100M tokens. Infinite compute. Lowest val loss wins.

⚒️ The open source dark software factory for expert engineers.

A MNIST-like fashion product database. Benchmark 👇

Bazel project template with many languages and tools

CIFAR-10 speedrun: Trains to 94% accuracy in 1.98 seconds on a single NVIDIA A100 GPU.

Minimal and scalable research codebase in JAX, designed for rapid iteration on frontier research in LLM and other autoregressive models.

Dion optimizer algorithm

Training transformers

Use Muon optimizer instead of AdamW.

A non-saturating, open-ended environment for evaluating LLMs in Factorio

a collection of lightweight TypeScript packages for AT Protocol, the protocol powering Bluesky.

Scripts for ozone

Official Implementation for NorMuon paper

WhisperX: Automatic Speech Recognition with Word-level Timestamps (& Diarization)

Cockpit: A Practical Debugging Tool for Training Deep Neural Networks

Muon is an optimizer for hidden layers in neural networks

Train to 94% on CIFAR-10 in <6.3 seconds on a single A100. Or ~95.79% in ~110 seconds (or less!)

The open source coding agent.

It's pronounced Model Control Padre, and it's your Swiss Army knife for MCP management.

script to get nvidia proprietary drivers on a fresh debian13 install

Make lists from Bluesky labelers!

Rust atproto crates and services for microcosm

Self-hosted PLC mirror

Basic ATproto labeler and a toolkit to make your own

Helpful code chunks for working with https://bsky.app

FOSS mass appraisal and automated valuation modeling toolkit

A simplified JSON event stream for AT Proto