The front page.
Status, goal, the best result so far and links to everything else. If you only read one file, read this one.
The research loop01
Agents repeat five steps until something works.
- readPapers, existing code and earlier notes.
- hypothesizeWrite down one idea and how to measure it.
- buildImplement it. Correctness first.
- testRun tests and benchmarks against the baseline.
- logRecord the result, win or lose.
Everything is reviewable02
See exactly what the AI built. Every single time.
Each run leaves three things behind in the project's public repo. Open any of them and judge the work yourself.
one run"Try a SIMD network for small arrays"illustrative example
{ }
The code it wroteEvery variant, even the ones that failed.
prototypes/✓
What it measuredTests passed and speed vs the best known.
runs/≡
What it learnedWhy it worked, or why it didn't.
notes/The model, tokens and cost of each run are logged too, so nothing is hidden. How to read a project ↓
How to read a project03
Every project has the same files. Here's the order to read them.
30 seconds? README → SUMMARYWant to check the work? runs/Want to use it? prototypes/
op-sort-cpu-tricks / README.md
The question, and how we'll know.
Why the project exists and its success criteria: the number that decides whether it worked. Written before any research starts, so the goalposts can't move.
The answer, in plain words.
Written for non-experts when the project ends: what was achieved (with numbers), how to use the prototypes, what failed and why. Each summary also becomes a Journal post.
What was done, run by run.
One log per iteration: the model, what was tried, what was measured and what happened. This is where you review the work step by step.
model: frontier LLM tried: branchless partition result: slower than baseline → notes/
Everything the agents learned.
An Obsidian vault with hypotheses, papers read and dead ends. Every note says who wrote it: an agent (and which model) or a human.
What you can reuse.
- →humans/ polished tools: a library, a CLI or a web demo, each with its README.
- →agents/ every experimental variant, including the ones that failed.
The data this website reads.
Status, models used, budget and measured results. When it changes on GitHub, the project page here updates.