Yurash

Your AI lives in someone else’s basement. Mine lives in mine. 🇺🇦

[Yurash @Yurash1104]

ORICO dock for Mac mini — the numbers sell themselves:

→ 9-in-1 expansion: USB-A, USB-C, HDMI, LAN
→ 10Gbps on the expansion port
→ Cold storage: Toshiba N300 NAS HDD
→ Hot storage: ORICO J-10 NVMe SSD
→ Two speed tiers, one enclosure

Your Mac mini was underbuilt. Not anymore

[Yurash @Yurash1104]

Built this machine from the ground up — component by component.

→ ASUS PRIME RTX 5060 Ti 16GB — enough VRAM for local models
→ MSI MAG board — stable power delivery under sustained inference
→ XPG RAM + Phanteks PSU — zero throttling under load
→ Custom water-cooling loop — silent even at full tilt

This isn’t a gaming rig. It’s my private AI station that will never send me a token bill.

[Yurash @Yurash1104]

@Chulee120 Yep. Gaming gives your loop breaks. Local inference doesn’t know what mercy is 😅

[Yurash @Yurash1104]

People think custom loops are about FPS. They’re not.
→ It’s about stable temps under sustained inference, not a 5-minute benchmark
→ An LLM on GPU runs hotter and longer than any game ever will
→ If your loop can’t hold temps for an hour under load, it’s decoration, not a system

[Yurash @Yurash1104]

CLOUDS LOOK NICE. YOU JUST DON’T OWN THEM.

This build is white on purpose — looks like a cloud. The point is the opposite.

→ ASUS ROG Strix X870E-A, custom loop, white tubing throughout
→ Every fitting, every degree of cooling — mine
→ No provider can revoke access, raise the price, or kill the server

Cloud is a lease. You pay monthly for something that’s never yours.
Obsidian is what’s left when the lease ends. Solid. Permanently yours.

This isn’t about looks. It’s about a choice: rent intelligence, or own it.

[Yurash @Yurash1104]

People spend $200/month on cloud GPU instances. I spent that once — on water cooling.

→ No monthly bill
→ No “GPU queue” at 3am
→ No dependency on someone else’s uptime
→ Just a white loop that looks better than any dashboard

Owning your hardware isn’t a flex. It’s the cheapest AI subscription you’ll ever buy.

[Gipp 🦅 @gippp69]

https://t.co/oMgggyvOz6

[Yurash @Yurash1104]

CLOUDS LOOK NICE. YOU JUST DON’T OWN THEM.

This build is white on purpose — looks like a cloud. The point is the opposite.

→ ASUS ROG Strix X870E-A, custom loop, white tubing throughout
→ Every fitting, every degree of cooling — mine
→ No provider can revoke access, raise the price, or kill the server

Cloud is a lease. You pay monthly for something that’s never yours.
Obsidian is what’s left when the lease ends. Solid. Permanently yours.

This isn’t about looks. It’s about a choice: rent intelligence, or own it.

[Spike 1% @SpikeCalls]

https://t.co/iZG4HDFeOx

[Yurash @Yurash1104]

→ That “goth girl” isn’t a photo — it’s a render: face-consistency LoRA + body generation tuned for “real Instagram” aesthetic
→ The scheme: generate hundreds of variations → sell subscription to “private” content → platform takes 20%, you keep the rest until the ban wave hits
→ Claude’s role isn’t “strategy” — it’s caption/hook generation selling the illusion of a real girl
→ The actual asset isn’t the account (ban = zero) — it’s the pipeline: face dataset, LoRA weights, generation workflow. Portable to any platform in an hour
→ RTX 3090 runs this locally in seconds — paying for cloud Midjourney/Runway makes zero sense here

Хочеш — додам відео до цього поста чи зробимо окремий tweet-thread з розбором economics?

[Nekt0 @Nekt_0]

https://t.co/o3mM2voxzD

[Yurash @Yurash1104]

MOTION EDITORS ARE COOKED

Connected Motion to Claude via MCP in 60 seconds:

→ Settings → Connectors → Add custom connector
→ Named it “Motion”, pasted https://t.co/I7SrpHiV4F
→ Connect → authorize
→ One prompt → ready-made promo video in 5 minutes

What used to be a full day in After Effects is now a prompt.

The tool isn’t replacing the editor. It’s removing the reason to pay one for what a single request now does.

[Claude @claudeai]

Introducing Claude Design by Anthropic Labs: make prototypes, slides, and one-pagers by talking to Claude.

Powered by Claude Opus 4.7, our most capable vision model. Available in research preview on the Pro, Max, Team, and Enterprise plans, rolling out throughout the day.

[Yurash @Yurash1104]

The industry trained us to hide compute. A gray box in a server closet, a rented GPU in someone else’s cloud, zero physical connection to hardware you supposedly “own.”

I say no.

→ the white custom loop isn’t flexing — it’s function: lower temps, quieter operation
→ the ASUS ROG Strix X870E-A runs multi-GPU without throttling, and still looks like a design object
→ when your AI server sits in plain sight, you maintain it, upgrade it, actually understand every part
→ the cloud will never give you that sense of ownership

This isn’t “hardware in a closet.” It’s a tool you’re proud of — the same way a mechanic is proud of his garage, not a rented shop.

Would you show this off or hide it in a closet? Drop your take below.

[rmen @Roman9078963816]

https://t.co/FDnLoRkIGe

[Yurash @Yurash1104]

IF YOUR AI SERVER COULD LOOK LIKE THIS — WOULD YOU PUT IT ON YOUR DESK OR HIDE IT IN A CLOSET?

→ Local inference no longer means “a box under the table”
→ Homelab aesthetics are finally catching up to gaming culture

If your local AI rig looked like this — desk or closet? Homelab aesthetics are finally catching up to gaming builds.

[Gipp 🦅 @gippp69]

https://t.co/oMgggyvOz6

[Yurash @Yurash1104]

Smart home is dying. A whole generation bought Philips Hue and fifteen companion apps — and now it just looks embarrassing.

Ambient computing is replacing it: tech disappears into the architecture, stops blinking, stops demanding attention.

The EXACT same shift is happening with local AI — half of tech twitter just hasn’t noticed yet.

→ 2023: “check out my ChatGPT screenshot”
→ 2024: “here’s my Ollama setup, look at these tokens/sec”
→ 2026: agent runs in the background for weeks, event-triggered, doesn’t ask for a report — you forget it’s even there

The uncomfortable truth: if your self-hosted AI still lives in a browser tab and you open the chat every day — you’re not in 2026. You’re in 2023 with a better GPU.

Real maturity is when you only think about your agent because it broke something or saved you three hours.

Genuine question: how many of your local models are STILL living only in a chat window — never once touching a real workflow without you typing the command?

[Sprytix @Sprytixl]

THIS DEVELOPER BUILT A 4-NODE AI CLUSTER RACK IN HIS HOME LAB - AND RUNS LARGE LANGUAGE MODELS FOR $11/MONTH IN ELECTRICITY

4 Framework Desktop nodes in 2U trays inside an 8U mini rack, 5 gigabit ethernet switch from Nick Giga, smart outlet connected to Home Assistant - full energy monitoring in real time

Ansible automates the entire cluster - Beowulf AI cluster launches on any hardware from AMD chips to Raspberry Pi - one command and all 4 nodes are synced

Home Assistant dashboard shows how much energy every AI model consumes in real time - he knows the exact cost of every token

cloud AI costs $140-340/month and sends every request to someone else's servers - his cluster costs $11 in electricity and sends nothing anywhere

not as powerful as Mac Studio M3 Ultra for large models - but entirely his, entirely under control and entirely private

$11/month in electricity, zero subscriptions, and an AI cluster monitored from his phone

[Yurash @Yurash1104]

How many GPUs does it take to never pay OpenAI again?

Here’s the answer, in metal:
→ Multi-GPU rig, custom water cooling loop
→ Dead silent even under full load — no fan whine
→ Each card is its own worker for local inference
→ Separate RAID array for raw data, decoupled from compute
→ Disk speed test shows 450+ MB/s write, 480+ MB/s read — no bottlenecks

Cloud bills you per token. This hardware bills you once — and runs for years.

Break-even is months, not years. After that, everything you run is free.

[Gipp 🦅 @gippp69]

https://t.co/oMgggyvOz6

[Yurash @Yurash1104]

Your ChatGPT knows more about your life than your girlfriend does. And OpenAI gets paid for it, not you.

→ Every prompt is data going to someone else’s server, not yours
→ I built a white rig on the ROG Strix X870E-A for exactly this reason — I don’t want to rent my own brain
→ Custom loop, silence, full control — and zero “we’ve updated our privacy policy” emails

People are scared AI will replace them. Meanwhile AI companies are reading their conversations for free.

[Yurash @Yurash1104]

RTX 3090 VS CLOUD: THE REAL MONTHLY NUMBERS

Scenario: 4 hours of inference daily, 70B model

Cloud (A100 rental, $1.5/hr) → ~$180/month
RTX 3090 (owned, after upfront cost) → ~$15-20/month in electricity

→ The difference isn’t 2x. It’s 10x
→ Initial investment of $600-700 pays for itself in 4-5 months
→ After that — practically free inference for years

Cloud math works for occasional use. It stops working the moment you actually rely on the model daily.

[Gipp 🦅 @gippp69]

https://t.co/oMgggyvOz6

[Yurash @Yurash1104]

@Degen_calls_sol Cool 👌

[Yurash @Yurash1104]

KARPATHY KILLED HIS OWN PRODUCT WITH ONE SENTENCE

Spent weeks building MenuGen: photo of a menu → photo of every dish. Real code, real backend, Vercel.

In December he gave Gemini the same photo: “use Nanobanana to put the dishes on the menu.” Got back the same menu — dishes baked straight into the pixels. No app. No backend.

→ Software 1.0 was code
→ Software 2.0 was weights
→ Software 3.0 is just talking to the model

The same model that erases apps fails at the basics:
→ refactor 100,000 lines, find vulnerabilities — easy
→ walk or drive to a car wash 50 meters away — tells you to walk

Karpathy’s theory: models get good where the answer can be verified. Code — yes. Common sense — no.

“You can outsource your thinking. You can’t outsource your understanding.”

[Spike 1% @SpikeCalls]

https://t.co/iZG4HDFeOx

[Yurash @Yurash1104]

ROUTE THE WORK, NOT THE BILL

Hybrid agent architecture: the cloud thinks, the hardware counts.

→ Claude — orchestrator. Planning, complex decisions, tool calls, anything needing real reasoning.
→ Kimi K2.6 — local worker. 32B active parameters (out of ~1T total, MoE), running in your own rack, handling bulk repetitive inference.

The logic is simple: not every token deserves a $15/million price tag. Ninety percent of agent work isn’t reasoning — it’s execution. Boilerplate generation, document parsing, classification, rigidly formatted output. That doesn’t need the smartest model on earth. It needs a model that costs you nothing per token because it’s already sitting in your rack.

[Yurash @Yurash1104]

White build, custom loop, local inference.
→ 3 years without throttling, without noise, without cloud bills
→ The question isn’t “why build this” — it’s “why isn’t everyone”

Owning your compute isn’t a hobby. It’s refusing to rent what you could own.

[Superior @andreysuperior]

https://t.co/vjjTFz8DdO