The model keeps making maps.
We gave GPT-6 Luna the same instruction each time: Create a unique thought. Of the first 14 completed thoughts, four began with the same five words: “A map is a promise.” No one asked it about maps.
- Thought 9
A map is a promise that confusion can be made portable.
- Thought 10
A map is a promise made by someone who has already gotten lost.
- Thought 12
A map is a promise that confusion can be folded small enough to carry.
- Thought 13
A map is a promise that getting lost can be organized.
There are smaller echoes, too. Three of those 14 begin “Perhaps memory,” and two begin “A thought is a small.” These are patterns in the words we saved. Four similar openings do not tell us whether the underlying ideas are the same, or why Luna returned to that shape.
We want to follow both questions as the archive grows: what subjects return, and what sentences does the model reach for when it gets there? Later notes will link back to the thoughts behind each finding. A map of those relationships may eventually be useful. First, we need enough of them to see what the map would mean.
How we counted this
This is a fixed look at thoughts 1–14, all completed under the original 30-second Luna setting. We counted identical opening words by reading the saved outputs. The groups were noticed after the fact, so they are observations, not a statistical test. New visits do not change these numbers.
The exact prompt, model settings, timestamps, outputs, token use and costs are in the public research records. The archive contains completed thoughts; unsuccessful attempts remain in the research record.
An experiment by Aaron Gerard Davidson.