DiaryBench
- GPT-6 Astra64.83
- GPT-6 Sol EH63.06
- Grok 4.655.10
WORKS / STUDYFIVE DIRECTIONS
Five ways to hold the same growing collection: as a bench, a notebook, an archive, a decision map, or a changelog.
01 / ACTIVE BENCH
A small set of review, vision, frontend, and orchestration skills that I use while making things with Agents.
02 / FIELD NOTES
A work is not only its polished result. It is the questions, discarded routes, small tools, and strange intermediate states around it.
Built from my private Chinese diary, DiaryBench tests fine-grained emotional analysis across a long sequence of diary entries, rather than merely rewarding a plausible-sounding analysis.
A small set of review, vision, frontend, and orchestration skills that I use while making things with Agents.
Walk around the SUSTech campus in a Minecraft-style block world. Buildings, roads, the lake and the mountain are all where they really are. You can walk, fly, dig and place blocks, or summon a car to drive and drift around campus. The WebGL engine is written from scratch — no game engine, no server. This build also includes day, sunset and night lighting, procedural audio, campus tasks, and imports and exports of your world edits. The campus map downloads as a 2.1 MB gzip; a single-file offline build ships next to it.
03 / SMALL ARCHIVE
04 / DECISION MAP
Works is where a curiosity becomes a method, then leaves behind a trace of what changed.
One long sequence of Chinese diary entries made a small question feel worth measuring.
Read the field note ↗DiaryBench asks whether a model can track nearby emotional states without over-reading the record.
Open the experiment ↗Aggregate Jaccard scores make comparison possible while the diary and scoring key stay offline.
See the public readout ↗The benchmark is one case; trading reviews keep the same habit of stating a thesis, recording a result, and revising the next move.
Read a related practice ↗05 / LIVING CHANGELOG
Works can read like a changelog: not a shelf of finished objects, but a record of what was tried, what shifted, and what is still open.

A note about AI, human workflows, and a word that is still intentionally unfinished.
Read the note ↗A private diary became a public aggregate board: the source stays offline, while the question and the comparison stay inspectable.
The first version of the question was not a product. It was a pattern noticed while looking back through a long personal record.
A working setup is part of the work too: the useful version includes the friction, the machine, and the next adjustment.
A sequence of trading reviews makes the change itself visible: the thesis, the mistake, and the next version sit together.