Shader eval puts Astra below GLM 5.3 flash on 3D spatial reasoning
teortaxesTex · x · 2026-09-06
A developer's shader-generation eval found Astra underperforming GLM 5.3 flash on what should be its strength, 3D spatial reasoning — prompting the remark "RL does not generalize."
teortaxesTex pushes back: Astra's water shader is mathematically deeper with a higher ceiling, physically integrating foam into wave crests, while GLM applies a trailing white glaze texture; describing issues plus arg tweaks gets perfect water from Astra.
More from Models
- GPT6 Astra reasoning tiers tested: Max takes 20 min vs 5 min on medium — bdsqlsz · 2026-09-06
- What comes after GPT-6 Astra: from chatbot to a coworker you leave running for a week — johnseach · 2026-09-06
- Claim: OpenAI killed Sora and Atlas to go all-in on Astra and coding models — sahilypatel · 2026-09-06
- Pangram reveals full technical details: training, dataset, architecture and evals — zainhas · 2026-09-06
- GPT Astra 6 powers an intubation training sim no other model could build — BraydonDymm · 2026-09-06
- "Astra is good but too token-heavy": user calls for a GPT-6-grade Sol — CtrlAltDwayne · 2026-09-06