Astra runs a 12-hour self-improvement loop on a 3D scene without prompts
Given one task, build the most realistic backside possible, Astra worked unattended for 12 hours, improving the scene's graphics, math and physics and redoing anything that looked plastic.
this is the GPT-6 Astra benchmark we actually deserved while the labs compete on ARC-AGI scores and olympiad math, one guy dropped the model into a self-improvement loop with a single task - build the most realistic backside possible and went to bed for 12 hours Astra worked without a single prompt, improving the graphics, math and physics of the scene on its own, comparing each result to what it had an hour earlier and redoing anything that looked even slightly plastic by morning the screen was spinning exactly what this was all for - and ARC-AGI wasn't even close
This post quotes another post: x.com/mikadzyki_nft/status/2097336663429046689