Small AI-model overlays could support per-user fact updates

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, though the experiments are on LongCat. The user says this could enable per-user continual learning, limited to subject-relation-object facts. In a separate post, the same user describes Internalizer as a “meta-learned hypernetwork” for DeepSeek V4-Flash, with projections on every layer’s shared-expert MLP.