Labs burn huge compute on 3D training data while users ask their model to boil an egg
theteknosaur · x · 2026-10-13
Quoting theo: the world's smartest code model won't be good at 3D without dedicated training. Labs have recently received large piles of 3D training data and RL environments to improve spatial reasoning. The jump from Fable 5.1 to Opus 5.5 shows at least 4x better 3D despite being only 5% better overall — Meta clearly didn't go hard on 3D this release, which doesn't make the model bad.
theo also admits fault for amplifying Fishslop demos: the 3D stuff is cool but a poor proxy for real-world coding capability — "3dmaxxing" is just another form of benchmaxxing. The poster quips that labs burn insane money baking spatial reasoning into weights while he uses the silicon god to ask how long to boil an egg.
More from Models
- New Paper Explains Mamba-3's State Rotations as Underdamped Second-Order Dynamics — tri_dao · 2026-10-13
- Developer floats using jev-like models to fill a frontier model's custom 256-parameter embedding space — tobowers · 2026-10-13
- StepFun teases open-source Step 5 Preview, a flagship model for agentic tasks — bdsqlsz · 2026-10-13
- OpenAI reportedly used internal math model to improve itself, halving inference costs — VraserX · 2026-10-13
- Reddit claims Meta shipped frontier model 'Muse Wonder'; users seek real reviews — Kingwolf4 · 2026-10-13
- Cursor's new pricing: Claude and other third-party models billed at API rates, plans include only Grok — iannuttall · 2026-10-13