A week of my own Claude Code usage, 2.5 billion tokens, broken down by what it would cost at list price. Output was under half a percent of the tokens; cache writes were nearly half the bill. What that means for terse-reply tricks, plugins installed on hype, and the instruction files that load into every session.

A news clustering job that compares every pair of article embeddings fell from 428 s to about 8 s at 30,000 articles with a hand-written WebAssembly SIMD kernel on every core, and the join decisions stayed identical. What failed first, what worked, and how exactness was kept.

A PostgreSQL query that runs in 11 ms spent 119 ms end to end, 108 of them compiling it with LLVM JIT. A reproducible experiment on PostgreSQL 16, 17 and 18: where the cost thresholds sit, what each tier costs, and how to switch it off at the right scope.

Scheduled jobs in launchd plists had no visibility, Jenkins fixed that and ate gigabytes of memory, and GitHub Actions ran out of minutes. So I built kipp: one Go binary, jobs as YAML, a DAG of stages, multibranch builds, and about 50 MB resident.

Under Bun, wrangler's createTestHarness() listens fine and then every request hangs with no error. The cause is Bun's fetch() ignoring the undici dispatcher Miniflare routes requests through - and wrangler 4.151.3 now fails fast with a message that says so.

Cloudflare cut 1.1.1.1's DNS cache from 953 to 420 bytes per entry with five Rust layout changes, freeing about 100 TB of RAM. A walkthrough of each change and the arithmetic that made it pay.