DataFramed #351 Will World Models Bring us AGI? with Eric Xing, President & Professor at MBZUAI
27 snips
Mar 16, 2026 
Eric Xing, President of MBZUAI and leading ML researcher, explains world models as simulators that move AI from book knowledge to physical and social skills. He discusses long-horizon planning, architectures like PAN that mix symbolic and generative components, evaluation beyond pixel fidelity, robotics and driving use cases, synthetic data and virtual cells for science, and the push for open K2 Think models.
AI Snips
Chapters
Transcript
Episode notes
Hybrid Representations Enable Long And Short Reasoning
- Representations must balance latent continuous reasoning and symbolic, discrete reasoning to maintain long-term plans.
- PAN uses a mixed backbone: LLMs for symbolic long-term inference and diffusion decoders for high-res short-range visuals.
Build World Models As Modular Pipelines
- Train world models modularly: separate encoders, reasoning backbone, and decoders before joint tuning.
- Eric describes using a visual language encoder, an LLM reasoning backbone with extended vocabulary, and a diffusion decoder assembled and tuned together.
Evaluate World Models With Task Consistency Metrics
- Measure world model quality by task-oriented latent inference, not just video pixels.
- Eric recommends testing long-term tasks like object sorting and minute-scale consistency via latent comparisons rather than generating hours of video.