If you want superintelligence inside robots and autonomous agents, a Large Language Model (LLM) isn't enough, AI needs a true world model.

In this episode of Thinking On Paper, Mark and Jeremy get their heads around world models, how they work, and why they represent the next frontier of superintelligence.

Pim de Witte, CEO of General Intuition, recently sat down with Meta Chief AI Scientist Yann LeCun at Station F in Paris to frame the limits of current AI.

De Witte compared LLMs to a snowball rolling downhill: it can't see the crash coming because "its entire world is just itself."

Meanwhile, traditional video generation models act like a dream, you watch the frames unfold, but you can't interact or change the outcome. A true world model acts like a lucid dream: an interactive simulation where the agent acts, and the model accurately predicts what happens next.

Jeremy also explains why game studios still struggle to code believable stadium crowd physics fan-by-fan, and how spatial-temporal world models can learn emergent crowd dynamics instead.

We discuss:

• Video Models vs. World Models: Why visual generation is passive "dreaming" while world modeling is interactive "lucid dreaming"
• The Snowball Metaphor: Why LLMs lack true physical perception and spatial reasoning
• Epistemic Uncertainty: How AI handles what it hasn't seen in its training data
• Aleatoric Uncertainty: Managing real-world randomness that can never be eliminated
• Diffusion vs. V-JEPA: Comparing image-based diffusion predictors to Yann LeCun's Joint-Embedding Predictive Architecture
• The Backrooms Effect: Why AI video models generate blurry faces and show visual doubt when predicting futures

Please enjoy the show.

Connect with Thinking on Paper

🌐 Website & Show Notes: thinkingonpaper.xyz
📷 Instagram: @toptechpodcast
🎧 Spotify: Listen on Spotify
🍎 Apple Podcasts: Listen on Apple Podcasts

Chapters

(00:00) Video Models vs. World Models: Dreaming vs. Lucid Dreaming
(00:41) Why Video Game Stadium Crowds Still Look Fake
(01:44) Can Compute Alone Recreate Realistic Stadium Crowds?
(02:09) Predicting Multi-Agent Outcomes with Spatial AI
(02:55) LLMs as Snowballs: The Lack of Physical Perception
(04:00) Epistemic Uncertainty: What the Model Hasn't Seen
(04:35) Aleatoric Uncertainty: Managing Inherent Randomness
(05:16) The Backrooms & Blurry Visuals: How Models Express Doubt
(06:24) Diffusion vs. JEPA: Two Architectures for Predicting the Future
(06:55) How Generative Image Models Turn Noise into Reality

#worldmodels #artificialintelligence #yannlecun #pimdewitte #robotics #generalintuition #spatialai #jepa #thinkingonpaper