‹ Culture 3.14 · What we ask ourselves
Question · Physical Intelligence
Which decisions should stay local?
The link drops at the worst possible moment, and whatever is decided then was already decided beforehand.
Why it matters
Moving the decision to a server gives access to more context, more compute and better models. It also adds latency and a dependency that fails just when the environment becomes difficult.
Every decision moved off the device is a bet on connectivity in the worst case, not in the usual one.
What we know so far
Safety and stopping stay local without argument. Long-range planning tolerates latency well. Between those two extremes, the criterion that has worked best for us is the time that can pass before the decision stops being valid.
A small local model that knows how to abstain turns out to be more useful than a large remote one that sometimes does not answer.
What remains open
It remains open how to reconcile state when the link comes back and both sides have gone on deciding separately.
And how to evaluate a system whose behaviour depends on a network we do not control.