AI Can't Learn The Way Humans Do - This Could Fix That
Y Combinator · 1:14:27 · Today
World models offer a pathway to Artificial General Intelligence (AGI) by enabling AI to simulate outcomes and learn from small amounts of data. Unlike current "model-free" systems that rely on massive datasets for simple tasks, world models encode an internal representation of how the environment functions—allowing for better planning, adaptation, and intelligence with fewer samples.
-
Sample efficiency gap — Humans learn new skills after a handful of attempts, while current models require tens of thousands of data points to achieve comparable results .
-
Defining the ideal — Newtonian physics represents a "perfect" world model, allowing for precise trajectory planning without needing to constantly collect experience from the environment .
-
Differentiable control — Systems utilizing physics-based models can solve complex navigation problems in closed form, whereas unpredictable, chaotic environments render this approach non-differentiable and difficult .
-
Robotic complexity — Controlling a robotic arm requires managing 16 degrees of freedom, which creates a massive action space that is significantly harder to navigate than the state space of games like chess or Go .
-
Synthetic training — Methods like Dreamer allow agents to train on "imaginated" rollouts, enabling them to master complex tasks, such as mining diamonds in Minecraft, entirely on synthetic data .
-
Latent space tricks — Joint Embedding Predictive Architectures (JEPA) compress high-dimensional pixels into latent states, which prevents model collapse and improves efficiency by focusing on abstract features rather than full image reconstruction .
-
Data distribution issues — Physics-informed neural networks often fail to handle edge cases—such as avoiding a crash when training data is dominated by safe highway driving—because of poor data mixing and lack of real-world constraints .
-
Real-time adaptation — Human brains constantly adjust to physical feedback like friction or terrain in real-time, a capability that current neural architectures struggle to replicate during test-time planning .