Simile's simulated shoppers: two 85 percents, one budget line

Two numbers in latent.space's write-up look like the same number and are not. The commercial figure is 85 to 99 percent agreement with human focus groups across tens of millions of simulations for Fortune 100 buyers ^[10]. The research figure, from Park's digital twins of 1,000 real people, is 85 percent as accurate as those people were at reproducing their own earlier answers ^[1]. The second is scored against human self-consistency, which is not a ceiling of 100. Anyone reading the first number as a straight hit rate against ground truth is reading a different measurement than the one the underlying work reports.

Then the band itself. At 85 percent agreement, roughly one answer in seven diverges from the human panel; at 99 percent, one in a hundred. The disagreement rate at the bottom of the quoted range is about fifteen times the rate at the top ^[9]. That is not a tolerance, it is two different products sold under one figure. A concept screen that only needs rank order survives one in seven; a claim test that decides packaging copy or a price point does not. The source does not say which question types land where in the band, and that mapping is the whole procurement question.

The mechanism is more interesting than the round. Park's position is that prompting frontier models does not get you there, and that capturing what he calls social physics may require changing model weights ^[3]. Models optimised to behave rationally make poor simulations of people who do not ^[11]. The inputs are long-form interviews, observational and transaction data, and post-training on randomised controlled trials ^[4], on the argument that web text records what people say rather than what they do ^[12]. If that is right, the defensible asset is the collection apparatus and the post-training run, not the inference call. A competitor with an API key and a persona prompt is building a different thing, and cheaply.

That also sets the honest limit. The interview lists evaluating simulations, rather than stacking model hallucinations, as an open problem in its own right ^[5]. A vendor whose accuracy claim is expressed against human panels needs human panels to keep proving the claim, which means the cheapest version of this business still buys some of the thing it replaces. Park frames market research as the starting point rather than the market ^[6], which is the usual signal that the near-term revenue and the long-term story are not the same slide.

One more thing worth naming: the write-up calls it a $2B Series B backed by GreenOaks and Index Ventures, with Fei-Fei Li and Andrej Karpathy among the backers ^[2], and does not say whether $2B is money in or the valuation attached to it ^[8]. Those are very different companies. Until that resolves, the number to hold onto is the fifteen-fold spread inside the accuracy claim, not the one with the B on the end.