Research & field notes

A look beneath
the surface.

The ideas we're investigating, the systems we're learning from, and the questions that keep us building.

4 notes

Hyper-personal Cognitive Engine

A proposed architecture for assistance that responds to how we learn.

A closer look at the tokenizer, attention, and training behind a language model.

Exploring preference learning and a more deliberate style of model response.

How a retrieval system searches, checks its sources, and knows when to try again.