Hex Horizon

Sharing life changing AI setups daily Don't miss my next post, prepared smth mind blowing

A GUY IS RUNNING THE MOST POWERFUL CODING AI IN THE WORLD FROM A $400 E-INK TABLET ON A BEACH He's on a Remarkable 2. No fans, no GPU, no thermal problems in direct sunlight. The tablet SSHs into his laptop at home. Claude Code runs on Opus 4.6 through the API. The beach device is just a keyboard and a readable screen. Every other laptop overheats in that spot. E-ink doesn't. E-ink also stays readable when the sun hits it, which is when regular screens become useless. The compute lives somewhere with power and cooling. The interface lives wherever the human wants to sit. The two are connected by an SSH tunnel and nothing else. This is what "AI workstation" actually means now. Not a $4,000 MacBook. A cheap reading device pointed at a rented model. The hardware race for developers is over. The winning setup is the lightest, coolest, most sun-readable thing that can hold a terminal session.

@whemohere the best guide I saw this month ngl

@starmexxx I didn't believe when I first saw this now I'm a bit shocked

A $300 STACK OF EX-CRYPTO CARDS IS RUNNING 10 SPECIALIZED AI AGENTS IN PARALLEL AND KILLING $459/MONTH IN CLAUDE CODE AND CHATGPT PRO, THE HARDWARE SIDE OF LOOP ENGINEERING'S OUTER LOOP 02:26 the plan lands mid-video, "my plan is to do an ai council, one node might excel at coding, another at math, another at creative writing" each node is a $25 amd bc 250 with 16gb of gddr6 vram, 12 of them ship in one chassis for around $300 total, ubuntu server on each, ollama with a different specialized model per node loop engineering is the industry pattern this maps to, agents running in parallel with shared artifacts, one loop's output becoming another loop's trigger, the system compounds as more agents write to the same knowledge base the software layer needs infrastructure that runs multiple models in parallel without a cloud subscription, this rack does exactly that at personal scale, 10 specialized nodes for the price of a single month of chatgpt pro your $459 a month claude code and chatgpt pro buys one generic frontier model, this setup runs 10 specialized ones in parallel and lets them read each other's work through shared state bookmark this and read the article below

SOMEONE IS BUILDING A 2026 SECOND BRAIN AND IT LOOKS LIKE A LIVING GALAXY A 3D wireframe globe labeled "2026" with a glowing core spitting out hundreds of connected nodes, stars and clusters. Every dot is a note. Every line is a relation. This isn't Notion. This isn't Obsidian's flat graph. It's a knowledge engine that renders your thinking as a navigable universe. The point isn't the visual. The point is what sits underneath: an indexable network where retrieval stops being "search" and starts being "ask the graph". Note-taking apps sold storage. The next layer sells recall. When 10,000 notes become 10,000 connected entities, your second brain stops being a folder and starts behaving like a model trained on you. The people still organizing notes in nested folders in 2026 are doing manual labor for a problem that's been automated. Every serious operator I know is moving the same direction: write less, link more, let the system surface the pattern. The tool above is just the prettiest version of that bet. The graph wins. The folder dies.

My gf makes $400k / year selling books she never writes I found out last night, and we dated for 3 years back then she used to type them out herself on her laptop, $21 a book now she hands ChatGPT a topic and a page count and gets 90 pages back Claude designs the cover, she uploads it and it sells for $21 reading one of her books takes longer than making one she's shipped 27 so far, each one an afternoon of work then it just sells, every single day she started 10 months ago with a single prompt that's the whole idea behind the 35 Claude workflows I put together, set it up once and let it run while you sleep

@ardchain $65k is right enough to pay the bills and live the best life good one

THIS OPERATOR BUILT A MASTER DASHBOARD TO AUTOMATE HIS $65K/MONTH AGENCY Here is the exact architecture of his visual operating system: - A unified content pipeline for all social platforms. - Live Facebook ad tracking mapped directly to sales. - Automated SMS & Email follow-up triggers. - Centralized finance and customer admin. Instead of switching between 10 different apps and losing context, every department is integrated into a single bird's-eye view. The ultimate flex isn't just the revenue. It's having zero operations living inside his head, allowing him to scale with absolute clarity.

@phosphenq 9 yo really?

A 9-year-old built a real online store in 10 minutes. No code. No engineers. No budget. She just told an AI agent what she wanted. Her dad runs AI for BNP Paribas, a 180,000-person bank, and he didn't tell it as a cute story. "My nine-year-old became a new node in this new agentic ecosystem that is the economy." The work was never the bottleneck. Intent is. And no one runs half a million agents by hand, so they run in loops that govern each other. Watch it, then read how to build the loops behind it, below.

@0xSlyth classic lol

this repo probably should not be free it teaches you how AI builders are replacing prompts with autonomous loops while everyone is trying to write the perfect prompt they're building systems that: discover -> plan -> execute -> verify -> improve over and over again inside you will find > production ready loop patterns > memory & evaluation frameworks > tool calling workflows > self correcting agent templates > human in the loop safeguards > starter projects you can run today this is the shift from asking AI to building AI that asks itself bookmark this before everyone catches on repo link - github.com/cobusgreyling/…

ONE GUY TURNED 2 OBSIDIAN FOLDERS INTO A SECOND BRAIN THAT TALKS BACK CarrotProximity, FoxProximity, HungerLevel, ThreatLevel. The thing decides when a rabbit flees. That's the joke. People are watching a prey animal make survival decisions and realizing their notes app can do the same thing. The setup is two folders inside Obsidian. One holds raw notes. The other holds the index the model queries against. That's the entire architecture everyone is paying 30 dollars a month for in some SaaS wrapper. A second brain is not magic. It is weighted connections between inputs you already wrote down. The rabbit does not think. It pattern matches against threat and proximity. Your notes do the same when you stop treating them like a graveyard. The people who figured this out 6 months ago are now charging 2k per setup to consultants who still email PDFs to themselves. The tools have been free for a year. The bottleneck was never the model. It was whether you actually wrote anything worth indexing.

@kocer_eth makes sense lol

CLAUDE CODE + OBSIDIAN IN UNDER 1 MINUTE The trick is not “AI note taking.” It is giving Claude Code a place where your thinking already has structure. The video’s setup is simple: download Obsidian, put your notes in one vault, keep them as local .md files, then connect that vault to Claude Code through the Obsidian CLI. Now your notes stop being a pile of old thoughts. They become files Claude can read like a codebase. Project notes can connect to meeting notes. Ideas can connect to goals. Open loops can connect to old decisions. Then the useful part starts: custom commands. A /today command can pull calendar context, tasks, recent notes, and active projects into one prioritized plan for the day. An /emerge command can scan across months of notes and surface the ideas you keep circling but never clearly named. That is the actual workflow: Obsidian stores the thinking. Claude Code operates on the thinking. Commands turn it into repeatable output. The caveat is obvious but important: this only works if the vault has signal. Random clipped articles, messy dumps, and half-written fragments will produce messy answers. But if your vault has meetings, decisions, goals, tasks, project context, and recurring problems, Claude Code gets much more useful. Stop asking it isolated questions in a blank chat. Give it your thinking substrate. Then make it run commands on top of it.

@thegreatest_sv is it at least legal?

THIS GUY IS USING WI-FI TO DETECT PEOPLE THROUGH WALLS. No cameras. No sensors. Just tiny changes in Wi-Fi signals. It can detect movement. Human presence. Even breathing. Five years ago this would’ve sounded like science fiction. Today it’s a weekend project. The crazy part? You could build a simple demo around this idea and put it online for $0. That’s exactly why I wrote a guide showing how to launch a real website with a custom domain - completely free. Full setup below

@0x_rody Karpathy as a beast

Creator of Claude Code, Boris Cherny: "While you're prompting your agent step by step, mine writes itself, reviews itself, and picks its own next task." In 26 minutes he and Spotify Chief Architect reveal how AI already runs their entire codebase, and what comes after. Watch it, then save the step-by-step guide below 👇

3TB HOME AI RIG IS RUNNING MODELS PEOPLE STILL PAY OPENAI TO ACCESS Same box runs DeepSeek 671B, Kimi K2 at one trillion parameters, and five models loaded at once without choking. This is roughly $25k-30k in hardware doing work that costs $2k-5k a month in API calls or cloud GPU rentals, depending on usage. The real shift is Apple quietly turning unified memory into the cheapest path to local inference. No NVIDIA tax, no H100 waitlist, no data leaving the room. Frontier models stopped being a moat the moment a guy in his office could run them off a power strip. The cloud AI business was priced assuming individuals couldn't host this. That assumption is now wrong by a factor of ten. Compute is decentralizing faster than the pricing models can adjust.

@kocer_eth I need a more advanced guide I guess

THIS GUY BUILT A COMFYUI WORKFLOW THAT TURNS ONE AI GIRL INTO A CONTENT PRODUCT The video is not just “make a pretty AI girl.” It shows the real money part: repeatability. ComfyUI lets you lock the character, load the model stack, use a reference image, batch prompts, and generate many shots with the same face, style, and vibe. That is the difference between random AI art and an asset you can sell around. The simple path: Pick a niche first. Fitness girl, fashion girl, gaming girl, luxury travel girl, skincare girl, anime-inspired creator. Then build a consistent face and visual style in ComfyUI. Not 1 image. A bank of 100+ usable posts. Then monetize the output in one of 4 ways: 1. Sell content packs to small brands that need ad creatives. 2. Build a synthetic influencer page and test affiliate offers. 3. Create custom AI model/look packages for creators who want a character but do not want to learn nodes. 4. Sell the workflow, prompts, and setup as a productized service. The mistake is thinking the model is the business. The business is the boring system around it: consistent character, content calendar, offer, distribution, and proof that the visuals convert. ComfyUI gives more control than simple AI image apps because the workflow can be repeated, edited, and improved instead of praying for one lucky generation. But the caveat is big: an AI girl with no niche, no offer, and no audience is just another folder of renders. If you want to make money from this, do not start by generating “hot girl” images. Start by asking: who pays for this visual attention? what product can she sell? what content can be posted every day? what makes the character recognizable after 3 seconds? That is where the money is.

@antisadh yeah and those money is just enough crazy

A $40 BC250 BOARD WITH 16GB GDDR6 GETS A GITHUB FIRMWARE FLASH, THE DEFAULT 8/8 CPU/GPU SPLIT REWRITES TO 0.5/15.5 AND OPENS 15.5GB OF UNIFIED MEMORY TO OLLAMA, DEAD CRYPTO HARDWARE JUST DOUBLED ITS AI CEILING 01:08 the operator points at the chart, "you only give 512 megs to the GPU part and that's reserved, that means all the rest of the RAM is available to either the CPU or the GPU" a BC250 ships from the factory with 16GB GDDR6 split 8/8 between Oberon CPU cores and the RDNA 2 iGPU, the iGPU only ever uses 4-5GB during inference, the other 3-4GB sits stranded SEC Bolt's firmware mod on the moth-enjoyer GitHub flashes a 0.5/15.5 split, the GPU reserves 512MB and 15.5GB falls into a unified pool that ollama treats as VRAM, qwen 3.6 14B at 4 bit fits with 2GB context headroom 15.5GB on a $40 board with a $40 CH347 flasher beats the $249 Jetson Orin Nano's 8GB by nearly 2x at 1/6th the price, the firmware is open source, the dump backup process takes 4 minutes per board your map's tier zero floor was the OptiPlex at $35-50 with iGPU only, the modded BC250 sits at the same $40 mark with 15.5GB of GDDR6 and an actual RDNA 2 GPU, the buyer who flashes once unlocks a 13B class local AI host for the price of a dinner the window is open, follow and bookmark before it closes

🚨THIS UPDATE KILLED CLAUDE GLM-5.2 by @Zai_org hits the same ~74% score as Opus 4.8 While burning HALF the tokens... Same brains, half the bill. Why would you ever pay frontier prices again? Full research attached below.

@51bodila when you say it this way, it just feels different but I agree mostly here anyway

Fidelity Vice Chairman told a room of investors: "if you can't explain to a 10-year-old why you own a stock - you shouldn't own it " - then gave them the system he used to turn $18 million into $14 billion "I skipped every tech stock on Wall Street - bought Dunkin' Donuts - made 10 times my money - because I could understand donuts " he averaged 29.2% a year for 13 years and no one has beaten that record, then retired at 46 - this is how he thinks for free "a fireman put $2,000 a year into one stock, no computer, no MBA - just an observant neighbor - became a millionaire " Warren Buffett called his house and asked to use his quote $1,000 invested with him turned into $28,000 - one out of every 100 Americans trusted him with their money bookmark & watch today ↓

ONE GUY MAPPED AN ENTIRE CITY'S ECONOMY INTO A 3D PARTICLE SIM IN 7 DAYS He took Bozeman's economic data and turned it into a spherical point cloud you can fly through. Zoom in, zoom out, render it as volumetric smoke, watch the motion trails in yellow and blue. One week. One monitor. One person narrating it like it's a weekend project. This is what Karpathy was pointing at. The bottleneck was never the data. It was the interface. Spreadsheets and dashboards compress 4D information into flat rows because that's all the tools allowed. Now a solo operator runs Houdini-grade simulations on a consumer GPU and treats regional GDP like a VFX shot. That used to be a 6-person studio quoting 80k for a 30-second render. Every "second brain" app from the last 5 years is still text in boxes. Notion, Obsidian, Roam. Linear notes about non-linear systems. Meanwhile this guy is flying a camera through the actual shape of a local economy. The note-taking era peaked. Spatial data is the next default.

🚨THE LOOP IS NOW INSIDE THE MODEL GLM-5.2 by @Zai_org doesn't stack 100 layers. It runs ONE block again and again. A loop, baked into the wiring. A small model thinking in circles to punch like a giant. Same trick the whole frontier quietly runs on. Full research below.