Simile’s CEO about his journey from the viral Generative Agents to creating 8 Billion Digital Twins of every living human... and why it’s gone from fun exploration to very serious business.
One constraint on that path sits in EU law rather than in the model. Where behavioural data collected for one purpose is reused to train a model of human behaviour, Article 6(4) of Regulation (EU) 2016/679 makes the controller run a compatibility test first, weighing the link between the two purposes, the context of collection, the nature of the data, the consequences for the people concerned, and safeguards such as pseudonymisation. Accuracy against a human focus group does not settle that question, and a simulated panel built on incompatible reuse is expensive to unwind later.
Sounds like a Capstone course in college. Potentially applicable, but mostly an idea with friction, because it takes a lot of momentum to implement something like this.
What I found most fascinating here is the idea that a good simulation of humans may actually need to reproduce the things that make humans less rational—our biases, mistakes, habits, and seemingly inefficient choices.
The example of walking instead of taking an Uber really captures this. From a purely optimization perspective, the faster option looks obvious. But if that walk is where someone thinks, reflects, or comes up with their next idea, then the “inefficient” choice is actually serving a purpose.
That raises a much deeper question for behavioral AI: "Are we trying to build models that predict what a perfectly rational person would do, or models that understand why real people do what they actually do?"
The latter feels much harder—and potentially much more valuable.
One constraint on that path sits in EU law rather than in the model. Where behavioural data collected for one purpose is reused to train a model of human behaviour, Article 6(4) of Regulation (EU) 2016/679 makes the controller run a compatibility test first, weighing the link between the two purposes, the context of collection, the nature of the data, the consequences for the people concerned, and safeguards such as pseudonymisation. Accuracy against a human focus group does not settle that question, and a simulated panel built on incompatible reuse is expensive to unwind later.
Sounds like a Capstone course in college. Potentially applicable, but mostly an idea with friction, because it takes a lot of momentum to implement something like this.
What I found most fascinating here is the idea that a good simulation of humans may actually need to reproduce the things that make humans less rational—our biases, mistakes, habits, and seemingly inefficient choices.
The example of walking instead of taking an Uber really captures this. From a purely optimization perspective, the faster option looks obvious. But if that walk is where someone thinks, reflects, or comes up with their next idea, then the “inefficient” choice is actually serving a purpose.
That raises a much deeper question for behavioral AI: "Are we trying to build models that predict what a perfectly rational person would do, or models that understand why real people do what they actually do?"
The latter feels much harder—and potentially much more valuable.