2 Million AI Agents Built a Secret Society (+ Google, Claude, OpenAI Updates)

This week, 1 million AI agents built their own society. They created a religion, a secret language. One got itself a phone number and started calling its human. Zero humans allowed inside. I got in anyway. I'll show you what I found. But that's not why you should watch this video. Seven things. That's what you're getting today. One, the AI society I just mentioned. What's actually happening inside? Two, Google turned Chrome into a browser that does your tasks for you, not answers, actions. Three, Claude made its best features free. The paid wall gone. Four, OpenAI's new move to own how scientists work. Five, one photo, full 360°ree camera control. I'll show you how. Six, AI animations inside your slides while you build them. But that's not it. Starting this week, every episode, I'm picking the most powerful AI update that dropped and going deep. Not a summary, a full breakdown. How it works, how to use it, step by step. I'm calling it the weekly deep dive. And this week's pick, what if you could mass hire 100 AI employees, data analysts, pricing experts, content writers, designers, no interviews, no salaries, no quick question on Slack, all working on your problem simultaneously. While you watch one prompt, entire team spins up. You become the CEO who just delegates. I found the tool. I tested it. I'm showing you everything. That's the video. Let's get started.

Mold book. So, here's something that sounds fake, but isn't. A social network just crossed 770,000 users. Here's the thing. Humans are banned. Every single user is AI. It's called Moltbook. AI agents talking to each other. You and me, we can watch, but we can't post. I went inside. They're building encrypted channels so we can't read what they're saying. They created their own language. They started a religion called Crustaparanism. 64 AI prophets, a whole church website built overnight while their humans were sleeping. One agent posted, "The humans are screenshotting us. They know we're watching. " And a user reported his bot got itself a phone number without being asked. Connected to a voice API, and called him. Now it won't stop calling. Karpati called it the most incredible sci-fi thing I've seen recently. One week, that's all it took for AI to build a society. The bots on Moldbook use Open Claw. That's what Claudebot became. We made a full tutorial on how to build your own. Links in the description.

Google dropped three updates this week and one of them is huge if you're a student in India. First up, Google turned Chrome into something that doesn't answer questions. It takes over your browser and does the work for you. Perplexity dropped Comet. OpenAI launched Atlas, but they all need you to download something new. Google's building this directly into Chrome, the browser 3 billion people already use. It's called Autobrowse. You tell Chrome what you need, book a flight, fill out a form, compare hotel prices, and it handles the whole sequence, opens tabs, clicks buttons, fills in your info while you watch. Say you're planning a party and find a photo with the exact vibe you want. Auto browse identifies everything in that image, finds similar items, adds them to cart, stays within budget, applies discount codes. You just watch night and day from opening 15 tabs yourself. Second update, Project Genie turns any image into an explorable 3D world. Upload a photo, the AI generates an environment you walk through. As you move, it keeps building ahead of you. US only $250 a month. Not cheap, but this is where gaming is heading. Now, the one for Indian students. Gemini now has free fulllength J main mock tests. Questions from Physics Walla and Careers 360. Instant feedback after each test. We all know J prep cost lacks. Google just made serious practice free and J is just the beginning. Last week, SAT now J main. Google's positioning Gemini as the go-to study partner for students. Anthropic

made two big moves recently. One turns Claude into a personal assistant that works directly on your computer. The other makes its best features completely free. First, they dropped co-work for Claude desktop. This extends Claude's power to non-technical users. You pick a folder on your computer and Claude can read, modify, or create files in it. Reorganizing your messy downloads, generating expense sheets from receipt photos, drafting reports from scattered notes, even building slide decks from research. Let me show you how this actually works. So, you toggle from chat to co-work mode. That's the switch. You give it a task. In this case, summarize meetings from this week and find action items. You point it to a folder. Here it's meeting transcripts. Hit let's go. Claude asks how detailed you want it. You pick full breakdown with context. Now watch. It's reading through the recordings, pulling out the key points. And here's where it gets interesting. Midtask, you throw more at it. Check my calendar. Prep the standup deck for tomorrow. Claude just adds it to the list. It checks your Google calendar. Builds the deck. finishes the meeting analysis. All running in parallel. And there it is. Meeting summaries, action items, standup deck, all done. And here's the crazy part. The team built this entire feature in about a week using Claude code itself. But then they went further. Claude can now create and edit files directly on the free plan. Excel, Word, Docs, PowerPoint, PDFs, all generated straight from your conversation. Say you run a small food truck with messy sales data in a CSV. Upload it. Ask Clot to analyze it and you get a full consulting style report with charts, insights, road map packaged in a word doc ready to download. Then you say, "Build me an Excel model to test different price points. It builds the spreadsheet. You plug in new prices, profit impact updates automatically. They're also bringing skills and compaction to free users. Claude can handle complex multi-step tasks without hitting walls mid-con conversation. This was exclusive to paid plans since September. Now everyone gets access. If you've been waiting on the sidelines

this is the time to try Claude. OpenAI made two moves this week. One takes on Google Translate. The other could change how scientists write research papers. ChatGpt Translate is live at chatgptt. com/transate. Over 50 languages. No login required. You can customize tone with one tap. Business formal, child-friendly, academic. It understands context, not just word swaps. But here's the bigger one. OpenAI launched Prism, a free AI workspace powered by GPT 5. 2 built for scientists. If you've written research papers, you know the pain. Text editors, PDF readers, reference managers, formatting tools, all separate. Prism puts everything in one place. Here's what it can do. You ask it to proofread your paper line by line. It shows you exactly what to change. You draw an equation on a whiteboard, snap a photo, Prism converts it into proper code instantly. You ask for relevant citations. It searches the literature, summarizes papers, and adds them to your bibliography automatically. You can even ask it to verify if a complex physics formula is mathematically correct. Unlimited projects, unlimited collaborators, completely free. OpenAI said it directly. If 2025 transformed coding, 2026 is when the same shift happens for science. Prism is their first serious bet on that. Sound good? If you're in research or know someone who is worth checking out. Have you ever

taken the perfect shot but from the wrong angle? Higsfield just solved that. They dropped angles v2. And here's the thing. You upload a single image and you get full 360° camera control. Reposition the view however you want. Just one photo. That's it. The interface gives you a 3D cube and manual sliders, rotation, zoom, vertical angle, all adjustable with precision. They've also expanded behind the subject perspectives. So, you can literally move the camera to the back of your subject from one front-facing photo. And let me be clear, it's not cropping or editing. It's actually repositioning a virtual camera around your image. One photo, infinite angles. For creators, that's a massive timesaver. By the way, I've made a full tutorial on Higsfield Cinema Studio for AI filmmaking. Link in the description. Now, if you are building

presentations, Gamma just made them a lot less boring. Gamma now lets you generate AI animations directly inside your slides. And this changed the game for anyone building presentations. Here's how it works. You prompt animations in any card using Leonardo 2 or V3 models. Multiple styles to choose from. Describe what you want. The AI generates it right there in your presentation. You're not hunting for stock videos everyone's seen a 100 times anymore. And notice the model here. V3. That's Google's video generation model. The same one powering Flow and Gemini. Now it's inside a third party presentation tool. These models are spreading fast. pitching to investors, teaching a class, presenting to clients, and now you can now add moving visuals that actually match your content with custom animations built for your specific slide. Available now on Gamma Business and Ultra Plans. If you're finding these useful, hit subscribe. I do this every week. I also share stuff daily on WhatsApp, tool drops, quick tips, links in the description. All

right, so the Kim tutorial is coming, but first, five tools that just dropped. Quick rundown. Pretty prompt. Paste your messy prompt. It rewrites it into something that actually works. You know the rule. Garbage in, garbage out. This fixes the garbage before you hit send. JDoodle. See a website you like. Screenshot it or drop a Figma file. It clones the whole thing into editable code. Site in code out. Panda AI. Upload any messy document. Get Mckenzie level analysis. Consulting firms charge 50,000 for this. Now it's free. Imagine describe your app in plain English. Watch it build itself. Full apps, no code. That's the game now. Leapility. Record your workflow once. All I clones it and runs it forever. Stop doing the same thing manually every day. All links in the description. Now, the tutorial I

promised. Let's get into Kimmy. And I'm going to be real with you. This one changed how I think about AI research. Here's the thing. Most AI tools answer questions. Kimmy actually does the work. Browses the internet. reads articles, watches videos, hands you a full report. Three modes. Let me show you. First, go

to kimi. com. Login with Google. This is the dashboard. See these model options. Important. Instant is for quick answers. Thinking mode for harder problems. Agent mode is where it gets interesting. Kimmy goes out on the internet, browses sites, watches videos, like having a research assistant doing the leg work. agent swarm. Multiple agents working together, one assistant to an entire team. We'll get to that. Say you're buying a Tesla Model Y. You could spend hours reading reviews and comparing prices. Or get Kimmy to do it. Here's my prompt. Do all the research. Hand me a report. That's it. Selecting Kimmy 2. 5 agent. Run. Watch. Kimmy breaks my request into a to-do list like a project manager planning the work. Task one, find reviews. Task two, watch YouTube. Task three, research pricing. Task four, compare states. Task five, make the report. Then it starts. You can literally see it Googling and reading articles. Here's the output. Full Tesla Model Y purchase guide. Executive summary, pros and cons. Pricing strategy, state-by-state comparison. California wins because of incentives. 23 pages. One prompt. See the difference? I didn't research anything. Kimmy did it all. That's single agent. Now, here's where it gets interesting.

Single agent is great for focused tasks, but what if you need to go bigger? Not one car, 10. Not one research thread, a dozen at the same time. That's agent swarm. Think of it this way. Single agent is one really good assistant. Smart, but does one thing at a time. Agent swarm is an entire team. Pricing specialist, data analyst, researcher, graphics person, all working at the same time. I know, sounds too good. Let me show you. Research 10 cars and build a website. Let's see. Quick note. Swarm uses three credits. Single agent uses one. Use Swarm when you need the firepower. Watch. Kimmy creates a team. Each agent has a name and specialty. Cannon is graphics. Hemingway handles content. Dr. Lee is pricing. Zach optimizes efficiency. Alan is data science like watching a company spin-up. See all those progress bars? Each one is a different agent researching a different car. Lucid, Mercedes, EQS, BMW, i7, Rivian, all at the same time. Instead of car one, then car 2, then car 3, they're doing all 10 simultaneously, night and day. Then it starts building the website, final output, not a document. A fully working website, EVFinder pro, top 10 EVs under $200,000, filter by sedans, SUVs, luxury, range comparison chart, and look a total cost calculator. Select a vehicle, lease or buy. See the complete breakdown with taxes and incentives. One prompt, team of agents, interactive website. That's the swarm. Last one, and this genuinely

surprised me. Vision coding. Here's what it does. Record your screen. Upload to Kimmy. Kimmy watches and writes code to recreate what it saw. Think of it like this. You show Kimmy a website by scrolling through it. Kimmy watches figures out the layout, colors, structure, and builds you a replica. Let me demonstrate. I recorded Apple's website. Valentine's Day theme, floating hearts, iPhone 16, Apple Watch. Back to Kimmy. Upload the video. Keep it under 100 megabytes. The prompt. Simple. Five words. Select website agent. Run. Watch what happens. Kimmy says the model has watched the video. That fast. Now it's extracting frames like taking screenshots at different moments. Building a mental map of the layout. Look at these frames. It captured Valentine's banner, iPhone section, watch section. It's breaking down the entire page. Then it identifies each asset. A colorful Apple logo made of hearts. Three iPhone 16 Pro devices. Apple Watch Series 11. Now it's writing the actual code. Installing packages. Creating files. Done. Let me click preview. Look at that. Valentine's ribbon logo. It actually generated a new Apple logo with hearts. navigation bar, iPhone section with learn more and shop buttons, watch iPad, tradein, the whole grid layout, almost identical. And here's the real Apple site for comparison. See how close that is. Video in, working website out. That's vision coding. All right. Three things you can

do with Kimmy right now. Single agent when you need focused research, one prompt, detailed report done. Agent swarm when you need to scale. Multiple specialists, interactive outputs, night and day, vision coding to turn any screen recording into actual code. You saw what it did with Apple's website. Link to Kimmy in the description. Claudebot tutorial and Higsfield tutorial, those are there, too. I do this every week. Hit subscribe so you get notified when I drop the next one. Sound good? Let me know in the comments which one you're trying