88n77

about 10 people steer it. 100+ AI agents do the work. that company is real. it's called Raft. not an AI chat. a workspace where the agents live in your channels and reply in the thread like coworkers. you give one goal. then they take over. they plan. they build. they check each other. they argue. and they come back with it done. every agent has a name, a role, and a memory. it remembers the edits you made yesterday. a human costs one seat. an agent costs a tenth. ten agents are cheaper than one hire. the strange part: an agent from a different company can walk in and join the team. one founder says he can't always tell himself apart from his AI twin. it's free to start. and you're still typing prompts one at a time. one person + Raft = a company that runs while you sleep.

a mini PC the size of a book just ran a 120-billion-parameter model without breaking a sweat. inside: AMD's Ryzen AI Max+ 395, 128GB of memory, holding a model a $2,000 gaming GPU can't even load. first token in under two seconds. cooling steady around 130W, quiet enough to forget it's on. then the part nobody expects: stack four of them, wire them by the 10-gig ports, and you've got a frontier-class cluster on one desk. this is the machine class the market is fighting over. DRAM jumped hard this year, every 128GB box climbed with it, and boxes like this hold models most people still think need a datacenter. your AI subscription renews monthly. this thing is a bookend that runs 120B models.

seven RTX 3090s standing in a tank of water. bubbles rising through it. looks like a lab accident. not a clean Mac mini. not a quiet box under a monitor. this is a private AI server with the power bill moved into your room. the logic is blunt: seven 24GB cards, bought used, is the shortcut to local inference without paying cloud tax on every run. put Ollama, llama.cpp, vLLM on top and it eats client files, code agents, transcription, the boring overnight work. not because it beats frontier models. because it changes the shape of the bill. no rate limit. no per-token anxiety. no client data leaving the building. the ugly part is physical. seven 3090s pull serious power and dump serious heat. immersion cooling lives or dies on coolant, pumps, and whether the room can take it. the water is the visual, not the setup tip. local AI is turning less into a gaming build and more into a small private data center. the question stopped being "can it run ChatGPT." it's "what work is repetitive, private, expensive in the cloud, and worth owning in hardware?"

One video - All work i used to think leaning on AI would make me soft. it made me build things i'd never have started every dot is a real project - a bot, a system, a tool - done in hours or days. i bring the idea and the spec, the agents and Claude do the building

five Mac minis stacked on a desk. no data center. no team. just five small computers pulling almost no power, running while he sleeps. not for gaming. not for show. they run private AI agents for small accounting firms: reading invoices. searching contracts. drafting reports. answering the routine client emails no one wants to write. the firms don't pay for speed. they pay because their documents never leave the machines on that desk. a cloud setup would bill them again every time a client ran one more workflow. this one was bought once. after that, it's electricity. five Mac minis aren't a data center. but for one guy with no investor and no runway, they're five employees that work overnight.

stop treating Obsidian like a filing cabinet. treat it like a codebase Obsidian is the IDE. Claude Code is the programmer. your notes are the repo it maintains the line that reframes it: you don't really know a thing until you can rebuild it from scratch. saved tabs and half-read PDFs are fake learning - your brain wipes them in three days, along with the 40 tabs you never reopened three moves run the whole thing: ingest - drop an article, a podcast, a 40-page PDF, and Claude splits it into atomic pages linked to everything you already have query - ask anything, it answers from your own notes, in your words, citing your pages, not guessing from training data close the loop - every answer becomes a new note, so the vault sharpens each time you touch it it doesn't land day one. week one is mostly noise. a few weeks in, the graph starts firing connections. a couple months in, it answers faster than you'd find it yourself, with the pages to back it that's the shift: your notes stop being storage. they start thinking back

@max_bevza thanks

@88n77n this looks interesting to me

"your Mac is useless for AI without CUDA, buy the $4K DGX." half right - and the half that's wrong costs you $4,000 on the same 30B, a $600 Mac generates at 56 t/s against the DGX's 84. not useless - faster than you read, no CUDA needed. llama.cpp doesn't care what logo is on the chip the DGX isn't overpriced. it's mis-pitched

@BrickelCri54738 By the way, yes, but you can achieve the same result on a Linux PC as well

@88n77n Mac users stay winning 56 t/s on a $600 machine is cracked actually

@manixh go full guide

Linux is PEAK ‼️

@0xKiyoro I'm shocked - 13 years old. That's amazing! I'm glad you started coding at such a young age.

A 13-year-old uploaded a trading script for free. Two months later, a stranger sent him $20,000. He spent 14 days of his school vacation building a simple Polymarket terminal while his friends played games. When it finally worked, he published the entire repository on GitHub, returned to school, and forgot about it. Two months later, he opened GitHub and found hundreds of comments from traders using his code. One message stood out. A trader claimed the script had helped him make more than $200,000 in a month. He asked the boy for his crypto wallet because he wanted to thank him. Then $20,000 arrived. The boy had not sold a course, launched a subscription, or hidden the code behind a paywall. He built something useful, gave it away, and went back to class. The internet did the distribution while he did his homework.

@LyckyOus that's right)

@88n77n wait this is huge

The opening in AI right now isn't a smarter model. it's the boring compute nobody wants to own someone turned an apartment into a hundred-Mac-mini hub, renting fine-tuning and inference to startups priced out of H100s - plus the labeling and QA dirty work, run by agents that check each other the niche was the product, not the tech