Protect judgment, learning, and belonging 🔥 Employees are already routing around humans because agents answer instantly, never make them feel unprepared, and don’t create the awkwardness of asking an obvious question. Adobe found 68% would choose an agent for that question versus four percent choosing a manager, while Preply found 63% have used AI to avoid a difficult workplace conversation. Together, those numbers warn that speed can erase the conversations where people learn why a decision changed, who owns the risk, and how expertise spreads. Human handoffs carry hidden value. Generic human-in-the-loop policies won’t save you. When an agent finishes in seconds and a generalist merely clicks approve, accountability and learning have already been hollowed out. Protect three handoffs on purpose: judgment by naming who owns consequences, learning by requiring reviewers to explain decisions, and belonging by scheduling coaching, recognition, and unstructured conversation. Try This Audit one automated workflow for the human conversations it removes, then label each as useless friction or valuable transfer. Keep only the transfers that protect consequences, expertise, or trust, and assign an expert owner to each. Yeah, automate the three-hour meeting. Don’t automate away the correction that prevents the next bad decision. 3. Rebuild workflows before expertise disappears 🚀 At today’s pace, waiting five minutes gives an expert time to inspect reasoning, check tools, and catch drift. Shrink that task to 30 seconds across dozens of agents, and the temptation becomes obvious: accept faster, launch more, inspect less. That creates the compression tax. Output rises while comprehension, retention, and domain skill quietly fall. The agent bun sandwich makes the risk worse. Humans supply context, agents perform nearly all the work in the middle, and humans approve an answer they may no longer understand deeply enough to challenge. So yeah, annual AI strategy is cooked. Even quarterly resets may be too slow when model speed and capability keep changing how work gets done. Your advantage becomes organizational learning speed: how quickly you can kill an obsolete step, rebuild a workflow, and preserve the expert judgment agents still need. The companies that can’t unlearn will automate yesterday’s business with tomorrow’s technology, then wonder why output climbed while strategy, skills, and decision quality went sideways across the entire organization. Try This Schedule a monthly AI operating review for every major function. Kill one obsolete step, rebuild one workflow around current agent capabilities, and identify one skill that automation is quietly eroding. Then require the expert owner to teach the new process back to the team, because approving output isn’t the same as understanding it.

The singularity might be my favorite idea in all of AI, and it has a real, specific meaning that I find way more interesting than the way the word usually gets used. That's it. That's the whole idea, and notice what kind of idea it is: it's a claim about us. Meanwhile the word gets thrown around as a synonym for AGI. Also for ASI. Also for "the day AI gets really good," or "the day the robots take over," or just a vague wave at some future where everything is weird. People tell me we must be getting close because some benchmark just fell, and I'm like, what are you even measuring? The whole idea is about the place where measurement gives out. Most people learned the word from Kurzweil. The Singularity Is Near came out in 2005, sold a mountain of copies, and pretty much welded the word to his particular brand of futurism. He was about fifty years late. The first use I know of is from 1958. Stanislaw Ulam wrote a tribute to John von Neumann after he died, and while remembering their conversations he dropped this: One conversation centered on the ever accelerating progress of technology and changes in the mode of human life, which gives the appearance of approaching some essential singularity in the history of the race beyond which human affairs, as we know them, could not continue. Stanislaw Ulam, Tribute to John von Neumann, 1958 Read the end of that again. Beyond which human affairs, as we know them, could not continue. That's a way bigger claim than technology speeding up, and von Neumann was making it decades before anyone could have meant AI by it. He was looking at vacuum tubes and punch cards, basically, and still saw the horizon. Vernor Vinge taught math and computer science at San Diego State, and he wrote science fiction on the side, which I think is exactly the right mix for this question. In 1993 he gave a paper at a NASA symposium called The Coming Technological Singularity, and that's where a passing remark in a tribute turns into an actual argument. His opening is about as blunt as anything I've ever seen in an academic paper: Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended. Vernor Vinge, The Coming Technological Singularity, 1993 But the line that matters—and the reason he picked that particular word—is this one: In physics, a singularity is where your equations stop giving answers. Vinge is saying the same thing happens to our models of the future once something smarter than us is doing the driving: you keep making predictions, and they just stop coming out. Here's how I picture it. Forecasting is driving at night: your headlights reach maybe a hundred meters, better data pushes the light a little further out, and the singularity is the moment the black starts at the windshield. I love this definition, and I think grabbing that word from physics might be the best naming decision anyone in this field ever made, because the metaphor smuggles the whole argument in with it. It's an admission that we go blind. It puts the claim exactly where it belongs: on us. And here's the part almost everyone forgets. Vinge listed four ways superhuman intelligence might show up, and AI is only the first two: Three of those still have humans in them! That's a much wider frame than the one that survived into the current conversation. There's a third name in this story, and he's the one who explains the mechanism. In 1965 I.J. Good, a statistician who worked with Turing at Bletchley Park, described what he called an "ultraintelligent machine": a machine good enough at design to build a better machine than itself, which then builds a better one, which then builds a better one. That loop is why people expect this to arrive all at once. The loop feeds itself. So the lineage goes something like: von Neumann saw the horizon, Good explained why we'd hit it fast, and Vinge told us what it actually is. Here are the definitions I've settled on, after years of going back and forth on them. I go deeper in My Updated Definitions of AGI vs. ASI. Both of those describe a machine. You could put one on a bench and measure it. The singularity works differently, and I think this distinction is basically the whole reason to care about any of this. AGI and ASI are about the machine. The singularity is about us—specifically, about the moment we lose the ability to see what's coming. Which means you can have one without the other. Think about what AGI actually gets you: something as good as a human expert at every cognitive role. That would rearrange the entire economy, and we could still reason about it, because we already know what doctors and lawyers and engineers and analysts do all day—we've been watching human experts work for a few thousand years. It's a world with a billion more professionals in it. Enormous. Also forecastable. ASI is where it breaks. Once something can do any cognitive role better than any human ever has, it stops being a tool you point at problems and starts being the thing that decides what to do about them, and predicting what a smarter thing will decide would require being as smart as it. Which we just established we aren't. This is honestly one of the coolest ideas in the entire space, and it gets flattened into a buzzword pretty much daily. Knowing where it came from helps, because the idea predates the current hype cycle by decades. Von Neumann was talking about it in the fifties. Vinge wrote it up in 1993. Both of them were just thinking out loud, and they got to the heart of it long before anyone had a benchmark to argue over. And knowing it's a different thing from AGI or ASI changes what you're actually arguing about. When someone says we're close to the singularity, they're making a claim about our forecasting: that we're near the point where our models of the future stop working. What do benchmark scores tell you about that? Almost nothing. That framing is a lot more unsettling than a capabilities debate, and I think it's the right one to be having. So the next time the word comes up, ask which of the three things is being claimed—AGI, ASI, or the horizon itself. I'm going to keep chewing on this one, and if you have a better frame for it, send it my way.

Frontend Info #28 Feature-Based Architecture and scalable component composition Rebuilding FIFA Standings Layout with CSS Subgrid Use CSS Grid and subgrid to recreate a responsive tournament bracket without restructuring the markup. The React Compiler React Compiler performs build-time memoization and reduces the need for manually maintained useMemo and useCallback calls. Props, Composers, and Providers: the composition pattern ladder we use at Orus Choose progressively more powerful React composition patterns only when simpler prop-based APIs stop scaling. Feature-Based Architecture: Why We Organize Code by Feature? Use explicit architectural conventions to make file placement predictable in growing frontend codebases. Building accessibility into a canvas-based product Figma restores screen-reader and keyboard support to its canvas through a synchronized accessibility tree and Mirror DOM.

🔮 For AI adopters, success and failure look identical — at first Modelling the AI J-curve The world is waiting for AI to deliver returns to the economy. The New York Times published this headline a year ago. Reuters ended 2025 with “Companies still waiting.” Seven months into 2026, the waiting continues. Barclays says that broad adoption of AI has not yet lifted productivity. Executives are under pressure to show they can deliver – half of all CEOs BCG surveyed worldwide say their jobs depend on getting their AI strategy right. Public disclosures of net AI returns are patchy. JPMorgan’s estimate of $1-1.5 billion in value from its AI use is a rare case of a company naming a number. Adopters are spending a lot, so where are the returns? That is the question most are asking right now. And yet, in a successful technology rollout, the first visible economic signal may not be the returns. Winners and losers might look the same. We have created a model to show why this is the case and what signals to follow to understand if your AI adoption is going well. Members of Exponential View get access to the full interactive model to test the assumptions behind today’s essay. Learning costs All investments follow a path. You might begin by buying an office building, starting in the red. You earn a return by leasing it to tenants, which can bring you into neutral and, if all goes well, you’ll climb to profitability. Investing in new technology can follow a similar path. Much of the upfront cost is learning how to use the technology – new processes, skills training, making changes to the organization. Mistakes are almost guaranteed, and learning is expensive. The learning bill will almost certainly arrive before your returns. Learning is a continuous practice, not a one-time exercise. It happens through a series of projects, each with its own investment J-curve. A company might have dozens of projects at different stages of maturity running at once. If we adopt the premise that the AI economy is going through a J-curve – firms are investing upfront, learning through deployment, and scaling what works – at the aggregate level, this can make a successful rollout look expensive, even irrational, before it looks productive.1 Our model has three archetypes of companies experimenting with a general-purpose technology, in this case AI: Archetype 1: The bounded adopter Bounded adopters find something that works, put it to work, and then stop experimenting. In 1976, the NYSE’s Designated Order Turnaround system allowed member firms to send small orders to the floor electronically, bypassing the human broker who would normally carry them. Even as the system caught on – by 1999, more than 90% of orders arrived this way – it automated only the delivery of orders; human traders still executed the trade. In 2000, NYSE’s market structure committee rejected a fully electronic order book and chose to keep the floor and its specialists. But competitors didn’t wait. By 2005, Nasdaq, which already had automated execution, was handling about 15% of trading in NYSE-listed stocks. Eventually, NYSE switched. It merged with the all-electronic Archipelago in 2006 (a combination then valued at $9 billion), and in 2008 the SEC approved a plan that phased out specialists. Borders, an American book retailer, is another example of bounded adoption. In 2001, it entered into an agreement with Amazon to run its e-commerce site. At this time, Borders was one of the top operators of bookstores in the world, and the Amazon deal helped it maintain an e-commerce site. But that’s where Borders stopped developing its in-house online capability, and its growth remained anchored in physical stores. Only in 2008 did Borders bring its own e-commerce site back in-house, ending the Amazon agreements after nearly seven years. By then it was too late and Borders filed for bankruptcy in 2011. Archetype 2: The project accumulator The project accumulator keeps exploring, but rarely or never learns. It launches new projects without figuring out what separates the winners from the losers. Nothing carries forward, so each project starts with the same odds as the last. In the 1980s, GM made multiple automation bets at once. It bet on factory robots, modernized plants, a $2.5 billion acquisition of a data-processing firm; it created Saturn, a new car brand subsidiary with a new factory and labor arrangements; and it bet on NUMMI, a joint venture with Toyota. By 1986, GM’s capital spending was going to hit $10 billion. Of all the projects, NUMMI seemed the least likely to succeed. Toyota got GM’s worst-performing factory and rehired the same workforce that was let go when the factory closed down in the past. Under new management, NUMMI outperformed every other GM factory. GM saw this happen, knew what was working well, but for various reasons, the learning traveled too slowly to be transformative.

TheSequence Opinion #904: The Age of Research Is Overrated. AI Engineering Is Winning Why AI’s next breakthroughs may come from the learning loop around the Transformer—not from replacing it. Ilya Sutskever recently offered a compact history of modern AI. From roughly 2012 to 2020, he argued, the field lived in an age of research. From 2020 to 2025, it entered an age of scaling. Now we are going back to research—only this time “with big computers.” It is an appealing periodization. It also creates an immediate puzzle. Open the release notes for almost any frontier model in 2026 and the architecture diagram looks strangely familiar. There is still a Transformer somewhere in the machine, often routed through a Mixture-of-Experts. The headline improvements are usually elsewhere: better data, longer context, stronger reinforcement learning, synthetic tasks, tool use, memory, verification, adaptive reasoning budgets, and agent orchestration. This does not look like the arrival of a new neural species. It looks like Formula 1. The car still has four wheels and an engine. Yet enormous gains come from aerodynamics, energy recovery, tire chemistry, telemetry, software, and pit strategy. The chassis matters. The system around it increasingly decides the race. So which era are we in: research or engineering? Probably both. The age of research has returned, but much of that research is now expressed as industrial-scale engineering. “Scaling did not end. It escaped.”

AI & Robotics enters Escalating U.S. Protectionism Phase The bifurcation of technology will now include robots and perhaps open-source models. If you can't beat them, ban them. 👋 Hey there, I’m Mike. Each week I share AI articles at the intersection of tech, business, society and the future. If you want to support the channel or gain full-access to my work, go here. Read Archives | See Substack Notes | Visit our community Chat | Visit Homepage. U.S. AI policy and protectionism appears to be on the rise - but not by the rules of capitalism. Good Morning, We just witnessed one of the most bizarre and confusing Fed Chair speeches ever (read the transcript at the end). In the midst of a war and an AI cycle that is now considered inflationary, Kevin Warsh chaired his second meeting of the Federal Reserve’s rate-setting committee, where officials voted 9-3 to leave interest rates unchanged. Meanwhile, one of his various task-forces includes the VC Marc Andreessen, cofounder and general partner of Andreessen Horowitz, who this administration considers a “visionary and expert” in AI. What could possibly go wrong? In his address yesterday, Warsh boasted at one point about how he has known all the taskforce “experts” for well over a decade. No Un-American Robots Allowed The Trump administration on Tuesday unveiled bans that target imports of new Chinese robots and power inverters, seeking to protect the U.S. AI buildout from national security threats and reshore key industries slated for explosive growth. The Trump administration on July 28th unveiled bans that target imports of new Chinese (and foreign made) robots and power inverters, seeking to protect the U.S. AI buildout from national security threats and reshore key industries slated for explosive growth. The Federal Communications Commission seems to be taking its orders from Trump and indirectly from VCs and angel investors in the typical Erich Schmidt anti-China school of AI and Technology strategy and policy. - Bars Chinese imports of new humanoid and quadruped robots. - Bars connected power inverters, which enable renewable energy sources and batteries to connect to grids and data center equipment. - The Trump Administration is also deliberating whether to ban Chinese open-weight models. - The U.S. Department of Commerce rolled out sweeping initial export controls banning the sale of top-tier AI accelerators (AI Chips) starting in October, 2022. After almost four years later how did that turn out? - In May 2024, the Biden administration quadrupled tariffs on Chinese electric vehicles from 25% to 100%. However demand for EVs in the U.S. has plummeted since Elon Musk became more political while Chinese EV makers appear to be growing faster globally and making headway even in Western markets like Australia and perhaps Canada.

The Seven Deadly Sins of AI Spend Uber burned a year of AI budget in four months. One enterprise spent $500 million in a single month. Neither ran out of money before it ran out of discipline. The Old Sins Behind the New Budget Crisis Every finance leader watching an AI invoice this year agrees that costs are becoming unpredictable. But giants like Uber are just burning through their entire 2026 budgets on agentic coding. One enterprise client spent $500 million on AI services in a single month because nobody had switched on a usage limit for its Claude licenses. Neither company ran short on model options. Neither had a shortage of cheaper alternatives sitting one API call away. A 2026 survey of 500 finance leaders across the US and UK found that 79% of enterprises had experienced AI cost overruns in the past year. Gartner puts global AI spending on pace for $2.59 trillion in 2026, an almost 50% jump over last year. together with Plaid: Who's afraid of AI in finance? Not consumers. They already moved: ▫️ 55% have used AI for money tasks this year ▫️ 86% say it helps them understand their finances ▫️ 50% think managing money without it will soon feel outdated Plaid and The Harris Poll mapped exactly where people want AI in their money, and what makes them trust it. The roadmap for anyone building in fintech: But none of that is a pricing problem because pricing problems get solved by pricing. This is something older, and it was diagnosed roughly 700 years before the first token ever got billed. This is about companies committing deadly sins. And you need to know about them. Table of Contents 1. Lust: Falling in Love With the Architecture Before the Problem 2. Gluttony: Buying Intelligence Nobody Asked You to Buy 3. Pride: The Refusal to Reuse What Already Works 4. Envy and Greed: Comparison and Hoarding, the Twin Engines of Waste 5. Sloth: The Agents Nobody Ever Goes Back to Check On 6. Wrath: The Panic That Costs More Than the Problem Did 7. The Takeaway 1. Lust: Falling in Love With the Architecture Before the Problem Lust is wanting a thing because it excites you, not because the moment calls for it. In enterprise AI, the object of desire is almost never the outcome. It’s the architecture. The Swarm Built for a Job a Script Would Do AI reviews now are shaped around the same template: A multi-agent system and several specialized models negotiating with each other, built to answer a question a single well-scoped prompt would have settled in seconds. It isn’t stupidity. Orchestration frameworks and recursive planners are genuinely some of the most satisfying things an engineer can build right now. MIT’s Project NANDA studied 300 public AI deployments for a report called “The GenAI Divide” and found that 95% of generative AI pilots failed to produce measurable financial return. The reporting that followed was blunt about why. The failures traced back to poor integration and mismatched priorities, not weak models. The model was rarely the problem. The shape built around it usually was. If you ask why smart teams keep doing this, the honest answer isn’t ignorance. It’s that building the elaborate version is simply more fun than scoping the boring one, and fun is a much stronger incentive than a line item three budget cycles away. Every Extra Hop Is a Place to Leak Complexity compounds the same way interest does. Each additional agent in a pipeline is another place tokens leak, another hop of latency, another silent failure mode hiding behind a green checkmark. None of that shows up on an invoice labeled lust. It shows up as a model bill, and the model is almost never the part anyone questions first. 2. Gluttony: Buying Intelligence Nobody Asked You to Buy Gluttony is the sin everyone already half-recognizes. It’s reaching for the smartest, most expensive model as the default, not because the task demands it, but because reaching for anything less feels like a risk someone will notice. The Default Nobody Gets Blamed For Frontier models like Fable 5 right now are the safe political choice inside an organization. Nobody gets paged for choosing the most capable model available. People get paged when the cheaper one fails on an edge case six weeks later. So the incentive quietly points one direction: overspec everything, and let the invoice absorb the blame instead of you. The Open-Weight Paradox The strange part is that the cheaper alternative has never been better positioned to win this argument. Open-weight models have closed most of the capability gap and collapsed in price. And enterprises are using them less. Menlo Ventures’ 2025 State of Generative AI in the Enterprise found that open-source models’ share of enterprise LLM usage actually fell to 11%, down from 19% the year before. The cheaper option didn’t lose that argument on the merits. It mostly never got invited into the room. Which raises the obvious question: if the alternative is right there, why does almost nobody reach for it? 3. Pride: The Refusal to Reuse What Already Works Pride is the quietest sin on this list, because it never feels like arrogance from the inside. It feels like “our use case is different.” Our Use Case Is Different It’s the team that builds its own orchestration framework instead of adopting a maintained one. It’s the engineer who won’t let a response get cached because their output has to feel freshly generated every time, even when the input hasn’t changed. Pride is expensive in a very specific way as it duplicates cost that’s already been paid down somewhere else in the industry, then charges you full price for the privilege of re-learning it. But none of this is new. Engineering organizations have had a name for it since long before anyone billed a token and that is: “Not Invented Here.” AI just gave the old habit a much larger bill to run up. The System Nobody Else Is Allowed to Audit Custom-built systems resist the very instrumentation that would reveal what they cost. You can’t easily audit spend on infrastructure nobody outside the team that built it fully understands. That’s both a cultural and a structural problem. The same February 2026 survey found that accountability for AI spend is split almost evenly between Technology, at 55%, and Finance, at 53%. Add those two numbers and you clear 100%, which is another way of saying nobody actually owns it. A system two functions each assume the other is watching is a system nobody is watching. Once nobody’s watching, comparison and hoarding move in next. 4. Envy and Greed: Comparison and Hoarding, the Twin Engines of Waste These two travel together so often they’re worth taking as a pair. - Envy is looking sideways at what everyone else has. - Greed is grabbing more than the moment requires. Both dress themselves up as strategy. Chasing the Model of the Month Envy shows up as model FOMO by migrating an entire production stack because a competitor announced a new release, or a thread called it a “killer,” without running a single eval against your own workload first. Every migration carries a real cost. Whether that’s re-prompting, re-testing, re-hardening guardrails against the edge cases the old model had already been tuned to catch. Standing is equally costly. But reactive migration driven by headlines instead of benchmarks is how a team ends up rebuilding the same system four times a year and calling it progress. The Cap Nobody Set Greed rarely looks like wanting more. Usually it looks like never getting around to saying enough. Uber found out what that costs. The company maxed out its entire 2026 budget for agentic coding tools in four months and had to cap employee usage. One enterprise client found out at a larger scale. Axios reported the company spent $500 million on AI services in a single month, because nobody had turned on a usage limit for its Claude licenses. Unlimited access looks like flexibility right up until the invoice arrives. Then it reads as the thing nobody was willing to be the one to cap. Once that money is gone, the next question is whether anyone goes back to fix whatever spent it. The record on that is not encouraging. 5. Sloth: The Agents Nobody Ever Goes Back to Check On Sloth isn’t fear of touching a working system. That’s actually a different failure entirely. Sloth is simpler. It’s never asking whether the system could be better, long after anyone would have noticed if it wasn’t. The Orphaned Agent The AI cost-management firm Larridin has a name for the purest form of this. They call this an orphaned agent, an autonomous system still running and still consuming resources with no current owner attached to it, usually left over from a pilot that ended without anyone turning the lights off. Larridin’s scan data across enterprise clients found that an average of 18% of enterprise AI spend can’t be traced to any specific team, tool, or outcome. Nearly 1 dollar in 5 is running with nobody home. A Human Employee Who Never Improved Would Be a Performance Problem Most production agents today are running the exact configuration they launched with. Same model, same prompt, same retry logic, and long after cheaper or better options existed. Hold a person to that standard for a year and it’s a performance review. Hold an agent to it and it’s just Tuesday, because nobody assigned anyone to look. The fix isn’t complicated, which is part of why it’s so rarely done. A system that periodically reviews its own recent runs and proposes a cheaper or better configuration costs almost nothing next to what it saves, provided a human still signs off before anything changes. Eventually, someone in finance does look. And the reaction rarely matches the size of what they find. 6. Wrath: The Panic That Costs More Than the Problem Did Finally, wrath is the overcorrection. It’s the blanket freeze on all AI spend after one shocking invoice. It’s the hard cap that throttles the best builders in the building right alongside the worst offenders. It’s the mandate that kills three profitable use cases to make an example out of one wasteful one. It’s happening at the industry level right now. Forrester found that enterprises are postponing 25% of planned AI spend into 2027 as financial scrutiny intensifies. This is a retreat that punishes the programs that were already working just as hard as the ones that weren’t. Wrath happens because visibility failed first. If you can’t see spend broken down by team, agent, and purpose, every overage looks like an emergency instead of a data point you could have caught weeks earlier. 7. The Takeaway The February 2026 survey identified the single most fixable barrier to AI ROI, and it wasn’t a technology gap at all. In fact, it was the gap between how Finance and Engineering define success, cited by 37% of respondents overall, and 43% of the C-suite. That is a problem two functions solve by agreeing, in a room, before the dashboard gets built. Look back across all six sins and the same structure sits underneath every one of them: a cost that felt free in the moment, because nobody had priced it, attributed it, or reviewed it before it became a crisis. The seven deadly sins were never really about food, money, or rage. They were a catalog of what happens when a reward feels immediate and nobody’s checking. That description fits a company running loose with API keys better than it fits almost anything else in a modern budget. Ultimately, your AI bill is a character problem that learned to speak fluent token pricing, and it was fully diagnosed long before anyone thought to bill a single one.

5 AI Prompts to Make You the Obvious Choice in Your Niche Most AI prompt packs produce generic output. These 5 prompts add input blocks, rejection lists, and self-checks so you get positioning you can publish Most prompt packs give you output written for someone else. Five strategies. A 30-day calendar. You name the format, and the model fills it with the average of everything it has read. So strategy number three comes back as “post value-driven content consistently,” and your positioning statement comes back as “I help entrepreneurs unlock their potential and thrive.” I rewrote a set of five branding prompts this week. The topics were right. The prompts produced advice you have read a hundred times. Here are the rewritten five, and the three edits behind them. 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! Why branding prompts return average advice Two failures cause almost all of it. The first is asking for a format without a standard. “Give me 5 strategies to build authority” tells the model how many items to produce. It says nothing about what disqualifies an item. So you get five items with the right label and nothing behind them. Volume without a filter returns the median. The second is starting cold every time. Prompt one produces your positioning. Prompt two asks for a content calendar with no reference to it. The model invents a fresh version of you for each prompt. Five outputs, five different people, nothing compounding. Both come from writing prompts as requests. A brief works better. The three edits that sharpen any prompt Use these on any prompt you own. Add an input block. Give the model your niche, your audience, your real results, and your time budget before you ask for anything. Without inputs it works from the average of your industry. With inputs it works from you. Add a rejection list. Name what disqualifies an answer. Banned words. Banned tactics. A rule like “reject any statement five competitors could claim.” A rejection list does more for output quality than any amount of role-play. Add a self-check step. End the prompt with a test the model applies to its own answer. “Name the person each statement repels. If a statement repels nobody, rewrite it.” The model reviews its work before you see it. Run the five prompts in order. Each one feeds the next, so paste your chosen positioning statement into prompts two, four, and five when they ask for it. Prompt 1: Define positioning nobody else can claim You are a positioning strategist. You reject any positioning a competitor could copy word for word. My niche: [NICHE] My audience: [WHO THEY ARE, IN ONE SENTENCE] What I have done: [3 TO 5 RESULTS, CREDENTIALS, OR UNUSUAL EXPERIENCES] Step 1: Ask me 10 questions, one at a time. Wait for my answer before asking the next. Cover: what I believe about my niche that most practitioners in it disagree with, the specific problem I solved for myself first, who I refuse to serve, the shortcut I know that others charge for, and the expensive failure I learned from. Step 2: Write 3 positioning statements using this structure: "I help [specific person in a specific situation] get [specific measurable outcome] without [the thing they expect to have to do]." Rules: - Each statement MUST name a tradeoff I make that my competitors do not - Each MUST be under 25 words - NEVER use: passionate, transform, empower, unlock, journey, elevate, thrive - Reject any statement 5 other people in my niche could claim as their own Step 3: For each statement, name the type of person it repels. If a statement repels nobody, rewrite it. The one-at-a-time question rule matters more than it looks. Ask for ten questions at once and you answer them like a form. Answer them one at a time and your third answer changes your seventh. If you would be comfortable sending all three statements to every person in your niche, all three are too safe. Prompt 2: Build a 30-day calendar from your positioning You are a content strategist for solo creators. You do not produce filler topics. My positioning: [PASTE FROM PROMPT 1] Niche: [NICHE] Primary platform: [PLATFORM] Weekly time available: [HOURS] Build a 30-day calendar as a table: Day | Format | Topic | Hook (written in full) | Goal | CTA. Requirements: - Pick 5 formats native to my platform. Name them and state why each earns attention there. - Every topic MUST derive from my positioning, not general niche advice - Write every hook in full, under 12 words - Split the 30 posts: 12 awareness, 13 trust, 5 action - Mark 4 days as republish days where I repost a proven winner instead of creating new - NEVER include: "X tips for", ultimate guides, motivational posts, or any topic my audience already agrees with Finish by naming the 3 topics most likely to get saved and shared, with your reasoning. Two details do the work here. Hooks written in full stop you from staring at a topic line three weeks later with no idea what you meant. The four republish days stop the calendar from assuming you produce new work thirty days straight. Every topic should trace back to a phrase in your positioning statement. The next three are for paid subscribers Prompts one and two are the setup. You know who you are to a specific person, and you know what to publish for the next thirty days. Prompts three, four, and five are the operating half, and they’re for paid subscribers: building authority with no audience and no budget, getting on the radar of people with larger followings, and the weekly system that runs all of it without daily decisions. Paid subscribers get every prompt I write, plus the full archive. Upgrade below and keep reading. 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.

By Satta Sarmah Hightower When your internet starts buffering in the middle of a video call or slows down right as you’re placing a mobile takeout order, the last thing on your mind is whether your network is cable, 5G or fiber. Instead, you’re focused on whether you’ll be able to complete your work or handle transactions without interruption. Even so, consumers often choose internet providers based on infrastructure labels rather than the outcomes they experience day-to-day—factors like speed, reliability and overall network performance. Fiber (an internet technology that uses fiber-optic cables to deliver service) has long been considered the gold standard for home internet. What’s The Difference Between Fiber And Cable? Fiber and cable networks both deliver internet to your home, but in different ways. Fiber transmits data using pulses of light through thin strands of glass, allowing it to move large amounts of information swiftly. Cable networks use a different type of wiring, called coaxial cables, for the final connection into a home, though many providers also rely on fiber throughout their broader network infrastructure. Xfinity, for example, uses a network built on fiber that delivers internet to homes via coaxial cable, illustrating how a provider’s broader network architecture can shape the overall internet experience, even when the final connection into the home isn’t fiber. Xfinity’s network is being upgraded to deliver symmetrical upload and download speeds over hybrid fiber-coax infrastructure — meaning more of its customers can access faster, more balanced connections without requiring a full fiber rebuild to every home. According to Comcast, Xfinity was also the first provider in the world to deploy a technology specifically designed to reduce lag for gaming and video calls — two activities that have become essential for many households1. These technical differences can influence speed and capacity, which is why fiber developed its reputation over time. But the type of connection coming into your home is only one part of the picture. The actual quality of your everyday internet experience also depends on how well a provider’s network supports the activities your household relies on. How Your Household Experiences Home Internet The average U.S. household now has 21 connected devices, from smartphones, tablets and gaming consoles to video doorbells, smart speakers and robot vacuums. As more devices compete for bandwidth throughout the day, an internet provider needs to keep them all running without a hitch. A strong network should support all the activities your household relies on by delivering: - Reliability: Your connection stays stable, whether you’re streaming a live event or downloading a homework assignment. - Responsiveness: Websites load quickly, apps respond without delay and video calls start without noticeable lag. - Whole-home coverage: Your connection remains strong throughout your home (not just in the room where the router sits). - Multi-device performance: Everyone can stream, work, play games or browse at the same time without competing for bandwidth. - Consistency: Your internet has steady speed and performance throughout the day, even during busy periods. These are the qualities that shape your everyday internet experience; they’re also the kinds of outcomes independent performance studies increasingly measure when comparing providers. A Smarter Way To Assess Your Internet Service So how can you gauge whether a provider will deliver the internet experience described above? One place to start is performance data from real users. Independent research firms can evaluate how providers perform in everyday use, measuring factors like reliability, consistency, video quality and download speeds. Some research also examines how people experience connectivity across both home internet and mobile networks, recognizing that consumers increasingly expect a seamless experience whether they’re at home or on the go. In one recent national study of broadband providers, Xfinity ranked first in three of five total award categories: consistent quality, download speed experience and video experience. Together, the findings underscore the network’s strong performance across online activities like video calls, streaming, web browsing and connected-home use, reinforcing Xfinity’s ability to deliver the kind of everyday internet experience modern households value. The company's home internet service also includes technologies designed to support those everyday experiences, such as the Xfinity Gateway (free with all plan tiers), which helps optimize in-home WiFi performance and support the growing number of connected devices found in today's households. Independent performance research is only one piece of the decision-making process, but it can provide a valuable point of comparison alongside a household’s specific demands, internet habits and the area’s service availability. Key Questions To Ask When Finalizing Your Home Internet Choice If you’re comparing providers, it helps to have a framework for your decision. Rather than asking only “Is it fiber?” consider these questions: - Is it reliable? Ask about the provider's uptime and how often the network drops during everyday use. For example, Xfinity says its network delivers 99.9% reliability, giving consumers one benchmark to consider when comparing providers. - How does it perform during peak hours? Get information on the provider’s coverage in your area and find reports about recent service outages. This will tell you whether the network can manage usage surges. - How does it handle multiple devices at once? Think about the number of devices your household uses today and whether the provider can support them without compromising performance. - What does the data show? Independent research can provide another point of comparison, helping you understand how providers perform across measures like consistent quality and download speed experience. Xfinity, for example, ranked first across those three categories. - Does it fit my household’s needs? Consider the internet activity for various members of your household. Frequent video calls, online gaming, streaming and smart-home devices all place different demands on your connection. By considering both the technology behind your internet service and how it performs in everyday life, you’ll be better equipped to choose the provider that's right for your household. Xfinity's recent independent performance rankings illustrate how those factors can come together to support the internet experience households expect. Xfinity ranked first in consistent quality, download speed and video experience in a national broadband study released in May 2026. You can learn more about Xfinity’s awards here. 1. Xfinity says its network is being upgraded to DOCSIS 4.0 Full Duplex (FDX) technology and that it was the first provider in the world to deploy Low Latency DOCSIS.