Vexo

@qkl2058 month 1 was $800, month 12 was $100k — the system just needed time to compound

一个17岁的小伙子，手里跑着12个YouTube Shorts频道，月入超过10万美金。 他每天只干3个小时，剩下的全交给AI。 不用出镜，不用拍摄，不用自己剪。 12个频道，每个占一个细分领域。每个频道月产出8000到15000美金不等。 他的技术方案是这样的： Claude每周给每个频道出30个选题点子。 ElevenLabs负责配音，40秒搞定一条旁白。 CapCut自动把画面和声音拼好。 一个Python脚本，掐着流量高峰自动上传。 每天24条视频，覆盖12个频道，人工参与度为零。 第一个月，赚了0到800美金。 第三个月，4000到8000。 第十二个月，突破10万。 一台笔记本，三小时，剩下的事代理全包。 这事儿跟天赋没关系。差别在于有没有搭好系统。

@cryptowluha everyone want to build AI apps but nobody talking about selling compute, smart move

This Nvidia GPU farm generates $18,000/month from a spare room. No crypto mining. Just AI companies paying for compute. 10 GPUs running 24/7. The setup cost $120,000. It paid for itself in 7 months. Now it rents processing power by the hour and demand keeps growing. The part most people miss: AI isn't just creating software winners. It's creating infrastructure winners. The people who secured compute early are already collecting the upside. Follow if you want to understand where the real AI money is actually going. 👇

@RoundtableSpace dorm room replaced a whole vfx studio, what are studios even selling now

A STUDENT BUILT A REAL-TIME VFX SYSTEM FROM HIS DORM ROOM Using TouchDesigner and Claude, he created hand-tracked visual effects that once required expensive studio workflows

@hey_madni no code written, $31 drone, one afternoon — engineers gonna hate this

HE SPENT $31 ON A TOY AND BUILT WHAT ENGINEERS CHARGE $50K FOR plugged a cheap drone into a laptop described the control logic in plain english Opus 4.8 built the entire flight interface by end of day: - calibration system - live controls - browser cockpit flying a real drone in real time not one line of code written the crazy part isn't the drone it's that he did all of this with one Claude session in one afternoon the article below shows how to run 1,000 of those sessions simultaneously while you sleep full breakdown below

@leopardracer ts not about being smarter, its about having better tools — always been like this

AN 18-YEAR-OLD RAKHAT FROM BISHKEK JUST OUTBUILT 80 DEVELOPERS IN 2 HOURS not because he’s smarter, because he had a system nobody else in the room had gstack turns claude code from a smart autocomplete into a full dev pipeline where architecture, qa, design and shipping are all connected 108k devs already found this and most people reading right now haven’t free repo, three commands to install bookmark this and drop a like ↓

@0xfinkus mac mini in a bag is actually insane, why didnt i think of this earlier

This is Li Ming's entire office in a shoulder bag Not a $200/month cloud setup A Mac Mini. An iPad A dock and SSD Full workstation on a café table in Bali Most people pay $200/month for the privilege of working anywhere He bought the hardware once And never looked back Bookmark this before portable AI offices become normal

@gippp69 been paying $180/month for separate tools that do same thing, this hurts

THIS 19-YEAR-OLD GIRL OPENED THE HERMES SETUP WIZARD AND SHOWED HOW A $0 INSTALL CAN TURN A $180/MONTH AGENT STACK INTO ONE LOCAL WORKFLOW she is not showing another AI logo. she is showing the boring screen that decides the whole game: provider, auth method, Claude login, Anthropic API key, local model, and where the agent actually runs most people burn money in tiny pieces. $20 on Claude, $30-70 on API calls, $25-50 on automation tools, $99 on an agent wrapper, then another $10-20 every time the workflow breaks and needs a new tool Hermes changes the math. set it up once, connect the model once, run the same type of work through one agent, and let repeated tasks turn into saved skills instead of paying tokens to explain the same process again that is where the savings actually come from. not from “free AI”, but from not rebuilding the workflow 40 times. research reports, pricing scans, customer reviews, competitor maps, client audits, all moving through the same agent layer the setup screen looks boring until you understand what it means. whoever controls the provider, the API, the memory, and the skills folder controls the whole workflow this is why Hermes is interesting. not because it looks cool on a laptop, but because it turns AI from something you rent into something you slowly build on your own machine

@Bober_smart people donated real money to an AI girl lol we are so cooked

A 28-year-old guy from Japan decided to just make a joke, but in the end, he earned $3,258 in 4 hours With the help of Claude, he created a Chinese girl and launched the first stream for a prank Claude changed the appearance, background, voice, and even language in real time He launched the stream simultaneously on 3 platforms: YouTube, TikTok, and Twitch Men liked the girl so much that they started donating just 20 minutes into the stream In 4 hours, they donated $3,258 to him The biggest single donation was $400 The guide on how to create this has already been leaked into the public domain

@ai_kisuke claude + elevenlabs + capcut is literally a full team for free

海外のYouTube Shorts事例たけどこれエグい。 17歳が12個のYouTube Shortsチャンネルを運用して月1,500万円超えしてた。 しかも本人の作業時間は1日3時間。 残りはAIエージェントが回してて 12チャンネルで1日24本投稿。 流れはこんな感じ。 Claudeが各チャンネルごとに週30個のネタ出し ElevenLabsでナレーション生成 CapCutで動画を自動生成 Pythonで伸びやすい時間に投稿 1ヶ月目：0〜12万円くらい 3ヶ月目：60〜120万円くらい 12ヶ月目：1,500万円超え もちろん海外事例だし、 そのまま日本で再現できるかはわからないけど見方としてはめちゃくちゃ大事。 これって、動画を頑張って作ってるんじゃなくて、 動画が作られる流れを作ってるってところは参考になるね。 AIをいかに作業の流れに組み込むかだよな。

@mikenevermiss $2400 once and $14k every month? bro just tell me the clients part

THIS DEVELOPER STACKED 4 MAC MINIS ON HIS DESK AND TURNED THEM INTO A LOCAL AI CLUSTER THAT GENERATES $14,000/MONTH 4 Mac Minis. one desk. no cloud GPUs. he connected them using EXO, an open-source framework that lets multiple Apple Silicon devices work together as a single AI system. a single Mac Mini can’t run massive 70B+ AI models at full precision. but when four of them are linked together, their memory is pooled into one cluster, making it possible to run models that would normally require expensive cloud hardware. most teams rent GPUs and spend thousands every month. he spent about $2,400 on the machines once. now the cluster runs locally with only a small electricity bill. he started with one Mac Mini. added a second when he landed his first client. then a third. then a fourth. the stack sitting on his desk slowly became an AI business. clients pay for model inference. data stays local instead of being sent to third-party servers. and when he needs more capacity, he doesn’t rent more GPUs. he just adds another Mac Mini to the stack. 4 computers on a desk. one local AI cluster. a business running in the same room where he makes coffee.

@carverfomo bro the hands on screen was the only thing AI cant fake lol

A Chinese developer spent 3 days testing local AI hardware. Mac Mini 16G against RTX 4070Ti. He posted the results to Bilibili. His conclusion: agentic workflows need $2,800 of hardware minimum. It looked like real work. A Mac Mini on one side. An RTX 4070Ti on the other. Both heating up. Charts running side by side. He ended the video with one number: anyone serious about AI at home needs to spend at least $2,800. Bro pause at 0:20. Look at the Ollama window his hand is hovering over. Look at the timestamps on the terminal logs behind it. Those timestamps are not three days apart. They are 27 minutes apart. The 3 days were the agent. The 27 minutes were him. A Claude agent ran every benchmark in 27 minutes. Mac Mini test. 4070Ti test. Tool integration check. The agent wrote the script. The agent generated the timestamps for the screenshots. The agent generated the 3 days narrative because Chinese tech Bilibili rewards visible effort. Someone pulled the upload metadata. The video was uploaded 31 minutes after the first terminal opened. Someone else compared the screenshots to public Mac Mini benchmark scripts on GitHub. The agent had generated everything except the hand on screen. Six months ago a 14 year old in Shenzhen pushed an AI agent to GitHub. Judges said no real world application. 3,100 forks later. The dev had been one of them. He had wired the agent into his content pipeline the week he stopped getting consulting clients. He had been a Bilibili tech reviewer since 2019. His best video had taken three months to make. The agent now does the work he used to do in 27 minutes. He still appears on camera because his audience comes for the hands. He wanted to show his audience the truth about local AI hardware. He accidentally showed them the only thing on screen his agent could not generate.

@MarikWeb3 I did same thing 2 months ago, saving like $300 every month now

my feed is full of people escaping AI subscriptions right now. the bills they're dropping: $200 claude, $200 chatgpt, cursor + copilot on top - $5,500 a year. the swap is one $599 mac mini, $3/mo, forever. is one-time hardware really beating $5,500/year? for heavy users, yeah all framework below

@0xfinkus Local AI starts looking very different when a $599 Mac Mini can run models well enough for real work.

Daniel Chen racked 1,000 Mac Minis into an AI data center He was burning $14K/month on cloud compute So he rebuilt it locally 1,000× Mac Mini M4 Same workload class Lower power draw than a single Nvidia server $599 once vs $200/month forever Ollama + Claude Code via local endpoint Zero API cost Apple didn’t market it like this Developers did the math Bookmark this before local AI infra becomes normal

WEI LI SPENT HIS LAST $599 ON A MAC MINI INSTEAD OF FOOD. 90 DAYS LATER HE MADE $12,400 FROM A RENTED ROOM WITH NO DEGREE AND NO ENGLISH. wei worked 12 hours a day at a factory. every evening he came back to a room with one mattress and a desk and watched youtube videos about ai he had no english. no degree. no connections but he figured out that ollama runs on a mac mini with one terminal command. that claude code connects locally with zero subscriptions. that the entire stack costs $599 one time he put the mac mini next to his laptop on the desk. at night he built workflows in a dark interface while his neighbors slept month one — market research reports for small businesses. $200 each. the mac mini finished them in 25 minutes. total: $1,800 month two he automated the entire process. raised the price to $400. clients from taiwan and singapore paid without asking why. total: $4,400 month three he hired two friends from the factory. total: $6,200 the mac mini is still on the same desk. the cables are still a mess. the room is the same he quit the factory on a tuesday bookmark this before your comfort decides for you.

@Flandermaxx Curious — is the Llama model fine-tuned on engagement data, or does it score purely on virality signals like share/save ratio? The scoring layer is usually where local AI setups break down

$9,400 a month. That is what a 24 year old in Porto pulls from pages he never appears on. A year ago he paid a social manager €2,400 a month to find trends and four SaaS tools to track them. All of them told him what worked three days after it stopped working. So he built one agent. It scrapes TikTok and Instagram through Apify, scores the outliers on a local Llama model in Ollama, drops the winners on his own board. Nothing leaves the building. No seats. No tokens billed. Pause at 0:30. The screen reads: 3 verified outliers found. Runs again tonight at 3AM. Every 2 days, as scheduled. He is asleep when it works. It flags a creator doing 46x their follower count, hands him the hook, his editor fills a Kling template by noon. Costs $19 in power. Replaced a €2,400 salary and five subscriptions. His girlfriend asked who he texts at 3AM. He said nobody. The box was working. He stopped renting trends. He stopped renting tools. He owns the machine that finds the money.

@SantiTorAI It's not the biggest that wins, it's the one who owns." That line is going to age really well. Renting every token was never sustainable

UNA STARTUP ALINEÓ 1.000 MAC MINI EN UN SOLO CENTRO DE DATOS PARA DEJAR DE PAGAR ALQUILER EN LA NUBE UN DESARROLLADOR pagaba 14.000 DÓLARES al mes para alquilar servidores H100 en la nube para su startup de IA. La calidad era increíble pero ese ritmo de gasto estaba acabando con la vida de la empresa. Se sentó, hizo los números y compró 1.000 MAC MINI M4. Los instaló todos en una sola instalación. Esta configuración consume menos energía que un solo servidor Nvidia pero produce el mismo resultado. Los números: MAC MINI M4 a 599 DÓLARES pago único, consume 10-20 vatios funcionando 24 horas, 3 DÓLARES de electricidad al mes. Una máquina Windows que hace el mismo trabajo consume 30-50 dólares al mes solo por estar encendida. Ollama añadió soporte para la API de mensajes de Anthropic en ENERO DE 2026. Eso significa que Claude Code se conecta a tu Mac Mini local CON UNA SOLA variable de entorno, la misma interfaz, CERO coste de API. La lógica es esta: el alquiler en la nube NUNCA termina, el hardware se compra UNA SOLA VEZ. El que ponga el hierro en su mesa en 2026, estará muy por delante en 2028. GUÁRDALO. No gana el más grande, gana el que ES DUEÑO.

@0xKnzo Ran this exact setup for 2 months. The thing you don't mention: the psychological shift is bigger than the math. When an inference costs $0 you just… try more stuff. That's where the real ROI hides

@AnthropicAI The key word here is "accelerating" — not "capable of." Claude isn't designing the next Claude yet, but it's compressing the timeline for the humans who are. That's a different kind of risk that most people overlook

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor. It’s happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recu…

@igus_ai Does the student verification work with non-.edu emails in Europe, or do you need to go through a specific university portal? Asking for anyone outside the US

CURSOR está regalando 1 año de Pro GRATIS solo necesitas conseguir un correo de estudiante, tuyo o de alguien que conozcas lo que incluye: → 12 meses de Cursor Pro (valor $240) → acceso completo a GPT, Claude y Gemini → sin tarjeta tras la verificación cómo conseguirlo: 1. entra en cursor.com/students 2. regístrate con un email universitario 3. pasa la verificación de estudiante en algunos países es con un email .edu, pero cada país tiene su propio sistema de verificación estudiantil mientras todos pagan $20/mes, tú puedes programar proyectos enteros GRATIS durante un año aprovecha antes de que termine el plazo

@dr_cintas How does the recall quality hold up compared to FAISS at that 16x compression — is there a sweet spot where you get most of the memory savings without meaningful accuracy loss?

Google's new algorithm just shrunk 31GB of memory down to 4GB 🤯 TurboVec is a new open-source tool that stores the data your AI app searches through, using 16x less memory. It runs on Google's TurboQuant, which skips the slow setup step every other tool needs. → Faster search than the popular alternative (FAISS) → Works on both Mac and standard servers → Narrow results to exactly what you want → Plugs straight into LangChain and LlamaIndex Your data never leaves your machine. Runs fully offline, works with Python out of the box. 100% Open Source.