Back-of-envelope: 1M context can hold 16+ hours of per-minute visual snapshots

YouJiacheng · x · 2026-11-23

YouJiacheng does the math: an image tokenizes to 256 tokens, and with modern LLMs' 1M context, allocating 256k tokens to visual snapshots yields 1000 minutes—over 16 hours of visual memory, far larger than any existing VLM's memory. This counters the claim that visual tokens are too expensive to scale context for robot history.

Related event: VLA Memory Design: LLM for Long-Term History, Sliding Window for Control(3 posts)→

Original post →

More from Embodied

Embodied channel →