Mark Poiler (@mpoilerfx)

A writer took a full AI library off-grid: DeepSeek R1, Qwen3 30B, Kimi, coder models, embedding models, all running from one machine on a patio table. No connection required. No subscription. No tokens billed. The video scrolls his ollama list like a bookshelf: deepseek-r1, qwen3:30b-a3b, kimi thinking variants, coder models, embedding models, some pulled 2 days ago, some 2 weeks. Dozens of gigabytes of intelligence sitting on a local drive. The wins: The screen shows what he uses it for. Verse in the style of Rumi about machines learning to dance. A short story about an intern named Alex wandering a computer history museum past vacuum tubes and a humming PDP. Every word generated on the spot, outdoors, by models he owns. The pain: A library this size eats storage and memory, and memory is the cruelest market of 2026, DRAM up 90% in Q1. The heaviest frontier reasoning still lives in the cloud he left behind. He writes with it anyway, in daylight, no wall of GPUs in sight. One machine, a chair outside, and every model on the menu. The muse used to bill monthly. His sits on the hard drive.

His AI earned him 8 hours of work last night for $0. He slept through all of it. The video is 17 seconds: one rig glowing red in a dark room, overlay reading "the rig never sleeps." A 26-year-old builder runs it next to a Claude Max subscription. Friends pay $200 a month and log off at midnight. His machine keeps going. Drafts done by 2am. Agents finished by 4. Fine-tuning wrapped before his alarm. The setup splits work in a way most people get backwards, and the split is the part he saves for the full breakdown. What he does show: the same rig costs more to copy every month. DRAM prices jumped 90% in Q1 and keep climbing. Last frame, dark room, one machine glowing: while you slept, I built. Your subscription reset this morning. His rig never stopped.