Researcher builds local GPU server to run 24/7 AI agent system after Mac mini runs out of memory
nodosenlared · x · 2026-11-21
Researcher Miriam González's Polaris agent system, which runs 24/7 on a 16GB Mac mini that can't be upgraded, is moving to a custom local server. Health data privacy rules out the cloud, so bigger models and more parallel agents require more local hardware.
The planned build (per the shared diagrams): workstation CPU, 8x32GB DDR4 ECC RAM on 8 channels, 2-3 GPUs with 24GB+ VRAM each, 4TB Gen4 NVMe, a 1600-2200W Platinum PSU, 10GbE networking to a 3x8TB NAS plus UPS. She and a hardware veteran will document the whole process in a YouTube series, from Mac bottlenecks to real Polaris load tests.
More from coding & agent
- Dev's take: Gemini beats Claude on UI/UX code, Codex 'the absolute worst' — HankYeomans · 2026-11-21
- PanelCAD 0.5.29 ships: open-source KiCad-style CAD for industrial control panels — burhop · 2026-11-21
- Agent estimates 'a few days of agent work' for a 10-minute job, sparking skill idea — jamesbrooksco · 2026-11-21
- Karpathy shares 70-minute screen recording of how he actually uses AI daily — shivsingh · 2026-11-21
- Indie dev says TanStarter boilerplate earned him tens of thousands of dollars; Black Friday lifetime deal drops to $99 — yihui_indie · 2026-11-21
- Second AI agent review doesn't equal verification: define acceptance criteria first — Anyone_0478 · 2026-11-21