THE WORLD
MODELS UPRISING
► A new family of AI models may be on the rise ◄
>> JUMP HERE <<★ join the waitlist ★
drop your email. we’ll keep you in the loop. no spam, just vibes.
🧬 WTH is a world model? 🧬
Okay, sit down. A large language model is a very confident parrot that has read the entire internet and guesses the next word. Impressive at parties. Has no idea what a cup does if you tip it over.
A world model is different. It builds an internal simulation of how reality works — objects, physics, cause and effect, consequences - so it can predict and imagine what happens next before it acts. Less “what word usually comes after this word,” more “if I do that, what does the world look like after?”
LLMs autocomplete the sentence. World models autocomplete the universe.
This is the direction folks like Yann LeCun keep yelling about (see: JEPA, “predict in representation space, not pixel space”). The bet: real intelligence needs a model of the world, not just a model of the text about the world.
📜 A BRIEF HISTORY OF WORLD MODELS 📜
- 1991 — Schmidhuber's "world models." Jürgen Schmidhuber proposes a recurrent controller + world model pair that learns to predict its environment. The OG idea.
- 2015 — Deep latent dynamics. Variational + recurrent nets start learning to predict future states from pixels, not just labels.
- 2018 — "World Models" (Ha & Schmidhuber). The famous paper: an agent dreams inside its own learned simulation of a racing game and trains there. Made the term mainstream.
- 2018–2023 — Dreamer (v1→v3). DeepMind's agents learn a latent world model and plan by "imagining" rollouts — eventually mastering tasks from Atari to Minecraft diamonds.
- 2022 — LeCun's JEPA vision. Yann LeCun publishes "A Path Towards Autonomous Machine Intelligence" — predict in representation space, not pixels. I-JEPA / V-JEPA follow.
- 2024 — Genie & Sora. Google's Genie generates playable worlds from a single image; OpenAI's Sora is pitched as a "world simulator." Video models start looking like world models.
- 2025 — Genie 2 & the land grab. Interactive, controllable 3D worlds from a prompt. Every lab suddenly has a "world model" team.
- 2026 — worldmodels.today more to come;