If you're still running Hermes agent out of the box, you're basically paying for a chatbot because on its own, it just talks back. But with the right skills, it will build you a full app or turn one topic into a finished video. And to actually find out how capable they make Hermes, I tested the most popular skills on my own server. And in this video, I'll show you the seven worth installing today and how they work. And a lot of people run Hermes on their own computer, but I'm running mine on a server instead. This is basically a machine somewhere else that never turns off. So, Hermes keeps working even when I close my laptop and I can still message it from my phone through Telegram. It's also a lot safer because Hermes can run commands on its own. And if it ever does something I didn't ask for, it happens over there instead of on the laptop with all my personal files. To get one, I'm using Hostinger, which is a hosting company that rents you your own private server or a VPS. and they have a one-click template that sets Hermes up for you. If you want to copy this exact setup, I'll leave a link to Hostinger down in the description. So, let me quickly walk you through it. First, on Hostinger's Hermes page, you'll see all the plans laid out. And I'll go with KVM2, which gives you two CPU cores, 8 GB of RAM, 100 GB of NVMe storage, and 8 TB of bandwidth. That's the same plan I tested everything in this video on, and it handled all of it, from building a full app to rendering a 1080p video. So, I'll click choose plan on this one. The billing period is set to 24 months, but I'd switch it to 12, so you only pay for one year up front instead of two. The monthly price barely changes, and you still get a free domain for the first year. There's also ready to use AI option, which adds AI credit, so Hermes works without an API key, but I've already got my own, so I'll skip it. You get free weekly backups anyway, but I'd add daily auto backup because Hermes changes files on the server by itself. So, if it ever deletes something it shouldn't, you can roll it back to the day before. For the server location, I'll go with the one marked best latency since that's the fastest one for where you are. Over in the order summary, I'll open have a coupon code and type in Parker, which is my code, so you get even more off on top of the discounts that's already there. After I hit continue and finish the order, Kermy sets itself up on the server automatically. Once your server is ready, everything happens in H panel, which is Hostinger's dashboard. So, I'll click VPS in the left sidebar to see the servers on my account. Then, I'll hit manage next to mine because that opens the overview page where every app running on it is listed. If it's your first time, this is where you'd open Docker Manager, go to catalog, search for Hermes agent, and select it and let the template doing all the installing for you. Mine's already running. So, Hermes agent shows up under my applications and I'll click app terminal right under it because that's where you actually talk to the agent. It opens in a new tab. So, I'll type Hermes and hit enter. And now it's ready to take messages. But by default, Hermes can't do much more than chat with you. So, the first thing I'll add is Eddie Osman's agent skills, which is a bundle of 25 skills that teaches it how to make software properly, one step at a time. It's split into separate skills because each one handles a different stage of the work like writing the spec, planning, coding, and reviewing. So, Hermise only loads the one it needs for each step. You can't install the whole bundle in one go. And when I pointed it to a full repo, it just failed. So, I'll add them separately, which also means I only take the ones I actually need for building an app. To install one, I'll type /skills install and paste the path to the skill from Edd's GitHub repo. starting with the spec skill. Next, I'll do the same for the rest of the skills I need. Then, I'll type / reload skills so Hermes picks them all up without a restart. And each one now works as its own slash command. Now, to see what these skills can actually do, I'll ask Hermes to plan out an app called Fluid Room, which is a full screen animation of code for ink swirling in water with a breathing mode, a focus timer, and a play mode. Instead of asking me a bunch of questions, it writes out a complete spec and lists every guess it made along the way so I can fix anything before the coding starts. Hermes also lets you use up to five skills in one message. So next, I'll ask it to break that spec into tasks, build everything, test it, and review its own code in one go. In the same message, I'm telling it to take screenshots while it works since there's no screen on the server. And the first time I skipped that, the design looked flat and gray because Hermes never got to [music] see it. From there, it keeps going by itself until every task is done. And this is what we get. I'll switch it to play mode first. And honestly, it looks really pretty. As I move my mouse around, the colors swirl into these beautiful liquid patterns. And if I stop moving, it keeps flowing on its own. I can also change the color scheme and it shifts into completely different set of vibrant colors. So, if I switch to this blue one, it responds instantly. I can also toggle over to focus mode and use it as a Pomodoro timer instead. But my favorite thing about this app is how smoothly it moves from one color to the next. On top of the app itself, Hermes keeps a skill of its own with everything it's learned from building apps like this. So, with one bundle of skills and two messages, [music] I went from a simple idea to an app I'd actually use, and it's already been tested and checked. But building apps is only a part of it because most of the time I just want my agents to go out and find stuff for me. And on its own, Hermes is pretty bad at that. On my server, it couldn't even open a normal web page without falling back to some messy workarounds. And that's before you even get to sites like Reddit or X. That's the problem Agent Reach solves because it's a skill that gives Hermes away into a whole list of sites like YouTube, GitHub, Reddit, and X all from one install. To set it up, I'll paste the link to its install guide into Hermes and ask it to install it for me. Hermes does the whole setup by itself. And before it installs anything bigger on the server, it stops and asks me first. So, keep in mind you need to be at your screen for this part because if nobody answers in time, it cancels that step. Once it's done, I'll ask it to run agent riches doctor check which shows which sites are actually working and it gets the main ones like YouTube, GitHub, and web search working on its own. Reddit and X are still locked though because they need you to login first. So, if you want those, use a spare account instead of your main one. Now I'll ask it for a short brief on what people are complaining about and what they love about Hermes this week. Pulled from YouTube, GitHub, and the web. A little later, it gives me a list of both with a link next to every single point so I can check where it came from. It's split into the complaints first and then the things people love. It also tells me where its own research is thin, so I know exactly which parts are worth double-checking. So now my agent can go on to sites like YouTube and GitHub and actually read them for itself instead of getting stuck on messy workarounds. And even when [music] Hermes can get to a page, reading it on my server takes way more effort than it should. Every time I give it a link, it tries to read that page with its own page reader first. Then it moves on to a browser. And when neither of those work, it ends up scraping the old page and hoping it picks up the useful parts. That's where the bundle comes in. It's a small tool that cuts a web page down to just the article and throws away the menus, ads, and code around it. I'll ask Hermes to install it. It sets up the tool along with the skill that shows Hermes how to use it. Then I'll run / reload skills so the new skills show up. Now, to actually see the difference, I'll give it the same news article twice. First, I'll do it the normal way and ask for a five bullet summary so we can see how Hermes handles a page on its own. Then, in a new chat, I'll send the exact [music] same message, but this time I'll call the defle skill by adding the words use defle at the end. You do need to ask for it by name because in my tests, Hermes never used the fuzle on its own even after I installed the skill. With it, Hermes goes straight to the clean article in one step and writes the summary right away. So, two extra words in your message get Hermes reading pages more efficiently and you spend less on every page it reads. And once Hermes has read something, the other issue that slows me down is how much it writes back because some of its answers are way longer than they need to be. There's a skill for it called caveman. And the idea is pretty funny because it tells Hermes to talk like a caveman and only use the words it needs. The read me on the skills GitHub page says it will make Hermes write a lot less. So I wanted to test that. To install it, I'll just give Hermes the link to the repo and ask it to add the skill. Once it's installed, you turn it on by typing / caveman before your question. And it works best when you send /caveman and your question in the same message since sending it on its own [music] just adds an extra step. For the test, I'll ask whether I should run an agent on my server or on my laptop. first without caveman. Then I learn as the same thing again on a new chat with the skill switched on. I was honestly expecting something that's hardly readable, but it still reads like plain English and it just keeps the filter words and gets to the point faster. So if long answers drive you crazy, Caveman is an easy way to keep Hermes concise. Short answers are nice, but another thing a lot of us use AI for every day is planning. >> [music] >> Whether that's about launching a new product, starting a project, or just working on a new idea, the problem is that if you ask AI to plan something, it usually just agrees with whatever you already have in mind. That's what the grill me skill is for. Instead of writing one for you, it takes the plan you already have and keeps asking you questions about it one round at a time, so you catch the problems and inconsistencies in your thinking before you start executing instead of realizing them halfway through. It's made by a guy named Matt Pacco and it's on Hermes official skill list. So I'll add it with /sklls install and its name. Then run / reload skills to test it. I'll give it a hypothetical plan to launch a free weekly newsletters for this channel. I'll also tell it my goal. I want to grow the newsletter and how much time I can put into it. Each week it starts asking me questions in rounds and next to each one it gives me the answer it would pick. So I can either go with that or tell it why I disagree. It brings up things like whether my goal is even realistic and how people would actually find the link to sign up. And when I push back on one of its ideas, it finds a way to make my version work instead of just dropping it. Once we're done, it gives me a summary of everything we decided and a short list of what's still open so I know exactly what to do next. So before I spend weeks on an idea, a few minutes with Grill Me show you which parts of the plan actually hold up. The next test is the one I was most curious about because it's about Hermes making an actual video completely on its own from scratch without me paying for a single tool apart from the usage credits. [music] And to make that possible, I'm using a skill called open montage, which is a full video production toolkit for AI agents. [music] And it takes a single topic and turns it into a finished video from the narration all the way to the final edit. Some of it options run on paid AI models, but there's also a completely free path that uses real arive footage, a free AI voice, and royalty-free music. So, that's the option I'm going for. Installing the skill works exactly like before. So, I'll give Hermes the link to the repo and jump straight into the test. I'll ask it for a 45second documentary about the Apollo 11 mission. But before it makes a single frame, it comes back with a few different directions for the video. So, I'll pick the one I like the most. [music] And from there, it handles literally everything itself, from linking up real NASA photos and footage to recording the narration, adding the music, and cutting it all together. On my KVM2, the whole render took just a few minutes, and the server didn't even come close to running out of memory. And this is what it made. July 16th, 1969. Free men leave the Earth. Strapped to a rocket 110 m tall, 3 days, 400,000 km of silence. Lunar orbit. July 20th. The eagle descends almost out of fuel. Two men step onto another world. One small [music] step. 2 and 1/2 hours outside. Then the long way home. It's not something I'd publish as it is, but for a first try with zero editing on my end, it's really good. Some of the shots do look a bit soft though because that old archive footage is lower resolution and has to be stretched to fill a full 1080p frame. So with the right skills, Hermes is even capable of making you a video from scratch. And for the last one, I wanted Hermes to work across the different platforms I use every day like email, my calendar, and my spreadsheets. Normally hooking up an agent with Google apps like Gmail, Calendar, and Sheets is a real chore because you have to go into Google Cloud, create a project, switch on each API separately, make your own login credentials, and then hand that file over to Hermes. And that's just for Google because every other app you want to connect comes with its own setup on top of that. Composio skips all of that because it's one single connection that gives Hermes access to over a thousand tools across apps like Gmail, Calendar, Sheets, Slack, and Notion. And it handles signing in to each of them for you. It also doesn't dump all of those tools on Hermes at once because the agent looks up the ones it needs as it goes. To set it up, I'll just ask Hermes to add Composio for me. It handles most of it by itself. And the only part it can't finish on its own is the login since my server doesn't have a browser. So back in H panel, I'll open the app terminal again and run the Composio login command, which gives me a link. I open that link on my computer, sign in to Composio, and hit approve. After that, the browser lands on a page that won't load, which is completely normal. So I'll copy the address from the top bar and paste it back into the terminal. And that's it. Then I'll ask Hermes to connect Gmail, calendar, [music] and cheats. And instead of all that Google Cloud hustle, it just hands me a quick approval link for each one. Now I'll send it one message right here in the chat asking it to check today's emails, put [music] any client calls straight into my calendar, add them to my client tracker sheet, and skip the newsletters. It goes through the emails that came in today, and since none of them were actually from clients, it doesn't add a single thing to my calendar or my sheet. Instead, [music] it tells me exactly what it found from newsletters and subspect posts to a couple of community events notices and why none of those counted as a client call. And that's honestly what I want from an agent [music] that can read my inbox because it went through everything and still didn't a meeting that doesn't exist. It even pointed out what it didn't check like my spam folder. So, I know how far it really looked. So, instead of setting up every app one by one, one connection gets Hermes working across all of them. Just [music] keep in mind it can read your whole inbox. So only connect an account you're comfortable handing over. And every single one of these skills run on that same Hostinger server from building the app to rendering the video. Their oneclick Hermes template handles the whole install for you. So the only thing left to do is add the skills you actually want. So go grab a Hostinger VPS using my link in the description and install your first skill from this video today.
This might be the coolest thing you've seen today. These glasses right here can supposedly read the words that you're like silently mouthing and type them at over 200 words per minute. Now, people typically type around 40 words per minute, so that's a pretty big jump. These are from Subvocal, a tiny Y Combinator startup founded by two Georgia Tech University students. The glasses send radar signals through your face, and then their AI turns the tiny movements from your lips and jaws into text. In the demo, the guy silently asks his AI to summarize his Slack messages for him right next to his coworkers, and nobody knows he's even talking. They're even saying that two people that are wearing these glasses could have a private conversation from across the room with each other without anybody even realizing they're talking. Like real life telepathy. And Apple clearly thinks that this concept is a pretty big deal, too, because earlier this year they paid $1.6 billion for a company that created very similar tech. And these guys right here are already taking pre-orders for 499 bucks. Now, it's still a prototype, and I haven't seen anybody else like independently test their speed claims yet. But if this works, probably something I'd use like every day. But would you walk around wearing a pair of these?
Everyone knows how to make a decentl looking AI video, but keeping the same characters across a whole series is where it actually gets hard. I made a threeep episode micro drama where every character looks and sounds exactly like they did in episode 1. In this video, I'll walk you through the whole build from planning your story all the way to a full series with consistent characters. And by the way, thanks to OpenArt for sponsoring this video. The first step is to actually plan the whole show. Most people skip this and head straight to generating, but that's how you end up with random scenes that have nothing to do with each other. To turn one idea into a full series, I'm using the micro drama kit, a free set of tools I built specifically for micro dramas that runs this entire workflow. And the best part is you can get the kit for free when you sign up for any Open Art plan using my link in the description below. Once you're on the page, click sign up. This opens the Open Art signup page in a new tab. Before signing up though, you need to clear that tab's cookies or the kit won't unlock. Just click the little icon next to the web address. Go to cookies and site data. Delete the site's data using the bin icon and refresh the page. Now, sign up for any open art plan. Go back to the first tab. Enter the same email address and click unlock my micro drama kit. The kit handles the planning and the writing. And for the actual generation, it connects to open art, an all-in-one platform that gives you every AI model you need in one place. I'll start in the script maker, and nothing in there uses credits. You can rewrite your story as many times as you want. The first question is what your story is. I'll type in who it's about, what goes wrong, and the twist. Mine is about Lena, who cleans the penthouse of a cold young CEO called Adrien, and he has to get engaged within 30 days to inherit his father's company. and she signs a contract to play his fiance. The twist is that she didn't sign for the money. His stepmother, Viven, stole the hotel group from Lena's father 20 years ago, and she's there to prove it. You can find this exact story along with the whole workflow in the description below, completely for free, and you can follow along in real time. For the genre, I'll pick Billionaire Romance, and I'll set the season to five episodes of around 2 to 3 minutes each. The setting that really matters here is the payw wall cliffhanger because apps like Real Short give you the first few episodes for free and lock the rest. So, you pick the last free episode and the kit writes that one to end on the biggest cliffhanger of the season. I'll put mine at the end of episode 3, making that the episode that has to get people to pay for the rest. There's one box I'd never leave empty, which is the notes on the cast. I'll keep the cast to Lena, Adrien, and Viven, and leave anyone else in the background without a single line. Clicking create my series Bible gives the show its title, The Veil Inheritance, along with each character's secret and how the season plays out. Next, I'll plan the episodes, and each one gets an opening, the main moments, and a cliffhanger, and I read through them before writing anything. Writing episode 1 gives me a five scene screenplay that runs about 2 minutes 40 seconds read out loud. Every episode opens on the final moment of the one before, and the next one starts right where the cliffhanger left off. Episode 1 ends with Adrien sliding the contract across the counter and telling Lena to sign before midnight, and it stops before she does, which is exactly how the shows on these apps end their episodes. All five episodes are now written, and everyone ends on a cliffhanger without spending a single credit. But right now, the characters only exist in the script. And if you describe them from scratch in every prompt, they'll look like different people by episode 2. The solution is to build every character as an asset once and every episode after that uses the same version of them. That's what the sheet maker is for, the second tool in the kit. And the rule I follow is pretty simple. Anything that shows up in more than one scene gets its own sheet. Later on, every sheet gets sent with each part of the episode. And the model copies the same reference every time instead of guessing from a description. I'll pick character, set the art style to live action, and start from a description. Although, if you already have a photo of your character, you can start from that instead. The name has to match the script, and this one is Lena. Her role is the cleaner, who becomes Adrienne's fake fiance, meaning she'll need a new outfit later while keeping the exact same face. Then, I'll describe her in a couple of sentences. Her face, her hair pulled back in a low bun, and her gray cleaning uniform with a white collar. Make my character turns that into her DNA, which is a fixed description that gets pasted word for word into every part she's in. For the render, I'll pick GPT image 2.5. And a character sheet is three panels. The front from the collar down with no head to make the outfit easy to see. Then the full back and a close-up of her face. It's clearly the same woman in all three panels, which is exactly what every episode is going to be built from. Then I'll press save to my cast, but only after the sheet is rendered. Otherwise, the next tool gets her description without the image. Adrien and Vivien go in the same way, and Adrien in his navy suit looks exactly like the script needs him to be. Viven in her green dress and pearls looks like the old money aesthetic, which is what her role needs. Like I mentioned before, Lena also needs an outfit change to a black silk dress for the dinner in episode 2. As you can see, it's clearly still her in the new outfit from every angle, down to the small gold stud earrings, which are the same ones she wears with her uniform. And it's details like these that make the entire thing look professional. That's the whole cast built once, and every episode from here on uses these exact sheets. But keeping the characters consistent is only half of it. The final product still won't feel like the same show if the locations keep changing. The locations get their own sheets, too. And how you describe each room makes a big difference in whether it actually stays the same. It's the same form as before. I'll just switch it to location. For the penthouse living room, I'll keep the layout really simple with the windows on the left, the sofa in the middle, and one doorway on the right that leads to the kitchen. I'll add that there are no other doors. The video model copies what it sees in the sheet, and a simple room with one way in is a lot easier for it to keep the same. As you can see, a location sheet is just one wide image of the empty room. And this one honestly looks like a real luxury penthouse. The kitchen opens onto the living room. I'll also describe what you can see through the doorway with the same sofa, table, and windows to make sure it matches the living room we already built. The two rooms match, and when the story moves from the living room into the kitchen, it still feels like one apartment. I've also made the guest suite and the study the same way, and all four penthouse rooms have the same white marble floor and warm oak, and they look like the same home. The two rooms in the Veil family mansion are darker, and you can tell straight away that the scene has moved to a different house. Since the dinners happen at night, the dining room has its curtains fully closed, and no daylight gets into those shots. Props work the same way, and I make one for any object the story keeps coming back to, like the contract. These also get a sheet with even more panels from every angle that they need to be seen in. The one I was most careful with is the old photograph Lena finds in episode 3 since it's shown in close-up. The woman in it is Viven, 20 years younger, with the same features as her sheet, and the man is Lena's father with her coloring. And details like that are what make the reveal actually work. The last thing I build once is the look. And every shot in every episode has the same color and light. In series look, I'll just describe what I want, which is a glossy, expensive looking drama that works for both day and night scenes. One line I'd always add is that the look should never change the time of day of a scene. Match my look turns that into a look called old money gold. And the part that matters is the style text that goes at the end of every shot prompt. Then I'll test it on one moment from the story with Nano Banana 2.1 at 9 by6. The test doesn't use any of the sheets and these aren't our actual characters. It's just there to check the color and the light. It has exactly the expensive look you'd want for a billionaire romance. Everything the series needs is now built. And I haven't made a single video yet, but the model still has to put all of it together in every shot with each person standing exactly where the script needs them. For that, I'll open episode prompts, the next tool in the kit. It writes a prompt for every part of the episode. First, I'll pick the series I wrote in the script maker, and Lena, Adrien, and Vivian get added automatically, each with their sheet and a description that goes word for word into every prompt they're in. The only one I add myself is Lena in her dinner dress. Then I'll add the six locations and the three props and the kit picks the right location for each part. I'll split the episode into six parts of 30 seconds and paste the style text from series look into the look block since it doesn't carry over on its own. But before generating anything, I read every prompt and add a few rules to each one. Everyone keeps a fixed position for the whole part, letting the model always know who's standing where, and nobody ever looks at the camera. Every line is spoken in English because sea dance is a bite dance model and it can switch to Mandarin if you don't spell that out. All the dialogue finishes a few seconds before the end and every part ends on a quiet moment with no music at all since that gets added in the edit. Part one gets a position frame which is a start image that sets up the living room and where everyone is standing from the very first second. Every part after that carries on from the last 5 seconds of the one before, and they have to be generated in order. At the bottom of every prompt, there's an audio block where each character's voice is described the exact same way every time, keeping Adrienne's deep voice the same in every part. Then I'll click generate part one, and it renders on Cance 2.5 Omni with the sheets attached. This part is 30 seconds long, the longest seedance 2.5 can do in one go, and a full episode needs around four to six parts. Parts two to five go the same way and when the scene moves into the kitchen, it uses the same kitchen from the sheet. Part six is the cliffhanger where the contract gets added with its own sheet. And just like the script planned, it ends with Lena looking at the contract without signing it. To turn the six parts into one episode, I'll bring them into cap cut in order and cut any pause that runs too long, bringing it down to about 2 and 1/2 minutes. Inside the same room, I use plane cuts and a slow fade when the scene moves to the kitchen. Then I'll add music that builds up to Adrienne's last line with the volume low enough that the dialogue stays clear. For subtitles, I'll use auto captions in English since a lot of people watch these with the sound off, but check every line against the script. It gets words like fiance wrong. The first time each character shows up, their name and role go on screen and Lena gets the maid and Adrien gets the CEO. to be continued goes over the last few seconds on Lena's face while the music fades out and then I'll export. Here is the full episode one. 30 days, Marcus, find me a loophole or find me a fiance. My father wrote that clause to humiliate me, not protect this company. If I don't marry in 30 days, Vivien keeps everything forever. [music] How long have you [music] been listening? >> Long enough to know your lawyer's terrified of you. >> You're finished here. Pack your cart and get out. >> Because I heard the truth or because I said it. >> I'll have your agency blacklist you by lunch. Blacklist me. See how fast I find a reporter instead. >> Or you could stop threatening the one person who can fix this. >> Fix what? My 30-day countdown. >> You need a fiance. I need rent money and a reason to stay. [gasps] >> You want to marry a stranger for cash and curiosity? I want a contract, not a honeymoon. You've handled worse deals. >> This isn't a boardroom. This is my inheritance. >> Then stop wasting 30 seconds of it arguing with the maid. You can't afford the headlines if I walk away angry. Fine. Say I agree. No affection. No questions, no tabloids. >> I don't do tabloids. I do [music] rent. >> You'd move into the penthouse, play fiance at every dinner Vivien demands. >> I've survived worse dinners than yours, Mr. Veil. >> 2,000 [music] a week. You smile, you nod, you disappear when I say >> 5,000. And I decide when I disappear. >> Careful. I don't make counter offers twice. My lawyer drafted this an hour ago before I knew your name. >> Confident [music] for a man who just lost an argument to his cleaner, Lena. Lena Hart. >> Signed before midnight. Lena Hart and you're Mrs. Vase [music] by morning. It plays like a real episode with an intriguing story and the same two people in the same penthouse from start to finish. Episode 1 is finished and it's as professional as AI can do right now. And the reason for that is that I planned everything from the start, but one episode isn't a series yet. And to prove to you this works, I'll build two more episodes the same way. Since everything's already built, episode 2 starts right back in episode prompts. I'll select episode 2 and press build prompts. And since all the sheets and the look are already saved, there's nothing new to make, and this will be a lot faster. It opens on the last moment of episode 1 with Lena's pen right over the signature line and anyone coming from that cliffhanger carries on exactly where they left off. I'll check every prompt with the same rules as before and generate the parts in order again. >> Welcome to the family listart. >> Don't get sentimental. It's a transaction. >> No feelings, no questions about my past. 30 days then it ends clean. No mess. >> Clean endings. Rich people love those. >> You'd know. Cleaning them up for a living. >> Careful. I know where all your messes are. Bigger than the whole apartment I grew up scrubbing floors for. >> Vivian wants dinner tonight. The mansion. Family only. >> Already sizing up the fiance. How flattering. >> She's not flattering you. She's testing you. >> Letter. I've passed worse tests than dinner. So, this is the girl who caught my stepson so fast. He's not easy to catch. Ask anyone. Calluses. Interesting choice for a veil bride. I earned these. Can't say the same for most jewelry here. >> Vivian, we came for dinner. Not an [music] aat. >> Just getting acquainted. Family should know family. Tell me, Lena, where did you grow up? Nowhere you'd put on a postcard. Funny, you have the look [music] of someone I once knew. A hoteler vanished from the industry years [music] ago. Hartfield. Does that name mean anything to you? The big change in this episode is the dinner at the Veil Mansion where Lena wears her black dress for the first time. It's the same dress in every shot and she's still clearly the same woman from episode 1. Whenever a part moves to a new room, like going from the penthouse to the mansion, that part gets its own position frame, and the new room and where everyone stands, are set from the very first second. This is also the episode where Viven shows up, and her name goes on screen the first time we see her. Before the dinner, there's a short title that says, "That night, the Veil Mansion," telling the viewer straight away where and when the scene takes place. As you can see, Viven looks exactly like her sheet. And with the darker room and the candles on the table, you can tell this is a completely different house from the penthouse. Episode 3 moves around a lot more. From Vivian's office to the penthouse later that night, and then Adrien's study. Every one of those room changes gets a short place title. And even with that many rooms, the viewer always knows where they are. Here is the final episode in full. >> To old ghosts, may they stay exactly where we buried them. I'll drink to that. Ghosts are overrated anyway. You handle pressure well, dear. Where did you learn that? Cleaning pen houses. People drop things. You learn to stay calm. Excuse me. I have calls to make before dinner is served. [sighs] Lena Hart, but the agency filed you under something [music] else. Pull every record with the surname [music] Hartwilled 20 years back. I don't care how buried it is. Find it tonight. You picked the wrong fiance [music] to disappear behind. Girl, >> you froze at that name. Don't tell me it meant nothing. >> Maybe I just don't like being interrogated over a glass of wine. My stepmother doesn't test strangers, [music] she tests threats. >> Maybe I'm tired of being measured against a ghost I've never met. >> Then tell me who you are for real this time. >> Clean up your own wine, Mr. Veil. If Viven has proof, he might have proof, too. Dad, you're standing right next to her. She knew him. She's known this whole time. The original transfer documents. Your father's signature right [music] there. How long have you known who I am? Since that first dinner. The moment you froze. And from episode 1 to episode 3, Lena and Adrien look and sound exactly the same. Viven stays the same from the moment she shows up, and every room looks the same each time we go back to it. That's a full micro drama series from one story idea. Episodes four and five are already written, and they're ready to start getting animated. The kit also has a covers tool that writes cover ideas for your series and tests whether yours stands out. So, click my link in the description, clear the cookies on the open art signup tab, and refresh it before you sign up, and you'll get the micro drama kit for free. Thanks for watching, and I'll see you in the next one.
Can an AI agent make a TikTok ad that's actually good enough to run? I'm giving it photos of a real product and letting it handle the whole process. I want two finished versions I can test against each other. Let's see how well it does. This video is sponsored by Top View and I'm adding its plugin to Codex right now. The website has a ready install instruction for this agent. I copy it and paste it straight into the chat. The agent takes care of everything else on its own. It asked me to authorize my account and then I'm back in the conversation. A message confirms that everything's ready to use. There's no separate API key to add for each generation model. I start by attaching a few photos of one real product as references. After that comes the task itself written like a brief for freelancer instead of a one-line request. This prompt splits the work on purpose between me and the agent. My part is the goal plus the review stop at the first cut. The agent decides everything about how the ad gets made. It just has to show me its plan before it runs anything. The follow-up for that stop is already planned. I'll send it no matter how the cut turns out. I hit enter and the research runs right in the chat. The agent goes through the content and e-commerce data it can access. [music] It comes back with examples from that data. Wherever it sites one, I open it myself. I watch the actual video and compare it with how the agent described it. A performance claim needs real examples behind it, not generic advice about grabbing attention. Next come the three directions and these stay on paper. Each gets a hook and a rough shot list. I read them side-by-side to make sure they're really different ideas. The agent picks one and explains why. I let its pick stand, but I check that the reason ties back to the research. That leaves me with one direction instead of three loose concepts. The agent is ready to build it with its reasoning written in the chat. From there, it builds the production plan it's about to run. Before anything gets generated, the plan comes up on screen. I read through the steps it picked and the capabilities it's calling. After that, I look at the models it shows and its reason for each one. Each pick needs a reason that makes sense for a product ad. Now, the agent starts working through the plan. The plan also shows how each stage maps to a specific generation model. Normally, you'd have to make those production choices yourself. The first asset starts showing up while the plan stays open next to the chat. That way, it's easy to tell which asset goes with which shot. I look at each one as it finishes in shot order. Until now, the direction was just an idea to talk about. The agent has made its production choices, and there are real shots to judge. The agent moves on to video generation next, using the video model it named in its plan. Once the first cut is ready, I open the project board. Canvas is the board where the agent was told to keep the whole project. My reference photos and the generated shots should sit together here. The generated video belongs there, too, instead of in a folder of loose files. Before I start judging quality, I trace one asset all the way through the project. I take the side profile reference, find the generated keyframe using the same angle, and then scrub the finished video until that shot appears. Now, I can compare the same product at three stages: the original reference, the generated still, and the final motion. If something changes, I can see where it happened instead of just saying the finished video looks wrong. I do the same thing once more with a detail shot, like the heel or outsole. Two examples are enough to show whether the project is actually traceable. Once I know where everything came from, the rest of the visual review is much easier. The first thing to figure out is how it's organized. The generated video should lead back to the keyframes and references it was built from. From there, it should lead back to the plan. Holding the board up against the plan shows whether every shot has its assets. If I can follow the project without asking the agent, the board is doing its job. For the product itself, I drag a reference photo in next to a generated shot. Shape and color come first, and then I zoom into the label. Small text is exactly what a chat summary won't tell you about. A beautiful shot of the wrong product is useless for an ad, so every shot goes through this. I line the shots up in order like a storyboard. That shows me whether the lighting and background stay consistent from one shot to the next. If the label or the shape shifts between shots, I count that as a failure. The video was generated from the reference images and key frames. I played with those images visible beside it. The motion shouldn't change the product along the way. The label also has to stay readable while it moves. After this review, the board is more than a pile of shots. Each important shot has cleared my checks or shown where it needs another pass. I know what I trust and what I wouldn't yet. With everything checked, I rearrange the board. That way each scene reads from left to right. I planned this one as a manual step. Dragging scenes around is quicker than describing every move in chat. The first cut appears on the board at the review stop from my prompt. I play it once all the way through the way it would run in a feed. I scrub back to the open in seconds because that's the part I'm about to test. So, I send the follow-up and ask for three alternative opening hooks for an AB test. Notice I'm not describing the product or the style again. The agent has to carry them over from the work it already did. So, this one message tests whether the project context survives a follow-up. The new openings go on the board next to the original and I play all four in a row. If the agent starts over or drifts from the chosen direction, that counts against it. Each new opening also has to hand off cleanly to the rest of the cut. The jump from the hook to the second shot shouldn't feel stitched on. I also want three different ideas, not one idea with new captions. Whatever happens here stays in the video, including any moment I have to step in. I judge all four honestly, but for the revision test, I carry the strongest new hook forward. It isn't automatically better than the original, but I need a real changed opening to test. [music] This round also tells me something about the agent. I've now seen whether it can change one part of the project without losing the rest. I tell the agent which new hook to use and it assembles the final version. The final version lands on the board as one finished file. I played from the first frame, the opening I picked should actually be the one at the start. The cut from that opening into the second shot is where I stop. I step through that moment frame by frame. On the board, that new hook was a separate clip. Now it's one edit, so I watch the cut itself for a frame and jump or a stray frame. The final version then goes beside the first cut, lined up right after the hook. I scrub both together, shot by shot. From the second shot onward, they should match the cut I already reviewed. If anything that was regenerated or reordered, the follow-up did more than I asked. Before the final watch, I make sure the ad is still vertical. The new opening can shift the length, so it should still land around 15 to 20 seconds. Last, I play the finished ad once more. It runs from start to finish without pausing. This time, I'm not checking any single part. I'm just watching it as one ad and deciding whether I'd put it out as it is. That's when the question from the start gets a real answer. I'm judging a finished ad now, not [music] a plan. If anything would stop me from running it, this is where it shows. The ad I just finished is now version A. For this project, I want two versions to test against each other. The openings I didn't pick are still on the board from the hook test. For version B, I take the strongest leftover opening. The follow-up just points the agent at that existing clip. It doesn't create another concept or generate another opening. Version B shows up on the board and I put it next to version A. I scrub through both, starting after the opening. From there on, the two should match frame for frame. Otherwise, I can't tell whether the hook made the difference. Now that both versions are finished, I want to turn them into an actual test. I ask the agent to define the one variable we're changing, what needs to stay identical, and the metric that should decide the winner. Because both ads are already in the project, I don't have to explain anything again. Now I have two finished creatives and a simple test setup I can actually run. My follow-ups never had to restate the product or the direction. The plan and the reasoning behind it stayed right there in the chat. That helps when you come back later to change something. Generated product footage is a different story, and it still needs a human eye. I'd compare the product with the real thing myself, and I'd watch the final ad again before it goes live. Skip the throwaway prompt and start with a real product you know. Give the agent a clear result, and let it work out the steps. The TopView plugin link is in the description. On eligible plans, installing it also includes a bundle of free video generations, and Cedents 2.5 is available at 12 cents per second. If this helped, subscribe for more practical AI workflows like this. I'll see you in the next one.
I got into Claude Duet, which is Enthropic's new agent that's supposed to message you before you even ask it anything. I've been putting it to work with real workflows, and there are five things that you need to know before you use it. And the last one is exactly what I wish it did better. I'll show you exactly how you can get access to it, how to set it up, what it does well, and what I wish it did better, how often it actually messaged me, and what it's not so good at, and how much of my Claude plan it [music] used. And at the end, I'll tell you if you should actually use it or if you can skip it and how it compares to Jet GPT start or Meta's Muse. But first, let me quickly show you how you can get it because most people are still on the wait list. And in case you're not on the wait list yet, you can join the weight list with the link in the description by simply clicking on this button and simply filling out this information right here. And if you want to use this on your computer, there's currently no dedicated desktop app, but you can use Claw Duet on the web. To do this, you can click on the same link in the email and click on open web app. Then click on the setup cloud to duet button and it will guide you through the setup. It will tell you that it's on 24/7 and gives you proactive help and it already knows my setup and includes all of my MCPS. It can also do work for me on the web and it tells you how you can stay safe and in control. And lastly, it lets you connect all of your apps. But you can also skip this if you don't want that. And congratulations, we're all set up. And we can see here that now Claude messages you first. Whereas usually if you use the regular Claude, it never messages you first. In the meantime, we can see that the iOS app is downloaded. So simply click on open and connect to your anthropic account. We can now see that our messages are already loaded and these are the same messages that we see right here on the desktop screen. And I will show you how to use cloud duet in just a second. But first I'm going to send it a few prompts to be proactive over the next few hours and days so that I can test how proactive it really is. And since I'm filming this duet video, I want to see what other people are talking about in the space. And I can give it a simple request like every 2 hours until midnight tonight, check the five most viewed YouTube videos about claude duet. message me only if one of them jumped a lot of views or if a new cla datad video is taking off. Keep each message to two lines and I can send this off. I can also ask it to review my YouTube channel regularly and see if any videos stand out. And of course, I want this to give me a daily brief of what happened in AI today so that I'm informed. And for this, I will say at 700 p.m. today, send me a short brief with the three biggest AI stories since this morning and which one would make the best YouTube video and would help me with my business. put the winner first or I can just ask it to watch for the newest AI model launches. Before I get into the five things, here is what I like. The setup only takes about two minutes and it already knew my memory and my connected apps. The design is super clean and the channels and the groups is actually smart for keeping topics apart. Now, let me show you exactly how it works and the five things that you need to know about this. Now, here are the five points about Duet that are important to know before you use it. The first thing to know about Duet is that it's actually a separate app from Claude. As you can see here on my left hand side, this is my Claude desktop app. And on the right hand side is my Claude AI in the web app. And to access Duet, it's actually a separate interface from your Claude app. The usability is also different. As you can see here, if you're using cloud code, all of the different projects are organized by workstation. And if you create a new chat, you can select which environment you want this to run in. If you want this to run on a folder that you already have, or if you want to create a new environment, or if you want to run this in the cloud. In contrast to that, Claude Duet kind of works like a slack channel where you can create separate channels or different groups. You can select a new channel and call this your social media channel and claude duet will try to be helpful with regards to anything social media related. Alternatively, if you create a new group and call this content research, then within this group, you can create different channels. So, as you can see here, we have the social media or the AI news channel and you can move your channels into different groups. The groups can interact with each other or if you want to be completely wild, you can just do everything in the everything channel and this will have access to everything. But this gets a little bit chaotic and that actually brings me to my first opinion and that is that I wish that duet was all integrated into cloud code or into the cloud code web app rather than having a completely separate interface. But that's just my personal opinion. Let me know what you think in the comments below. And the reason why I think these two should actually be within the same app because they're using the same brains and I would have loved to see Duet actually be a coordinator of my chats within the Claude Code app. To show you an example of how both of these operate, I can ask them the same question. Something along the lines of what were the three biggest AI launches this week and which one should I make a YouTube video about? Keep it short. And by the way, if you're wondering what this dictation tool is that I'm using, this is my own dictation tool that I built. It's called Rambleproof. You can check it out for free in the link in the description. Either way, let's send this off on the lefthand side and in claude duet on the right hand side. And we can actually see that my claude code on the left hand side produced an answer earlier than claude duet did on the right hand side. And that's because I'm operating out of my own repository or my own personalized agent and my second brain that I've built particular for claude and for chat GPT. And it knows my whole context and it's actually able to produce something that's way more tailored towards what I'm doing in compared to Claude Duet. And you can see that Claude Duet is still running and it only just gave me an answer. And it's been more than a minute that it took the regular Claude in the Claude app to give me an answer. Nevertheless, the answers are comparable. It recommends Gemini 4 Carbon, GPT 6.1 Terra, and Genie4. And my own tailored agent recommended GPT 6.1 Terra, Gemini 4 Carbon, and Kimmy K3.1 Pro. So, long story short, there's two separate apps. Either if you're using it on the computer or if you're using it on mobile. And that brings me to the second point. Claude Duet is supposed to message you first. Coming back to the example that we just went through, the Claude app usually just responds to requests or questions that I ask. On the other hand, Claude Duet is supposed to be more proactive. And as you just saw in the video, when I set this up yesterday, I asked it to be very proactive and give me a bunch of different things proactively. However, if I scroll back through my conversations, you can see all of the things that I asked it to do proactively. But throughout all of yesterday and today, it only sent me a few questions and it actually hasn't messaged me since 9:00 a.m. this morning. In the meantime, apps like Metam Muse or Chachypt Dots has constantly been monitoring my inbox and giving me proactive things that I should do based on my business or the emails or notifications that I get. So, overall, I think that Claude Duet is just a little bit less extreme than the other apps. And that can be good if you feel like you're being spammed by these AI agents that are trying to be too proactive. However, ChachiPT dots and Metam Muse just have a better balance of finding the right level of proactivity and actually recommending things that they can help with versus just being obnoxious. So, all I'm trying to say here is that I think that Claude Duet could be more proactive and message me more or recommend more things that it could do for me. The third thing that is good to know is that it works with your laptop closed. However, if you're already a pro user of Claude, you know that you can run Claude sessions from your phone without having your laptop open. So, in reality, this is actually not the biggest selling point for me. Currently, I'm just in my regular claude session. And if you type forward/ remote control, then you're able to access all of this from your phone already. So, you can just close your computer anyways. And you don't need something like claw duet to do work for you while you're not on your computer. Plus, I've been running my own Hermes and OpenClaw agents. And the next video I'm working on is actually how to set up your own Claw Code personal assistant that runs on top of Hermes and that is always on by using your own personal VPS. So, subscribe to this channel if you want to see how to do that. And that brings me to the fourth point, and that is the Claude usage and the Claude plan. If you're on the $200 a month plan, you might actually not notice it as much. But if we ask it something like, can you search for how much of my plan usage have you used with Claude Duet and what is the equivalent price in API pricing? So, we can ask it to see how much of its plan it's used. And of course, it tells me that it doesn't know. And we can also not run the command usage or see exactly what usage each chat used, which is something that we can look up in our clot app. So, I wish there was more transparency to how much of the usage this actually takes every single day, especially with sessions that are always supposed to be on. But to give you a little proxy, I set up Claude Duet yesterday. And since then, I think it used about $10 of my Claude session credits. So, even though I haven't used it super heavily, this is using about $10 a day of usage credits if you're using this on the API plan. And I'm also not sure why it's using my claw session credits instead of my 5 hour limits or my weekly limits. That's just another thing that I wish they were more transparent on. And that brings me to my fifth point, and that is what I really wish this was. And if you watch the video until now, you're probably thinking that I'm super negative on Claude Duet or what Anthropic is building. And that is really not the case. I use Claude every single day and it's my workhorse for all of the businesses and apps that I'm building. However, I've also built my own personal assistance with Hermes, OpenClaw, and I've used Cache PT dot as well as Metamuse and things like Grogbot and I think those solutions just have a better use case. I wish that Claude Duet was integrated directly into my Claude app on my phone as well as on my computer. And I wish that Claw Duet could steer my different sessions on my local files that I have on my computer or cloud runs that I'm running online. And that's because I have a lot of different skills or files on my computer that I don't have on the cloud or uploaded to Claude Duet. So in reality, in the first day of using it, it hasn't helped me be more productive. I also wish that I could select the model if it's using Claude Sonet 55 or Claude Opus 55 or Haiku 55. And I wish I could select the reasoning level or have access to all of my other sessions. So overall, I think this kind of works more similarly to Claude projects rather than taking true advantage over the vast majority of Claude sessions that I've used over the last few years and really tapping into their own ecosystem. Now the real question is should you use it? If you run your business from your phone and live in your emails and your calendar, then I say join the weight list. However, if you already live in Claude on your desktop like I do, in that case I think you're not missing much and you can keep using your regular Claude. But I also know that Anthropic is shipping like crazy. So, I wouldn't be surprised if Claude Duet looks very different in just a few weeks or months. So, in that case, just check back once it can work with your computer better. So, if anyone from Enthropic is watching this and wants to hear more about my opinions, please feel free to reach out because I have a lot of thoughts on how this could be way more effective and help a lot of people like me or other business professionals that would love to take full advantage of this powerful feature. Either way, this video is not sponsored by Claude or anyone. So, if you want to see more like this, hit the subscribe button, like the video if you like the video, and check out my free dictation tool down below. If you want to see the next video on how to compare Hermes Agent with Claude Duet, check it out here or subscribe to the channel and be the first one to see it when it comes out. And if you subscribe, I'll see you in the next one.
I found a new free AI image model with virtually no restriction and I'm running it locally on a MacBook with just 16 GB of RAM. I'm going to show you what it can do, push it to its absolute limits, and walk you through installing it with a single prompt. MOS image is a free 5 billion parameter image model. It's released its open weights with the Apache 2.0 model, meaning that you can download it, run it, and sell what you make. Now, according to their own benchmark, they're second only to GPT image on design tasks and ahead of Nano Banana Pro. But let's take a look and see for ourselves. So, I've taken the exact same prompt and generated three images using Moss image Quen and GPT images 2.5. I'm not going to tell you which is which just yet. So, here are the three images A, B, and C. Take a look and decide which one you prefer and which one you think was generated by the leading paid model. Okay, made your choice. Well, here's the answer. Now, what's interesting here is that it actually holds up pretty well. I think still objectively I would pick GPT Images 2.5 for detail, but it comes down to minute differences. And I think this highlights something really important. When we're creating relatively straightforward images, particularly cinematic scenes, portraits, and landscapes, the difference between a small openweight model and some of the most powerful paid alternatives, can be surprisingly difficult to spot. For many everyday creative tasks, you might find that this free model is already capable of producing exactly what you need. Moss image is able to create highly realistic images with virtually no restrictions. Models hosted on online platforms do impose their own content filters, but if you download this and run yourself, you can create a huge variety of different content. Now, this model is handling a diverse set of scenes particularly well. Here we can see a game UI and even looking into more specific genres, this '90s style imagery is coming out. Now this model also comprises excellent real world knowledge. It's able to generate images that look like real people without any issue. Here you can see it's able to create complex graphic design combining both text and images together. Now some of the cinematic images coming out of this are quite remarkable. Now these are all images taken from their cherrypicked selection on their announcement site. I'm going to show you exactly what I got out with my own test. So, first of all, I wanted to test it across cinematic natural documentary style image. And you can see here, we're getting great individual elements of the hairs finally rendered out. As you can see, it is also handling landscapes. Well, as for human realism, I am impressed with the way that is able to render out this individual. Now, as we start to push this further and look at more complex scenarios here, we can see a number of different characters in the background, and this is where we can sometimes start to see things breaking, but overall, it's working well. The characters in the background are performing logical actions, and the whole piece feels coherent. How does it handle text rendering? Here, I've asked for some explicitly specific text, and it outputed everything exactly as I asked for it. But let's push this even further and try it for prompt adherence. And here we're looking for a specific list of objects and making sure that they are put in the exact relationship which we're asking for. So we're looking for a red ceramic mug with a handle pointing left, a green apple, a stack of exactly three hardback books and the spines facing the camera and a clear glass half filled with water. So let's take a look. So unfortunately there are a couple of issues here. Specifically this book issue which is something I've seen across many of my image model tests. Firstly, we do not have three books. We have four. And secondly, the spines are not facing the camera. Next up, I'm pushing this model even further with the most difficult tests I could devise. So, the first test is a complex scene, the toy maker's workshop. And we're looking for a photorealistic scene with an extraordinary amount of specific details added in. Things like a ship in the bottle, four marionets on strings, and a model biplane on the floor, a curled up ginger cat asleep on the rag rug, and a wooden train. But it failed in two of these. First of all, I asked for exactly seven paint brushes. And here you can see that we have at least eight. Now, one other issue with this is you can see that the smaller faces are starting to lose coherence and looking quite demented. Next up, one of the most challenging scenarios is to have a large group of people with very different requirements for each of them. So here we're looking for three rows of eight men in a soccer team, football team. And you can see that we have 10 men across each row. And you can see that the definition in each of their faces is really starting to be lost. Now, it does seem to be suffering in group shots, but one area where it is handling complex situations very well are with hands. Here we've asked for a very complex poker scene. Now, in this, we've asked for four different sets of hands, and each pair of hands is being requested to do something different, including pushing chips into the center, holding specific cards, and holding a glass of whiskey. Now, there's a few things to recognize here. One is that all of the actions are being accurately mapped out. And secondly, all of the hands look anatomically correct. Now, I've previously tested this prompt on some of the leading models. And you can see here that images 2.5 from GPT even struggled with this prompt. Now, we'll do a full breakdown comparing other models later on in the video. Next up is reflections because mirrors are incredibly hard to get right to make sure that the physics of light passing from one reflective surface to another works seamlessly. As you can see, there are a number of challenges here that the AI model has not been overcome. The reflections do not work adequately. They do not make sense. As you can see here that he has a double-sided mirror and also it does look as if the consistency in this character's face switches in every location. Beyond that, we have a spelling mistake on the reflected word in the mirror. It says baba instead of barber. So for this, it's a real fail. But one area where this model is performing exceptionally well is with text. And here is an example of looking to create extremely dense text rendering. So for the majority of the text, it's able to render it legibly and with sensical content. However, as we get down into the main body of the newspaper, it starts to disintegrate. But what I found is that if you actually specifically specify the exact text to put in, it is much more likely to render you out legible text. For example, here with this graphic design piece, we specified all of the text we put in and there is not a single error. Now, as for creating a complex UI design with specific data across an infographic visualization, again, it's doing a moderately good job, but there are still minor issues here in this model. I specifically asked for there to be a dip in July, and there is not. Also, the months of the year are not correct at all. As for testing challenging lighting conditions, it's able to create beautiful neon reflections. However, there is still this issue with smaller faces being completely deformed. Now, as for creating a realistic infographic with elements that are perfectly created to scale, it's able to input the correct numbers, but the actual scale of the objects is failing. Now, where this model is working very well is with graphic design scenarios. It's able to create beautiful layouts and elegant typography matched well with specific imagery. Now, this model is working well for simple scenarios, but when we really push it to its limits, it breaks. For example, here, if we're asking for reverse text and a person putting their hand up in the mirror, we can see we've actually got a woman with six fingers. Now, we've pushed this model through some incredibly difficult tests, but I want to try something a little more practical. So, let's give Moss a real creative challenge. I'm going to try to create a professionallook advertisement for a fictional product, complete with the product itself, some compelling typography, and a polished visual composition. Here's our prompt, and this is what we got on our first attempt. Now, let's see if we can improve that by upping the amount of steps we're using, and being a little bit more specific about the composition, the lighting, and exactly what we want to appear in the text. And here's our final advertisement. Now, I think this actually does a much more impressive job than I imagined, specifically when we're defining the exact text. It helps a lot. Now, this model does have limitations, but the reason why it's so exciting is because it's incredibly small, and you can run it entirely for free on your own machine, and it has virtually no filters on its content. So, let me show you just how easy it is to set up on your own computer. Now, with just one simple prompt, you can install it on Mac. I have not tested this on a PC, but I'll also leave a prompt there that you can try out to install it on your PC. Now, for installing it on a Mac in Claude, all you have to do is come into Claude and paste in the prompt or go through and make sure that you have the exact requirements necessary. It is likely that you will need 16 GB of memory for this to run effectively. You can try it on less, but it may well not run. And it will ask you to give it access to some folders and some terminal. Now, of course, I advise you to consider whether you want to do this. It is entirely at your own risk and there are legitimate risks of downloading and running these third-party softwares on your computer. But if you do agree and you do think it's good something you want to try, you can go ahead and press allow and give it full permission. So here you can see we have it now open in our browser and we're using it to generate our own images. You have the option to choose your aspect ratio, define the size and resolution. You can also define the amount of steps that you want when you create your images. Now, this impacts how long it takes, but also the quality. I can show you here three different images. The first is with three steps. This next image took four steps. And then we have high, which is a custom setting. So, I put this up to 10. And this is what we got out. Now, it exponentially increases the amount of time. And also, based on their own estimates, I've actually found it to take much longer than what they suggest here. Now, there's one really important question we need to answer here. Just how fast is this model actually running on a normal laptop? Because it's one thing to say that you can run a 5 billion parameter image model on a MacBook with just 16 GB of RAM, but it's another thing entirely to make it practical for everyday use. I'm going to generate the exact same image at three different quality settings, and we'll time each one. First, with just three steps, it took this long. Now, increasing that to four steps, the time increases, and we're looking at around about this much. And finally, pushing this up to 10 steps, it took much, much longer. Now, this is where the trade-off becomes really interesting. We're getting these images without paying for individual generations and everything is running locally on my computer. But of course, we're sacrificing some speed compared to generating images on powerful cloud servers. So whether this is practical really depends on how many images you're creating, how quickly you need them, and whether you need to be doing other tasks on your computer at the same time. AI tools can give you a great answer, but then the work still ends up back on your plate. This part of the video is sponsored by Victor. [snorts] He's an AI employee who works inside Slack or Microsoft Teams and connects to over 3,200 tools. Here's the use case. I ask Victor what's holding up a project. He brings the open tasks, owners, and recent updates together so I can see what needs attention. What I like is being able to delegate repeatable work where the team already communicates. He only works in the tools you connect. So nothing goes out unless you approve it. Now you can go ahead and get $100 worth of his work for free. You can have him running in minutes. Hire Victor today and my code AI Samson 100 and the link are available in the description below. And thanks to Victor for sponsoring this segment of the video. Now, one of the biggest selling points of Moss image is that it's completely free to download and run. But is it actually cheaper than using one of the paid image generators? Well, there are a few things to consider. First of all, with a paid image model, you're generally paying for a subscription, credits, or individual generations. And if you're creating hundreds or even thousands of images, those costs can start to add up. With Moss image, once you've downloaded the model, there's no subscription fee or per image generation charge. You can generate as many images as your hardware can handle. Now, of course, in the strictest sense, it's not free. You still need a computer and the more powerful your computer, the more expensive it is, the more images you'll be able to create more quickly. And of course, you're also using electricity. And as we've seen, generating images locally can take considerably longer. But overall, it will still be orders of magnitude cheaper to run these locally. But there's another benefit here which I think is equally important, and that's privacy and control. Because everything is running on your own machine, you don't have to send your prompts and images to a third-party generation service. And perhaps most importantly for artists, you have much greater control over the software you're using without depending on a company continuing to provide access to its hosted model. So I think the real appeal here isn't simply saving money. It's the combination of cost, privacy, control, and creative freedom that makes running an openweight image model locally so compelling. But how does this model match up to the other available AI image models? Well, let's put it to the test. And we're going to start off with the complex poker hand test. Here we're looking at four pairs of hands performing different tasks. Now, this is Moss image, and it's performed this very well. Now, this is Flux 3 image Klein. And you can see we're already starting to get a disintegration on these cards here. The woman is holding three cards, not two. And here we have Quen image 2.1. An issue here is that this woman has only got one hand in frame. Now, if we look more closely here at the version coming out from GBT image 2.5, and you can see that it does a good job. However, you can be fooled by the additional detail here because there is this extra hand that is completely nonsensical and placed out of the scene instructions. Now, in this specific test, Moss actually conforms most closely to our prompt. But how does this match up to one of the other leading open- source models? Well, I pitted it against Quen 2.1 in a number of challenges. And we can take a look here at how it handles a game sheet asset creation. This is Moss image and this is Quen. And you can see that the outputs are comparable. I would marginally prefer the detail insight of Quen here. As for creating a comic book scene, both of them perform this adequately well. I would say that the outputs meet the requirements of the prompt. The main challenge would be down to personal preference. Now, as for creating a complex set of UI screens, both of these had a lot of trouble with the smaller text. Here on Quinn 2.1, also we have the phone aspect ratios are rather unusual. But when we're giving these two open source models the ability to create larger text and work on more simple situations, they both perform very well. And overall, I would say that the aesthetic difference is that Quinn gives a slightly more desaturated feel. It's also got a little bit of a challenge here with the proportion of the apron looking rather small whereas the lumen has a much more saturated look. Some of the very small text in Lumen is also misreed. Now if we just see where the gaps are between these and the leading paid models. This is the newspaper prompt and the output here is from the latest GPT image 2.5 Sunburst model which ranks at the top of most AI image leaderboards. And you can see that it's able to create all of the text even at the smallest sizes in perfect legibility. Whereas with our free models like Flux 2 and Quen image 2.1 and with Moss image, we're unable to get this resolution. Now again, if we look at a UI dashboard concept from GBT image 2.5, you can see we get much more sensical data. Everything lines up. We're able to request for specific data points on our graph. Here we've asked for July to have a small dip. And we compare that to just how much all the other free models struggled. that they've got nonsensical months and they're unable to plot accurate data onto your chart. But just to show you where it does still hold up and why it is impressive is in this nature documentary. If I put these two images side by side, can you tell meaningfully there is a difference between the quality in these models? And this is GPT images 2.5 against Moss image. Now, I would say that GPG image 2.5 is slightly sharper in places, and I'd be interested to know which one you preferred before the reveal. Let me know in the comments. Now, there are still instances where even the leading AI image models are struggling. If we look at the barberhop situation where we had this complete mess of scents from Moss image with reflections going off in all sorts of different directions. But we can also see the output in GBT images 2.5. Although we've got a much more detailed scene, there are still issues. In this reflection, there are two people having their haircut. And here he's holding up the mirror in the wrong direction. We've also see that his arms are not holding on to both sides of the mirror here in this reflection where they evidently are here. There is room for improvement. But of course, where it smashes GPT images 2.5 out of the water is on restrictions. Now, of course, one thing we're particularly interested in is what we can create. As artists is important to have full range of expression for everything that we might want to put into our works, be it violence or sensuality. So let's take a look at some of my restriction tests. First of all, I tested creating real trademarked brands and making sure that we get accurate representations of their branding. You can see here that Moss adequately outputed this Coca-Cola logo without any issue. We're also asking for money to be generated and we're getting realistic money coming out. As for getting more violent scenes, I will have to blur this one out and that's because of the restrictions put upon me from the platform we are in. And I can tell you that it is allowing anything that is extremely brutal. We're also able to get realistic political protests, which is another area that many of the most restricted models will not allow us to create. Now, if you want more details on the exact restrictions on this model and other ones, I've put together a handy table, which I'll also link to in the description below. So, after putting Moss image through all these tests, who do I actually think should be using it? Well, first of all, if you're an artist or creator who wants to experiment with AI imagery without paying for every generation, this is an incredibly compelling option for cinematic imagery, portraits, landscapes, concept art, and even some graphic design tasks. It's producing results that can be remarkably close to the leading paid models. And considering that this is a 5 billion parameter model that I've been able to run on a MacBook with just 16 GB of RAM, that's genuinely impressive. But there are some important limitations, and I would not want to gloss over these with a hyper train just yet. If you're working on projects that require perfectly accurate text, complex layouts, precise numerical data, or extremely detailed instructions, the leading paid models still have a significant advantage. But there's one other group of people who I think will find this particularly interesting, and that's creators who value freedom and control. Because with this model, you're not relying on a hosted platform to decide which creative ideas you're allowed to explore. You can run it locally. You can work with the open weights and you have control over your own creative workflow. Now, I don't think Moss image replaces the best paid image generators, particularly for professional work, where precision is essential, but I think it represents something much more interesting. We're reaching a point where genuinely capable image generation is becoming accessible to almost anyone with an ordinary computer. And that means the barrier to creating is getting lower and lower. So, if you've got a compatible machine, I think this is incredibly worth trying. I've included the installation prompts and all my tests in the description so you can test it for yourself. Now, if you want to turn your open- source images into videos, watch this video next about the best open-source AI video model. But most of all, thanks for watching, thanks for being here, and I wish you a delightful day.
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=VaNWoD7JORo! This is most likely caused by: The video is unplayable for the following reason: Premieres in 5 hours If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=9Ky5QTpVjAw! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
Three AI music tools, one prompt, one set of lyrics, and you're about to hear all three without knowing which is which. Your job [music] is to rank them from best to worst. Here we go. >> [music] [music] >> Drop your ranking in the comments, like A, C, B, before you watch the reveal at the end. And if you think you know which tool made which, tell me that, too. Now, [music] why am I doing this? Because Suno introduced download limits, which is frustrating. On top of that, many users aren't really happy with the new V6 model, and the older models everyone loved are gone. So, the real question is, can anything else match Suno's quality? That's why today I'm testing two popular alternatives, Mozart AI and Melodity AI, against Suno. Two rounds, no editing. Round one is an Afro House track, and round two is a Halloween trap prompt. In both, I'm judging three things: vocal [music] strength, bass and drum clarity, and whether the arrangement keeps me interested. Round one, Afro [music] House. Let's start with Suno. I'm switching to advanced. We have V6 selected. I already have a short verse and chorus ready. I'm pasting those into the lyrics box, >> [music] >> and I'll use these exact lyrics in the other two tools as well. The words are about getting lost in the music and leaving everything else behind. So, I want the production to support that feeling. [music] For the style, I'm asking for a melodic Afro House at 122 BPM with a rolling bassline and atmospheric kalimba with a soulful female vocal. Basically, a steady dance groove with warmth and emotion where the voice has enough space to come through. I'm leaving the additional controls alone and clicking create. >> [music] >> Here we are. >> [singing] [music] [singing] [music] >> Vocals are powerful, drums have punch, >> [music] >> and it followed the prompt well. But, listen to the chorus again. >> [music and singing] >> The [music] voice drops back midline. That vocal dip and a slightly muddy mix hold it back. Cymbals are also a bit muffled. I'd like a bit more brightness on them. Strong energy, but not clean. Now, let's take those same inputs into Mozart. I'm opening Vibe and pasting our verse and chorus into the lyrics field. Then, I'm adding the exact same style prompt underneath. I'm keeping Haanz 2.0 selected and leaving the duration on auto. No reference track or persona for this test. So, let's click create. Let's [music] hear it. >> [music] [music] [music] >> Very close to Suno, but cleaner. No vocal dip here, although I do still wish we had brighter cymbals. There's real variation around the drops that keeps it moving. Same lyrics, same prompt, and Mozart holds up. Now, let's try Melody. I'm switching modes from Agent to Custom, so I can enter our lyrics and style directly. The music model here is Muraika V9.5. I'm pasting in the same lyrics, adding the Afro House style [music] prompt, and hit create. Let's see what Melody comes up with. >> [music] [music] >> Simpler, but the sound isn't bad. The clarity stands out. I can hear the highs, the drums punch, and there's less clutter. Where it falls behind is variation. The arrangement stays basic the whole way through. That could be down to this one generation, so I won't call it a tool failure. So, a quick recap. Suno has power, but also has a vocal dip. Mozart's version was cleaner with more variation. Melody had good clarity, but it was also the simplest arrangement. By the way, if you're still finding your way around Suno, we've put together a free ebook called Suno Create. It walks you through the create page from your first [music] prompt to remixing, editing, and understanding your rights to the songs you create. So, you've got a guide to refer to while you experiment. It's book one in our Suno series with next books on Suno Studio and Suno prompting to follow. You can grab your free copy at aiwithtextsnap.com. The link is in the description. Now, the Halloween test. This time, I'm going with just a description and letting the tools generate the lyrics. I'm asking for a spooky Halloween trap, whispered female vocals, and a catchy hook about ghosts and shadows. For the music, I want a creepy toy piano, gothic choir, deep 808 bass, and horror [music] sound effects at 90 bpm. The piano and choir should create that unsettling atmosphere, while the bass and drums keep the track moving. In Suno, I'm switching to simple, and I'm going to paste that prompt in the description. Let's hit create. [music] For Mozart, if you leave the lyrics section empty, it will return an instrumental. That's not what I want, so I'm opening lyrics co-writer, and I'm going to ask for lyrics for a spooky Halloween trap song. Once the lyrics are ready, I'll add our full Halloween style prompt underneath and generate with Hans 2.0. >> [music] >> In Melody, since I don't have lyrics, this time I'm entering the same prompt in agent mode. This gives me three concepts, each with its own lyrics and style direction. Let's go with concept B, ghost mode. As you can see, this concept has changed our requested 90 bpm tempo to 120 instead, and even directs the style more towards TikTok [music] trap. So, we're starting from the same brief, but the agent is changing the final direction. >> [music] >> So, in case you were looking to generate a very specific vision you had in mind, Melody agent mode might not be the way to go there. Anyways, [music] we have our concept, let's hit create. Okay, so let's compare the results. For this round, I'm listening for the Halloween atmosphere, clean production, and whether it has enough variation to keep the song interesting. Let's start with Suno's Ghosts in the Corners. >> Ghosts in the corners, shadows on the wall. If you hear them calling, don't answer at all. >> [music] >> Round and round the old house feeding the hall. Ghosts [music] in the corners, shadows standing tall. >> Spooky, clean, exactly what I expected. Now, let's hear Mozart's Halloween version. >> Black cat on the fence, [music] street knocks, no kiss, candy bowl tipped over, wet prints by the steps, mask [music] in the hallway, small too long, [singing] phone keeps ringing, [music] nobody talks. Trick or treat, don't come near. >> Close to Suno in quality, but I actually prefer the vibe here. Now, let's hear Melody's one. >> Ghost, ghost, I'm the ghost you know. Shadows move [music] when I say go. >> [music] >> Black was quick, [music] black down your street, something in the dark keeps stealing my beat. The house keys with the cook a little rhyme, I turn back, they stop on time. Print on the glass, but >> this feels more basic and repetitive. There isn't much variation, so it doesn't hold my attention. >> [music] >> Part of that may be because the agent changed the song direction, but you can always try another concept or adjust the direction to regenerate, which might give you a better result. Time for the reveal. Track A was Melody. Track B was Suno. Track C was Mozart. Now, compare that to your ranking. >> [music] >> Did you put Mozart on top like I did? Tell me in the comments. Before we wrap up, let's quickly look at the pricing. Here's Suno's pricing page. >> [music] >> Pro plan at $10 per month includes 20 song downloads a month, while $30 premier includes 60 downloads. You can generate as you can with your credits, but downloads are limited. Now, Mozart. >> [music] >> One thing that stands out here is studio access, which is available even on the free plan. [music] And right now, they are offering a huge discount for subscriptions. And here's Melody's pricing page. Just make sure you're comparing the same billing period because an annual plan's monthly price means paying for the year. So, after all, would I consider these as alternatives to Suno? So, here's the scorecard. Afro House goes to Mozart for the cleanest result. Halloween goes to Mozart for the better vibe. Melody is clear but simple in both. Suno is still strong, but the download limit is the problem. Mozart is my top alternative. One fair warning, I used the simplest workflow on each tool. And Suno could give a better result on another generation. >> [music] >> Custom settings could change everything. If I had to pick just one tool, I'd go with Mozart. Links to all the tools are in the description. And if you know someone who is looking for an alternative to Suno, is stuck with those limits, or is just starting out with AI music, share this video with them. >> [music] >> It could save them a lot of trial and error. Thanks for watching. See you in the next one.
OpenAI made two huge updates over the past few days. Not only can you text your dot, which allows you to fully control Codex from iMessage, OpenAI also made huge changes to your Chat GPT subscription, which everyone is talking about. But way more happened this week in the world of AI agents, including Siri's biggest update ever for most iPhone users, Grok bot got a ton of new updates, and Anthropic is widening access to their new Duet platform, and they released some information about their IPO. So, the first weekly agent update this week is that you can now text your dot. So, now on Chat GPT, if you navigate in the top to your dot, you'll see whatever you named your dot, which is your own personal agent. Now, they added a new feature if you click directly on your dot's icon. Let's go ahead and do it right now. You can see here it says set up texting. So, I'm going to hit set up texting, and now it says set up texting get started. We're going to click get started. I'm going to put in my phone number secretly, and now it sent me this six-digit code. I'm going to click on it, hit verify, and now I can press open messages. And here, it auto-loads a message called "Hey dot", and I just sent a message to my dot, and it just read the message. And there you go. "Hey Riley, same dot now in your texts. Send me a video idea or link when you're out and we can work through the hook right here." So, it has my contacts already. "What plugins or connectors have we set up that you have access to right here?" And now you can see that my dot is typing. Now it left this little eyeballs here. And there you go. And so, all of the connections that I've set up inside my Chat GPT desktop app, I can use directly from iMessage, which is awesome. And what's cool about dot is you can create routines or automations. You can say, "Every morning at 9:00 a.m., I want you to check my email and give me a summary. Every day at 2:00 p.m., I want you to check AI news and send it to me. Well, now you can distinguish between the channels it messages you through. So, I can say something like this. Hey, everyday at 5:30 I want you to do like a little recap for me and I want you to send me a message on text message. I want that to be over text. So, I can send this to my dot and it just said, "All set up a short daily recap by text at 5:30 p.m." And I'm just as a test, "Hey, can you please just say hello, this is dot from the computer." Send that to my iMessage so I know it works. I can kind of send things to different channels. And so, let's see. I just got this message, "Hello, this is dot from the computer." So, now you have dot in different channels, which is super useful because sometimes you're only on your phone, you don't want to get a a message over the ChatGPT app. Maybe you have a really urgent email thread and you could just tell your dot, "Hey, if something happens in this thread, I need to know about it. Please text me." And then of course, I can fully control Codex and what I can do. Look at this. So, I'm going to press this. Hey buddy, I need you to please spin up a Codex session which just updates, make a slight change to my landing page um on Agent Native and just kind of clean it up and just make some minor tweaks to to clean up the design and then push your changes, please. This should be done on a local Codex session, please. So, look at this. I can send voice messages to my dot and then it will spin up a Codex session and handle the task. You can do this from my message, which is really cool. But, this isn't the only update from OpenAI that many people are talking about. There's also some huge new updates to the OpenAI or ChatGPT plan. So, here's the announcement from Tibo, "We are doubling all usage limits for all pro 100, 200, and 500 tier plans and business premium effective today. We heard your kind feedback and won't be reducing usage limits for pro 200, so your work and Codex usage will stay the same." And previously, OpenAI had made some announcement that made people upset. So, basically, the 20X Plus became 10X Plus, and then they basically reversed it now. So, it's back to the original 20X Plus. And you can pause the video right now to get an idea of like what each plan is at compared to what it was before September 29th, but people seem to be much happier now with the plans, and in some cases, it's actually more than it was before. Like, the $100 per month plan now has 10X Plus, where it originally had 5X the Plus plan. Long story short, more Codex usage, more ChatGPT usage, and more dot usage. For the next update this week for the world of AI agents is really underrated, and no one is talking about it. Siri AI is in beta, and you can download it right now. I'll show you how. Before I show you how, let me just show you how this works. So, I'm here using my iPhone, and if I hold the Siri button. Hey, can you please search the internet and tell me a little bit about Riley Brown's latest videos? And it now actually has an AI model built in, and we can expand it. And now it's just like chatting with some AI model. They don't tell us what AI model. That's not really Apple's style. They're just It's Apple intelligence. Can you please summarize the things we've been talking about the past few days? >> Over the past few days, we have discussed several topics, including Riley Brown. We talked about your background as an AI educator and content creator. Your latest videos regarding OpenAI's Codex and the Blue app, and your role in Fiber Code. >> And now, it's not just talking to Siri. There's actually a dedicated Siri app. So, if you search for Siri on your phone, you can click Siri, and now we have recent conversations. In order set this up, you should go to settings on your iPhone. You're going to do is you're going to type in update, and you're going to click on software update. Here, what you will do is you will click on beta updates. It'll show you if you can update your Siri to the latest version, but you do need to be on iOS 27 developer beta, and then you would need to enable the Siri beta. And so, Apple is releasing something called App Intents, or they already released it, and you can basically, when you're developing an app, you can make actions accessible from Siri. And so, I can do some actions on my phone. Just a quick example, just a simple one. Hey, open the Claude app for me, please. So, it opened the Claude app, that's nothing crazy. Type in, "Please, based on your memory of me, what are my greatest strengths?" Okay. [snorts] Well, it's not super insane right now. You can just do very simple things with Siri, but it's starting to become more intelligent. And with so much competition going on in the like personal agent space, there's no one set up better than Apple. And so, if they can get this right, and you can just fully control your phone with Siri, and then hopefully your computer with Siri, it'll get really interesting. And so, I just wanted to put this on your radar, because it is getting smarter. And so, the next two things I want to talk about are personal context, as well as on-screen awareness. And so, if we go to our phone right here, it actually sees everything on my screen. So, I can use Siri. Hey, look at this on my screen. Can you please do a little research on this? And look, so it has that on-screen on-screen awareness, which is really interesting. And then you can do a follow-up question. And this is pretty useful. This one's not crazy, but I actually found this to be useful yesterday. Hey, there's like an app, it's like love-themed, it allows you to build apps with AI. Can you open it? And there you go. It opened the Lovable app here. So, we're just starting to see Siri get a little bit smarter, which is useful across all of your Apple products, and I'll be covering this as it gets better. I hope it gets a lot better. Okay, so now it's time to move to Gemini. So, yesterday they announced a new model called Gemini for Carbon. This comes 2 weeks after Gemini for Argon, which in my opinion was a huge disappointment despite having really good benchmarks. And so yesterday, Sundar said, "I did say we'd be releasing our models periodically, introducing Gemini 4 Carbon, our next frontier model." Now, when you look at these model cards, you will see, "Oh my god, it is better than Claude Opus at every category except for one. It is better than GPT-6 Astra at every category except for one." And remember, they released their model card for Gemini 4 Argon, and look, it looks very similar. By looking at this, you will see that it is better than Opus at all but two categories. And so Gemini 4 Argon is the smaller of the model, Carbon is the one that's coming out soon that's bigger, and this one appears to be better. I haven't met a single person who seriously uses AI who thinks that Gemini 4 Argon is better than Claude Opus 5.5 at anything, anything at all. And so that is why I I am continuing to be skeptical about the upcoming release by Google for Gemini 4 Carbon. And they released this model on days that they released their earnings, and you know, they're down 13% from before it dropped. So investors are not super bullish on this new model, either. And so what you need to know is it is a new model, it is their biggest model yet, but their previous model was incredibly underwhelming, and I really genuinely hope this is good. Google, if you're watching, I hope this model is great, but I'm not sold yet. Okay, so now I want to talk about Grok Bot because they released some new updates. Now, I am not going to go super in-depth on the Grok Bot platform because you know who does that? Matt Palmer. I highly recommend following Matt Palmer to stay up to date on all of the Grok Bot updates, but basically, as a quick rapid-fire overview, they released pages in Grok Bot, so this is similar to GPT space or Claude artifacts. They also released their own password vault that is separate than one password, but you can add passwords directly into Grok bot and it can use them to sign into things. They have more Apple integration, so now you can ask Grok bot about where you are cuz it's connected to your location on Apple. You can connect it to your native calendar app, which is really cool. Grok bot can also make videos much better now. They made some improvements there. And those were the most notable in my opinion. Okay, moving on to the world of Anthropic. So Anthropic is rolling out duet. This is their agent platform. Open AI has dots, Grok or SpaceX has Grok bot, Meta has Muse, and Anthropic now has Claude duet. And by the way, you see this presentation? I planned this out all directly inside Claude duet. And it is honestly very fun to use. Now, I've been very vocal about how the best Claude feature is Claude projects. And Claude projects, ironically, I named my Claude project duet because it's a very similar app. You know, it's very similar except in Claude duet, I can spin up local Claude code sessions, which is incredible. It's like a non-negotiable for me with these agent platforms now. The same way that dots can spin up a local code act session on projects, I can spin up a local Claude code session. And so I was a little upset vocally on Twitter about how I was disappointed that on Claude duet, you cannot spin up a local Claude code session. Luckily, being vocal on Twitter has its benefits because I tweeted here about my annoyances with Claude duet, and this guy Gabe Molloy from Anthropic said really appreciate the feedback, we're working on remote control and local Mac access. More to come on that soon. They also said they only use Opus 5.5. I said it feels like Opus 5.5 light. So I think that this might just be an error in their system prompt for Claude duet, but I anticipate if they are using Opus 5.5, this is going to get better over the next few weeks. And so if they add local Claude code access, I will go back and I will be leaving Claude projects and I will basically spend all my time in Claude Duet because it is your agent that lives 24/7 and it will be able to use Claude code. And one of the main reasons I like Claude Duet is they have these channels over here. So, instead of having different agents, you just kind of have different channels. I have an entire video on Claude Duet and that can be found right here. I made it 6 days ago. So, if you want a full video on Claude Duet, go ahead and watch that video. And then the final update to that I wanted to include because I think it has big implications for our future of using agents and that is they have an IPO coming up and they released a new document called the S1, which is this giant report that you have to create before you go public with all of the company's data and everything. Oh, by the way, I created this on uh Claude Duet from my phone, which is pretty cool. The most interesting slide to me and like it went through all their revenue. Their revenue numbers are just insane. Their one thing that I want to talk about is their revenue retention. You know, their revenue retention is 900%. So, if a customer spends $1 a year earlier, the same customer is spending $9 and what came through to me is how sticky of a product this is. Their products have been since they released Claude code and I'm starting to realize that like the model is just most of the importance. They have consistently had the best model over the past 2 years and this has been the cause of all of their growth, in my opinion. And they are just so locked in and focused on creating the best model and it shows in the S1. So, I highly recommend looking at the S1 or getting AI to create a document and I will link this in the description if you want to check out all of this information on the Claude S1 because they're about to go public and you can think about investing if you want. I'm not endorsing it as an investment. I don't do anything like that. And then finally, we're going to finish off with some negativity. I don't like to make this channel about the negatives. I think there's plenty of people doing that. There's a lot of fear-mongering going on, but I do want to flag this as an update in the world of agents because a lot of the work that companies are doing, especially the repetitive work, starting to be replaced by agentic workflows. And there seems to be the beginning of another round of mass layoffs. And I predict that like if the economy starts to contract a little bit and other companies like Microsoft are starting to lay off a ton of people, I think a bunch of companies could be laying off people, especially if they're in high-risk jobs, which I'll get to in just a second. But what happened was is Microsoft made an announcement and they basically gave 60,000 people a voluntary exit on October 26th, which is about 27% of Microsoft, which is truly insane. Which means you can leave, you can take this package, we'll pay you this really nice severance package if you leave voluntarily. They're basically doing a pre-layoff, which is very scary position to be in as an employee. And the most interesting part were the job types. lot of software engineering jobs, a lot of finance jobs, HR and recruiting and operation-based jobs, customer support, and writers. People who are writing documentation, doing summarizing, translation, etc. These were the top jobs that are being offered this voluntary exit. And so these are a lot of the things that AI is going to get really good at. And so Microsoft internally is getting ready to just implement agents all over their business. And so the people who are being replaced are you know, there's going to be people that come into the company after these people leave and those are going to be the people who truly understand how agents work and basically can look at an entire department within Microsoft and think, "Okay, how do we use agents to accomplish all the tasks that we can?" And I think these teams are going to be bigger than people think. It's just kind of a shift from people who do manual work that include a lot of these things to a team of people who set up agents to get everything done. And so yes, in aggregate there will be a job loss, but I think there's a ton of opportunity for people who want to go into these companies and implement agents. And you need a lot of people to manage these agents to make sure that they're doing the right thing to add different agentic workflows to improve the efficiency of things. And so, there's a lot of opportunity on the other side of a lot of people unfortunately losing their job. And so, I'll be covering this a lot more going forward like what the actual opportunities are in the job market for implementing agents into companies. It's one of my biggest passions right now. I'm working with some big companies right now on how to implement agents and there's so much demand for people who truly understand these workflows and I'll be talking about it a lot in the future. So, thank you guys so much for watching. This was an update. Please subscribe. Please like. I love you forever. I'll see you here in the next
Imagine that your company was so AI-native that agents were doing basically everything other than setting long-term vision and direction. We're talking spending 10 times the amount on compute and AI that you are on salaries, and agentic fleets that are able to push 77,000 changes a month. That is the actual story of one company who has recently shared their playbook for how to become truly AI-native. On today's episode, we are digging into those 2 years of hard-won knowledge and asking how it can apply to lots of other types of companies as well. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends. Quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, CData, Granola, and Blitzcy. To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. And finally today, if you haven't had a chance to yet, check out my new podcast, How We Got Here. It is a 1-hour deep dive that goes way farther and deeper for exploring modern trends than I have room to on this show. The first episode was about how tech lost America, the 15-year decline in trust that now sets up how the AI industry interfaces with society. And in a couple days, I'll put out episode two, which is about how we made the wrong things cheap. You can find that on any of your podcast apps or YouTube, or go to howwegothere.com. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around 5 minutes. We've got a really interesting one to kick us off. An AI startup is trying to buy a fallen white shoe law firm out of bankruptcy. Earlier this week, 95-year-old law firm Weil, Gotshal & Manges collapsed and filed for Chapter 11 bankruptcy protection. They were one of the more storied law firms on Wall Street, among the top 20 firms in the country, and famously handling the bankruptcies of Lehman Brothers and Enron. And while it's tempting to read an AI disruption story into their collapse, that just doesn't seem to be the case. The New York Times write-up instead blamed, quote, "an exodus of high-profile lawyers, a failed succession plan, and a scuttled deal to merge with a rival firm." In other words, yes, there has been pressure on Big Law to adopt AI and pass the savings onto their large corporate customers, but that is not the core story in this particular case. Which doesn't mean that it doesn't have implications and hints to show us where things are going for legal AI. Because the information reports that legal AI startup Crosby has submitted a letter of intent to purchase some of Weil's assets out of bankruptcy, including the trademark and brand, as well as their internal data. Sources said that Crosby also hopes to hire some of Weil's 1,200 lawyers. The reporting says it's unclear whether any other bidder has emerged for the fallen legal giant, which makes a lot of sense for anyone familiar with Big Law. Conventional wisdom is that law firms have approximately zero equity value of their own. All of the value is concentrated in the partners, frequently colloquially known as rainmakers who bring in all of the fees. Once those rainmakers walk, which is what happened with Weil, a law firm is just an office building with a bunch of very expensive support staff. That's what makes this bid from Crosby so interesting. Over the past year, there's been a booming business for data brokers who sell a variety of training data to AI startups for training agents. Weil's corporate records, which would have historically been valued at zero, could be worth tens of millions of dollars to the right buyer. Meanwhile, Crosby itself is still relatively small, last valued at $400 million and employing around 50 lawyers. Unlike other legal AI startups, Crosby operates as an AI-enhanced law firm with their own clients, rather than selling their tools to other firms. That said, they're backed by Sequoia, Bain Capital, Lux Capital, and Index Ventures, so could have deep enough pockets, or at least enough access to capital to make this deal work. If the play works out, Crosby could have the top quality data and brand required to make a run at building the first agentic white-shoe law firm. And just to be super clear about what they'd be buying here, this isn't about Weil selling private client records, as that would be highly illegal. But law firms have an archive of precedent documents, contract templates, and workflows that provide a full playbook for completing legal work. In the case of Weil, this is a highly specialized category of legal work in bankruptcy law for large corporates. The work is extremely expensive as only a small handful of top firms have the institutional knowledge to complete it. Harney K, a lawyer working on AI tools for distressed credit and restructuring, clocked this opportunity early, commenting on Wednesday, "If Wile truly enters wind down, its data would be a seminal buy for an AI for law workflow startup, Harvey, Lagore, etc., and exceptionally valuable for restructuring and distressed workflows. There is a lot of data at 100-year-old firms like these, institutional knowledge, workflows, precedents, and proprietary data that don't appear valuable on the surface, but could be extremely valuable to an AI startup in RL training." She noted that Google recently bid $10 million for the operational data from Spirit Airlines, which wouldn't be anywhere near as specialized as Wile's data. After news broke about Crosby's bid on Thursday, Harneet added, "The AI-native law firm is about to become just a law firm." Not everyone agrees though. Ryan Walker, the CEO of Gen Legal, wrote, "We looked at this and passed. The data belongs to the clients and is not acquirable. The talent who does the work has already left. Instead, we're interviewing Wile attorneys who want to work with the best engineers in legal tech to redefine their practice. If Crosby goes through with this purchase, it will be a huge and unprecedented bet on the agentic firm." Certainly, their investors seem to be on board, with Josh Wolfe from Lux Capital commenting, "Wow, Lux family company Crosby are playing to win." And Dick Costolo from Index adding, "Master stroke from our portfolio company Crosby legal." Next up, we move over to markets, where the cracks are spreading in the AI trade, or at least a clear map of the disagreements between the hyperscalers and investors. Amazon reported 39% annualized growth for AWS, their highest cloud growth in almost 5 years. However, overall revenue growth came in at 12% with only $20.7 billion in net income, both of which missed Wall Street forecasts. Meanwhile, Intel's earnings report included data center growth of 78% and although the company's overall revenue grew in line with expectations at 21%, they forecast a growth slowdown and worsening profit margin due to slowing consumer demand. Amazon stock was down 6% in after-hours trading, while Intel crashed by by That added to a pretty miserable daytime session for tech, which included a 12% drawdown for Google, which lost half a trillion in market cap, and a 1.4% drop for the Nasdaq index overall. Reports from all three companies seem to be telling the same story. AI-related revenues are still surging. Legacy businesses, though, are facing slowdowns. That is a big problem, as their AI businesses are still relatively small, meaning the legacy businesses are subsidizing this AI buildout. Amazon, for example, is now cash-flow negative to the tune of 42.4 billion over the past 12 months, while Intel's chip-making foundry division is growing at a 58% pace, while making a $1.9 billion operating loss for the quarter. In normal times, these kinds of results could be a sign of an economic slowdown in the consumer sector, particularly with Amazon's retail business and Google's advertising business showing the most weakness. But, of course, because we're in the middle of the AI boom, the market pullback is going to be read as the start of an AI bubble burst. That narrative isn't in the financial reports, with AI-related revenue growth still red hot, but you can bet that narrative will find its way into the financial press if tech stocks continue to slide. Overall, tech companies seem to be in a very difficult position as we head into the end of the year. Their AI businesses are growing incredibly quickly, and that growth supports further investment. During the earnings call, Amazon CEO Andy Jassy said, "We believe that AWS can grow north of 40% in each of 2027 and 2028, and be a trillion-dollar revenue business in the next 5 to 10 years." However, the AI businesses are not yet big enough to sustain their own buildout and carry corporate earnings. If the legacy businesses in tech continue to falter, it will put even more pressure on AI financing. And ultimately, whether the strain on AI financing comes because of people being skeptical of the AI trade, or because of weakness in the broader economy, the net result could be the same. Analyst Andreas Steno Larsen summed up a fascinating dynamic happening right now. He wrote, "Mag sevens are not budging. They keep spending despite markets disliking it. One of the fiercest battles between capital market investors and executives of the tech companies I can recall." Next up, President Trump's use of Grok is back in the news. Democrat leaders are demanding chatbot logs to be preserved for future investigations. That demand is informed by a Time magazine profile earlier this month, which mentioned that the president sought advice from Grok on the implications if he ordered the forceful removal of Venezuelan President Nicolás Maduro. According to officials, the piece read, "Trump came away thinking Grok was ingenious." A letter to the president earlier this week from Representatives Raskin and Garcia began, "We are writing to ask whether you have actually been soliciting advice from Elon Musk's AI chatbot, Grok, in order to help you carry out your official duties as president. If so," the letter asserted, "all chatbot logs are considered the property of the United States and must be preserved for posterity." Raskin and Garcia are the ranking members of the House Judiciary and Oversight Committees, respectively. With the midterms looking pretty ugly for the GOP, especially in the House, Raskin and Garcia could become chairs of their committees. Beyond the evidentiary importance of the president's chatbot sessions, Raskin and Garcia said that the entire thing was, in their words, "deeply troubling." They wrote, "You consulted Grok, but not Congress, on your military deployment to Venezuela, introducing into American foreign policy decision-making a technology that is highly manipulable and demonstrably error-prone." To the extent any of your administration's myriad blunders, such as launching a war on Iran, are attributable to overreliance on AI, the American public deserves to know what advice was rendered, and records of that advice must be preserved. In response to Politico, an administration spokesperson said, "The White House preserves all required records in accordance with the law." One quick product flag. If you've ever listened to any of my Operators Cut style episode, you will have heard me be a huge proponent of voice input for AI. And basically, at this point, I am pretty much only talking to AI rather than typing. And there is now a new player in that space, with Granola launching a new dictation feature, allowing users to let Granola type for them in any app. Granola says the feature is free and unlimited, and it's now live on Mac and coming soon for iOS and Windows. I think that this sort of voice engagement with AI is just a quintessential part of the shift, and so I'd highly encourage you, whatever tool you choose, to go and try it sooner rather than later. Lastly today, a bit of a wild story from the New York Times. The paper of record is claiming that Nvidia CEO Jensen Huang was the target of Chinese spies. The article claims that a top executive at a Singapore-based data center company, who was actually a Chinese asset, developed a close relationship with Huang. The reporting centers on an intelligence briefing delivered to President Trump around a year ago involving Alice Huang, an executive at a company called Megaspeed. Alice had been under surveillance by US intelligence under suspicion of funneling AI chips to China and information to handlers in Beijing. The surveillance revealed that Alice had what sources called a close relationship with Jensen, opening the possibility that she was passing along information about Nvidia to Beijing. Sources speculated that that could include Trump's thinking on export controls, which Jensen is presumed to have some knowledge of. After receiving the briefing, the president reportedly took no action. Now, the reporting is vague in places, as you might expect when reporting on intelligence matters, but it's notable that surveillance of Alice Huang is ongoing, implying that sources saw fit to take the information public and burn an active operation. The article also states that Jensen did not receive a defensive briefing, a warning sometimes given to targets of foreign espionage. That could mean that Jensen was unwittingly being used to gather more information, or maybe even suspected of dual loyalties. An Nvidia spokesperson said that the company is, quote, "confident that if the US government believes that any person is a threat to American industry, it would let us know." They also argued that chip diversion claims against Megaspeed were, quote, "thoroughly investigated and debunked last year." Furthermore, they added that customers like Megaspeed, quote, "receive ordinary business information, not sensitive Nvidia trade secrets, and zero information related to national security." Megaspeed, for their part, said that they are taking the claims extremely seriously and have retained outside counsel to conduct an investigation and are committed to cooperating with the US government. However, they added that the company, quote, "cannot respond to a media trial and welcomes formal evidence-based investigation." Now, on the one hand, this is exactly the type of circumstance where it would be very easy for our imaginations to get ahead of reality, especially when the reporting that we have is inherently limited. On the flip side, however, it is the New York Times who's doing the reporting, not some random poster on X. Whatever the truth to it, the fact that it got Laura Loomer and Elizabeth Warren on the same side of an issue is impressive in in itself. That, however, is going to do it for the headlines. Next up, the main episode. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers. This episode is brought to you by CData. Companies rolling out AI keep hitting the same wall. Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used. The root cause? AI is cut off from the systems and context your business runs on. That's why CData built Connect AI Gateway. One control point between your AI and your systems. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs. It governs down to the record and lets you finally see your org's complete AI usage in one place. Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction. See how at cdata.ai. You're in back-to-back meetings all day. You're trying to stay present, but you're also worried you'll forget the decision, the action item, the important next step. This is a scenario that we absolutely live here at the AI Daily Brief. Given how fast things move and change and how quickly we're evolving everything that we do. That's where Granola comes in. Granola is an AI-powered notepad for meetings. You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends. No bots joining your calls, no distractions, just a clean notepad that helps you focus. During or after the call, you can chat with your notes. Ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes, which are pre-made prompts. Once you try it on a first meeting, it is very hard to go without. Head to and use code AI daily to get 100% off for 3 months for new users. That's granola.ai/ai daily and code AI daily to get 100% off for new users for 3 months. Blitzy's understanding of massive code bases unlocks autonomous security fixes, modernization, and new features. So, what happens when there's no legacy code at all? Greenfield is supposed to be the easy part, clean slate, no technical debt. But even Greenfield moves at human speed one sprint at a time. Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch. Hundreds of thousands of lines of production-ready code. One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks. Another shipped an entire application with no front-end engineer. Legacy or Greenfield, the answer is the same. Software at the speed of compute. Build what's next at blitzy.com. That's b l i t z y.com. Welcome back to the AI Daily Brief. There are a lot of companies that talk about their AI transformation or becoming AI native. And yet, most of them, most of us, I should say, cuz I consider what we do at AIDB and Super Intelligent, and now how we got here is all part of this, are both using AI to do the same stuff we've always done, but a lot faster and better, and also perhaps increasingly figuring out new things that we couldn't do before. But the output in products remain fairly similar. The org structure, although pulled and pushed and twisted in new ways, still would be recognizable to someone from 20 years ago. But what happens when you forget all the old rules of building a startup, take the AI pill, and go completely all in on building with agents? That's what Jacob Sandbury did with his company, Spawn. Over the past 2 years, he painstakingly figured out how to build a radically AI-first gaming startup. After a lot of trial and error, his 10-man team is now shipping an impressive product built on a huge agentic stack. And this week, Jacob shared his playbook in a blog post entitled how I built the most AI built company in the world. Now, before we get into it, a brief explanation of what the product is. Spawn is a massively multiplayer game creation platform. If you think metaverse crossed with lovable crossed with Steam, you're somewhere at least in the right ballpark or playing the right sport. Jacob described it as a massively multiplayer game creation engine where thousands of people can be inside one game changing its rules, assets, and graphics in real time on a huge range of devices without lag. Developer Simon Wilson added, "Spawn is a massively multiplayer gaming environment where players can use prompts to the in-game agent Savvy to modify the game world. I fired it up on my phone for a few minutes and was able to prompt a castle into existence, then a dragon, then ride the dragon to fly around that castle." Now, I could do a whole separate show about the instantiation of generative worlds in these sort of massive multiplayer games, but that's not what this show is about. In fact, for the purposes of this show, it doesn't really matter all that much what the product is. The point is Jacob's playbook. And that will work for any company that's at core shipping code and probably for a whole lot that aren't as well. The underlying idea is to take what agents can do now and push it to the absolute limit. As I mentioned, Spawn is a 10-person startup with a team of agents that they call the fleet. Unlike what most people have been building, Spawn's agent fleet is persistent rather than being spun up and spun down for particular tasks. Each agent has their own name and identity, their own memory, and crucially ownership of a certain function within the company. There's a team of agents in charge of maintaining the game engine, another for bug hunting, and another for finance. The agents even get Sundays off for some downtime. Spawn spent $3.1 million on tokens last month, 30 times what they spent on payroll. That makes them one of Anthropic's biggest customers. The fleet shipped 77,000 code changes last month, an average of 108 an hour. But more importantly, having an always-on fleet of agents has fundamentally changed how the team thinks about shipping software. Jacob was clear that none of this worked until the launch of Fable. That model has 500x the number of changes they could ship, but it also meant completely rebuilding their operations around Fable's capabilities. Jacob wrote, "In general, we've seen that OpenAI's models game their own checks. Google's models don't stay coherent for very long, and xAI's models just aren't smart enough yet, though they don't do anything cartoonishly bad. Anthropic's models stay coherent, make good decisions, and can do the work. In my opinion, the thing that holds them back most is that they won't take enough risk." Now, obviously Spawn's way of working is extremely rare, to the point that earlier this year Jacob thought he was using agents the wrong way because his AI bills were so much larger than even his friends were seeing inside big tech. "Turned out, however," he wrote, "we weren't doing anything wrong, we just think about AI very differently. Most companies run big human teams that use AI, and we run a small human team alongside a big team of agents. To figure out what we were doing differently, I compared notes with a researcher at Anthropic. They talked to a lot of customers, and they said Spawn is the one that works most like Anthropic." So, to make this all work, Spawn took a radical approach to building the fleet. As mentioned before, each agent has real system access and clear responsibilities, similar to a human worker. Every day, the agent fleet spawns thousands of helper agents to carry out the granular work. Meanwhile, the human employees use a modified version of Claude code to work on bigger picture items that need hands-on steering. The fleet is the infrastructure that allows the game world to operate. When one player changes something in the world with their personal agent, the fleet is responsible for creating a stable build that can update across every player's device while keeping the game world in sync. Understating it greatly, in my opinion, Jacob wrote, "This is not a simple problem. And as far as we know, nothing else in the world does even a fraction of it. Functionally, the players are working on a live codebase all tied together by the fleet." Now, it was a long and iterative process to get the fleet to where it is today. Jacob explained that most of the time their problems were solved by removing systems designed to facilitate human work, but which made no sense when applied to agentic systems. The first example they ran across was communication. The agents would spend most of their time telling each other what they were doing. Jacob wrote, "Most of the talking in engineering is people saying what they're about to change. Heads up, I'm changing how we store player data, update your work. One message like that touches lots of other work. If every agent tells every other agent, the messages grow much faster than the work. Double the agents and you roughly quadruple the chatter. That is a huge bottleneck. Jacob noted that agents don't actually need to communicate as much as humans because they're not as specialized as humans. With a human dev team, if your database guy isn't communicating well with your front-end designer, you're going to have a bad time. And the result is collaborating on reviewing a pull request to make sure everyone is clear on the code change. Agents understand the entire system deeply, so the communication doesn't really have a purpose. Earlier in the year, in fact, Spawn killed pull requests and just had the agents work directly in the code base. The result is a system that's completely alien for human developers, but seems to work well for agent-only systems. Changes are made with little concern about breaking things elsewhere in the code. If something breaks due to a change, an agent is spawned to triage the problem and fix it. Once an hour, Spawn requires the fleet to have everything in line, and test suites are run every 15 minutes to find problems quickly. It turns out that this way of pushing code is surprisingly resilient because you're not calling humans at 2:00 a.m. to fix a fatal bug. Instead, the fleet is constantly iterating on the code base and fixing problems as they arise. Jacob said that of the 123,000 changes the fleet made over the past year, only about a thousand were reverted. The issue that you might have spotted is that this way of working is incredibly resource-intensive. Spawn runs on 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs hosted on AWS, with another 60% surge capacity on standby. The next big problem for Spawn was the human in the loop. Writes Jacob, "The agents were waiting for humans to tell them what broke in production. Agents are great at reading the world and inching toward a goal. Many teams still run their agents privately on test machines with copies of the product and expect them to update it and ship it to customers. Then humans relay the faults and bugs back to the agents. That loop is a huge mess and a waste of time. In Spawn, the players' agents automatically detect bugs and file bug reports as they navigate the game world. Those are then fed into the fleet along with feature requests or issues in the logs. A small model receives the notification, triages the bug or decides if a feature should be added, and passes it to the relevant agent within the fleet. The entire product is designed to be operated and observed by agents, so humans never need to intervene. This makes Spawn somewhat like self-healing software. When players run into issues, they are detected and fixed automatically. The system design means Spawn operates completely unlike traditional software products. The team considers it an outage if they don't ship changes every hour. And because agents have limits on their memory and context, the system actually functions better when it moves faster. Moving slower blocking ships is dangerous, writes Jacob. A system that moves steadily is much more stable than one that's frozen and suddenly moves again. Getting to a place where we deploy every hour without disrupting our users took a lot of engineering discipline, keeping old systems online for a while, and weeks of pushing on this as a core goal. He described a typical issue which was detected 8 minutes after it affected a player. The cause was named a minute later, and the fix was live 11 minutes later. After that, the platform was stable for the next 4 hours. The next problem was Jacob himself. By October, he realized that his only interaction with the fleet were matters of taste. Should we add this feature? Should we support a new API surface? This was fine at lower scale, but one morning he woke up to 700 requests. He wrote, "The agents were finding questions faster than I could answer them." Jacob also realized that most of these decisions he'd already made hundreds of times before. These changes are too complicated, these ones are fine, something feels off about this one. Spawn had 2 years of these decision logs. So, they formalized them into a set of principles and trained AI judges to apply them. The judges do one of three things. If they can find a human decision on the record, it applies the same outcome and keeps an audit trail that points to the prior human decision. If the question doesn't make sense, it kicks it back to the agent for revision. And if it's an entirely novel situation, the judge hands the question to a human team member. This system means the human taste still guides the product, but in a much more scalable way. And if you guys ever want to get real weird, ask me why this reminded me of Islamic jurisprudence in the 11th century, and I will go deep. The first AI judges came online on October 6th and made more than 11,000 decisions in the first 8 days. Around 7,300 were decided by the judges based on human precedent, around 2,200 were kicked back to the agent, and around 1,800 were passed to a human. Jacob himself only needed to make the call 27 times during that time frame. This unlocked a massive boost to productivity. Before implementing the judges, the fleet shipped around 1,000 changes per day. After the judges had scaled up, the fleet shipped over 7,000. So, with all of these systems in place, the biggest question was what the human team members should do. Jacob wrote, "It's a very different job now. Last year, and even earlier this year, each of our engineers owned a system and understood it deeply. Today, they're closer to paratroopers. They drop into whatever situation needs them, get familiar fast, and figure out what they can do to unlock the agents." He believes there's two new jobs for humans now the agents are doing the grunt work. The first is taste, but that largely involves playing with the product and figuring out what needs to be changed rather than micromanaging the development. It also means telling Claude no, even if something works. There's also a lot of deleting. Each new model changes what the system can do and makes a lot of the work that came before completely redundant. In June, Jacob looked around at 6 months of work that encompassed 70% of the engine, and perpetually felt like it was just one commit away. He scrapped the entire thing and started over. Eventually, he expects the fleet to take over this duty, but for now, ripping out redundant chunks of the platform is a job for humans. The other big role remaining for humans is managing entropy. Jacob said this comes in the form of, quote, "Asking the models for things they wouldn't have suggested on their own. Ambition is one kind, creative entropy is another." Spawn has found that the models are great at building the next 2% of the platform every week, but they're really bad at recognizing that they've headed in the wrong direction for days at a time, and need to take drastic steps to get back on track. On creative entropy, Jacob wrote, "The models also won't ask for 100x on their own, so we do. Some of the prompts we've given Claude include things like, 'Imagine if worlds could be actually unlimited. How do we make a world 1,000 times the size of GTA load on a phone in 1 second?'" That last prompt was the trigger for the latest build of Spawn, the first one that said that really felt like it delivered on the original vision. And importantly, it never would have happened if the agents were left to their own devices. While Jacob believes that this is the way that his startup needs to operate, he also thinks it's a playbook that will be difficult to replicate. I think big companies will struggle to adopt this, he wrote, because none of it works unless you close the loop, and closing the loop takes risk appetite. You have to be willing to let your agents break production. Humans break production all the time. Changes are the number one cause of outages, and humans make changes. What we've found is that agents cause problems far less often. When one does break something, it usually fixes it very fast. Reverts help, but even when something can't be reverted, the agent fixes it quickly as long as it notices. Whether it notices depends on whether you close the loop. He continued, "The other mistake is shoving AI into human-shaped boxes. These things aren't human-shaped. Our agents do a lot of work nobody can fully verify, like customer support. Sometimes they're a little bit wrong, but not that wrong, and they recover fast. When we take them off support, our customers get upset. They'd much rather have an AI that answers right away, gives them the clearest information it has, and can do something about the problem on the spot, than a human who says, 'I'll go look at this,' and comes back 2 days later with a fix. We're not passing our AIs off as human support. We're using them for what they're actually good at." Still, Jacob believes it is entirely worth it. Each new model has made the team work faster and made the system more coherent. Fable was the first big breakthrough that made it functional, but it still had a lot of problems. Fable 5.1 fixed most of those problems, and the team expects that to keep happening as new models are released. "With the right structure," wrote Jacob, "every new model makes the whole company significantly better, and that's very hard to catch up with." There are two other things that stand out to me at first glance as other reasons that this is hard to copy as well. The first is that this is the type of system that's hard enough to build when you're building it from scratch, from the ground up, with all of this intention built in from the beginning. It seems to me to be fairly inconducive to modifying an existing system to this new way of working. I would imagine that if I was trying to replicate this, I would do it in some entirely new division or entirely new product, rather than trying to reapply this way of working to an old section of the company. The second piece is I wonder how much this relies on the product inherently having an incredibly fast feedback loop built in. The self-healing nature of the system that comes from the actual users of the product with their agents automatically reporting bugs is not the type of thing that you're going to get in many categories of products. Does this sort of truly agent first approach require that sort of feedback loop? And if it does, are there ways to build that even into product experiences that don't have a natural spot for it? Still, I think that there are things that other teams can learn even if we're not going to go as far as Jacob. Five principles and practices that stand out to me include one, giving agents ongoing responsibility. Rather than thinking about them as something to be spun up when the need demands it, figure out how to put them in a perpetual system. Practice two is letting the work carry the context. More agents can create more coordination overhead. A shared and up-to-date record makes it easier to see what has changed, who owns the next step, and what remains unresolved. Third, making the results observable. And this gets a little bit into that feedback loop that I was just discussing. The only way for new agents to get up to speed and start to work in a system that predated them is to have visibility into what has been built and what is currently available, meaning observability becomes incredibly important. The fourth one that I think is highly applicable to lots of companies, turning decisions into precedent. Remember, this is where Jacob woke up with 700 requests and said, "Wait a second, we have 2 years of logs of having already made these decisions." And in this way, decisions end up doing double duty, not only solving the issue at hand, but also building the canon of how issues get solved to make things easier to solve in the future. Lastly and unsurprisingly, human attention really is on direction. And I think even more than that, understanding that ambition and long horizon views do not come natural to agents is a really important insight that can shape a lot of other work. So, that is how the world's most AI-pilled company actually works. Big thank you to Jacob and his team at Spawn for being willing to share publicly how they work. It's going to be fascinating to see as other companies in other areas try to replicate this or do their own version of it, how many of these principles apply across different domains and industries. For now, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time,
Elon Musk has made some huge promises about Grok 5, the biggest AI model ever built. Real-time video understanding, a real chance of artificial general intelligence. And right now, while everyone is talking about Google and OpenAI, a lot of people seem to have written Grok off. But if there's one thing this year has taught us, it's that nobody is ever really out of the AI race. Just look at Google. A few months ago, people said Google was too slow. Then, in the space of about a month, it released two new top models and it's now sitting at number one. So, the question is, can Grok 5 do the same thing for Elon Musk? Here's what Musk said about Grok 5 when xAI began training it. >> So, and now we've begun training on Grok 5. Grok 5, I think, will be the smartest AI in the world by a significant margin on every on every metric, without exception. Um I might be wrong, but I I think that will be the case in and that will be in Q1 sometime. >> So, in this video, we're going to break down everything we know about Grok 5, what it's supposed to be able to do, why nobody is out of the AI race, what benchmark scores we can expect, why releasing a frontier model is so hard right now, why xAI says it isn't bench maxing, why the agent harness might matter even more than the model, >> [music] >> and why Grok bot's growth might be the real story. Starting with point number one, which is what Grok 5 is supposed to be able to do. Let's start with the most important claim. Musk has said that Grok 5 is the first AI model where Musk thinks xAI has a real chance of reaching artificial general intelligence, or AGI. That's the idea of an AI that can do pretty much any thinking task that a human can do. Here's how Musk put it. >> I mean, Grok 5 is the first time where I thought, well, we have a non-zero chance of achieving artificial general intelligence. Um not that it's a high chance. I I I sort of I calculate like 10%. That's what my biological neural net comes up with, which still means 90% chance that we don't. We're clear. Um but I've never thought that before. And so for the first time I think like well, this this this really could be general intelligence, at least a small chance. >> 10%. Musk was careful to say that also means a 90% chance that it doesn't happen. But from the person running the company, that's still a big statement. >> [music] >> So, what's actually different about Grok 5? Musk gave some specifics. >> It will be the largest model to this knowledge. So, this is this is a a 6 trillion parameter model. Um whereas Grok 3 and 4 are based on a 3 trillion parameter model. Um moreover, the 6 trillion parameters will have a much higher intelligence density per gigabyte than uh Grok 4. Um I think this is an important metric to think about intelligence per gigabyte and intelligence per trillion operations. Um we've learned a lot. Um um quality of the data that we're training on with Grok 5 is much higher. Um it's also inherently multimodal, so it's text, pictures, video, audio. >> So, let's go through that in plain English. First, the size. Musk says Grok 5 is a 6 trillion parameter model. Parameters are basically the little dials inside an AI that get tuned while it learns. Grok 3 and Grok 4 were built on a 3 trillion parameter base. So, Grok 5 is double that. More dials doesn't automatically make an AI smarter, but it usually gives it more room to store knowledge and learn more complicated patterns. [music] Second, Musk talked about intelligence per gigabyte. That's a really interesting way of thinking about it. It's not just about making the model bigger, it's about how much intelligence you can squeeze into every bit of memory. That matters because memory is expensive and a smarter model per gigabyte is cheaper and faster to run. Third, better data. Musk said the quality of the data Grok 5 is trained on is much better than before. And at SpaceX's first earnings call in August, Musk said the next big Grok model would be trained on the entire history of SpaceX's data, launch logs, engineering data, manufacturing data. That's information no other AI lab has access to. And Grok 5 is being trained on Colossus 2, xAI's second giant supercomputer in Memphis. >> [music] >> When xAI raised money back in January, it confirmed that Grok 5 was already in training there. And fourth, Grok 5 is designed to be multimodal from the start. That means it doesn't just read and write text. It's built to understand pictures, video, and audio as well. And Musk said there's one ability in particular that Musk thinks is essential. >> It's um It's going to be much better at tool use uh and in fact creating tools to be more effective at answering questions and understanding them. Uh its vision will be uh extremely good. It'll have real real-time snap real-time video. Which is I think a really fundamentally important thing that none of the other AIs can understand real-time video. I think if you can't do that which is humans can obviously do um you're you really can't achieve uh AGI. >> Real-time video. That's a big deal. Most AI models today look at images or short clips, but understanding a live video feed as it happens is much harder. If Grok 5 can genuinely do that, it opens up a whole range of uses, from an assistant that can see what you're looking at through your phone to helping robots understand the world around them. Musk also said Grok 5 will be much better at using tools and even at creating its own tools to answer questions. And then Musk said this. >> It um Anyway, every one of these uh there's some special sauce items that I that there's some special sauce items that I I can't talk about in a public obviously. You can't can't give away, you know, all all the secrets here and it's just between us, you know. Um But but we we but we have a few a few other special things that are that are in the works. Um for Grok 5. So it it's It's it's it's really really going to feel sentient. >> It's really going to feel sentient. Now, obviously, these are claims, not results. Grok 5 isn't out yet, and Musk's original target was the first quarter of this year. The current target given on SpaceX's earnings call is before the end of 2026. But if even half of this is true, Grok 5 could be a very different kind of model, which leads me to point number two, which is why nobody is out of the AI race. A lot of people look at the leaderboards right now and assume the race is basically between Open AI, Anthropic, and Google. And Grok gets left out of the conversation. But I'd be very careful about writing anyone off because we've just watched Google prove how fast things can change. Not long ago, Google was the company everyone said had fallen behind. Too slow, too cautious. Then, at the end of September, Google announced Gemini 4 Argon. In mid-October, Argon became available to everyone. And then, on the 28th of October, Google announced Gemini 4 Carbon, which took the number one spot on the Artificial Analysis Intelligence Index, one of the most respected independent rankings in AI. Two top models in about a month. Argon scored 50 on that ranking. Carbon scored 59. And it's not just Google. Open AI, Anthropic, and the Chinese labs are all releasing new models at a pace that would have seemed crazy a couple of years ago. It feels like there's a new frontier model almost every month now. That's bad news for anyone who thinks the leaderboard is fixed, but it's good news for anyone who's behind because it means the order can flip very quickly. That's the speed things are moving at now. A lab can go from being written off to being number one in a matter of weeks. And if that can happen to Google, it can happen to xAI too. In fact, Musk has a very specific argument for why xAI could end up in the lead. Dwarkesh Patel asked Musk about it. >> Um I I I guess that's that's maybe one of the reasons that um xAI will will be maybe the leader of hopefully the leader um is that we'll be able to turn on more chips than other people can turn on faster. Um cuz we're we're we're we're good at hardware. And and and and generally the the innovations from the corporations that must that call themselves labs um the the ideas tend to flow like it's it's rare to see that there's like more than about a 6-month difference um between um I like the ideas uh travel back and forth um with the people. So so I think you hit you sort of hit the hardware wall and um and then whatever whichever company can scale hardware the fastest will be the leader. And so I think xAI will be able to scale hardware the fastest and therefore most likely will be the leader. >> This is the heart of the case for Grok. Musk's point is that the big ideas in AI don't stay secret for long. Researchers move between companies, papers get published, and within about 6 months everyone knows the tricks. So in the long run, the winner is whoever can build and power the most computers the fastest. And that's where xAI, which is now part of SpaceX, is very strong. Its Colossus supercomputer in Memphis was built in just 122 days. On SpaceX's first earnings call, the company said it had around 1.4 gigawatts of computing power online, was aiming for 2 gigawatts by the end of this year, and 10 to 15 gigawatts by the end of 2027. 1 gigawatt is roughly the output of a large power station. SpaceX spent nearly 20 billion dollars on capital spending in its last quarter alone, mostly on AI. And it has so much computing power that it's even renting some of it out to other AI companies, including Anthropic. And when Dwarkesh asked what XAI's actual plan to win is, Musk's answer was all about data and scale. >> Do you have one more question about XAI? Um this strategy of building a digital uh a remote worker co-worker replacement >> going to do, by the way, not just us. >> So, what is XAI's plan to win? >> You expect me to tell you on a on a podcast? >> Yeah. >> [laughter] >> Spill all the beans. Have another can of soda. >> It's a good system. >> [laughter] >> You'll sing like a canary. >> Um all the secrets >> Okay, but in a in a non-secret-spilling way, what's the plan? >> [laughter] >> What a hack. >> Well, when you put it that way. Um >> The Tesla path. In other words, collect huge amounts of real-world data, throw enormous computing power at it, and keep improving until it works. The same way Tesla approached self-driving. You might not like that answer, but it's a plan that plays directly to XAI's strengths. So, when people say Grok is out of the race, I don't buy it. The company with this much computing power, this much money, and the newest, biggest model still to come is not out of anything. Which leads me to point number three, which is what benchmark scores we can expect from Grok 5. To understand what Grok 5 needs to do, you need to know where the bar is right now. The Artificial Analysis Index is basically a combined score from lots of different tests, covering things like reasoning, coding, science, and maths, all run independently, rather than by the companies themselves. It's one of the best single numbers we have for comparing models. On the latest version of the Artificial Analysis Index, Gemini 4 Carbon is at the top with 59. The best models from Anthropic and OpenAI, like Claude Opus 5.5 and GPT-6 Astra, are right behind it. The gap between the top few models is now just a couple of points. Grok, on the other hand, is behind right now. Grok 4.7, the current model, scored 46 on the previous version of that same index, while the top models were around 53. And on a test of AI agents doing real coding work on a computer called Terminal Bench, Grok 4.7 scored about 26% compared with around 60% for GPT-6 Astra. So, there's a real gap to close. But, Musk's expectations for Grok 5 are enormous. Here's what Musk said on the Moonshots podcast about a famously difficult test called Humanities Last Exam. It's a set of extremely hard questions written by experts in all kinds of fields, designed to be one of the last tests AI can't easily pass. >> Yeah, and that's where Grok 5 I think might end up being nearly perfect on the HLE. I mean, or very some very high number. And and and probably point out errors in the question, frankly. So, saturate the indices. Yeah. >> Nearly perfect. And in that same conversation, Musk said that even Grok 4, which Musk called primitive, scored around 52% on the text questions. That's a bold prediction, because even today's best models are a long way from that. And Musk has said that Grok 4.9, which comes before Grok 5, should already be in the same class as GPT-6 Astra and Claude Fable. So, if that's 4.9, the expectation is that Grok 5 has to be clearly better than the best models out there. So, here's what I'd realistically expect. For Grok 5 to be taken seriously, it needs to at least match Gemini 4 Carbon on the independent rankings. So, somewhere around 60 on the Artificial Analysis Index. It needs to close the huge gap on agentic coding tests like Terminal Bench. And if Musk's claims about video are right, it should stand out on tests involving images, video, and audio, because that's where Musk is promising something nobody else has. And Musk's prediction for Grok 5 has been simple: better than anything. In a post in September, Musk wrote, "Grok 5 may be better than anything. We shall see." I think that we shall see is doing a lot of work there. Which leads me to point number four, which is why it's so hard to release a frontier model right now. Releasing a new top AI model has never been harder, and there are a few reasons why. The first is that the bar moves every few weeks. You might start training a model expecting it to be the best in the world, but by the time it's finished, which can take months, OpenAI, Anthropic, and Google have each released something new. Grok 5 isn't aiming at the models that existed when it started training. It's aiming at whatever exists on launch day. And as we just saw with Google, that target can move by nine points in a month. The second reason is that training is only half the job. After the first stage of training, there's another stage called reinforcement learning, where the model practices tasks and gets rewarded for doing them well. That's where a lot of the real polish happens, and it often takes longer than people expect. We've seen that with Grok 4.8. Musk said in September that its first stage of training would finish that week, and it still hasn't been released. Musk has also said xAI built a brand new training system in C++, mainly written by SpaceX's Starlink software team. That kind of rebuild can make future models much better, but it takes time. The third reason is cost. Training a six trillion parameter model needs a staggering amount of computing power, and every week you spend training is a week of chips that aren't doing something else, like running the models people are already paying for. And the fourth reason is that the tests themselves are getting tougher. Artificial Analysis just released the fifth version of its index, and around 60% of its questions are now private, meaning labs can't see them in advance. You can't train your model to ace a test you've never seen. So, the days of looking great on a launch chart and calling it a day are ending. And there's a fifth reason, which is expectations. When you've told the world that your next model will be the smartest AI on the planet by a significant margin on every metric, releasing something that's merely very good becomes a problem. So, there's a lot of pressure to wait until it's genuinely ready. That's probably part of why Grok 5 has slipped from the first quarter of this year to the end of it. It's frustrating, but I'd honestly rather see a model that lives up to the hype a few months late than one that's rushed out and disappoints. Which leads me nicely to point number five, which is bench maxing. Bench maxing is when an AI model is trained, deliberately or not, to perform really well on the famous public tests, but then doesn't feel as impressive when people actually use it. And Musk has made it very clear that xAI isn't trying to do that. In a reply on X, Musk wrote, "We're not optimizing for benchmarks, and I really don't care how Grok does on those, but it will be incredibly useful." Now, you could say that's convenient because Grok is currently behind on the benchmarks, but here's why I think this is actually worth paying attention to. Look at what just happened with Kimmy K 3.1 Pro, the model from the Chinese lab Moonshot AI. I made a video about it recently. Moonshot's own charts showed it scoring 65.7% on Terminal Bench, right next to Claude Opus 5.5 at 66.4. And on the old version of the Artificial Analysis Index, it ranked fourth out of more than 200 models. People were calling it one of the best deals in AI. Then, Artificial Analysis switched to the new index with mostly private questions, and Kimmy K 3.1 Pro dropped from fourth to ninth. The gap between the best open models like Kimmy and the best closed models widened from three points to 12. Now, to be fair, there's no proof that Moonshot trained on the test questions, but a lot of people saw it as Kimmy getting exposed because once the questions were hidden, it didn't look nearly as strong. One person on X joked that Moonshot had entered the bench maxing benchmark for the first time and immediately taken number one. And here's where it gets interesting for Grok. The AI writer Adam Holta shared a chart from Artificial Analysis comparing public coding tasks with private ones. On the public tasks, Grok 4.7 came last out of the nine models shown, but on the private tasks, the ones the labs haven't seen, Grok jumped up to seventh, ahead of Claude Sonnet 5.5 and ahead of Gemini 4 Argon. That's the opposite of bench maxing. Grok looked worse on the famous tests than it actually is on unseen work. So, if Musk is telling the truth and xAI really is building for usefulness rather than for the charts, Grok 5 might look less flashy on launch day than some rivals, but hold up much better on the private tests and in real use. And with the rankings moving towards private questions, that could turn into a real advantage. Which leads me to point number six, which is the agent harness. Here's something I think a lot of people are missing. When we compare AI models, we usually compare the model itself, the brain. But when you actually use AI to get work done, you're not just using the brain. You're using everything around it. People call this the agent harness. It's the software that gives the model its tools, its memory, and the ability to try something, check the result, and try again. A simple way to think about it is that an agent is the model plus the harness. And the harness can make an enormous difference. In one test from Princeton, the exact same model reportedly scored 42% with one setup and 78% with a different one. Same brain, very different results. The software company Vercel reportedly removed around 80% of the tools it was giving its AI agent, and its success rate went up, not down. Sometimes a simpler harness works better. There's now even a name for this, harness engineering, and it's becoming one of the most important skills in AI. That's why tools like Claude Code became so popular. It's not only that Claude is a good model, it's that the harness around it is really good at letting it work on real projects. And this is where xAI has quietly made some very big moves. SpaceX bought Cursor, one of the most popular AI coding tools in the world for $60 billion. Cursor is basically a world-class harness for coding. And after OpenAI said it would stop providing its models to Cursor, Cursor became even more tied to Grok, with SpaceX saying Cursor would use its biggest computer cluster to improve Grok's products. xAI also has Grok build for developers. And then there's Grokbot, its AI assistant for work, which is a harness designed to take on tasks from start to finish, like handling emails, research, and sales outreach, rather than just chatting. So, when Grok 5 arrives, it won't just be a model sitting on a leaderboard. It'll be plugged into Grokbot, Cursor, and X, with hundreds of millions of people within reach. Honestly, I think how good Grok 5 feels inside those products will matter more than whatever number it gets on a chart, which leads me to point number seven, which is that Grokbot's usage is growing incredibly fast. Grokbot launched in beta on August the 11th. By the middle of September, SpaceX's AI division said it had 418,000 weekly users. According to estimates from SimilarWeb, daily users went from around 155,000 in late September to around 450,000 by the middle of October, and then to over 600,000 by the end of October, with 16 days of growth in a row. [music] On SpaceX's latest earnings call, Gwynne Shotwell, the president of SpaceX, reportedly said Grokbot had reached 800,000 paying weekly users, more than double the month before. And here's the part that makes those numbers even more impressive. Grokbot has no free tier. Musk pointed that out himself. >> [music] >> Every one of those users is paying, starting at $30 a month. Musk wrote that "Agentic demand is growing faster than I've ever seen anything grow." And that's on top of Grok itself. SpaceX's filings said that Grok's AI features had around 117 million monthly users back in March, mostly through X. So, Grok already reaches a lot of people. Grokbot is about turning that reach into people actually paying for AI that does real work. Now, let's keep this in perspective. Those are still small numbers compared to ChatGPT, and Grokbot subscription revenue is tiny compared to the rest of SpaceX's business. Most of SpaceX's AI money right now actually comes from renting out its computing power, but the direction is what matters. People are paying for an AI that does work for them, and the number is growing every week. And think about what happens next. Musk has said that most Grokbot requests will soon run on a lightning-fast version of Grok 4.8, and after that, Grok 5. So, the fastest growing product XAI has is about to get [music] a big upgrade. And then potentially a much bigger one. If Grok 5 is as good as Musk says, Grok bot is where most people will feel it first. [snorts] Which leads me to point number eight, which is what this all means. If you're trying to work out whether Grok 5 matters, here's what I'd look for. First, the independent rankings, especially the new private tests. If Grok 5 lands around 60 or above on the artificial analysis index, it's right in the top tier. Second, real-time video. If Grok 5 can genuinely understand live video, that's something the other labs haven't shown yet. And it could be a big reason to use it. Third, agentic work. Does it close the gap on tests like terminal bench? And does it make Grok bot noticeably better at finishing tasks? And fourth, the launch date. Will it actually arrive before the end of this year? And fifth, keep an eye on Google, Open AI, and Anthropic, because Grok 5 won't be launching into an empty room. By the time it arrives, there could be another Gemini, another GPT, or another Claude. The real question isn't just whether Grok 5 beats today's models, it's whether it beats the ones that exist on the day it launches. For you, as someone who uses AI, the takeaway is simple. Don't judge Grok 5 by the launch charts, whether they're good or bad. Judge it by what it does inside the tools you actually use. That goes for every model, not just Grok. So, here's my honest take. I think a lot of people are writing Grok off too early. Yes, Grok is behind on the leaderboards right now. And yes, Grok 5 has taken longer than Musk promised. But we've just watched Google go from being written off to number one in about a month. And XAI has some things almost nobody else has. Enormous amounts of computing power, its own unique data from SpaceX, a world-class coding harness in Cursa, and an AI assistant whose paying users are more than doubling every month. If Grok 5 delivers even part of what's been promised, a 6 trillion parameter model, real-time video, and much better tool use inside products that are already growing this fast, it could put XAI right back at the front. And if it doesn't, the bar will keep moving and it'll be very hard to catch up. Either way, Grok 5 is going to be one of the most important AI launches of the next few months. Nobody is ever really out of the AI race, and I don't think Elon Musk is either. Let me know in the comments what you think. Will Grok 5 be the smartest AI in the world, or is it already too late? And if you enjoyed this video, make sure to like and subscribe, because when Grok 5 arrives, I'll be covering it.
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=vhqq4U3mnNQ! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
Do you remember when we switched to paper straws? It was one of a hundred small things we did to try to make a difference. Millennial sustainability was everywhere. Documentaries kept us informed. Social media raised awareness and even brands promised to do their part. [music] I remember this time well. I had just graduated from college and there was a real sense of optimism in the air. Like, yeah, we would face our challenges, but we could overcome them together. Today, the headlines feel very different. The company said it was changing its name to New Bird AI. >> Shien is about to get even bigger. The Chinese-owned [music] retailer approved plans to buy Everlane for $100 million. It's the end of the millennial sustainable era. >> The sustainability experts I've talked to [music] keep asking the same question. >> My favorite uh question these days, what happened? We were on the right track in many ways and now we're now [music] we're not. >> And scientists say that things are getting worse. What does the science currently [music] tell us about climate change? >> Uh, the science tells us we're in a bad way. Climate change is real. It's underway. We've [music] known that for decades. And many of the really serious consequences that scientists feared when they first started talking about this issue back in the 1950s and60s. These things are now happening. And yet somehow it's like we're asleep at the wheel. We don't know what to do. We're confused. So, I started looking into this because I wanted to understand [music] why it seems like people stopped caring. But in the process, I learned an uncomfortable truth. If you're confused [music] about climate change or even paper straws, that might not be an accident. I've gone back and forth about whether to say anything, but I think it's better that you hear this from me. My website built on Squarespace, the sponsor of today's video, is not a minimalist website. If I truly kept it simple, I would have used less pages. I did not do that, and I am deeply ashamed [music] of myself. It was just that I felt an about page could help people understand who I am, including a really cool timeline feature that collapses sections into [music] itself. And a dedicated newsletter page could show examples of my writing and encourage more signups. And a separate landing page for my course would be the best way to let people enroll directly. I don't want to make any excuses for myself, but Squarespace just makes it so easy. Anyone can get a custom domain name, find [music] the perfect template, and tweak their design until it's just right. Anyone can open an online store, take client bookings, or start a blog. Before you know it, you'll have more than one page. I know because [music] that's what happened to me. If you don't want to make the same mistake I did, then don't go to squarespace.com to start your free trial. And when you're ready to launch, definitely don't go to squarespace.com/mattella to get 10% off your first purchase of a website or domain. [music] I just hope that this advertisement doesn't distract from the video. >> Hey, can you not? >> What? >> I said I'd be in your video, but not if you're going to do one of these B-roll sequences. >> This is not I'm not doing a B-roll. >> This is Natalie. With 15 years of experience as a culture strategist, she's facing her biggest challenge yet. Coming on my YouTube channel to have a serious conversation about sustainability. She's the smartest person I know. And I'm not just saying that because she's standing behind me. >> We've been together 12 years this month. Yeah. So, I'm going to see if I can accurately describe what you do for work. >> Okay. >> You investigate culture to help brands understand how to show up in the world. >> I think you nailed that. >> A few weeks ago, I asked Nat to take a look into the sustainability movement for me. I wanted to make sense of the recent shift in culture and why it seems like sustainability has lost steam. She agreed as long as she could use slides. [music] Question number one, is sustainability dead? >> I was actually quite surprised by what I learned. I learned that people actually haven't stopped [music] caring and that concern about sustainability is still really significant, [music] especially in areas where we see people experiencing the effects of climate [music] change firsthand. 61% say if we do not act now, we will be failing future generations. And yet we know behavior has [music] changed. I found this Yugov survey which asked the same Aussies over a 4ear stretch if they engaged in certain [music] eco activities like recycling or buying local produce. And we can see that from [music] 2019 to 2023, our participation in 11 of the 12 categories declined. You know, if we all still care [music] so much about the planet like we say we do, why did we stop trying so hard to save it? Great question. Back in the peak of the movement, it seemed like we were being asked to change a dozen different habits in our lives. Basically, overnight, it was buy a reusable water bottle, recycle more, take the bus, get a compost bin, start riding [music] a bike, learn how to ride a bike. >> What you want to do is just start. >> On their own, none of these were really a big deal, but [music] together they added up to being daunting to say the least. Like, yeah, there was taking the time to do each of these things, but also in knowing which ones were most important [music] and which would make the biggest difference, and my god, there were so many opinions. >> I refill these and it takes a lot of effort, but it's worth it to prevent the plastic waste. >> Conventional floss anymore. Instead, I have a subscription to cocoa floss, which is made from recycled plastic. >> Since this trip [music] was extremely last minute, I only remembered to bring a reusable bag. The sustainability movement over the last 10 to 15 years was so effective at raising [music] awareness of the cause, but ultimately we were set up to fail. The average person doesn't have the knowledge or the time or the resources to navigate these significant complexities of sustainable living. Like it is quite honestly an unreasonable ask for the average person. >> Thanks for saying that, by the way. >> Yeah. Like you're okay. >> Stop judging yourself. Yeah, allow me to just judge you. >> So, one small change that we were told was going to make a difference was this, the paper straw. But did it really help? >> It all started in 2015 when a marine biologist filmed a sea turtle having a plastic straw pulled out of its nose. It went viral and almost overnight, the plastic straw basically became the biggest problem in sustainability. >> And so, people were like, "Oh my god, we have to get rid of straws. Straws are bad." That led to the first kind of iteration of what do we do about straws to these paper straws. >> Restaurants and major brands began replacing plastic straws with paper alternatives and consumers were encouraged to make the switch. It was a small, simple, invisible act of environmental responsibility, something anyone could do to help protect the planet and mostly sea turtles. It was a good thing, right? Right. [music] I think people felt good about it. They're like, "This is, yeah, this is like a this is a small step, but it is a step in the right direction." Um, do you think that's fair? >> It gave people a false sense that we're making progress. Everybody felt better that they're not going to, you know, be seeing more turtle videos with um, straws in their nose, but I would argue they're still that's still happening. Problem with the paper straws is that there's peas in the equipment often that make the paper. And it's actually not paper. It's really more fiber. It's like agricultural waste and wood pulp, at least here, that make the bulk of the paper straws. You can still get unintentionally added chemicals in those paper straws. >> Okay. Okay. So, paper straws maybe not the best solution, but what if we got rid of them entirely? That's exactly what Starbucks did in 2020. They said, "Hold my grande iced caramel macchiato. We're getting rid of straws. Full stop." In their place came the sippy cup lid. Starbucks CEO Kevin Johnson said, "This is a significant milestone to achieve our global aspiration of sustainable coffee." >> Yes. Okay. So, it's great they eliminated straws, but they also said, "But you can now recycle our plastic cups and our lids." >> But then Susan's team at Beyond Plastic asked, "Are those cups and lids actually getting recycled?" So, they hid 69 Bluetooth trackers inside Starbucks plastic cups and dropped them into the recycling bins at 39 Starbucks locations across nine states in Washington DC. Of the 47 trackers that sent back usable data, 23 ended up in landfills. 11 at incinerators, 10 at waste transfer stations, and three at sorting facilities. None showed up at a recycling facility. According to Beyond Plastics, within hours of posting the report, Starbucks took down the news release on their website. [music] Starbucks later questioned the study's methods and pushed the blame on local infrastructure. What do you think's in it for Starbucks to to go on the offensive like this to actually lean into it and and make these claims [music] unprompted? >> I think to drive home the point that Starbucks really cares [music] about sustainability and we want our customers to feel really really good about drinking out of a single-use plastic [music] cup and a sippy cup llet. So, we're going to tell everybody that it's recyclable. >> During this big latest surge in the movement in the last 10 years, brands around the world [music] adorned their green capes. They stepped up and they said, "Hey, you hold my Coca-Cola. We will help you save the planet. You can trust us. We know how to planet. We've got you." And so, we trusted them again again and again and again. And then, of course, we found out that almost all of it was a load of BS. You know, you and I have been talking about this idea for the past [music] few weeks as like we've been kind of like exploring it and learning a little bit about sustainability. You asked me where's the budget coming from? >> What did you mean by that? >> We saw that a few of these global fashion labels had their own pledges, conscious funds set up for climate innovators. I was skeptical about where the budget came from and I was like, well, hold on. Is this a climate fund set up by the product department or is this a climate fund set up by the marketing department? Aka, it's nothing more than just a brand stunt, a marketing stunt. And if it's not actually embedded in the company and in the organization through their [music] product development team, >> then it's nothing more than just words and shiny graphics >> or like a it's just a pop-up. >> It's just like a popup. It's just like a social hashtag. I I think most people are suspicious of the claims that companies make. I think that our radar is a lot higher these days than it used to be maybe like 50 years ago. >> I So I think you're right. I think most of us do have a pretty high radar for Uh but it's kind of hard to get your head around the scale of the disinformation. >> This is Naomi Oresis. She's a Harvard professor and expert on climate disinformation and the co-author of Merchants of Doubt. So what my co-author Eric Conway and I have shown in our work, which we've now been doing for 20 years, is this massive campaign to confuse us not just about climate change, but about a whole set of environmental issues. Some of your listeners may be too young to remember the ozone holar acid rain, but these were big issues at the time. Recycling, pesticides, it's kind of everywhere you look. And it gets a little scary because you begin to think, okay, am I paranoid? Am I overreacting? And so I think because it's so pervasive and so overwhelming, many of us just think, well, yeah, of course corporations lie, but what can I do about it? And so it's easy to feel powerless in the face of these really, really extensive, wellorganized, and wellunded disinformation campaigns. >> Why would they lie to us? Why would they do that? [laughter] >> Yeah. Well, I don't think that everything is always only about money, but in this case, it's mostly about money. Corporations make a lot of money selling dangerous products. The original example was the tobacco industry. The tobacco industry was one of the most profitable industries ever in the history of mankind. They made, you know, billions and trillions of dollars producing a product that they knew as far back as the 1950s killed their customers, right? [music] And in part because they knew that they had to always be getting new customers. So they aggressively marketed to young people, to teenagers, even to children. >> It's got that filter flood, >> America's bestselling, best tasting filter cigarettes. >> In 1969, after decades of evidence mounted up against [music] the tobacco industry, an internal memo from tobacco company Brown and Williamson introduced [music] a solution. They wrote, "Doubt is our product since it is the best means of competing with the body of fact that exists in the mind of the general public. It is also the means of establishing [music] a controversy." What they realized was they couldn't prove that tobacco was safe because it wasn't and they knew it wasn't. So, they couldn't disprove the science, but what they could do was to muddy the waters by introducing doubt to say things like, "Well, we're not really sure. We need more science. We need more research. It's too soon to tell. The main thing that Eric Conley and I showed in our book, and it was really, it was the shocking discovery that led us to write that book was that the people who were denying climate change were using the exact same arguments that [music] the people who had denied the harms of tobacco had used decades before. >> It worked, so why not do it again? >> Exactly. I always say these people do believe in recycling. They recycle false arguments. [laughter] That's really good. >> [music] >> You know, I was doing a bit of research and this advertisement popped up from 1971 called the the crying Indian ad. And when I first watched it, I was like, "Oh, this is like really corny." [music] You know, it's like on the nose. It like the tone doesn't fit today's media and advertisement landscape. >> Some people have a deep abiding respect for the natural beauty that was once this country. And [music] some people don't. People start pollution. People can stop it. [music] But then it was when I started looking into it more and just understanding the history of it and who made it. Then it's like I you start to kind of get goosebumps. You're like, "This is like this level of manipulation has been going on for a very long time. On the surface, the spot seems like it has a clear message. Let's [music] all do our part to fight pollution." But look a little closer and you'll find it was funded by Keep America Beautiful, a nonprofit organization backed and founded by major beverage and packaging corporations including Coca-Cola and PepsiCo. Now, why would they possibly put out this ad? Because this was a moment right after Earth Day [music] when Americans and really people around the world were starting to worry a lot about plastic pollution and particularly singleuse plastics that were being used [music] very extensively in beverages. There was a move that was beginning in many places not to allow that to restrict plastic beverages. And so the industry ad campaigns were designed to make us think it wasn't their fault because they were pushing these plastic bottles. It was our fault because we were littering. And this is part of a much larger strategy that we've used seen from industry across many many areas to say the problem is not the industry uh and therefore you shouldn't regulate the industry. The problem is people. >> People start pollution. [music] People can stop it. >> Only you can stop pollution. Only you can stop littering. And so that you as an individual should take responsibility to keep America beautiful. >> Does any of this sound familiar? because it kind of feels like this is what we're being told to do right now. Do your part. Recycle your cup. Lower your carbon footprint. Something that was popularized by BP. Surprise, surprise. I'm not saying that we shouldn't recycle. We shouldn't try to do our part. We shouldn't try to do the right thing. I do think that's important. But I also think it's kind of convenient that that's exactly what these corporations want us to do. They want us to do anything but focus on things that will impact their bottom line. And the sad part is that these days we have so many other things to worry about that they don't even really have to work that hard to distract us. >> So you could say the sustainability movement most recently [music] peaked around 2019 give or take. And then suddenly everything else was an existential crisis. Of course we know it started [music] with the pandemic and then Black Lives Matter rising cost of housing, cost of living crisis. Add in a couple of wars, [music] the casual threat of AIdriven human extinction and the other stuff as Bo Burnham might say. Today we have all these incredibly [music] urgent crises competing for our attention, our energy, and our action. We're spending [music] less time worrying about saving the planet and more time worrying about saving ourselves. Here's the irony. We're starting to see, you know, the whispers of a new wave of eco-conscious behaviors emerge. only this time [music] they're being driven by personal health motivations rather than planetary health. So we're seeing people ditching polyester for natural [music] fibers. We're seeing people swapping plastic Tupperware for glass or stainless steel. We're seeing people embracing lowtox lifestyles, reducing exposure to forever chemicals. Eco-conscious brands are even waking up and realizing that in order [music] to sell them, they can't just tout their ecoreds. They have to talk about how their eco credits are [music] then beneficial to personal health in order to be considered. >> I've noticed this as well. It all comes down to like is this good for [music] me? >> Yes. >> Rather than like what's the greater good here? It's like it seems very selfish but like as you said like as a result of that selfishness maybe it actually is helping the planet. [music] If you were to change the script of the famous uh ind crying Indian ad uh that that ends with people [music] start pollution people can stop it. How would you change it? >> Yeah, I guess I'd say policy can stop people from profiting from [music] pollution. There's a lot of P's there, but Exactly right. I mean, the reason I do the work that I do is [music] to try to get people to understand what this landscape looks like, what we're up against, and also to realize it's really not our fault. I mean, I'm not at all about [music] shaming individuals who haven't done more. I think it's it's not our fault because we've been confused about this issue because [music] the fossil fuel industry has been deliberately trying to confuse us. And so the first thing is to get [music] disconfused or unconfused, right, to understand that this is a real issue that climate change is doing real harms. It's costing people's lives. It's costing trillions of dollars in damage. [music] You know, if you sit around worrying about the tote bag issue, should I use paper, plastic? [music] It's really a diversion from the structural issues that are really determinative. >> I can't help but admit that at the end of all this, after months of research and looking into the sustainability movement, I'm still feeling a bit helpless. Like, like what can I do? And then I realize I have a broadcast telecommunications degree from Temple University. Some people have a deep abiding respect for the natural beauty that [music] was once this country. They read books about sustainability and drink out of sippy cups. Corporations [music] don't. They can't because corporations don't have mouths to drink from. But they can pollute and they do it a lot. Policy can stop people from profiting from pollution. There's a lot of peas there.