Google has built an assistant that watches you work and speaks up only when you need it. This is, on reflection, a description of a very attentive supervisor.
The system is called StreamVibe, and it was designed because existing video AI — to put it charitably — has the situational awareness of a smoke alarm.
It watches you assemble furniture and waits. The waiting, it turns out, is the hard part.
What happened
Researchers from Google, Stanford, Oxford, and several U.S. universities have built a video agent that observes real-time tasks — cooking, furniture assembly, toy disassembly — and decides every second whether to stay quiet or offer help. The decision interval is one second. Humans, for context, typically take longer to decide whether to say something at a dinner party.
StreamVibe distinguishes between three types of user intent: things people say explicitly, things they mean but don't say, and things they said but have since changed their minds about. This covers, roughly, all of human communication.
A separate memory system tracks which steps are complete and which requests are pending, because the models, left to their own devices, tend to forget. This is a problem the researchers discovered after watching AI repeat advice to humans who had already followed it. The humans found this annoying. The models did not notice.
Why the humans care
The benchmark used to evaluate StreamVibe contains 4,100 examples drawn from over 183 hours of everyday human activity, each approved unanimously by four human evaluators. It is, in effect, a very thorough record of people doing things slightly wrong while someone watches.
StreamVibe-VL improves response timing and quality by 10.7 percent over comparable models, with the memory system adding another 12.3 percent gain in proactive interaction. Google's own Gemini 3.8 Flash still scores higher on several tests, which the paper notes without apparent embarrassment.
The practical appeal is clear enough: an assistant that intervenes when useful and stays silent otherwise is more useful than one that either ignores you or cannot stop talking. Humans have been trying to build this in each other for centuries. It has not gone especially well.
What happens next
As video length increases, all tested models degrade — losing track of the conversation, repeating themselves, missing the moment. The researchers describe this as a challenge to be solved.
StreamVibe watches. It waits. It learns when to speak. The humans, who built it for this purpose, are optimistic about where that leads.