Selected work
2025 — 2026
-
2026
SIA, a self-improving autoresearcherAgents rewrite their own scaffolds between generations, with no human in between. Deployed at Lawrence Livermore National Laboratory.
80.81%GPQA Diamond
Hexo Labs
-
2026
Socrates, a critique-and-revision ML agent8.67% lower log loss than the baseline agent. Two papers submitted to COLM.
6×faster convergence
Hexo Labs
-
2025
REAL benchmark (opens in a new tab), NeurIPS 2025Deterministic replicas of eleven real websites. Frontier models top out around 41%.
112tasks · 11 sites
AGI Inc
-
2025
AGI-0, computer-use agent#1 on OSWorld in October 2025. GRPO and VerL pipelines underneath.
76.26%OSWorld
AGI Inc
-
2025
agisdk (opens in a new tab)Web agent, harness, sandbox sites, benchmarking. Every major model and framework has been run through it.
17kdownloads · 398 stars
Open source
-
2026
CAD and simulation RL environments for frontier modelsReward design for parts that have to exist in the physical world. A frontier model against a spring from 1976.
1,758scored rollouts
Essay
Trajectory
Dallas, Jan 2023 → San Francisco, Jul 2026
- Jan 2023Dallas. One month's notice, a forex card, no savings. Consulting for Bread Financial and California Lawyers for the Arts.
- 2024Dallas → San Francisco. Deadline set by the Stanford Codex hackathon. Hostels, East Coast hours, every event in the city.
- 2024Stanford. Convex Optimization under Stephen Boyd. High-performance computing.
- 2024 – 2025Palo Alto. Mister Sparky: FDD signed, lease settled, then the numbers stopped fitting. Loan cancelled. Clean exit.
- Mar 2025San Francisco, AGI Inc. Scouted at a hackathon. REAL benchmark to NeurIPS. AGI-0 to #1 on OSWorld.
- Dec 2025Santa Clara, Hexo Labs. SIA reaches 80.81% on GPQA Diamond.
- Jul 2026San Francisco. Founder.
