GPT 6 Astra is a freak
The video showcases GPT6 Astra’s ability to build a highly interactive, physics-based ray tracing simulation from scratch, using iterative critic-agent feedback to improve realism across multiple rounds until it produces a convincing bullet-through-water-balloon app with adjustable parameters and multi-angle viewing.
MAIN POINTS FROM TRANSCRIPT
- GPT6 Astra is presented as a powerful new model, especially when used through the desktop app for persistent, multi-project workflows.
- The demo challenges it to create a realistic bullet-through-water-balloon simulation without external libraries like 3JS.
- A critic-agent loop scores screenshots, identifies issues, and drives repeated revisions to improve realism and physics accuracy.
- The final app includes sliders, timeline freezing, camera angle changes, and visual controls for lighting, gravity, and spray effects.
TAKEAWAYS
- Iterative self-critique can significantly improve generated software quality compared with a single-pass prompt.
- Complex visual simulations are now feasible for frontier models to build from scratch in a desktop-agent environment.
- Multi-agent workflows help refine both functionality and realism by separating building and evaluation tasks.
- Interactive parameter controls make generated simulations more useful for testing, exploration, and presentation.