History
Ten years of
the same question.
Does the result hold when someone else runs it? Everything below is one answer or another to that, in order.
2016–2017
Reading, then reproducing
Before any of it was an institute, it was a habit: read the paper, then make the paper run.
- Oct 2016
Mah Paper Notes
Notes on papers, kept in public. The reading habit, written down.
- Dec 2016 ★163 · 37 forks
Study: Reinforcement Learning
Started as CS294 notes. Still being pushed to, nearly ten years later. Foundations through RLHF, DPO, GRPO, RLVR and agentic RL, with citations that resolve.
- Jan 2017 ★198 · 53 forks
One Hundred Layers Tiramisu
A Keras implementation of Fully Convolutional DenseNets for semantic segmentation, from the paper by Jégou, Drozdzal, Vazquez, Romero and Bengio. Fifty-three people forked it to run the paper themselves. It remains the most-used thing here, and it is a reproduction.
2018–2020
Congruent AI, first pass
The habit became a thesis: if reproducing a paper is this hard, that is the problem worth working on.
- 2018
Congruent AI founded
A research institute for independent research and reproducibility in machine intelligence. No venture capital then, none since.
- Apr 2018
Mora
Finding the implementation that belongs to a given paper. Papers With Code covered this ground for six years; Meta sunset it in July 2025 and the problem is unowned again.
- Jan 2020
Program Synthesis
Datasets and notes on program synthesis and proof checking, six years before the tooling arrived.
- Feb–Apr 2020
Bourbaki Space
An engine, a models API, a wiki and a search index, all aimed at machine-assisted mathematics. It went quiet for six years, until the tooling caught up.
- May 2020
Piplio
Conversation discovery across podcasts and recorded talks. Indexing ideas rather than episodes.
2026 onward
Agentic engineering
The tooling arrived. Bourbaki came back, and the reproducibility question turned into an engineering one.
- Feb 2026 v0.9.2.3
Pacenote
A framework for building coding agents; seven CLIs in the box. Shipped as Chimera, renamed on the road. Every benchmark traces to a receipt in the repository.
- Feb 2026
Bourbaki, again
The 2020 idea, rebuilt on tooling that can carry it: SymPy for computation, Lean 4 for proof, OEIS and arXiv for lookup. If it makes a claim, it can check it.
- Feb 2026 ★5
Erdős Navigator
Agents pointed at Erdős's 1,217 unsolved problems, with Lean 4 to settle whether anything produced is real.
- Apr 2026
staxtrace
Local-first observability and memory for coding agents. Twenty providers into one SQLite store on your own machine. An agent with no record of yesterday is a very fast intern on its first day, every day.
- Aug 2026
Novalis
A workspace terminal agents can drive. Pure Rust, GPU-rendered, sessions that outlive a closed lid.
Dates and counts are from the public record. Star counts are a snapshot and will drift.