Amazon knows what you buy but not why you buy it. Google knows what you click but not what it means. Every major consumer internet platform is a Mechanical Turk: a correlation engine with no comprehension. Benedict Evans argues that LLMs represent a structural break from this limitation, because a model that has ingested the written and visual record of human behavior can connect products, actions, and intent in ways that purchase graphs and click streams never could. The example is concrete: Amazon can infer bubble wrap from packing tape, but an LLM can infer you are moving house and serve you home insurance and broadband ads.
The cold start problem changes shape under this model. Historically, Amazon and TikTok had to build their own Mechanical Turks by watching their own users. Evans argues that a sufficiently general LLM collapses that requirement to an API call. You rent the cold start. The human-in-the-loop shifts from the live user base to the training data accumulated over centuries of human output. Tinder is already testing a version of this logic by reading users' camera rolls to infer identity before a single swipe is cast. The argument about how Google, Meta, Amazon, and an agentic LLM each hold a partial view of the user, and why none of them alone holds the full picture, is the sharpest section of the piece.
Evans does not pretend to know how this resolves. He places the current moment alongside the web in 1997 and mobile in 2007: the scale is legible, the mechanics are not. The real question he leaves open is not whether AI replaces Google Search, but whether it becomes a new kind of filter for a world that already has infinite product, infinite media, and infinite retail, and still no reliable way to find what you do not yet know you want.
[READ ORIGINAL →]