Michael Bernstein, Stanford professor and Simile cofounder, says products must show users they are superpowers rather than replacements, citing Doug Engelbart's intelligence augmentation.
Simile, which has raised over $300m from Index and Bain Capital, says its bespoke data gives it a durable moat that a thin API wrapper lacks.
Bernstein says AI cannot replace talking to real people in discovery, though it lets founders test far more ideas and take bigger risks.
Stanford Professor Builds AI That Simulates How People Behave
Simile is betting that a foundation model of human attitudes and behavior can replace slow, costly research into how people will respond. Its defense is proprietary data that cannot be bought, plus reliability signals that earn user trust. Bernstein argues the wider AI application layer is still early, so thin wrappers around models will not last.
Takeaways
Simile is a what-if machine for human behavior. It builds a foundation model of attitudes and behavior so companies can test how people would react to a product, price or policy. A CEO had it predict seven of nine earnings call questions, and ten-week consulting work now takes an afternoon.
Bespoke data is the durable moat. Bernstein calls Simile a frontier lab with a product. It collects data from tens of thousands of people every week, and that data is not online and cannot be bought. A company offering only an API has nothing differentiating.
Trust decides whether users adopt AI products. Users need a way to absorb moments when the model surprises them. Simile gives green, yellow or red reliability lights. Buyers trust live demos over benchmarks, and Bernstein suspects accuracy claims such as 97% are made up.
Position AI as augmentation, not replacement. People fear AI will displace them, so products must show users they gain superpowers, following Doug Engelbart's intelligence augmentation. AI cannot replace talking to real people in discovery, but it lets founders test many more ideas and take bigger risks.
Simulation needs the tails, not the average. Post-training pulls models toward the average, while businesses such as a high-end fashion firm need outliers. Simile checks its training data coverage, and customers build trust by asking questions they already know the answers to.
AI applications are still in their MySpace era. Models have been transformed but the application stack has not, and Bernstein expects that to change over the next decade. To escape the pull of Gemini, Claude and ChatGPT, a product must be much better than the defaults.
How Simile came together
Bernstein's group built simulators of online social networks, then a town of generative agents that went viral after it was posted to arXiv.
A follow-up built digital twins of 1,000 Americans from interviews. They replicated people's answers about 85% as well as people replicate themselves two weeks later.
Simile was spun out in 2025 as June Park finished his PhD. Bernstein, Percy Liang and Lainie Allen, the chief operating officer, joined him as cofounders.
The company has raised over $300m from funds including Index and Bain Capital.
Tap a name to go there · tap a line to see why they are linked