General Instinct | Any frontier model. Any edge device.

The deployment layer for physical AI. Any frontier model on any edge device, sub-100ms.

Pixel-level video prediction is the most expensive thing a World Action Model (WAM) does. This post is really about whether it is worth it. We start where the case for generation is strongest, DreamZero, which turns a 14B video-diffusion model into a robot policy, then follow the evidence from Fast-WAM and ImageWAM that the generated future is a training scaffold you can throw away at inference. From V-JEPA onward, one thread runs through all of it: maybe the latent is all you need. 1DreamZero: the strongest case for imagining the future A Vision-Language-Action model (VLA) maps what a…

saved by

• nicolas

related reading

• Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models | NVIDIA Technical Blogdeveloper.nvidia.com
• The flavor of the bitter lesson for computer vision - Vincent Sitzmannvincentsitzmann.com
• The First Fully General Computer Action Model | blogsi.inc
• [2603.16666] Fast-WAM: Do World Action Models Need Test-time Future Imagination?arxiv.org
• Generalist - GEN-0 / Embodied Foundation Models That Scale with Physical Interactiongeneralistai.com
• 1X World Model | From Video to Action: A New Way Robots Learn1x.tech
• Causal Video Models Are Data-Efficient Robot Policy Learners | Rhoda AIrhoda.ai
• [2506.09985] V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planningarxiv.org
• [2509.22407] EMMA: Generalizing Real-World Robot Manipulation via Generative Visual Transferarxiv.org
• The Model That Dreams the Worldmoe-capital.com
• Video models are zero-shot learners and reasonersarxiv.org
• [2602.10098] VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Modelarxiv.org