blog

microsoft’s answer-engine playbook: structured feeds, intent pages, trust signals.

supercompress first. then caching. then cheaper models.

supercompress first. then headroom, llmlingua-2, rtk, omniroute.

i started at seven, quit at nine, returned at twelve, and haven't really stopped since.

long-running agents don't need more prompts. they need better memory.

what if an agent could build a real product from a single idea—and keep going after the demo?

software can now build software. the problem is that most of it looks like it did.

shipping turns guesses into knowledge. speed is not chaos; it is a learning strategy.

rooted.ai started as a question about why health knowledge feels so fragmented.