Your Time Series Is Already Training Data
Interleaved telemetry and commands from IoT and energy devices are already world model training data. What world mode...
What world models actually are, built up from code rather than equations. Action-result pairs, statefulness, the equations read as code, a model fitted from logs that you can train in the browser, and a latent state that finds a door the logs never recorded.
Everyone explaining world models starts with the maths. This series doesn't. It starts with runnable TypeScript, builds the intuitions from code, and reaches for the equations only once they describe something you've already seen work. The posts so far cover why telemetry is training data, a world model in forty lines, the two equations the field rests on read as a filter call, a model fitted from logged state and action pairs, with a playground where you can record logs and train it yourself, and a latent state: one private field that works out a door the log never mentioned. Next come how a world model differs from an LLM, the open-source landscape, and what all of this means for hardware in the field.
Interleaved telemetry and commands from IoT and energy devices are already world model training data. What world mode...
What a world model is, built in forty lines of TypeScript with no equations. State in, action in, predicted next stat...
The world model equation explained in code. Conditional probability, latent variables and subscripts, shown to be the...
Replacing a hand-written world model with one fitted from logged telemetry. Four weights, gradient descent in TypeScr...
Giving the heater model one private field and watching it work out a door that, according to the data, is not there....
I gave Claude the heater room from the last two posts and expected a plausible story. It fitted the physics and beat...
The physics-shaped heater model from 106 built properly, then a planner on top of it: every schedule, random guesses,...