A workbench for taking interesting ideas seriously enough to try to break them.
We read a lot: books, papers, talks, interviews. Most interesting claims in AI are stated in one place and tested in none. This repo is where we pull those ideas apart, combine them across sources, and build a small proof of concept to find out which parts survive contact with a GPU.
Each project ends in a written verdict. "It did not replicate" is a publishable result here.
| Path | What lives there |
|---|---|
docs/ |
The method, the toolchain, how to start a project |
templates/ |
Copy-paste starting points for notes and projects |
library/ |
Cross-project knowledge: sources, atomic ideas, syntheses |
projects/ |
Self-contained PoCs, one directory each |
tools/ |
Shared scripts |
The pipeline is deliberately linear:
source ──► idea ──► synthesis ──► claim ──► experiment ──► verdict
(what was (atomic, (combine (falsifiable (preregistered, (honest
said) reusable) across + a metric) seeded, with a writeup)
sources) baseline)
See docs/method.md for what each step demands.
| # | Project | Question | Status |
|---|---|---|---|
| 001 | geometric-reasoning | Is reasoning a property of a model's internal geometry rather than its size? Can a small model be engineered into robust reasoning? | 🟡 active |
Everything here is designed to run on one workstation: RTX 3090 (24 GB), Ryzen 7 5700X3D (8c/16t), 64 GB RAM. This is a constraint we like — an idea that only works at datacenter scale is an idea we cannot check, and a good chunk of this repo exists to ask whether that scale was ever necessary.
# one-time: uv manages Python and every project venv
irm https://astral.sh/uv/install.ps1 | iex
cd projects/001-geometric-reasoning
uv sync # creates .venv, installs deps
uv run python -m geomreason.doctor # verifies CUDA, prints hardwareFull details in docs/toolchain.md.
- Claims are tagged by evidence class (
[source],[lit],[ours],[open]) — a company's claim about its own product never silently becomes a fact. - Falsification conditions are written before the code that could satisfy them.
- Every number is backed by a seeded run in
results/with a git SHA. - Baselines are matched on compute budget, not on convenience.
More in CLAUDE.md.