Musk explains Tesla's in-car world model: UVE features + ~1B diffusion model
mayfer · x · 2026-11-22
Elon Musk detailed the architecture behind Tesla's in-car world visualization:
- Tesla has had a pixel-perfect world model since last year, critical for synthetic data generation.
- The much smaller in-car version is a conditional mean-flow model over Tesla Universal Vision Encoder (UVE) features — the same features feeding the 3B encoder that produces driving actions.
- A separate 1B diffusion model forks off these UVE features, noising the previous reality-view generation and denoising to the current visualization.
- It outputs small video chunks (8 frames at 36fps, UI-screen resolution), using half a TRIP engine per UVE frame at 18Hz, spread over 4 cycles.
- Latency is currently compute-bound, but interleaving newer UVE features into the diffusion transformer layers should cut 35% of it.
Quoter mayfer adds that since the video shares the driving vision encoder, it acts as a fully open-ended "latent interpretation" arguably better than vector representations.
More from Embodied
- China unveils a bionic fish robot that swims like the real thing — TinfoilTricorn · 2026-11-22
- Moore Threads rumored 64GB consumer GPU "Lushan" would double RTX 5090 VRAM — CeFurkan · 2026-11-22
- Musk: Optimus will be Tesla's finest product ever and 'the best product of all time' — CyberRobooo · 2026-11-22
- Musk amplifies FSD V8 vs V15 comparison: bounding boxes evolve into photorealistic Reality View — elonmusk · 2026-11-22
- Beijing's Hokmind unveils lifelike bionic humanoids with emotion AI that can't walk yet — TinfoilTricorn · 2026-11-22
- Musk claims Tesla AI4 can play Diablo, already at Torment II — elonmusk · 2026-11-22