Math and CS at UChicago. Agents, robots and games, written up with the results.
The highlight / agents / October 2026
It dodged almost everything and still won only 8% of its fights. The fix turned out to be where it stood.
The agent reads the game's exact internal state, simulates 16 short moves every two frames, throws out the ones that get hit and picks from the rest. Its predictions were right more than 99% of the time. It lost anyway, always in the last phase. The logs showed it dealt its damage from under the boss and took its hits at the walls, and that every trip to the wall was its own choice. So I tested a position term on the same 100 seeds.
Also in this issue
markets / co-founder and CEO
We set out to let mid-size companies price business risk with prediction-market contracts. The research changed what we think the product is: one of our findings was that 1,440 simulated trading desks behaved like about one. Here is what went well, what didn't, and what we learned.
robotics / world models
Two SO-101 arms, each run by its own agent, coordinating only through a shared, predictive picture of the table. In 710 simulated episodes they matched a central planner and kept their speed under 300 ms of network delay. Next, the real arms.
From the lab
lab / shared world
Arrive on an island with one photograph. Take it and vanish; a hiker walks your route, and the island is rebuilt from everyone's photos as Gaussian splats.
lab / non-Euclidean
A garden walk where each lap is a new season and the cabin is bigger inside than out.
lab / notes
The database, the one-memory rule, moderation, and how the world could keep growing.
From the archives
robotics
Flood fill, real-time control code and the IEEE Micromouse competition at UCLA, with UChicago Robotics.
games
DEADFALL, TRAINING DATA and MODEL EMPLOYEE, plus the games still in the workshop.
game jams
A student studio of 5+ people, entries to Ludum Dare and the GMTK Game Jam, and what each one taught us.
hackathons
I started a yearly hackathon and game jam at Trevor Day. It grew past 50 people and spread to other NYC schools.
tools
Atlas, a personal context engine; OpenClaw, five agents in one Discord; NightShift, a coding loop that works overnight.
About
I'm a second-year at the University of Chicago studying mathematics and computer science. I co-founded Marketry, co-lead Micromouse software at UChicago Robotics, and started a game studio and a hackathon in high school. I care about agents, robot learning and world models, and about measuring them honestly.