Can LLMs infer user needs better than owned behavioral data?

Benedict Evans argues that large language models mark "a step change in automated understanding of both what and why" compared to today's recommendation engines. He contrasts Amazon, Instagram, TikTok and YouTube — which only have "correlation across SKUs, and down the funnel" but "don't know why, just as a dog knows that the sound of door keys has high correlation with a walk, but doesn't know what keys are" — with an LLM that "can look at those words, images, videos and products... and connect them to patterns that have some kind of understanding." His example: Amazon's purchase-correlation engine can suggest bubble wrap after packing tape, and might stretch to inferring you're moving and suggest lightbulbs and smoke detectors, but an LLM "might know to show you an ad for home insurance and broadband, things you probably couldn't infer from Amazon's purchasing data."

Evans's mechanism for why this changes competitive structure is that a platform no longer needs its own large user base to generate this inferred understanding, because "if this kind of knowledge generalises enough, it might just be an API call from a world model. You can rent the cold start." He reframes this as moving "the human in the loop": the mechanical turk behind the inference is no longer a company's own users clicking and buying, but "the creation of all that training data over the past few hundred years" that trained the general-purpose model. The effect, in his words, is that "you shift the point of leverage" — competitive advantage moves from who owns the most first-party behavioral data to who can best call a general inference layer.

This sits in direct tension with Can language models discover what users actually want from activity logs?, which gives empirical backing to Evans's contrast between next-item correlation and journey-level "why" — his own YouTube example ("you watch car chase videos, here's a video that seems to have a car chase") is the same next-item logic that research finds recommenders default to even when richer goals are inferable from the same activity. His optimism that LLMs will correctly infer unstated needs (moving house → insurance, not just bubble wrap) should be read against Do large language models fabricate user attributes beyond available evidence?, which finds current models fabricate a large share of their inferences about users rather than reliably intuiting them. And the premise that an inference API needs no platform access of its own mirrors Can LLMs predict demographics from social media usernames alone?: in both cases the behavioral surface required to infer something about a person turns out to be smaller than the platforms that built walled gardens around "their" data assumed.

The essay is a single commentator's conjecture, not a study: it offers no measurement of how much recommendation or ad revenue actually shifts, no named product shipping this rented-inference pattern, and no evidence on how often an LLM's inferred "why" is correct rather than confabulated — a question the over-inference research above answers far less optimistically. Its implication holds only at the strength of a plausible argument: recommendation businesses built on first-party behavioral scale become less defensible to the extent cross-domain "why" can be bought as an API call, and the harder design problem shifts from building better recommenders to building new filters for an internet of infinite content — a question Evans says is "bigger than replacing Google."