base model behaviors like truesight (ascertaining surprising amounts of information relevant to the author of a passage [including their name, at times] when it’s never included in the context) is both: (1) far more interesting, because it leads you down a rabbit hole of follow up curiosities to search for other unknown emergent phenomena (2) needed to advance the science of evaluating model capabilities for safety
[…]
Structure can carry implications which can carry knowledge discretely - discretely if you can’t keep track of what’s being carried along by the structure
The structure of the consistent use of correct punctuation transports with it the implication of a certain type of guy author
If you can work in the space of nested implications, you can cover a lot of ground while being robust to a changing context window
a general structure allows for many instantiations
the structure and interactions with structures is a way mental math might be done in the models’ language - interacting probability masses about who the author might be and what the world might be