Speculative decoding doesn't prove perception stays fresh during long robot reasoning

krishnan · x · 2026-11-18

Part 3-4 of krishnan's thread on edge robot inference: SiMa offers its competitive benchmarking details only by email, and those are vendor throughput claims — not robot-safety or task-completion results. Containers don't change the fact that shared silicon is shared, so vision, depth and reasoning compete for the same compute and memory. His core point: speculative decoding speeds up accepted tokens but doesn't show perception stays fresh while a long reasoning request runs — the first question a buyer needs answered.

Related event: krishnan questions SiMa.ai Palette 3.0's single-chip robot AI claims(8 posts)→

Original post →

More from Embodied

Embodied channel →