A dissertation in which the agents held parties
Joon Sung Park built a small town. In his Stanford doctoral work he populated a simulated place called Smallville with AI agents that carried on lives, formed routines and, unprompted, held parties. It was a beautiful result and it was, in the way of doctoral results, a demonstration rather than a product. On 30 July his company Simile raised more than 200 million dollars at a 2 billion dollar valuation, led by Greenoaks with Index, Hanabi, Bain Capital Ventures, A*, Factory, Definition and CVS Health Ventures alongside. That is five months after leaving stealth with a 100 million dollar Series A led by Index.
What Simile sells is the town, pointed at your customers. It builds agentic twins, generated stand-ins for real consumers that a company can survey instead of running the study on people. The stated ambition is to simulate all eight billion people on earth, accurately and honestly. Set the ambition aside; every founder needs one. The thing that needs deciding in your own building is narrower and it arrives sooner than the eight billion do.
What 300 million dollars in twelve months is actually pricing
A roughly one-year-old company has now taken about 300 million dollars in total and carries a 2 billion dollar valuation. In December, Aaru, another synthetic-research startup, raised a Series A at a 1 billion dollar valuation. Inside eight months the category has attracted at least 3 billion dollars of paper value on a single premise: that asking a model about people is a usable substitute for asking people.
That premise is not absurd. Panel research has been degrading for years on response rates and professional respondents, and a well-calibrated simulation can beat a badly-run survey. But notice what the money is priced against. Market research exists in the first place because people are inconsistent, and the value of a real respondent is not that they are accurate, it is that they are outside your model. A twin generated from your data cannot surprise you with something your data never contained. Investors are not paying 2 billion dollars for accuracy. They are paying for the removal of a cost, and the cost being removed is the friction of contact with reality.
The check you stop being able to run
Here is the trap, and it is a loop rather than a mistake. A synthetic panel is validated by calibrating it against studies you already ran on real people. Good. You then use it instead of running new studies on real people, because that is the entire economic proposition. Six months on, your market has moved, your customers have re-sorted, and the calibration is stale. Nothing in the system can tell you that, because the only instrument that could detect the drift is the study you replaced.
This is the same structure as a monitoring system whose alerting depends on the service it monitors. It works beautifully until the day it matters. The failure is silent by construction, and it will not present as a bad answer; it will present as a confident answer that used to be right.
The countermeasure is unglamorous. Keep a real study running in parallel, small and on the same question, at a fixed cadence. Two hundred real respondents a quarter is not a research programme, it is a calibration instrument, and it costs a fraction of the licence. Ground it against something outside both systems where you can: national statistics offices publish the population structure you are supposed to be modelling, and a twin population that drifts from the real one on age, region or household size is telling you something before any answer does.
A marquee customer sitting on the cap table
CVS Health is identified as a flagship Simile customer, and CVS Health Ventures took part in this round. There is nothing improper in that; corporate venture arms routinely back the tools their operating businesses adopt, and an investor-customer often gets better support than anyone. It is still a fact that changes the weight of the reference. A customer with equity has two reasons to be satisfied, and only one of them is the product.
The instruction is simple and it generalises well beyond this vendor. When a supplier's headline reference appears on its own cap table, ask for a second reference that does not, in your sector, at your scale, willing to talk about what did not work. If the second reference is hard to produce, that is the answer to a question you had not asked yet.
Where sunk cost enters, and how to keep it out
The dangerous moment is not procurement. It is month seven, when the licence is paid, the turnaround has dropped from six weeks to two days, and someone proposes spending real money on a real study. In that room the validation looks like paying twice for an answer you already have. That argument is nearly always the one that wins, and it is exactly backwards: the licence is the sunk cost, and the study is the only thing generating new information.
Take the decision out of that room in advance. Fund calibration from a separate budget line that is not owned by the team using the tool, define the cadence and the divergence threshold in the contract, and agree now what result would make you stop. A tool that cannot be falsified is not a research instrument, it is a house view with a subscription fee, and the difference only becomes visible on the decision that goes wrong.
Read next: The One Bet That Could Not Be Margin Called | A Former CEO Paid 3 Million Of His Own Money
