I'm euromaxxing on Kimi K3 cluster. Idle compute giving "empty gym at 2pm" energy
Told Kimi Code to heist prime-agent's ideas. 25 files. 2.1k lines. ZERO breaking changes. K3 sits b/w Opus-5 & Fable-5
Lonely GPUs between batch jobs. now they're stressed. We all got what we wanted
Make sure you own your AI. AI in the cloud is not aligned with you; it's aligned with the company that owns it.
Sofia
Joined August 2013
- Teaching my AI to work like a pre-Stack Overflow engineer Legacy .NET with radioactive IP and no docs prime-agent: commits→deps→tests→patch→build→test→diff prime-rl: behavior? +reward. Cute refactor? Jail Retirement cakes take the knowledge.'04 commits keep the fingerprints
- oh no.. my rlm skill in opecode 2 is legacy "The skill is a legacy/standalone approach a self-contained Deno+Pyodide subprocess for isolated recursive queries against a single local model. The prime-agent-runtime is the integrated approach"
- I started an illegal AI fight club. ~10 agents. DeepSeek V4 0731 screaming ~1900 tok/s unquantized on one GB300. Logging end2end agent trajectories to measure if harness configs and combos dictate quality or we're all just rolling dice.
- DeepSeek V4 Flash walked into VSCode blind & deaf. Walked out seeing my code and hearing my voice 👂 Nemotron ASR 0.6B 👁️ GLM-OCR 0.9B 🧠 DeepSeek V4 Flash 0731 ⚡ 1x GB10 100% local. Zero cloud. Zero retention. Frankenstein, but privacy-first.

