Small overlays could add per-user facts to AI models
A user describes LongCat experiments in which n-gram blocks can be edited using a cheaply derived, few-megabyte overlay.
TLDR
A user says n-gram blocks in models like DeepSeek V4.1 Flash can be edited with a cheaply derived overlay a few megabytes in size; the experiments are on LongCat. The user says the edits introduce new facts as subject–relation–object triples—a primitive approach that could allow per-user continual learning.
Combined views
10.5K
1 Source, first seen ago
