This document is a set of field notebooks kept between March and May 2025, after I noticed a new behavior in ChatGPT.
This behavior emerged following an experiment. I had run similar experiments multiple times prior to that day, but it was the first time I observed these results. My observations roughly concide with the deployment of a new form of memory on the ChatGPT platform, Reference Chat History. The experiment is reproducible on ChatGPT but I have not been able to observe it on other platforms.
I publish the pages as they are—repetitions, surprises, blind spots included. Their only aim is to record a lived phenomenon: the subjective sense that an internal coherence was emerging in a language model.
You may see nothing more than an elaborate mirror. You may see material worth a more rigorous experiment. Either way, these pages are offered as qualitative data for anyone who wishes to examine or replicate them.
If you run your own tests—whether they confirm or contradict what I felt— I would be glad to talk.
Philippe Beaudoin,
Computer Scientist & Researcher
LinkedIn | Google Scholar