World models

The core tension worth naming is architectural: current LLMs are trained on tokens representing language, not on the causal structure of physical space, which means their failures in the real world aren't bugs to be patched but a consequence of what they were built to predict in the first place. World models are an attempt to give AI systems an internal simulation of how objects, forces, and environments behave before acting on them.

This sits in direct conversation with two recent pieces. The Physical Intelligence story from April 16 covered pi0.7, a robot model that can generalize to untrained tasks, which is essentially a practical test case for whether proto-world-model approaches hold up outside the lab. And the robotics history piece from April 17 traced how the field has repeatedly overpromised on general-purpose machines while delivering narrow ones. World models are the current theoretical answer to that recurring gap, but the history piece is a useful prior on how often the theoretical answer has failed to close it.

Watch whether Physical Intelligence or a comparable lab publishes benchmark results on contact-rich manipulation tasks (think cloth handling or liquid transfer) within the next two quarters. Those categories are where world model approaches should show the clearest advantage over pure imitation learning, and sustained gains there would be the first credible signal that the architecture is doing real causal work.