Harvard and MIT have created a "Matrix", where 8.3 billion intelligent agents accurately mirror the real human beings around the world.
If you were about to be reincarnated, which country would you most likely be born in?
What kind of career, income, values would you have, and even tiny preferences for the button color of an APP?
In the past, countless sociologists and statisticians would have to work themselves to the bone to figure out this series of probabilities spanning sociology, behavioral economics, and psychology.
The sci-fi scene in the 1999 film *The Matrix* has just become a reality!
Led by PhDs from Harvard and MIT, with more than 40 participants from OpenAI, Anthropic, Google DeepMind, xAI and other institutions, and a total of over 200 top scientists, MatrAIx is released, which builds 8.3 billion agents to simulate the real behaviors of humans all over the world.
In the fierce battle for the throne of large models, the giants who are fighting tooth and nail have this time tacitly joined hands to weave a net —— a "Digital Curtain" that imprisons all human behaviors.
Paper: https://arxiv.org/abs/2608.04205
GitHub Repo: https://github.com/MatrAIx-ai/MatrAIx-Persona-8B
They endow AI agents with personalities:
1. 8.3 billion persona records covering the size of the global population are created, covering 1290 dimensions including background, psychology, ability, behavior and lifestyle.
2. It supports the evaluation of persona agents in four environment types: questionnaires, AI chatbots, web pages and applications.
3. It provides more than 1000 evaluation tasks covering 25+ fields, including business, software, finance and healthcare.
The core skills of traditional user researchers and product managers have depreciated instantly.
At the same time, everyone has to face a question:
When your "personality" is just a self-consistent probability distribution in a 1290-dimensional space, how do you prove that you are still the unique one that cannot be completely simulated?
AI Has Calculated All Human Personalities!
The "Oppenheimer Moment" for Product Managers
Traditional User Research ushered in its "Oppenheimer Moment" on this day.
In the past, when tech giants wanted to launch a new app or new strategy, they needed to spend months recruiting hundreds of thousands of real volunteers of different skin colors and backgrounds, and conduct countless gray tests and offline interviews.
In the silicon-based world of MatrAIx, this process with high barriers, high latency and high cost is compressed into an instant.
The size of the Persona-8B database is 8.3 billion, which achieves a one-to-one precise mirroring of the real total population on the physical Earth at this moment.
They don't need to pay social insurance and housing fund, but they know better than you how to gracefully reject a bad UI design on macOS.
With these 8.3 billion digital natives that can be dispatched at any time, the next step is to put them into social life.
The MatrAIx team has customized four full-path simulation interaction environments for it, called MatrAIx Playground:
Surveys: Test concepts, price sensitivity, and willingness to pay of different groups.
AI Chatbots: Fully record the conversation trajectory, and track the real emotional fluctuations and questioning habits of virtual users when facing AI making mistakes and talking nonsense [1.1.5].
Websites: In this sandbox network, agents search, compare prices, read reviews, and finally make purchase decisions just like ordinary netizens.
Applications: This is no longer a confrontation at the text level. Agents can directly manipulate the desktops of Linux, macOS and iOS, and perform various complex daily software operations through virtual mouse, keyboard and touch. The system will record their file changes, permission changes and operation trajectories without missing a single detail.
In this sandbox, the research team has deployed 1,010 complex evaluation tasks for 25 different fields (covering business, software, finance, healthcare, etc.).
In total, they ran 18,189 large-scale simulated user interaction experiments.
In 400 extremely strict control experiments, these virtual agents driven by the underlying large model achieved an astonishing 91.5% consistency rate with the specified persona!
When evaluating the consistency of personas extracted from real humans, human experts gave a high score of 4.135 (full score is 5 points).
The underlying large models that drive these virtual humans have performed infinitely close to the gold standard of human evaluators:
Claude Opus 4.8: In 93.8% of cases, its evaluation error and the score deviation of human experts are controlled within 1 point!
GPT-5.5: In 79.2% of cases, the error is within 1 point.
This shows that today's top large models have not only mastered logic and common sense, but even mastered the "empathy simulator" of sociology.
They can accurately calculate what subtle anger and disappointment a "50-year-old housewife in the American middle class, conservative and introverted" will show when facing a bug in a tech product.
How on earth are these 8.3 billion digital phantoms "created"?
Behind each "person" is a precise attribute matrix containing 1290 persona dimensions. These dimensions are divided into five core areas: background information, psychological characteristics, professional abilities, behavioral interactions, and daily life.
Data sources include UN demographic statistics, General Social Survey, Wikipedia biographies, Amazon real consumer reviews, Stack Overflow developer surveys...
If it is just random combination, AI will only create "cyber monsters" with broken logic. For example, an existence living in a rural area of Kenya, with only primary school education, but only understands Icelandic, has a Harvard doctorate degree, and earns millions a year.
To solve this problem, the research team built a huge Directed Acyclic Graph (DAG). There are strict conditional dependencies between attributes:
When the system determines the "English proficiency" of a virtual person, it must first calculate the joint probability of the two parent nodes "main language" and "region".
Then add a strict compatibility filter: as long as there is an unreasonable conflict between the parent and child attributes, the judgment is immediately reset to zero, and the combination is erased with one click.
Under the baptism of this set of formulas, the synthesized virtual people not only maintain broad diversity, but also maintain unbreakable logical self-consistency within individuals.
Coupled with hundreds of millions of "real soul slices" extracted from real human historical remains such as Wikipedia, Amazon consumption history, and Stack Overflow developer surveys, every phantom in the Persona-8B database is like a flesh-and-blood person who has lived a real life in a parallel universe.
Have you ever thought that your preference for a certain APP color can actually be reduced to the product of the probabilities of parent nodes?
When personality is accurately measured by 1290 scales, the so-called "uniqueness" of human beings is just a self-consistent probability distribution in the eyes of AI.
The Nihilistic Möbius Strip
This is a technical miracle, but if you peel off the outer coat of efficiency, you will see a creepy truth.
Large models (such as GPT-5.5, Claude Opus 4.8) act as "consumers" and "members of society" in MatrAIx to evaluate and test another "virtual human" driven by the large model.
This is like a person playing the role of a customer with his left hand to buy the bread made by his right hand, and then the left hand gives the right hand a five-star praise.
In this closed loop, real human beings are gone.
If virtual users give a 91.5% high score to a new drug or a social software in the sandbox, can it definitely please the physical human beings in reality who will cry, make unreasonable troubles, and be influenced by weather and hormones?
The most terrifying side effect of this "self-circulating ecosystem" is the complete disappearance of "black swans and souls".
The greatest art, the most subversive business models, and even the most amazing scientific breakthroughs in human history often do not come from the "self-consistency" calculated by 1290 probability scales.
They are often born out of unreasonable obsessions and occasional logical chaos —— those outliers that are regarded as "incompatible" by the DAG algorithm filter and erased with one click.
In the future, if all digital products, policies and content can be tested and optimized by "8.3 billion AI-simulated people", the world will become extremely smooth.
But at the same time, it will become extremely empty and boring.
This is a dull world tailor-made to cater to the "preferences of digital phantoms".
At the end of the paper, the research team maintained the unique restraint and sobriety of scientists:
Virtual users can never completely replace real human beings. For those high-risk decisions and major scientific conclusions related to the fate of society, the personal participation of real users is still irreplaceable.
References:
https://matraix.ai/
https://arxiv.org/abs/2608.04205
https://github.com/MatrAIx-ai/MatrAIx-Persona-8B
https://x.com/MatrAIx2026/status/2085217711781564492
This article is from the WeChat official account "AI Era", author: ASI Revelation, editor: David, published with authorization from 36Kr.