three weeks in one: a lunch with lio that made me ask whether my optimising was worth doing at all, a stats midterm, and gym buddy shipping through it all
A research note on System One models, TypeSafe's Jev and Convai's open-weights Laya: what a typed-judgment primitive returns, what Laya's benchmark claims do and do not show, why fusing detection and generation into one LLM call forces the cheap job to inherit the expensive job's cadence, and the trigger rate at which splitting them stops paying.
a week packed enough to feel like the old ones, friends as the first VIP users who are too nice to roast it, and the one number that decides when this becomes an app: 100 daily actives
A research note on deciding whether to adopt a library: six advertised features turned into six greps against my own codebase, why the biggest storage surface was out of the library's reach by type signature rather than by quality, and when to take a wrapper's dependency instead of the wrapper.
ten years of beatboxing snuck up on me halfway through writing this, and the next wildcard gets terms i can be held to: ready two months out, recorded one month out, zero flaws
A research note on choosing a terminal git client: why the speed benchmark everyone quotes is worth 0.18 seconds at real repo sizes, why lazygit and tig are a mutator and a reader rather than rivals, and why the biggest win turned out to be a pager one layer down.
four days of diarrhea taking my strength away, catching my own plan lying about what i actually trained, and slowly admitting we are just souls living in bodies that are counting down
A research note on DeepSeek Harness: 272k lines of TypeScript in which the model adapter is 3 per cent of the code, what 'everything is a plugin' buys and costs, and four concrete ways a software engineer can use an open-source agent harness.