I build 山乇丨尺ᗪ small fast projects.
DigDice: a first-person dice game where the dice are blocks in the world. You dig them out, carry them, and build poker hands with them.
|
|
|
|
|
|
Warning
All benchmarks performed on a 2021 MacBook Pro with 16GB RAM
Some benchmarks and claims change as I revisit projects and may become out of date
| Compression | ||
|---|---|---|
| Metric | Description | Project |
| 0 files | Texture and map assets shipped with DigDice - every texel and every world is generated from a seed | dig-dice |
| 1 bit | A wall in the raycaster's chunked world map (a vertical stack of 16 walls fits in 2 bytes) | chunked-z-level-raycaster |
| 1 byte | Snibble's custom event protocol size | snibble / snibble-bench |
| 5 keys | DigDice's entire control mapping scheme (←↑→↓␣) handling 10+ actions and movement | dig-dice |
| 18 bytes | Maze-solving Neural Network solving 99% of unseen mazes | maze-solver |
| 20MB | Total size for a 100×100×10 chunk ray-casted world with O(1) lookup performance | chunked-z-level-raycaster |
| 44 bytes | Tether's entire memory: a maze walker with no map and no coordinates, on any grid under 67 million cells | tether |
| 100 bytes | Typical Snibble replay size for a 5 minute game (entire state) | snibble / snibble-bench |
| 369 bytes | Microlex's entire syntax-highlighting engine; each language adds a 37-178 byte state machine | microlex |
| Performance | ||
| Metric | Description | Project |
| 2ms | Snibble's average total frame time with 256 simultaneous players (logic AND rendering) | snibble |
| 30x | Speed of Microlex JavaScript lexical classifier vs PrismJS under equivalent benchmark (99.6% per-character agreement) | microlex |
| 2048 | Simultaneous agents performing complex spatial/lexical/adversarial tasks at 60FPS | snibble / snibble-bench |
| Accuracy | ||
| Metric | Description | Project |
| 89.9% | Unseen mazes (up to 161×161, plus 1,600+ step game-map routes) Tether solves within 10× the shortest route, from four wall bits and one goal-direction bit - ahead of every Bug algorithm tested, all of which are told their position | tether |
| 97.3% | Blink link navigation prediction accuracy with a 10 byte Neural Network (best hand-written heuristic: 66.2%) | blink |
| 100% | Exact replay accuracy for Snibble replays | snibble / snibble-bench |