Building camera-motion measurement to diagnose open world model failures: measure before training
PrateekVJoshi · x · 2026-11-21
When post-training open world models, reliably measuring instruction-following and world consistency is critical. The author built a system that estimates camera motion from generated videos to locate failure modes.
Key findings: some failures don't require more training — a prompt rewrite improved spatial placement, flipping a video reversed the apparent camera orbit (with limitations), and enabling built-in retrieval improved scene consistency.
But none fully solves the underlying problems; the core argument is that measurement should come first.
More from Models
- User says ChatGPT's ultrafast mode is impossible to go back from — TheMoonMidas · 2026-11-21
- House report on PRC models urges national US open-weight AI initiative — evijit · 2026-11-21
- FrontierBioML benchmark: 25 long-horizon comp bio tasks, Claude Mythos 5.2 leads at 40.3 — afurgs · 2026-11-21
- 2M canary keys seeded into cloud LLMs were stolen and abused within a day — AccBalanced · 2026-11-21
- Google ships Neon, its third Gemini model in 8 days, plus expanded Tasks in the Gemini app — GeminiApp · 2026-11-21
- Gemini CLI is actually Antigravity CLI, notes dev; unverified 'Gemini 4 Neon' access questions arise — Angaisb_ · 2026-11-21