World Models 101 — How Machines Learn Inside Their Own Dreams
— by Vivax
A short intro to what we teach live in World Models 101 — the opening course of Vivax Academia, running 26 August 2026 at Boğaziçi Teknopark in Istanbul.
This post is a small intro to what we cover live in World Models 101, the opening course of Vivax Academia — a single evening on 26 August 2026 at Boğaziçi Teknopark in Istanbul. Read it as a taste of the curriculum: the arc we walk through together in the room, compressed into a few minutes. Start where the course starts, with a question that sounds like philosophy but is really engineering: how do we experience the world? As the cartoonist Scott McCloud observed in Understanding Comics, our minds fill the gaps between panels — we do not process every frame of reality, we predict it. NVIDIA's glossary puts the modern framing plainly: a world model is an AI system that builds an internal representation of an environment and uses it to simulate future states. Google, introducing Project Genie, describes the same idea for a general audience. The engineering question that follows: can we build neural networks that predict rather than merely process — and can an agent learn inside of its own dreams?
The idea has a long lineage: Schmidhuber's 1990 RNN model-controllers, AlexNet in 2012, the variational autoencoder in 2013, and David Ha & Jürgen Schmidhuber's 2018 "World Models" paper, which combined all three into one elegant agent. Its architecture has three parts: V, a VAE vision model that compresses each frame into a 32-dimensional latent; M, an MDN-RNN memory model that predicts the next latent as a mixture of Gaussians; and C, a tiny single-layer controller trained with CMA-ES. On CarRacing-v0 the full V+M+C world model scored 906±21 — beating a spatial-only V+C agent (632±251) and deep-RL baselines like DQN and A3C, and "attacking" sharp corners with foresight. The boldest experiment trained an agent entirely inside its own VizDoom dream, using a temperature parameter to stop it from exploiting the model's hallucinations — seeding the Dreamer lineage and today's world foundation models like NVIDIA Cosmos and Google Genie 3.
So why does Vivax Academia teach this? Because medicine is the ultimate predict-rather-than-process domain: a clinician runs an internal model of how a patient will evolve and acts on the forecast. The 2018 V-M-C lineage — a latent representation of state, a learned model of dynamics, and a small policy that plans against it — is exactly the shape of a clinical world model, and every limitation the paper named (artefacts, no long-horizon planning, the need for hierarchy) is a live problem in healthcare AI. This article is only the intro — we build the rest up hands-on in the room. Come do it live: World Models 101 runs on 26 August 2026 at Boğaziçi Teknopark in Istanbul. Register to go deeper with us in person.