Alex LongDistance @CodeJourneyAJ

A MIKROTIK CRS804 CAN TURN FOUR GX10 BOXES INTO ONE LOCAL AI CLUSTER, IF PCIe 4.0 STORAGE DOES NOT BECOME THE BOTTLENECK For an operator running larger local models, that changes the constraint. The goal is no longer one workstation fitting one job. It is keeping several jobs moving on owned capacity. A MikroTik CRS804 links the four DGX Spark/GX10 nodes. NVMe-oF over RoCE keeps shared models and datasets close. Then the small 1 TB PCIe 4.0 drive becomes the weak link. It can read around 7, 8 GB/s, but the cluster still has to load weights, swap layers, and serve more than one job. The Corsair MP700 Micro PCIe 5.0 upgrade is not cosmetic. More pooled memory only matters when storage can keep the stack fed. A local AI cluster is not just about fitting a larger model. It is about turning parallel work into capacity you control. Bookmark this before your next cloud bill. Follow for local AI builds that actually change the math.

kocer @kocer_eth

3 weeks ago

x.com/i/article/2071…

4 5 61 496K 61

THE REAL KIMI K3 PANIC IS NOT “CHINA CAUGHT UP” it is that frontier intelligence might stop being a 90% margin subscription business. Kimi K3 is a 2.8T parameter, 1M context, multimodal model from Moonshot that is now close enough to GPT-5.6 Sol and Fable 5 for people to ask the uncomfortable question: why are we paying closed-lab prices for every task? the economics are the whole story: > closed models = OpenAI and Anthropic control access, pricing, quotas, APIs, usage rules > open-weight models = enterprises can self-host, fine-tune, switch providers, and avoid platform lock-in > Kimi K3 pricing = about $3 input / $15 output per 1M tokens > GPT-5.6 Sol pricing = around $5 input / $30 output per 1M tokens > but Kimi may use more tokens per task, so the real metric is cost per completed task the winner is not “China.” the winner is every layer above and below the model: > startups get cheaper intelligence > inference providers get more demand > Nvidia sells more chips > data centers use more compute > software builders get less platform risk > closed labs lose pricing power that is why Washington is suddenly talking about Chinese open models. a real ban is hard. weights can be downloaded, mirrored, quantized, fine-tuned, and run privately. so the likely move is softer: make companies afraid to use Chinese models through procurement rules, audits, disclosures, national security warnings, and regulatory FUD. the risk is real: cyber capabilities, removed guardrails, distillation attacks, foreign dependency. but the counter-risk is bigger: if you kill open-weight AI, you hand the entire intelligence layer to 2-3 closed labs. bookmark this before “AI safety” becomes the excuse for protecting AI margins 👇

s1rozha1 @s1rozha_

a month ago

x.com/i/article/2066…

6 2 51 135K 79

American Open-Source Labs Think They Can Beat China’s Best AI Startups forbes.com/sites/iainmart… #digitaltransformation #ai

UPDATE: #Cardano $ADA Founder Charles Hoskinson says Input Output Group is "aggressively pushing for decentralization of open-source development, and having many companies, not just a few, work on the core of Cardano. The Haskell node is now the product of many companies."

SOMEONE FROM THE ANTHROPIC TEAM LEAKED HIS OBSIDIAN VAULT STRUCTURE. ONE CLAUDE.MD FILE AT THE ROOT CHANGED WHAT 474 DAYS OF NOTES COULD DO. He did not rebuild his workflow from scratch. The funniest part? Nothing looked dramatic from the outside. He did not write new notes. Did not reorganize folders. Did not tag everything by hand. He imported what already existed and created one CLAUDE.md file at the root of the vault. That was it. 474 days of accumulated work sitting in Obsidian. Meeting notes, research threads, half-finished decisions, project context that never made it into any ticket or doc. All of it already there. The graph started as a small cluster. A few dozen nodes loosely connected. Then clusters branched. Research linked to decisions. Decisions linked to old projects. Old projects surfaced context he had completely forgotten writing. Want to know what was inside? Active work. Archived research. Dead-end experiments that mattered later. Decisions without outcomes. Outcomes without the reasoning behind them. Isolated pages became connected context. Notes in separate folders started finding each other across time. Then he went deeper. Context layers for the current project. A research cluster. A decisions thread. One folder Claude could walk through before answering anything. That is the actual mechanism. Claude opens the relevant context layer instead of starting from zero. Instead of searching an undifferentiated pile of notes. Every other note app is where notes sleep. Obsidian is where they talk to each other. He turned 474 days of forgotten work into an external memory Claude can actually walk through. The full graph is in the video. Bookmark this before your own notes become invisible. Follow for more Claude setups worth studying.

kocer @kocer_eth

4 weeks ago

x.com/i/article/2069…

18 22 124 463K 271

Most founders have no idea how ChatGPT talks about their brand — or that it's recommending competitors first. Magister runs your prompts across ChatGPT, Claude, Gemini, and Perplexity in fresh sessions, scores your share of voice vs competitors model by model, and shows exactly which pages to publish to win the answer. Not a score. A playbook.

The Self-Driving Company Is AI’s Next Business Model. forbes.com/sites/johnsvio… #ai #digitaltransformation

AN AI SECOND BRAIN THAT NEVER SLEEPS. THE SAME ARCHITECTURE ALREADY BILLS $9,000/MONTH. Pause at 0:10. That's the setup — two folders, raw sources and wiki, built on Karpathy's own prompt for turning notes into a self-organizing knowledge base. Drop in articles, book notes, meeting transcripts, podcasts. Ask a question. Claude answers with citations and saves the gap it just found back into the wiki — the system gets smarter every time you use it. A weekly health check scans for contradictions and outdated notes automatically. This is the free, personal version of what six companies are already paying $1,500/month for. Same two folders. Same Claude-reads-markdown trick. The only difference is whose knowledge lives inside it. $0 — this setup, today. $2,200 — what a business pays to build the same thing. $1,500/month — what it bills once it's live. $11,200/month — six clients in. One person built this to remember their own research. Someone else is charging companies to remember theirs.

Superior @andreysuperior

2 weeks ago

x.com/i/article/2056…

8 12 74 142K 255

Serial entrepreneur Phil Hollingdale has founded, scaled and exited 6 companies. He sold everything, put his family home on the line, slept in his office for weeks at a time. He’s employed thousands of people and paid millions £ in tax. Increase CGT rates again and he’s leaving the UK. There are thousands of Phils across the country.

$69 is now enough to put a real x86 home server on your desk. ZimaBlade keeps the parts that make a small server useful: > Replaceable DDR3L RAM up to 16GB > Native PCIe expansion > Two SATA ports > Gigabit Ethernet > USB-C power > CasaOS, ZimaOS, Proxmox, Ubuntu and Debian support The board-only price does not include RAM or the power adapter, but once assembled it can run Docker apps, backups, Home Assistant, a personal cloud, media services and lightweight AI agent gateways from one fanless box. The useful part is not running the largest model at home. It is owning the always-on layer around your models, files and automations. A home server used to look like a rack. Now it looks like a $69 board.

RetroChainer @RetroChainer

4 days ago

x.com/i/article/2076…

11 12 98 516K 246

HUMANOID ROBOTS ARE AMAZING. BUT WHERE IS THE ROI? Watching a humanoid robot work at BMW is genuinely impressive. As an engineer, I love it. But I also can’t ignore one question. For this task, why not use a 6-axis robot on a linear axis? It would likely be: * Faster. * More precise. * More stable. * Much cheaper. So why the humanoid? Because BMW isn’t optimizing this task. They’re training Physical AI. Every successful grasp, every failure, every unexpected situation becomes data that teaches robots to handle the real world. That’s the real investment. The demo is impressive. The business case? I’m not convinced… yet. Are we looking at the future of manufacturing—or the world’s most expensive AI training program? Best regards Ulrich #HumanoidRobots #PhysicalAI #Robotics #Automation #Manufacturing #Engineering #AI #IndustrialAutomation

A 20-YEAR-OLD MICROSOFT AI ENGINEER SAYS MOST PEOPLE ARE LEARNING AI THE WRONG WAY Instead of jumping between every new model and viral tool she recommends spending months mastering Python machine learning fundamentals PyTorch and real projects because every advanced AI skill is built on top of those foundations She argues that too many beginners rush straight into prompting LLMs and AI agents without understanding how these systems actually work making it much harder to build products when the technology inevitably changes again The engineers who stay valuable are rarely the ones chasing every weekly release because they understand the principles that remain useful long after today's trending models disappear Strong foundations are still the biggest advantage in AI

Insomnia @insomnia_vip

a week ago

x.com/i/article/2076…

28 13 133 122K 186

SOMEONE JUST OPEN-SOURCED A FREE ALTERNATIVE TO CLAUDE DESIGN WITH NO LIMITS u can now build UI prototypes and animated layouts without hitting paywalls > works with any AI - your ChatGPT subscription, APIs, or local models running offline > generates animated websites, mobile app prototypes, slides, and motion graphics > full interactive canvas lets you prompt iteratively to refine the exact layout Most platforms force you to buy another expensive subscription to use their visual canvas. This lets you plug in the models you already pay for (or run for free) repository is free on GitHub: github(.com)/nexu-io/open-design Once you have your layouts, here is how to turn them into a live, animated site in Claude Code in 2 hours Full walkthrough in the article below

monokern @monokern

4 weeks ago

x.com/i/article/2070…

60 287 3K 2.9M 12K

Ask a single AI model to evaluate a hard decision and you get one averaged, cautious view that smooths over exactly the tension you actually needed surfaced! The fix isn't a smarter model. It's several experts, each locked into a hard bias, money, technical risk, the user, arguing independently and never seeing each other's conclusions until a separate merge step reconciles them. Force each expert to push its angle to the extreme instead of trying to be balanced, and you get a real spectrum instead of five identical cautious opinions. This article shows you the exact build you need. Bookmark so you do not lose it! Follow @neil_xbt for more content like this!

h100envy @h100envy

a week ago

x.com/i/article/2077…

19 58 450 984K 1K

HE BUILT A 7-AGENT AI COMPANY WITH ITS OWN MISSION CONTROL DASHBOARD. Real-time revenue, net profit, YTD numbers, savings, ad spend, all live on one screen at the top. Below that, an actual team. Cortana runs operations 24/7 and watches the inbox. Jarvis handles client work. Boss manages the content calendar. Aura runs marketing and growth. Cash does the financial analysis. Forge and Titan handle product. Every agent sits idle until there's work, then springs into action. The Operations Board shows what's queued, in progress, waiting on approval, or done. The Neural Network view maps how agents actually talk to each other in real time. And the War Room streams every single action live as it happens. To prove it's real, not just a mockup, he pings Jarvis for a status check. Instantly it flips to Active, a live connection line draws straight to Cortana on the map, and it shows up in the activity feed in real time. This isn't a chatbot. It's a running business dashboard for a team that never clocks out.

LUMENX @Lumenix0

6 days ago

x.com/i/article/2077…

4 2 30 36K 16

One prompt produced a playable Mario clone with three original levels. In ten days the model behind it becomes a public download. A gaming channel ran Kimi K3 through Kimi Code in YOLO mode, which auto-approves safe actions. Every game started as a single prompt asking for Three.js and 60 FPS. Star Jump 64 came back with wall jumps, red-coin puzzles, a retro graphics toggle, and a snow level with a working slide. The reviewer called it the most impressive single-prompt build he had seen from any model and scored it 9 out of 10. Black Ops II returned as an arena shooter with a sniper tower, a tomahawk you can throw and pick back up, and a UAV. His estimate for the whole build was under five dollars in API cost. Natural Disaster Survival became an island cycling acid rain, lightning, tornado, earthquake, flash flood, and meteor shower. K3 initially reached for the Godot engine and had to be redirected to a web build. The article below covers the machinery. 2.8 trillion parameters with 16 of 896 experts active per token, a 1M-token window, and cached input at $0.30 per million with hit rates above 90% in long coding sessions. On Arena's independent front-end tests K3 finished ahead of Fable 5 and GPT-5.6 Sol. Reasoning is fixed at max, so one trivial SVG burned 13,241 thinking tokens. On July 27 the weights land under a Modified MIT license, the first open release at this scale. What stays with me is the Godot detour. K3 had its own plan for the stack until the reviewer pointed it at the browser.

obssnnn @0xObssnnn

5 days ago

x.com/i/article/2078…

12 7 77 429K 158

Systems thinkers will crush everyone else in this AI era. As the cost of intelligence keeps dropping, the bottleneck stops being raw capability and moves to how well you can architect everything around it. Which model orchestrates the task, which models do the workhorse tokens, how data flows across the workflow, where the evals sit, how permissions and security get handled across functions. None of that is a single-point problem. It's all systems design. The people who only think about one layer - the model, the prompt, the app - will get commoditized as the layers shift underneath them. The people who understand how the full stack fits together (frontier intelligence as the manager, cheaper or tuned models for volume, enterprise context and data feeding the whole thing, evals to know when something broke or improved) are the ones who will compound. This is true at every level. Inside enterprises, the teams that can model workflows end-to-end will drive the actual transformation. In the applied AI layer, the harness and the routing logic become the differentiation. Even individual careers now hinge on whether you can reason about how work moves through a system, not just how to do one step of it. AI keeps getting better at the individual tasks. The durable advantage is understanding how the pieces connect. Net net: intelligence is becoming abundant, but systems thinking is still scarce. That gap is where all the leverage is.

2 years ago I was just a guy who decided to try indie hacking Moved to Barcelona, shipping days and nights Woke up at 6am for Spanish class, came home and shipped till 11pm Startup #1 - failed Startup #2 - failed This went on for 8 months On month 8, I built @mediafa_st Simple question: how do I help others avoid Reddit bans after getting suspended 6 times myself? First month: $363 Now: $7k/month and hundreds of users The point... Anyone. Anywhere. With an idea and some willpower can make a small dent in the universe

David Siegel And The Open Source Movement forbes.com/sites/johnwern… #ai #digitaltransformation

🛑 IF YOUR AUTOMATION STRATEGY RELIES ON BIPEDAL ROBOTS, YOU'RE BURNING CASH. We’ve been brainwashed by pop culture. When we hear "robot," we instantly picture two legs, two arms, and a glowing head. But let’s be brutally honest about the physics of a factory floor. I recently sat down with a group of automation experts, and we dissected a common client request: “We need a humanoid that walks.” When I asked “Why the legs?”, the room went quiet. The reality? 80% of industrial environments don’t have stairs. They have smooth, flat surfaces. So why are we forcing robots to learn how to balance, wasting precious battery life just to stand still, and risking catastrophic tipping accidents? The Twist: The smartest players aren't building terminators. They are putting advanced, dexterous torsos and arms on wheeled bases. The advantages aren't marginal—they're exponential: ⚖️ Stability – Zero risk of tipping over, even with a 40kg payload. 🔋 Energy Efficiency – Standing still costs nothing. No constant servo-jitter to maintain balance. 🚀 Speed – Wheels are significantly faster and more precise on the production line. 🛡️ Safety – Drastically fewer points of failure mean drastically lower maintenance costs. We don’t need humanoid bodies. We need humanoid capabilities—reach, grip, and intelligence—packaged in the most efficient chassis possible. Here is my Call to Action: I want to settle this debate right now. Drop a W (Wheels) or L (Legs) in the comments. Tell me: If you were scaling a logistics operation tomorrow, which would you bet your budget on? And if you are currently evaluating automation for your own floor, comment "SMARTER" below. I will personally DM you my 5-point checklist to help you choose the right morphology for your specific ROI—no sci-fi, just pure engineering. Let’s hear it. Are we building for Hollywood or for the shop floor? Best regards, Ulrich #Automation #Robotics #humanoid #Engineering #Innovation