AI's Next Scaling Law Isn't Bigger Models. It's Simulating You.

Frontier AI can write a sonata, pass the bar exam, and ace the medical boards. Ask it whether a thousand real shoppers will switch brands next quarter and it shrugs. The gap between what large language models do well and what decision-makers actually need is the market that Simile is trying to close.

On July 30, Simile closed a $200 million Series B at a $2 billion post-money valuation, co-led by Greenoaks and Index Ventures, per the company's announcement and TechCrunch's reporting. The round landed five months after a $100 million Series A led by Index and pulled in participation from Hanabi, Bain Capital Ventures, A*, Factory, CVS Health Ventures, and Definition. CVS Health is also a customer, alongside Wealthfront, Deloitte, and Gallup.

The bet is a category Park and his co-founder Liang call simulation AI, or foundation models of human behavior. Predicting how people decide, the pitch goes, is a different scaling problem than teaching a model to summarize text. Frontier LLMs scale by adding compute to language. Simile argues that human behavior scales by adding the right kind of data: long-form interviews, observational and transaction histories, randomized controlled trials, and post-training on the causal mechanisms behind decisions rather than correlations in text.

In a long interview with Latent Space, Park, who led the 2023 Smallville study (formally the Generative Agents paper, arXiv 2304.03442, with roughly 7,200 Google Scholar citations), argues that "social physics" demands changing model weights rather than better prompting. The paper showed AI agents that remember, plan, and socialize inside a simulated town. The company is the attempt to run the same trick against real human populations.

The early numbers are narrow but specific. Simile says it has built digital twins of roughly 1,000 people whose survey responses reproduced the original humans' answers about 85% as accurately as those humans reproduced their own answers when re-tested. The company frames a broader 85% to 99% accuracy band against human focus groups. Both figures come from Simile's own research page and Park's own framing in the Latent Space interview, and no independent replication has been published. Park's stated mission, lifted directly from the company page, is to "simulate all eight billion people on earth, accurately and honestly."

That is also where the source itself flags the hardest problem. Like other AI assistants, Simile's models are trained to produce coherent, helpful answers. Real people are biased, moody, and self-contradictory. Park's own argument is that "good simulations need to reproduce human biases and mistakes," the very properties that post-training tends to smooth out. Whether Simile's stack can model irrationality without inheriting that smoothing problem from its base models is the open technical question. Park has not published a benchmark that resolves it. The company says it has a "first-of-its-kind confidence model that predicts the accuracy of every simulation," a useful product claim but not a published one.

The commercial traction is real, if thin in the public record. Simile says it has run "tens of millions of simulations for Fortune 100 enterprises" and grown revenue 5x since launch. Named customers, CVS Health, Wealthfront, Deloitte, and Gallup, are notable logos, not measurable ROI. CVS is the only customer that is also an investor, through CVS Health Ventures. The closest public benchmark for the category is Aaru, a synthetic-research competitor that closed a Series A at a reported $1 billion valuation in December 2025. Adjacent to the bet, Shopify published SimGym in April 2026, a traffic-grounded testbed for evaluating vision-language agents in e-commerce A/B tests. It is the same thesis, narrower in scope, and it does not have Park's accuracy numbers behind it.

The falsifier for Simile's story is the gap between the 1,000-person digital twins Park can show today and the 8 billion-person society-scale simulation he says is the goal. Park himself estimates a society-scale run could require a dedicated data center. The intermediate steps, 100,000-person product tests, million-person election forecasts, region-scale policy simulations, have to land in public, with held-out evaluation, before the company's own "second summer of simulation" framing earns its keep. The headline stands: a $2 billion company, a new category, and a $200 million bet that AI's next scaling problem is better humans, not bigger models.