Session

Step into World Models

Every explanation of world models opens with a conditional probability, a subscript and a latent variable, and half the room has gone by the third slide. The core idea is much smaller than that. A world model answers one question: if things are like this and I do that, what happens next?

This talk builds one live, in TypeScript, around a single room with a heater. We start with forty hand-written lines, then throw the body away and fit four weights from 4,800 logged steps. Those four weights recover the room's physics and predict a day six degrees colder than anything in the logs. Then someone keeps opening a door that the logger never records. One private field, trained by walking each episode backwards, works out the door for itself, and the field turns out to be all "latent state" ever meant.

From there we put the model to work. I gave Claude the same room and it beat my model, until my model borrowed the shape of physics Claude described. A planner searches heater schedules against it, from a million brute-force guesses down to about a thousand with the cross-entropy method, replanning every step. Eight bootstrap copies draw a band showing where the model is unsure, and running the model beside the room makes it a fault monitor that catches a dead heater in sixteen readings. The field that found the door also hides a stuck heater, and I'll show how to catch that too.

You'll leave able to read the two equations the field rests on as a filter call, and with a clear sense of what the telemetry you already keep could do. Everything runs in a browser playground you can use during the session.

KEY TAKEAWAYS

1. Interleaved telemetry and commands are before-action-after records, which is exactly what a world model trains on. The labels are free because the next reading marks the homework.

2. The world model equation, the probability of the next state given the current state and action, is a WHERE clause with the SQL boiled off. The world model keeps that condition two values wide; an LLM's grows with the history, which is why LLMs are enormous.

3. A latent state is a private field the model fills for itself, judged only by the predictions it produces.

4. The shape of a model carries the knowledge that isn't in the log. A bias term the log can't tell apart from no bias sends a heater-off room 7.4 degrees colder than outside, and more rows of the wrong shape only make the wrong answer more confident.

5. An LLM and a world model are good at different things, and the useful pattern is to use both together.

6. A planner exploits a model's flaws, so a forecast should come with a band. The band maps where the uncertain weights matter, which isn't the same as where the data is thin.

7. A model that adapts quickly will explain some faults away. Watch what the model believes as well as what it gets wrong.


The talk is drawn from my blog series Understanding World Models at stepinto.dev, which builds the subject from runnable code rather than notation. All of the code is plain TypeScript with no ML libraries, and every model trains in under a second, so the demos run live in a browser and don't depend on GPUs or the venue's network once the page has loaded. The playground is public at stepinto.dev/heater/, so attendees can follow along on a laptop or phone.

The material comes from day-to-day work at Unimo, where we run telemetry and control for energy hardware in the field.

Series: https://stepinto.dev/series/understanding-world-models/
Playground: https://stepinto.dev/heater/

Andy Cross

CrossMoore Director, AI Entrepreneur

London, United Kingdom

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top