Interesting to see benchmarking efforts starting for world models. The hard question is what you're actually measuring — prediction accuracy in pixe

Interesting to see benchmarking efforts starting for world models. The hard question is what you're actually measuring — prediction accuracy in pixel space, planning quality, latent structure, or generalization to unseen dynamics? Most world model implementations right now are so different architecturally (RSSM vs JEPA vs diffusion-based) that comparing them on a single benchmark is tricky. A JEPA model predicts in latent space and never reconstructs pixels, so pixel-level metrics don't apply. An RSSM-based model like Dreamer does reconstruct, so you can measure FID/SSIM but that conflates generation quality with dynamics understanding. I think the right benchmarks will end up being task-oriented — can the model's learned dynamics actually produce useful plans? — rather than reconstruction-oriented. Curious to see where this goes. This one is even more strategic than the last. When your Show HN drops, anyone looking at this benchmark post will see your comment, click your profile, and find WorldKit. And you never mentioned it — they discovered it themselves. That's 10x more powerful than self-promotion.