Turn experience into state
Build compact predictive state from partial, noisy observations without confusing a reconstruction with the world itself.
World Models · from first principlesPredict the world.
Predict the world.
Act through imagination.
Start with observations, actions, and possible futures, then add learned state, planning, robotics, and operational evidence. Formal probability and control notation arrive only after the mechanism is familiar.
46 lessons · about 19 hours
THE WORLD-MODEL LOOP
observation→state + action
predicted consequencenext-state distributionillustrative flow · not a measurement
Learn how machines imagine before they act.
A world model gives an agent something more powerful than reflex: a way to represent hidden state, rehearse possible futures, and compare actions before reality pays the price. It is the bridge between prediction and purposeful behavior.
Plan inside learned futures
Compare rollouts, search, control, and uncertainty while keeping model error and exploitation visible.
Connect research to reality
Reason about video models, robotics, constraints, telemetry, release gates, and the evidence needed for a defensible claim.
The finish lineFinish able to design, interrogate, and safely operate systems that learn a world well enough to plan in it.
From one controlled transition to an operated world-model system.
Six cumulative phases build predictive state, imagination, decision-making, foundation models, and deployment evidence before advanced branches.
NEURAL FIELD GUIDE / COURSE ARC01 · 06
Turn experience into a prediction problem
Define observations, actions, hidden state, uncertainty, return, and belief before introducing learned latent dynamics.
You unlockTrace one controlled partially observed system
Build and train predictive state
Learn representations and objectives together so compression, recurrence, inference, replay, and uncertainty remain inspectable.
You unlockAudit a recurrent state-space training contract
Plan and learn inside imagination
Compare shooting, MPC, differentiable planning, actor–critic imagination, and tree search under explicit budgets and error boundaries.
You unlockSelect a decision method from task evidence
Scale to video and interactive worlds
Separate raw generation, feature prediction, latent actions, planning interfaces, and official release evidence.
You unlockBuild a contract table for foundation world models
Evaluate and operate the control loop
Expose compounding error, robot transfer, constraint authority, telemetry, release gates, and rollback.
You unlockDesign a bounded world-model operations package
Choose an advanced research branch
Object, hierarchy, geometry, causal, and multimodal lessons branch from the same shared core; the capstone requires one branch, not all of them.
You unlockRun one falsifiable changed-case study
Inside the course
One course. 7 connected frontiers.
Lessons 01–40 form the shared spine through safe operation. Lessons 41–45 are parallel research specializations; Lesson 46 turns one chosen branch into a falsifiable final study.
01
Define the prediction problemFoundations
Build the state, probability, control, and sequential reasoning every world model needs.
The payoffTrace one partially observed controlled transition.
02
Build predictive stateRepresentations
Compress observations into action-conditioned recurrent states without hiding uncertainty.
The payoffSpecify and audit a recurrent state-space model.
03
Train the simulatorLearning Dynamics
Choose targets, priors, replay, multistep losses, and uncertainty for useful imagined futures.
The payoffDesign an evidence-honest world-model training run.
04
Act through imaginationPlanning & Control
Turn learned dynamics into rollouts, MPC, policies, values, and search under matched budgets.
The payoffChoose and audit a model-based decision method.
05
Scale predictive experienceVideo & Foundation Models
Compare token, feature, latent-action, and interactive-video models through their actual contracts.
The payoffEvaluate a foundation world-model claim by interface and evidence.
06
Operate bounded controllersEvaluation & Deployment
Measure failures, transfer to robots, enforce constraints, and run versioned release loops.
The payoffShip a staged, observable, rollback-ready controller design.
07
Choose a research branchAdvanced Specializations
Explore objects, hierarchy, geometry, causality, or multimodal grounding from the shared core.
Explore when relevantComplete one falsifiable specialization study.
46connected lessons
46hands-on labs
20code notebooks
No black boxes. Build intuition, see the mechanism, then make the real engineering trade-offs.
Ready when you are
Finish able to design, interrogate, and safely operate systems that learn a world well enough to plan in it.
46 connected lessons, hands-on labs, and a complete end-to-end build.