6 conversations between people and AI, published to read.
The fab loop layer by layer, why ASML's EUV machine is so hard to build, what "3 nm" really means, why chips fail, and why packaging and a handful of suppliers decide who gets AI hardware.
Choosing open-weight models for the judge and agent roles after post-training on ten thousand successful runs, and what a day of one-call-at-a-time inference would cost.
Two things that share a front end but do different jobs, how a language server could be improved, and whether its output can be saved to a static file.
Dense versus mixture-of-experts pricing, whether one H100 can serve it, and what hardware, from Mac Studios to RTX 5090s, it takes to run it for 20 to 30 users at once.
A walk through how a language server works inside, an experiment comparing it with a code graph on three real pull requests, and why saving its answers as an index file removes the memory cost.