The Unboxing Cliff

The new moat in consumer AI isn't model power — it's Day 30. Call it the Unboxing Cliff: the moment a product's novelty collapses and habit has to carry it.

Capabilities are commoditizing faster than founders can refresh demos. The interesting question stopped being "can the robot do a backflip" and became "will someone still open the app four weeks in." A new product now lives or dies on scenario-specific evaluation data — the labeled, narrow feedback loops a device builds in your home, your routine, your kid's bedtime. That dataset cannot be scraped from the public web. It has to be earned, one mundane interaction at a time, by something you do not put in a drawer.

The implications cut two ways. Investors who still underwrite "large model + cool demo" are pricing the old cliff. The ones who already learned the lesson now ask a single question of any companion pitch: what does the product know on Day 31 that it did not know on Day 1, and what makes the user come back to teach it? Whoever owns the answer owns the next category.

The Huxiu roundtable, translated in ChinAI #366, surfaced a 30% return rate for some Chinese companion robots and named the problem plainly. The robots aren't failing. The demos are.