Abu Dhabi:The Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has launched PAN 2, an open-source world model designed to simulate reality. The model was unveiled on Wednesday and is now accessible through the institute's website. It can be applied in several sectors, including robotics and education, by predicting possible outcomes in video form when provided with an initial scene and a descriptive action input.

According to Emirates News Agency, PAN 2 employs a technique known as 'simulative reasoning' that allows AI agents to learn and strategize without the necessity of performing real-world experiments, which are often costly or complex to reverse. This approach contrasts with language models that predict the next word, as world models like PAN 2 predict subsequent events or outcomes.

Eric Xing, Founder of IFM and President of MBZUAI, stated that the launch of PAN 2 marks the institute's initial step towards achieving practical or physical intelligence. This form of intelligence enables systems to engage with the world and foresee the results of their actions before executing them. Xing highlighted that while language models derive knowledge from textual content, practical skills, such as swimming, require more than theoretical understanding.

The PAN 2 model is constructed on the Generative Latent Prediction (GLP) architecture, which was developed by Xing and other IFM researchers. This architecture forecasts changes in an internal world representation and reconstructs observed outputs, allowing for verification of predictions against real-world events. PAN 2 has demonstrated advanced performance across several benchmarks that measure physical intelligence, including physical consistency, causal coherence, and constraint satisfaction.

The IFM has released PAN 2 under the Apache 2.0 open-source license, providing the video model weights, full training and inference code, and a technical blog for researchers. Hector Liu, Director of IFM's Silicon Valley research lab, emphasized that by making the model weights and code available, they aim to enable further testing, understanding of limitations, and enhancement of the model's capabilities.