Small-model Apprenticeship
A research direction: can a small local model "apprentice" under a larger cloud model - learning preferences, patterns, and style from CSM memory - and eventually replace it for routine work?
Small-model Apprenticeship
The economics of local-first AI favor small models (7B, 13B): they run on consumer hardware, have low latency, and cost nothing per token. But small models lack the capability of frontier models. The question: can CSM memory bridge the gap?
The hypothesis
If a small local model has access to the same CSM memory that a frontier model built - preferences, decisions, corrected mistakes, procedural lessons - it may perform closer to the frontier model on tasks within the scope of the memory. The memory is the “distillation”: not weights, but context.
What would test it
- Run the same task on a frontier model + CSM and a local model + CSM; measure capability gap
- Measure whether the gap narrows as CSM memory accumulates for the task domain
- Measure whether the local model can handle routine tasks autonomously after sufficient memory accumulation
Status
Research direction. No implementation yet. Infrastructure (CSM memory) exists. The experiment does not.