Work Projects in the open.
stoa
RL post-training with Apple-silicon Macs as the rollout fleet. The Macs generate rollouts in int8 with MLX, a GPU trains, and weight deltas sync through Cloudflare R2. Our July run used 14 Macs in four countries.
Covenant-72B
Pre-trained a 72B LLM on ~1.1T tokens with permissionless peers over the open internet, the largest run of its kind at its release.
Grail
Geo-distributed, incentivized RL post-training for LLM reasoning, where untrusted nodes generate rollouts that are verified before training. Covenant AI's open research stack.
q Evaluation Harness
The first open-source evaluation framework for LLMs on q/kdb+, built with KX. Top models score 96.2% on Python's HumanEval; on the same problems in q, the best one, Grok 4, scores 43.4%.
Draw Your Circuit
A Meta Quest 3 app that turns a hand-drawn circuit sketch into a schematic.
Deep-RL-CS285-Pytorch
PyTorch solutions to Berkeley's CS285 deep RL assignments, my most-starred repo (144 stars).