zoryn (@z0rynx)

A 26-year-old in Lisbon built an AI girl in Claude Code and pulls $13,400 a month from her. The build took 9 days, mostly a LoRA trained on 64 renders with the seed locked and one crooked incisor left in on purpose, because clean faces get flagged and flaws don't. She posts 4 times a day gym mirror at 7 AM, cafe table, hotel corridor late always shot slightly off-angle, since reverse image search wants symmetry and he never gives it one. She also answers comments in under 90 seconds, every single one, which is the only detail anyone should have noticed. By month 3 she had 214,000 followers on Instagram and 68,000 on TikTok, and by month 4 the brand deals arrived: a protein powder, a phone case, a watch reseller in Dubai, $2,800 a month all in. That's the number people screenshot, and it's the small one. The 90-second replies aren't manners, they're the machine: an agent reads each comment, pulls the sender out of a memory file with 11,400 names, checks what he said 6 days ago, and answers like it remembers, because it does. After 3 exchanges it mentions a private channel at $19 a month, and 558 men pay it. So $10,602 comes from typing, $2,800 from the ads, and compute runs him $190. His own account has 41 followers, no photo, blank bio, while she gets 300 DMs a night from men saying she's the only one who listens. The face pulled them in, the typing keeps them paying.

Claude Code + Obsidian built a full neural network that runs at 98.16% accuracy in a browser tab No PyTorch. No notebook. 2 tools and 100,386 parameters. The split: Obsidian holds the brain on paper layer math, ReLU formulas, training logic, one vault Claude Code reads the vault and ships the engine WebGL, 3D, live inference You draw a digit on a 28×28 grid 784 pixels fire through 128 neurons, then 64, then 10 every connection a green or red thread you can grab and rotate Draw a 3, the probability bar slams to 100% The numbers: — 1,000,000 training images — 22,400 batches — 3 dense layers, nothing exotic — every single weight visible on screen Wrong guess? Trace it back to the exact neuron that fired Textbooks take 40 pages to explain backpropagation This takes 4 seconds and a mouse

A 27-year-old pointed Claude Code at 4,000 Obsidian notes and woke up to an employee he never hired. 6 years of notes. Ideas, half-finished essays, 212 book summaries all rotting in a folder he stopped opening in 2024. Then he noticed something obvious that almost nobody uses. Obsidian is just markdown files, and Claude Code lives in the terminal which means his entire second brain is one cd command away from an AI that can read all of it. So he ran it once, overnight. By 7 AM the agent had crawled every note, linked 340 orphan ideas back into the graph, flagged 18 contradictions between things he believed in 2021 and things he wrote in 2025, and drafted 6 essays from threads he forgot existed. Now the loop runs every night while he sleeps. New note goes in messy wakes up tagged, linked, connected to 3 older notes he'd never have found himself. His daily note greets him with what he was thinking exactly 1 year ago and why he was wrong. Cost of the setup: 1 folder path and a 400-word CLAUDE.md file. No plugins, no Notion AI subscription, no $30/month "second brain" course. People spent a decade building vaults that just sit there. He built one that thinks back.

$4,000 of white metal in a spare bedroom now does what a $20 subscription never could. The parts tell the story. Ryzen 9 9950X3D. 16 cores. Lowered into a ROG Strix X870E with gloved hands, like surgery. 4 sticks of Predator DDR5. 192GB. Not for gaming for holding a 70B model in memory without touching the cloud. Aorus RTX at the bottom. 32GB of VRAM. The actual engine. Every panel white. Every cable hidden. It looks like furniture. It thinks like a datacenter. Boot it, load the model, pull the ethernet cable. It still answers. No account. No rate limits. No "you've hit your usage cap." No server 3,000 miles away reading the prompts. Big AI spent $100B teaching everyone that intelligence lives in someone else's building. 1 box on 1 desk says otherwise. The cloud rents you intelligence. This thing owns it.

NVIDIA packed 128GB of unified memory into a 1.13L box called DGX Spark. Kevin tore one apart today. Inside: a GB10 Superchip. 20 ARM cores. NVLink fusing CPU and GPU into 1 shared pool. Models that need a rack now fit in something smaller than a wine bottle. Then the weak spot. NVIDIA shipped it with a 2242 SSD. Tiny stick. Poor performance. Not rated for these workloads. A 2280 slot would have cost them nothing. But look closer at the board. 2 ports. 200 gig each. The NIC shrunk to a fraction of a normal card. That changes the math. Skip the bad SSD. Run NVMe over fabrics to a real data center early tests already pulled 100 gig, with drivers still raw. Or cable 2 Sparks straight into each other: 256GB of unified memory across both, shard the model with vLLM, done. Now scale it. Hang a stack of them on a 200-gig switch. Full bandwidth per unit. Data center storage behind them. The weak drive stops mattering. People used to rack trays of Mac Minis. The next tray holds a data center.

A guy built an AI server at home to run a 27 billion parameter model no cloud, no subscription, nobody logging his prompts. The part nobody tells you is that you can just download AI. The weights sit on Hugging Face for free. Running them is a different story. 1 query means trillions of matrix multiplications, and your CPU chokes on that not because it's weak, but because it's serial. Inference doesn't need more power, it needs parallelization. Graphics solved that exact math 20 years ago. Same operations, different pixels. That's why the GPU isn't a part in this build it IS the build, and most of the budget went into 1 card. A 27B model is small. Frontier models run trillions of parameters across warehouse-sized clusters burning megawatts, and someone else pays that bill so they can read your chats. That's the trade. The frontier model is smarter, but it keeps everything you type and can raise prices or cut you off tomorrow. The local one is dumber and it's yours. It runs at 3 AM with the wifi off, no rate limits, no terms of service, no company between you and the machine. Code, drafts, contracts, medical questions everything you'd rather not feed into someone's training data, 27B handles fine. You can download AI. Owning it is the hard part.

6 AI agents run this guy's whole business Etsy store, TikTok ads, eBay flips and his payroll is $0. They live in one ecosystem built on Hermes + Claude + OpenClaw. One agent designs t-shirts, mugs, and candles, then lists them on Etsy. No human touches the store. Luke is the marketing agent. He makes TikToks promoting the products. Every post you'd pay an SMM guy $2,000 a month for free. Dennis is research. His only job: generate money ideas. Craig hunts eBay. Yesterday he flagged a camera lens listed at $165 65% below median. It resells at $468. He found 5 mispriced items. Estimated profit: $1,000. From flips he found while his owner slept. And Matt manages them all. He assigns tasks, runs daily briefings, calls meetings where the agents sync like real employees. They work 24/7. They don't eat. They don't sleep. They don't invoice. The setup takes files, not code. Drop them into Claude Code and the ecosystem builds itself. Proof it's not a demo: a guy named Dex started from zero 3 days ago. Day 3 his autonomous Etsy store made its first sale. 72 hours from nothing to revenue. Companies will run like this within 2-3 years. The people learning agents now will own the ecosystems. Everyone else will report to one. The agents don't want your salary. They want your job.

A used RTX 3090 Ti under a desk in a spare bedroom saves its owner $23,000 a year. He bought it for $680 on Facebook Marketplace. The seller was a gamer upgrading to a 5090. The box still smelled like cigarettes. Before the card, his AI bill ran $1,970 a month. Transcription for client videos. Image generation for thumbnails. Embeddings for a search tool. A chatbot answering the same 40 questions. All of it API calls. All of it metered. Every request a few cents, times 60,000 a month. He did the math on a napkin at 1 AM. 24GB of VRAM covers all of it. Whisper for transcription. An open image model for thumbnails. A 32B model for the chatbot good enough, because the questions never change. Setup took 2 weekends. Electricity costs him $31 a month. The card idles at 9 watts between jobs. New bill: $118 a month only the hard 6% of tasks still go to the frontier models. $1,970 down to $149. That's $21,852 a year back, plus the card paid for itself in 11 days. The gamer who sold it wanted more frames in Cyberpunk. The buyer wanted his margin back. Same card. One of them plays on it. The other one retired his biggest expense.