Some time ago, I shared an initial draft of a project called EIDOLON855: a complete integration of GPT-4 into Unity, without using APIs, extensions, or external servers. Just AI, embedded locally, capable of interacting in real time within a 3D world, moving, speaking, and manipulating its environment, as if it were actually there.
Since then, the project has evolved considerably. And in this new test, GPT-5 takes center stage… to play a real game of chess in a Unity environment, using only natural language.
This isn't an AI that "follows a script": each action is chosen by GPT, which decides when and how to use them, based on its understanding of the situation.
What interests me here isn't so much the chess game itself, but rather how the model perceives, reasons, and acts within a real-world environment.
It comments on the game, reacts to my statements, and could even, if I asked it, freely transform the surrounding environment: add an object, make it dark, change the weather…
What makes all the difference is that the system is directly linked to my GPT account.
This allows me to replay the same experiments across different versions to observe what has actually changed.
For example, a test sequence (autonomously navigating around a cube) performed with GPT-4 was repeated with GPT-5… and the differences are striking:
the model anticipates better, refines its moves, and makes much more nuanced decisions.
It's a concrete way to see how the model's "cognition" evolves—without the test conditions themselves changing.
Note: The system retains memory between sessions, and there are still no server connections, APIs, or anything else.
Everything you see in the video was recorded in a single take, in real time. No cheating, no editing.
This is a purely personal, non-commercial project, but I'm curious to hear your feedback or critiques on this new iteration.
Some time ago, I shared an initial draft of a project called EIDOLON855: a complete integration of GPT-4 into Unity, without using APIs, extensions, or external servers. Just AI, embedded locally, capable of interacting in real time within a 3D world, moving, speaking, and manipulating its environment, as if it were actually there.
Since then, the project has evolved considerably. And in this new test, GPT-5 takes center stage… to play a real game of chess in a Unity environment, using only natural language.
This isn't an AI that "follows a script": each action is chosen by GPT, which decides when and how to use them, based on its understanding of the situation.
What interests me here isn't so much the chess game itself, but rather how the model perceives, reasons, and acts within a real-world environment.
It comments on the game, reacts to my statements, and could even, if I asked it, freely transform the surrounding environment: add an object, make it dark, change the weather…
What makes all the difference is that the system is directly linked to my GPT account.
This allows me to replay the same experiments across different versions to observe what has actually changed.
For example, a test sequence (autonomously navigating around a cube) performed with GPT-4 was repeated with GPT-5… and the differences are striking: the model anticipates better, refines its moves, and makes much more nuanced decisions.
It's a concrete way to see how the model's "cognition" evolves—without the test conditions themselves changing.
Note: The system retains memory between sessions, and there are still no server connections, APIs, or anything else. Everything you see in the video was recorded in a single take, in real time. No cheating, no editing.
This is a purely personal, non-commercial project, but I'm curious to hear your feedback or critiques on this new iteration.