https://img.inleo.io/DQmbF8xuyFxG4KZ5kjUkKcsoTHduRfkzCABXR9AGhiTGWCk/2024-11-30_02-09-31_1067777777.png 

On 15 November 2024, a study titled -  [Generative Agent Simulations of 1,000 People](https://arxiv.org/abs/2411.10109)  was released.

The essence of the research is the construction of individual consciousness of a person, with the help of large language models based on AI. Scientists created 1,052 simulacra of ‘typical’ people. Real Americans selected to represent the US population by age, gender, education, political views, sexual orientation.

The main tool of the experiment was the generative AI of large language models ChatGPT-4o.

# What's the point of the experiment:

- More than 1,000 ‘typical’ Americans are selected according to certain criteria.
- An in-depth 2-hour interview (approximately 6,500 words) was conducted with each.
- Each interview was loaded into the memory of a separate ChatGPT-4o-based AI agent, thereby turning it into a simulacrum of the personality (individual consciousness) of a particular ‘typical’ American or American woman.
- Then they administered to each of more than 1000 people several canonical social psychological tests: the General Social Survey (GSS), the Big Five personality questionnaire, the well-known behavioural economic games (dictator game, public goods game) and five sociological control experiments.
- In parallel, they ran all the personality simulacra of ‘typical’ Americans through the same tests, games, etc.
- The responses and behaviours of the simulacra and real people were then compared, from which the specific simulacra were based.


# Results
- Simulacra predicted the responses of real people on the GSS test with 85% accuracy (which is significantly more accurate than AI agents that used only basic demographic information).
- Of five experiments involving both humans and their simulacra, in four the simulacra produced results that were almost indistinguishable from the responses of their human prototypes (correlation 0.98).
- Simulacra made more accurate predictions across different political ideologies and ethnic groups. They also showed more balanced performance when analysing responses between different demographic categories.

In this experiment, the obtained simulacra of individual consciousness are narrowly focused and do not model all features and aspects of personality. But the beginning has been made. It should be understood that in this research 1 two-hour interview was conducted, but it was enough to get a socially predictable model of individual's behaviour. What would happen if a series of several dozens of interviews were conducted?!

What do you think this technology (creation of typical simulacra) could potentially be used in? - marketing, political companies, social modelling, creation of religions, other options?

In fact, this study is another kick under one soft spot of mankind. If to predict the actions of a typical person, it is enough to conduct 1 two-hour interview with a language model, maybe it's time to finally become smarter, smarter and work on pumping up your own intelligence! Quite an interesting and enlightening experiment, don't you think?

Posted Using [InLeo Alpha](https://inleo.io/@detectblock/potentially-gamechanging-experiment-frb)