I write when I have something to say, not to keep a cadence. Every post is something I worked through — and had to explain to truly understand.
Writing since 2026
By the numbers
Researchers from UIUC, Stanford, NVIDIA and MIT built multi-agent systems that skip text entirely: agents exchange latent vectors through a tiny trained module, cut token usage by up to 75.6%, run up to 2.4× faster — and score higher. A full breakdown of the paper.
MCP, command-line interfaces, and Agent Skills solve different parts of the same problem. A research-grounded analysis of what each boundary contributes to LLM tool use, where context and tool-space interference creep in, and how to choose the right one.
A source-grounded analysis of DeepSeek V4 Preview: V4-Pro, V4-Flash, 1M-token context, hybrid sparse attention, KV-cache economics, API migration, benchmarks, limitations, and what developers should actually test.
Jul
May
Apr