Notes

Technical notes and working thoughts by Albert Chang.

• Three DGX Sparks in a ring, and the two places NVIDIA's guide fights your LAN — A practical field guide to adapting NVIDIA's DGX Spark ring setup to a real office network
• I read the system prompt before I read the benchmarks — One prompt across ten models, a glimpse of my own memory in GPT-4.5's output, and a later look at Gemini Storybook
• Kimi K2 starts answering, then the message vanishes — Answers that began streaming, then disappeared—and why I suspected an output filter
• Model Selection When Everything Changes Weekly — Internal leaderboards beat model fandom
• Distilling omni-moderation into ShieldGemma 2B on one RTX 4090 — LoRA on a 24 GB card, an uneven mix of training data, and a moderation model that flagged a person's name as violent