JST-2, a full-stack physics-native world foundation model
The Hook
“For the past four years, we've been focused on one question: If the world we live in could be simulated, what would its architecture look like? That question is what our company Viggle was built to answer.”
Problem Agitation
“Most video models only learn what the world looks like through pixels. Our team built JST to model what the world actually is.”
Product Reveal
“Persistent 3D atoms with learned vectors encoding their physical attributes and dynamics over time. Everything was built around this physics-native principle: from tokenization, architecture, pre- and post-training, and our data engine—the full stack.”
Feature Teaser
“The result is JST-2. The world's first full-stack, physics-native world foundation model. One model generates characters, motion, and scenes. JST-2 outperforms every specialist model we benchmarked across multiple tasks. It generates a full scene in about a second for under a cent. And every JST-2 world is real 3D. Generated once, then render locally in the browser from any angle without rerunning the model.”
Feature Teaser
“The same efficiency changes what's possible in robotics. In this demo, JST-2 generates the environments, people, and actions, pointing to a scalable path to robot training. JST-3 will further replace rule-based physics with learned physics.”
Call to Action
“Four years ago, we asked what it would take to simulate the world. JST-2 proves the architecture works with real quality, efficiency, and scale. Try JST-2 for free today at viggle.ai/jst2, and read the architecture and benchmarks in our research blog.”
Synthesizable in Remotion · key components of the keynote-foundation-model-reveal template.
Simulates chat input with prompt text typing and auto-expanding system status chips.
Use spring interpolation on height and opacity; render staggered text slices based on frame index.
Presents an isometric schematic where sub-labels illuminate sequentially with connecting dotted spline lines.
Render SVG paths with strokeDashoffset animations synchronized to interpolate active layer badges.
Displays a 3D wireframe camera cone with dynamic tracking vectors inside a 2D floor grid.
Project 3D Euler angles to 2D canvas/SVG points, rendering dynamic coordinate labels.
Purpose. Establish company mission and ground the technical problem.
Execution. Presenter speaking straight to camera with staggered typography overlays.
Purpose. Explain technical differentiator over pixel-based diffusion models.
Execution. Diagram showing persistent 3D tokens and full-stack pipeline.
Purpose. Prove speed, cost, and cross-style spatial fidelity.
Execution. Rapid multi-environment flythroughs with speed/cost data badges.
Purpose. Demonstrate practical robotics and foundation model scalability.
Execution. Multi-view interface showing real-time text-prompted robotics simulation.
Purpose. Direct viewers to hands-on testing and technical blog.
Execution. Animated gradient wave settling into URL and API badge.