LESSON wm.0.3 · Code · 140 min
Build the smallest working world model
Build the smallest working world model. Learn GridWorld, action-conditioned dynamics, rollout, and model-predictive control and complete: Train a next-state predictor and plan acti
Learning objectives
- Explain what problem “Build the smallest working world model” solves without hiding behind terminology.
- Trace the variables and causal links across GridWorld, action-conditioned dynamics, rollout.
- Complete “Train a next-state predictor and plan actions through imagination” and judge the result with evidence rather than intuition.
Core concepts
GridWorld
GridWorld sends candidate actions through a learned transition model to predict later states or observations. Model-predictive control scores complete trajectories, executes only the current action, and replans after the next observation; report both one-step error and multi-step failure rate.
action-conditioned dynamics
action-conditioned dynamics sends candidate actions through a learned transition model to predict later states or observations. Model-predictive control scores complete trajectories, executes only the current action, and replans after the next observation; report both one-step error and multi-step failure rate.
rollout
rollout sends candidate actions through a learned transition model to predict later states or observations. Model-predictive control scores complete trajectories, executes only the current action, and replans after the next observation; report both one-step error and multi-step failure rate.
and model-predictive control
and model-predictive control sends candidate actions through a learned transition model to predict later states or observations. Model-predictive control scores complete trajectories, executes only the current action, and replans after the next observation; report both one-step error and multi-step failure rate.
Build and verify
Train a next-state predictor and plan actions through imagination
- Predict: write the expected output, trend, or failure before running code.
- Build: implement only the minimum components needed to answer the question.
- Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
- Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.