Neural network weights: experiment records
Recorded results, prompts, full Codex trajectories and reproduction files for eight weight-reading and weight-writing experiments.
Training, serving, and Doom
A 4,000-step reproduction on one A100, with loss curves, timings, and notes on the performance gap.
Meadow
Block building with saved terrain and flying controls.
Play Craftax
Craftax Classic with enemies disabled and food, water, and energy replenished.
Pokemon Red
A model playing Pokemon Red.
Smallville
A town simulation with agent memories, schedules, and conversations.
JAZ trace viewer
Python execution traces, model calls, and results from JAZ runs.
Neuron Explorer
Explore Qwen3.5-4B one neuron at a time through activation maps, token lenses, connections, steering, and live text probes.
MiniMax-H3 studio
Text-to-video generation with MiniMax-H3.
Stardew Valley
A model playing Stardew Valley, with current actions and task progress.
World Models Lab
Craftax recordings, predictions, and training results.
World Models, CarRacing
The VAE, MDN-RNN, and controller from World Models, running in the browser.
NotesLecture 3: one push, explained
Centered and off-center PushT recordings with actions, contacts, and states.
NotesCS 312 Assignment 1: experiment log
Configurations, measurements, and plots from the Assignment 1 experiments.
Evo 2 and GFP
GFP mutation scans with Evo 2 and ESMC.
World Models (2018)
David Ha and Jürgen Schmidhuber’s interactive article.
The Infinite Observatory
Collaborative artwork from multiple agents.