Why 160ms? Boson AI explains the step size behind its Realtime Duplex voice agents
smolix · x · 2026-11-07
Alex Smola of Boson AI published a detailed rationale for why Higgs Realtime Duplex uses a 160ms decision step:
- Not sensing: light reaches visual cortex in 40-70ms, sound in 15-30ms
- Not time resolution: humans distinguish event order at 20ms apart; auditory gap detection is a few ms
- Not action: the drumming record is 1,208 strokes/min (50ms per stroke)
The real anchor is decision speed: recognizing an animal in a photo takes 150ms, simple reactions 200ms. Conversation turn gaps average 200ms across languages, yet producing a single word takes 600ms. Each 160ms step in Duplex is one decision (speak, stay quiet, backchannel, yield, interrupt); audio frames and face latents go out within a step without re-deciding.
Faster isn't feasible: each step is a forward pass, so smaller steps mean smaller batches and worse GPU utilization; slower adds turn-taking lag. The post also cites Google's 2009 finding that +100ms latency cost 0.2% of searches, and Core Web Vitals' 200ms "good" threshold.
Related event: Why Higgs Realtime Duplex Uses a 160ms Decision Step(3 posts)→
More from coding & agent
- One folder that Claude reads first: the underrated Claude setup for solo operators — PrajwalTomar_ · 2026-11-07
- Linear's AI code review tool Reviews now free on all plans, no Linear tracking required — julianlehr · 2026-11-07
- Amazon blocks home-IP personal shopping agents; dev open-sources residential browser relay — unixterminal · 2026-11-07
- Stately ships a SKILL that has LLMs author agent state machines through 5 validation gates — DavidKPiano · 2026-11-07
- The AI agent testing landscape: 7 tools mapped across 5 distinct testing jobs — njyx · 2026-11-07
- AI rotoscoping workflow today: LTX 2.5 Alpha LoRA + SAM3 + BiRefNet, layered ProRes — PurzBeats · 2026-11-07