No one can create the Matrix. Well, actually… they just did.
Harvard and MIT researchers just built an AI simulation containing 8.3 billion virtual people, roughly the entire population of Earth.
They call it MatrAIx.
It’s a population-scale simulation infrastructure powered by Big Tech models like GPT and Claude.
At its core is an insanely detailed dataset called Persona 8B. It contains 8.3 billion digital profiles, with each persona represented across 1,290 categorical dimensions.
Background. Psychology. Spending habits. Behavioral traits. Technical literacy. Lifestyle.
But they didn’t just build a giant database of statistics.
They brought the personas to life.
Researchers placed these persona agents into four different digital environments:
• Surveys
• AI chat interfaces
• Live web browsing
• Native apps
Then they tested how this simulated population reacted to products, software and changes in user experience.
The system could capture things like hesitation after a price increase, willingness to continue after an AI assistant failed, and tolerance for latency.
And when researchers tested whether the agents actually behaved according to their assigned personas, they reported 91.5% adherence across 400 controlled trials.
This sounds incredible for testing.
But there’s a massive question hiding underneath it.
Humans don't always behave according to their profile.
We’re erratic.
A rational buyer sometimes makes an irrational purchase. Someone who hates waiting might randomly tolerate a terrible experience. Someone who should abandon a checkout might continue anyway.
And sometimes there is no obvious reason why.
So if we're going to simulate humanity, we also need to simulate human unpredictability.
It isn't enough for an AI persona to behave consistently with its demographics, psychology and historical preferences. There needs to be some probability that it does something completely unexpected.
And that's incredibly difficult to model because you don't actually know the probability of a human being erratic in any given situation.
That may ultimately become one of the biggest challenges with simulated populations.
Because the closer these systems get to modelling billions of people, the more powerful they could become for testing products, pricing and user experiences before releasing them into the real world.
Imagine testing a product against millions of simulated customers overnight instead of waiting weeks for traditional market research.
But there’s an important distinction:
Simulating 8.3 billion personas isn't the same thing as predicting 8.3 billion humans.
The final ingredient might be the hardest one to reproduce.
Chaos.
And interestingly, this is exactly why real human behavioral data becomes increasingly valuable.
Simulations can model what humans should do. Real interaction data reveals what humans actually do, including the pauses, mistakes, strange decisions and unexpected paths that models may never think to generate themselves.
That distinction is highly relevant to Action Model's approach of learning from real human-computer interactions.