World Models: Engineering AI Reading Group

​This Sat we’ll be reading about World Models.

​Last week, we read about Karpathy's Autoresearch and landed on a key insight: agents excel at hill-climbing well-defined metrics in constrained action spaces where iterative experimentation is possible.

​But, pushing the frontier of science and discovery is a different problem. Many real-world domains lack the tight feedback loops that make iteration tractable. E.g.

• ​Drug discovery: we have sparse rewards, no gradients, and have limited monkeys

• ​⁠Robots: Actions have delayed, non differentiable feedback so credit assignment is hard.

• ​⁠Games: We have non stationary opponents and reward landscape shifts

​All of these become tractable with a good world model. A world model is the learned transition function: given a current state and an action, it predicts the next state.

​We'll be reading works from:

• ​World Labs

• ​JEPA

• ​Dreamer

• ​Genie

​Engineering AI RG comes together to read and code on tools, libraries and codebases that are used frequently in the modern AI stack.

​All are welcome.