New Meta paper shows that deciding on its own when to interrupt, not just what to say, is where proactive assistants fail

Rohan Paul retweeted a new Meta paper on Oct 21, 2026: proactive assistants, such as AI glasses, get no user prompts and must decide on their own when to speak and what to say, and deciding when to speak is where they fail. The paper introduces ProPerMem, a benchmark built on first-person video plus year-long personal memory averaging 618k tokens with distractors and human-verified trigger windows: a model only passes if it speaks at the right moment inside a window with grounded content, and the best baseline model scored just 15 percent. This matters because proactive assistants are central to next-generation interaction like AI glasses, and this paper is the first to formalize interruption timing as a measurable problem, exposing a major weakness in current models.

• Rohan Paul2026-10-20 20:06