2025
Scaffolding to Superhuman: How Curriculum Learning Solved 2048 and Tetris
Training gaming agents that beat massive search-based solutions on 2048 using a 15MB policy, and discovering that bugs can be features in Tetris.
Shedding Expertise: What Matters?
I'm living this, questioning what I thought was my expertise. AI models provide internet-scale possibilities, and motivated humans can crystallize one thread with their own preference and taste.
The US–China AI Bet: Monopoly vs. Commodity
Deep insights on why China is building open-weight models and their bet on energy and action over intelligence alone.
If Robot Hands Can Make Robot Hands Can Make Robot Hands Can
Robot hand assembly could serve as an excellent benchmark for robot manipulation, similar to how coding benchmarks work for LLMs. Who knew LLMs would code so well?
Reinforcement Learning Is Hitting Its Inflection Point
Reflections on technology maturity cycles, from missing the eye-tracking wave to catching the RL wave with PufferLib.
Puffer-PHC: Simplifying Perpeptual Humanoid Control with PufferLib
Making the Perceptual Humanoid Control repo simpler and training 60k+ SPS on a single 4090 with PufferLib.
2024
Bring Me a Beer: A Turing Test for Home Robots
Starting to hack the Hiwonder Armpi Pro robot arm. The 'bring me a beer' task is still hard.
Building Card Table with Claude Artifacts
I had never used React before, but with Claude Artifacts I built a web app for literature reviews. A strange but increasingly common experience.
Thoughts on Picking Games Worth Playing
An interesting analysis of how luck and skill influence rankings across chess, sports, video games, and hierarchies. Makes me think about which games I'm playing.
mg2hfbot: Testing LeRobot on MimicGen Tasks
Converting MimicGen datasets to train and evaluate LeRobot & RoboMimic policies on complex manipulation tasks without a physical robot.
Spatial Intelligence: Research or Product?
Thoughts on Fei-Fei Li's spatial intelligence venture and whether transformers are the right architecture for this problem.
Reflections from Actuate 2024
Key reflections from the 2024 Actuate Robotics & Embodied AI Conference on unified models vs specialized collaboration and interpretability.
Crafting Review Criteria for RL Environments, with AI
Developing review criteria for RL environment papers with Claude's help. Emphasizing reproducibility and usability over novelty.
Meta MMO: A Massive Update to Neural MMO
Introducing Meta MMO - 3x faster training, diverse minigames, and successfully trained generalist agents. Open-source under MIT license.
2023
Neural MMO Talk at NeurIPS 2023
Excited to speak about Neural MMO at NeurIPS and discuss how NMMO can support open-ended multi-agent RL research.
NeurIPS 2023 Neural MMO Challenge
The NeurIPS 2023 Neural MMO Challenge is live on AICrowd with $20K in prizes. Train agents that generalize to new tasks, maps, and opponents.
In Support of Libraries as AI Research
Supporting Joseph Suarez's open letter to NeurIPS D&B organizers about the importance of libraries and infrastructure in AI research.
2020
Web-based Cognitive Tasks: Interactive jsPsych Experiments
A collection of web-based cognitive tasks built with jsPsych for research. Try them yourself - from decision-making to working memory tasks!
2019
Math Anxiety & Avoidance: We Can't Even Pay People to Do Hard Math
Higher levels of math anxiety were associated with a tendency to select easier, low-reward problems over harder, high-reward math problems, suggesting that we cannot even pay math-anxious people to do hard math.