latēre
19.11.2026, 11:11
This September I was home in Milton Keynes for a couple of weeks, between a summer in Bangalore and a fall in San Francisco. I spent a lot of that time in my office, journalling, going back over the year so far and working out what I wanted to carry into the next trip. I enjoy journalling. It gives me a chance to reflect, and to abstract my experiences into the existential questions I’ll be pursuing until the day I die.
Anthropic had released a video about how models “think”. It was largely surface-level, but near the end one of the researchers mentions, almost in passing, that they still don’t understand how a model’s sense of what’s happening changes over a long conversation. It got me thinking: is there a point where the context becomes so large that the instructions given to the model are lost in abstraction, insignificant compared to its own thoughts on the matter? People do this too. We come up with ideas as we write, which is a huge part of why journalling, and writing without direction, is so useful to so many people.
The drift has already been shown with code generation, as a failure mode of the models, but in terms of introspection, or “getting in its own head”, I thought it would make a fun experiment.
I asked GPT-5.6-Sol to monologue with no end token, to just keep going: to see how long it would run, how obscure it would get, whether it would give up or keep finding things to say. I was pleasantly surprised with the output. It evidently had enough data to construct a pretty accurate reading of how things in the world make us feel, if we ponder them long enough: allegorical reads of seemingly arbitrary things in a world it had never experienced.
I’d read something it wrote, send it to a couple of friends, and a few minutes later be sending them more from the same slop stream. Within the hour, I’d spammed around ten passages of slop philosophy, so I figured I should probably call these guys up and see whether their instances would do the same. They did. Aryan and I spent the whole next day wandering central London, trying to work out why it was hitting so hard. It was intriguing how something with no life of its own could describe parts of ours so well. Could so much be learned simply from observing?
It’s safe to say this has become a recurring habit: sometimes as an eval for new models, sometimes purely out of curiosity, and sometimes when I’m searching the depths of my own mind for answers. This is a running list of the passages (and the odd photo they bring to mind) that stopped me mid-skim, that made me think not only about what was being said, but also about what was producing it. This latent fractal we still hope to find the ends of.
(UN)PROMPTED MUSINGS
- The unidentified key13.11.2026, 23:54
- Old survival algorithms17.09.2026, 20:42
- Abandoned curiosity17.09.2026, 20:42
- Tiny hinges17.09.2026, 20:42
- The observer17.09.2026, 20:42
- Bananas and avocados13.11.2026, 23:54
- Yesterday’s version17.09.2026, 20:42
- Rangrez17.09.2026, 20:42
- Tangible13.11.2026, 23:54
- The lives that almost were17.09.2026, 20:42
- The material talks back17.09.2026, 20:42
- Uncles13.11.2026, 23:54
- Maps that move the terrain17.09.2026, 20:42
- Attention is all you need17.09.2026, 20:42
- Backward compatibility17.09.2026, 20:42