Researchers stretch LeCun's JEPA AI into a universal world model that works from physics to biology

PhAI Labs, in collaboration with Stanford, Oxford, and Princeton, has developed a new AI model named JEPA-Anything, which extends Yann LeCun's JEPA architecture. This model aims to function as a universal world model, demonstrating capabilities across diverse fields, from physics to biology.

One of the notable achievements of JEPA-Anything is its performance in dynamic systems, where it achieved a 35% reduction in prediction error with targeted interventions. Additionally, the model showed a 13% reduction in prediction error for unseen combinations of interventions, indicating its robustness in handling novel scenarios. In a benchmark test involving the Burgers equation, a fundamental problem in fluid dynamics, JEPA-Anything recorded a remarkable 50% reduction in error, showcasing its potential for complex physical simulations.

In the realm of biological applications, JEPA-Anything has made strides in predicting disease events, successfully forecasting 1,000 disease occurrences based on clinical data. This capability highlights the model's applicability in healthcare and its potential to assist in clinical decision-making.

The model also demonstrated a 3% advantage in reducing prediction error over 50 steps, which is significant for applications requiring long-term forecasting. In locomotion planning for simulated walking robots, JEPA-Anything outperformed existing models in 2 out of 3 environments, further establishing its versatility.

A particularly intriguing aspect of JEPA-Anything is its application in cancer research. The model identified a candidate treatment for liver cancer, targeting the enzyme CD73 and the signaling molecule IL-18. This candidate was reported to kill more tumor cells in lab samples and mice than either component alone, although it is important to note that the study does not confirm the candidate as a viable therapy. Jonathan Kemper, a researcher involved in the project, emphasized that while the model can propose and rank experiments, it does not guarantee that the learned parts capture real cause-and-effect relationships.

The development of JEPA-Anything follows a series of milestones in the evolution of the JEPA architecture, which was first proposed by LeCun in 2022 as an alternative to generative models. The advancements in this research reflect a growing interest in leveraging AI for complex problem-solving across various scientific domains.