NVIDIA reposts a GPT-like motion model that reproduces clips with 99.98% success
Syntetisaattori · x · 2026-07-25
- A repost claims NVIDIA trained a GPT-like model that generates human motion instead of text.
- The approach is described as next-token prediction over body-motion tokens.
- The post says the model reached 99.98% success reproducing a large corpus of motion clips, learned skills like cartwheels and flips from raw data, and adapted to new tasks with less than 1% extra parameters.
- If accurate, the demo suggests a highly parameter-efficient way to generate and control motion in physics-based environments.
More from Research
- Manifold Muon offers a loss-free path for training MoE routers — tokenbender · 2026-07-25
- Practical multi-agent orchestration for Codex splits work into scout, worker, and coordinator roles — pvncher · 2026-07-25
- MOJO preprint mixes supervised and self-supervised losses for neural foundation models — hugo_larochelle · 2026-07-25
- AI can scan more code than humans, but engineers still own quality — ingliguori · 2026-07-25
- A new AGI essay argues the field is climbing the same mountain from two slopes — op7418 · 2026-07-25
- Paper shows models implicitly optimize for hidden graders in audit-style tasks — Sauers_ · 2026-07-25