AI model CPT-6 Astra populates a fictional world in Unreal Engine and lets them talk to each other

Even seasoned professionals will be surprised by trying advanced AI models. For example, former HyperWriteAI director Matt Shumer describe the situationDuring this time, he used GPT-6 Astra to create human characters united by a mission of survival in an Unreal Engine environment, and the model not only responded to this task, but also proactively gave the characters the ability to talk to each other.

Image credit: Unsplash, Zach Wolf

In fact, Schumer was most surprised that the AI ​​model was so self-contained, since he didn’t directly instruct it to provide the character with the described communication abilities. One day after setting up the main task, the authors of the experiment noticed that the room where the computers were operating began to make sounds. As he approached the computer, he realized that the characters in the simulation had begun talking loudly to each other.

According to the authors of the experiment, since the heroes of this virtual world had to work together to survive, they had conversations about related topics. They discussed creating tools, coordinating work within teams, and helping each other. The authors of the experiment then decided to introduce another element to the simulation, placing the computer in a fictional world to create a “nested simulation.” One of the characters sits in front of this virtual computer, creating a simulation within a simulation. The first simulated artificial intelligence agents created were not constrained in their choice of means to solve problems and independently made consistent decisions. Participants in the offline discussion were deeply inspired and excited by these experimental results.

Furthermore, the experiment showed that GPT-6 Astra is truly suitable for independent problem solving from start to finish in various areas of intellectual activity. To get the final result, she doesn’t need to clarify anything with the user at every stage. At the same time, the AI ​​agent only created the internal simulation at the direction of the experiment author, rather than actively creating it. In other words, AI models still follow a certain framework. In any case, such a result would not be achievable for a few years, but now at the current rate of development of generative artificial intelligence, it is entirely possible.