Dima T.

@RoundtableSpace This is wild. One video. AI watches it and boom, whole scene in 3D. You don't know where camera is, where objects are, everything in space. And it's free on GitHub. That's money on the table.

CHINA JUST OPEN-SOURCED A MODEL THAT REBUILDS ENTIRE REAL-WORLD SCENES IN 3D FROM A SINGLE VIDEO AT 20 FPS.

@Vettan0 F#ck !)) This breaks my brain. A year ago: "this model only lives in the cloud." Now: guy ran it on his desk. Works like a server. Commands from Mac like it's the cloud. But it's just his desk.

A 397B parameter model is running on somebody's desk right now. Not a datacenter. A desk. Qwen shipped their new 397B flagship and the usual take was instant: nobody runs this at home. One creator took that personally. He chained 4 DGX Sparks into a single cluster. 128GB of memory each, linked with the kind of networking you normally see in server rooms. The clip is 76 seconds. Watch the memory readouts around the halfway point. All boxes pinned at 112 of 119GB. The model barely fits, and it fits. Then he opens a terminal on his Mac and benchmarks the whole thing like it's a cloud API. Same commands, same endpoint, except the endpoint is 3 feet away. I paused twice on the monitoring screen just to count the memory bars. A year ago this class of model lived behind an API key and a waitlist. The datacenter lost its monopoly. It just doesn't know yet.

@0xKnzo This is peak. Guy built GPU on Raspberry Pi. Walks home, turns it on, off. Pays for electricity like a kettle. Anthropic charges $50k/quarter for what this guy does once for $480.

CHINESE DEV JUST WIRED A FULL-SIZE GPU TO A RASPBERRY PI AND MADE $190K LAST YEAR WHILE HIS EX-COFOUNDERS BURNED SERIES A ON ANTHROPIC BILLS Silver Raspberry Pi on a floral tablecloth. Full-size GPU balanced on top. Blue Ethernet cable. Corrugated cardboard base. Build cost $480. Power $18/mo. Four clients pay $3,950 each per month. Pause at 0:04 — the rig sits on a floral tablecloth. Anthropic just billed his ex-team another $50K. Anthropic bills $50K/quarter for enterprise API. This rig cost $480 and runs 8B models offline. Same output. No signup form. Every "AI infra" pitch deck this month is one delivery box and a soldering iron away from being replaced. Full setup in the video.

@v_nefodov The people who move first will own the future🔥

@DimasikUSDT Yes, we need to adapt to it in time

DRONE DELIVERY: HOW TRADITIONAL LOGISTICS IS DYING RIGHT BEFORE YOUR EYES 90-second delivery? This isn't utopia — this is 2025 reality, and it's about to bury everything you knew about shipping MAIN TEXT: Remember when "30 minutes or it's free" sounded like a revolution? Funny. That's museum history now. 90 SECONDS. That's the new standard. While your regular courier is still remembering where he parked his car, a drone has already landed on your balcony with your hot pizza. THE NUMBERS THAT ARE DESTROYING THE INDUSTRY: Speed: Drone delivery: 90 seconds - 5 minutes to your doorstep Traditional delivery: 20-40 minutes (if you're lucky with traffic) 70% of delivery time is wasted on the road in megacities. Drones fly over all that. Money (this is what really matters): Cost reduction for delivery: 65-72% decrease Average drone delivery cost: $0.50-2 per order Traditional delivery: $5-15 per order Savings on driver salaries: 3.2 million workers in the logistics sector are facing retraining by 2030 Scale: Drone delivery market: $1.5B (2024) → $8.3B (2030) — 450% growth 47 cities already using commercial drone delivery In Dubai, drones deliver 35,000+ packages monthly Google Wing serves 100,000+ orders daily in the US alone 92% of young users willing to pay for drone delivery (vs 67% for traditional) Ecology: One drone replaces 4-5 vehicles on a single route Route CO2 reduction: 80-90% thanks to electric motors By 2030, drones will prevent 2.1 million tons of CO2 emissions WHY TRADITIONAL DELIVERY IS ALREADY DEAD: Problem #1: Traffic kills profits In Moscow, New York, London — average courier speed: 12-15 km/h Drones cruise at 60-80 km/h in a straight line Result: one courier delivers 10-15 packages daily, one drone handles 50-70 orders Problem #2: Complex routes = losses Island delivery? Expensive, requires ferries, hours of time High-rise building on the 50th floor? Courier sweats, broken elevator, 20 minutes wasted Rural areas? Unprofitable Drones solve this in minutes, ignoring terrain and obstacles Problem #3: The human factor Package loss rate: 2-3% of all deliveries Damage during delivery: 5-7% Drones? Controlled system with GPS, cameras, insurance tracking REAL EXAMPLES ALREADY WORKING: 🔹 Amazon Prime Air — already delivering in Texas, California, Arizona. Plans: cover 50% of the US by 2026 🔹 Google Wing — serving Australia, Finland, France. Increased volume by 320% in one year 🔹 Flytrex — delivering in New York (!) directly to residential areas. Made 50,000+ flights in 18 months 🔹 Dubai — government project: drones deliver medicine, documents, food. 99.7% successful delivery rate PROBLEMS TRADITIONAL DELIVERY CAN'T SOLVE ANYMORE: ✗ Expensive labor (salaries + taxes + benefits) ✗ Fuel costs (depends on oil prices) ✗ Environment (society demands green solutions) ✗ Scalability (every city needs new couriers) ✗ 24/7 operations (people sleep, drones don't) ✗ No route logic (human error kills efficiency) Drones solve all of this. WHO SHOULD PANIC RIGHT NOW: Courier services — will lose 40% of routes within 5 years Logistics companies — adapt or disappear Delivery drivers — need retraining immediately Parking lots and gas stations — unnecessary for urban delivery Traditional taxi services — next in line after couriers And who's winning? Tech giants (Amazon, Google, Apple) Well-funded startups Cities ready for change Customers SUPPORTING STATISTICS THAT SEAL THE DEAL: Efficiency: Drone operational cost per delivery: $0.40-1.50 Van delivery operational cost: $8-12 Delivery success rate (drones): 99.2% Delivery success rate (traditional): 94-96% Market Adoption: 68% of urban consumers prefer drone delivery if available Cities planning drone infrastructure: 156 globally Expected job creation: 2.4 million new roles in drone maintenance, dispatching, regulation by 2030 Initial investment by major companies in drone tech: $15+ billion (2024-2025 alone) Speed Comparison: Last-mile delivery (traditional): 15-40 minutes Last-mile delivery (drone): 2-8 minutes Peak hour efficiency gain: 300-400% This isn't a revolution. It's the quiet death of the old delivery world. If your service isn't investing in drones RIGHT NOW — in 3-4 years you simply won't exist on the market. Competitors will show 90-second delivery times, and customers will just forget you ever existed. 2025 is the year when the side dish becomes the main course. The old guard didn't see it coming because they were too busy fighting each other. Meanwhile, the future flew right over their heads. 🚁 Welcome to the future. It arrives at 70 km/h. HASHTAGS:#DroneDelivery #DroneRevolution #Logistics2025 #FutureTech #TransportRevolution #AutonomousDelivery #Innovation #Technology #LastMileDelivery #SmartCities #LogisticsTech #DeliveryDisruption

@v_nefodov Real moment. You checked Street View before a trip. A guy made that by just walking around in a backpack for $400. And there's the whole street in 3D. He got $400, Google got millions of users. He doesn't even know you used it.

YOU HAVE USED THIS MASKED GUY'S WORK A HUNDRED TIMES AND NEVER KNEW A HUMAN MADE IT. HE GETS OVER 400$ A WALK. Every time you drop into street view to check what a place looks like before you go, someone walked that exact path holding a camera over their head. On this street it is him. The case on his back says Apple Maps. He is a human mapping unit. Look at the rig. A pole with a ring of lenses at the top shooting 360 degrees, a laser scanner reading distance, GPS locking every frame to a real point on the map. He walks and the whole street gets rebuilt in 3D behind him. No car can do the narrow parts. So a person does. That person gets paid. Here is what nobody watching him realizes. He is not working hard. He is walking. The backpack does the thinking. His only job is to move at a steady pace and cover the route Apple assigns. A finished route pays around 400 and the gear is not even his. He shows up, straps in, walks a city, hands it back. Look at the label, not the face covering. Apple. Maps. That is the reveal. The weird man everyone is avoiding on the sidewalk is quietly building the map on the phone in their pocket. He walks a street for a living and the street ends up in everyone's phone.

@v_nefodov You don't even need to write anything. Just a photo. AI sees everything. Where you are, who's with you, what you're looking at. This isn't theory anymore, it works right now.

SOMEONE UPLOADED ONE PHOTO WITH NO LOCATION TAG AND AN AI FOUND THE EXACT STADIUM AND STREET IN UNDER A MINUTE. He runs a single screenshot through an AI geolocation tool. No GPS data attached. No caption. Just a guy at a game with a blurry crowd behind him. The model scans the whole frame at once. The shape of the stands. The letters on the field. The mountains on the horizon. The light and the reflections your eye skips right over. Then it searches the entire planet for a match. A few seconds later it lands on Salt Lake City at 100 percent confidence and narrows down to a street. It pulls the real stadium up next to the photo so you can check it with your own eyes. Same stands. Same field. Same skyline behind it. It even spits out a PDF report of the whole thing. Here is the part most people have not caught up to yet. You do not need to tag your location for a photo to reveal it. The background already does that job. A stadium, a storefront, a ridgeline, one recognizable tree line is enough to place you within a few blocks. Pause at 0:30 and look at the split screen. The photo on the left

@v_nefodov Whether you're ready or not, the future has already arrived.

@DimasikUSDT That’s brilliant, though it does look a bit creepy, like they’re calling for help from an infected area.

@0xkerazcity I just got it. You don't need the physical world to train anymore. Simulation + AI = result. Robot never fell once because it made all mistakes in the computer.

We are officially bypassing the physical universe to train the next generation of labor 🤯💻 Historically, teaching a bipedal robot to walk, stabilize, and recover from impacts took years of grueling, real-world hardware trials. Now? Look at this simulation loop. UK startup Humanoid AI took their Alpha Bipedal hardware from raw structural assembly to completely stable, independent physical locomotion in exactly 48 hours. How? They compressed 19 months of real-world physics training into 2 days by running massive reinforcement learning pipelines inside parallelized NVIDIA synthetic simulation engines. The physical world is becoming a mere afterthought for AI deployment velocity. See the data matrix mapping this rapid evolution

@Flandermaxx Bro that's another level. He didn't write an app. Took actual physical hardware (printer), plugged in Llama, now prints houses. No software, just "machine, print what I describe." $540k/month.

A 32 YEAR OLD AUSTIN DEV PIPED LLAMA 3 INTO A CONCRETE PRINTER AND NOW PULLS $540,000 A MONTH SELLING AI-GENERATED HOUSES TO TEXAS BUILDERS tyler is 32, oak hill garage stuffed with PERI concrete printer parts he bought used from a bankrupt california construction startup for $84,000, ex-solidworks engineer at trilogy PERI COBOD BOD2 printer · custom llama 3.3 70B fine-tuned on 12,000 texas building code PDFs and 40,000 CAD floor plans · text-to-Gcode pipeline he wrote in a weekend on a mac studio in the garage pause at 0:20 on the nozzle laying concrete ribbons, that is a 3-bedroom house being extruded from a paragraph of client requirements 3 texas home builders pay him $180,000 per house · 8-day build time · 3 completed houses a month since june, $540,000 MRR $84,000 for the used printer, $340 a month in texas power, first house paid the printer back, no architect fees no permit consultant, the design pipeline never leaves the garage while the texas housing code still allows automated permits, follow and bookmark

@0xKiyoro Bro that's not video production anymore. You just write "FIFA final, night, stadium" and Claude generates the finished clip. No camera, no team. One guy doing what used to need a studio.

Claude + World Cup just became a one-tab YouTube studio worth $6K a month No camera, editing team or second AI platform open in another tab. Connect Higgsfield to Claude through MCP and the same model that writes the prompt can now generate the finished football clip inside the chat. The setup: > Open Claude and go to Settings → Connectors > Add mcp.higgsfield.ai/mcp > Claude receives Generate Image and Generate Video tools > Describe the World Cup scene and tell it to run the generation “FIFA World Cup Final, packed stadium under bright floodlights.” Claude expands that sentence into a cinematic prompt, sends it to Higgsfield and returns a realistic six-second clip generated at Kling 3.0-level. One person can turn matches, players and viral narratives into original YouTube Shorts while the attention is still moving, building a World Cup channel generating $2,000 to $6,000 a month. The important shift is not the video quality. Claude no longer tells you how to make the content. It ships the content itself.

@chewadot This is hilarious. NVIDIA sold the cards, got paid. Then watches entire factories disassemble them and remake what NVIDIA can't sell cheap. And can't do anything about it 😂

CHINESE FACTORIES ARE BUYING RTX 5090s IN BULK, DESOLDERING THE CHIPS, AND REBUILDING THEM AS 128GB AI SERVER CARDS WITH 6X THE BANDWIDTH OF A DGX SPARK. THE WHOLE BLACKWELL LINEUP - 5090, 5080, 5070 Ti, 5060 Ti - NOW SHIPS FROM WORKSHOPS THAT DON'T OFFICIALLY EXIST $4,156 for a 32GB blower 5090. $573 for a 5060 Ti with a rack cooler. 0 official partners workers pull the GB202 and the GDDR7 off a gaming PCB. Reflow both onto a custom board with side power connectors and a blower shroud. Test. Pack. Ship to AI farms automated arms handle most of it. The high-end frankencards go further - a 128GB modded 5090 sells for $13,200, still cheaper per gigabyte than an RTX 6000 Pro NVIDIA got paid for every one of these cards at full price. Then watched its entire consumer lineup fork into a parallel supply chain it doesn't control the whole 50-series. Every SKU no partner board. no warranty. no support line. no way to put this back in the box the most interesting AI hardware of 2026 isn't being announced at keynotes. It's being reflowed in a workshop you'll never visit

@0xKnzo The moment you realize cloud is just marketing. Hardware isn't expensive. Electricity isn't expensive. Only expensive if you pay a middleman.

CHINESE DEV JUST STACKED 4 INTEL ARC PRO B70s INTO A WORKSTATION AND MADE $410K LAST YEAR WHILE HIS EX-COFOUNDERS BURNED SERIES A ON ANTHROPIC BILLS Four blue Intel Arc PRO boxes on a wooden bench. Tan Noctua cooler behind. Server rack in the back. Build cost $5,200. Power $95/mo. Five clients pay $6,800 each per month. Pause at 0:04 — four blue Intel fans spin up. His ex-team's Anthropic autopay just cleared for the ninth month. Anthropic bills enterprises $80K a month for GPT-4-class API access. Intel dumps Arc PRO B70s at $800 with 24GB VRAM each. Break-even in 12 days. The moment Intel started dumping Arc PRO cards for $800, every AI SaaS pitch became a rounding error. Full setup in the video.

@Tserawho Wait, I just realized organizing folders is just theater. You spend 40 hours making everything pretty but search it anyway. Just let Claude organize it, he'll figure it out.

ARCHITECTING FOLDER HIERARCHIES IS A WASTE OF TIME. Most people spend 40+ hours per quarter manually tagging notes that will never be actionable. You are treating your second brain like a digital landfill, not a processing engine. The shift: From manual P.A.R.A. curation to autonomous agentic workflows. The stack: -> Obsidian -> Claude Code -> MCP Servers (Gemini Vision + custom tools) The technical breakdown: 1/ Initialize the vault with pre-configured CLAUDE.md files. 2/ Map your 3-year vision directly to folder-level execution. 3/ Use Claude Code to automate daily reviews and goal alignment. This turns your note-taking from a digital junk drawer into a recursive feedback loop. Stop hoarding information. Start building a system that executes against your goals autonomously. Build your stack before the market saturates.

Watch what happened: A year ago: everyone learning "how to write good prompts" Now: Hugging Face, Microsoft, DeepLearning - all teaching "how to build agents" This isn't just a course update. This is a SIGNAL. When Hugging Face drops free courses - it means "this is no longer exotic, it's baseline."People learning THIS RIGHT NOW will earn 3x more than "prompt experts" in a year.History is flipping in front of you.

TOP 5 FREE AGENTIC COUDE COURSES: 1. Hugging Face – AI Agents Course huggingface.co/learn/agents-c… 2. DeepLearning.AI – AI Agents in LangGraph deeplearning.ai/courses/ai-age… 3. DeepLearning.AI – Multi AI Agent Systems with CrewAI deeplearning.ai/short-courses/… 4. Microsoft Learn – AI Agents for Beginners microsoft.github.io/AI-For-Beginne… 5. DeepLearning.AI – Building Code Agents with Hugging Face smolagents deeplearning.ai/courses/buildi…

This is it. Anthropic themselves just came out and said: "forget GPT-5, forget best models. Winners will have the best LOOP." 40 minutes. Three agents. No prompt engineers. No "perfect prompt." Just architecture: plan → build → judge → repeat. App works. Not "works well." WORKS. Fully. This means everyone still thinking "I need Claude Pro" instead of "I need a system that fixes itself" - they're 2 years behind.

Anthropic engineers just showed how they build a full app from scratch, using a loop of agents. 40 minutes from the team behind Claude Code. They used three agents: one to plan, one to build, one to judge, cycling until the app actually works. The winners won't have the smartest model, they'll have the best loop. Watch it, then read the full guide on how to actually use loops below.

@0xAurexx AI doesn't create leverage by writing code. It creates leverage by changing what you charge for.

michael from california made $9,700 last month with a $599 mac mini. the server rack he's gutting in this video is what it replaced. for years michael did freelance dev the "serious" way. a rack of secondhand enterprise servers in the garage. dual xeons, 256gb of ram, a fan wall you could hear from the kitchen. california electricity did the math for him: ~$120 a month just to keep the rack humming. plus hourly clients who paid for his time, not his output. then he ran one test. $599 mac mini m4 on a shelf. claude code on top. $60/month in tools. total stack: less than one month of the rack's power bill per year. claude code took over the repetitive engineering - boilerplate, crud, tests, refactors. the mini hosts every client app and automation, sipping 10 watts where the rack drank 500. and he changed what he sells. not hours. finished saas builds at $5,000 flat. the math: month 1: one build, $3,500. underpriced it. took 5 weeks. month 3: $5,000 per build, two weeks each. claude code carries ~70% of the code. month 6: two builds plus hosting retainers running on the mini. $9,700. honest slope: $9,700 was his best month. normal runs $5-7k. still 3x what hourly ever paid him. the catch nobody posts: claude code saves him 20+ hours a week - but it doesn't close clients. those 20 hours go straight into sales calls and scoping. the bottleneck was never the engineering. the rack sold for parts on ebay. the mini sits on a shelf next to the router. silent. printing.

@beamnxw One agent gives answers. Multiple agents give confidence.

This paper explains why I stopped using single-agent verification loops in my builds: Self-reflection creates confirmation bias → Single models suffer from Degeneration-of-Thought → Wire up agents to debate "tit for tat" → Use an independent judge to filter signal from noise Here's the problem with almost every self-improving prompt right now: we treat LLMs like humans who can objectively catch their own mistakes In reality? Once an agent outputs a wrong answer with high confidence, it hits a cognitive wall called Degeneration-of-Thought. It basically blinds itself to alternative solutions. You can tell it to "think again" 10 times, and it'll just find ten new ways to justify the original mistake The Multi-Agent Debate framework solves this by turning prompt orchestration into a structured cross-examination: > independent generation: multiple agents tackle the same prompt separately without seeing each other's initial bias > "tit for tat" cross-examination: agents read conflicting answers and explicitly attack the logical flaws in the other side's reasoning > controlled divergence: the debate forces models out of their comfort zones, generating novel solutions that a single agent literally cannot access alone > the judge's blind spot: they proved that using one LLM family as a judge introduces favoritism toward its own "sibling" outputs → a critical detail most frameworks completely miss Building reliable systems is about designing architectures where agents keep each other honest through structured friction Check out the full paper quoted below. If you're designing agentic workflows, this will change how you approach verification Read this, then check the article below

@0xkerazcity The biggest opportunities always appear before the crowd notices them.

A 20-year-old is already making money from GTA 6. The game hasn't even launched yet. While millions are waiting to play, a small group is already building products they'll eventually sell to those players. Rockstar quietly opened an official marketplace for paid mods, and it's still early enough that almost nobody is competing there. Claude can turn plain English into working Lua, making game development faster and dramatically cheaper than it was a year ago. By the time GTA 6 reaches tens of millions of players, the best creators could already have customers, reviews, and recurring revenue. Most people will play GTA 6. A much smaller group will build businesses around everyone else playing it.