Experiments Mapping the "Primitive Layer" in Language Models
Semantic primitives show up in activation patterns across Qwen, Gemma, LLaMA, SmolLM2.

Tests 13 LLMs on 291 people to reveal what's actually baked into model weights.
ML researchers, AI engineers, people curious about model knowledge
Hugging Face Open LLM Leaderboard · LMArena
Semantic primitives show up in activation patterns across Qwen, Gemma, LLaMA, SmolLM2.
Ditches the rendering engine entirely to let a VLM hallucinate the pixels from HTML.
Streams LLM weights from CD-ROM during inference to fit 77MB models in 32MB RAM.
Direct weight editing for persistent memory—MEMIT meets LoRA consolidation with null-space math.
Deterministic fingerprinting for model structure without loading weights.
Uses Google's CLD3 neural net to match Chrome's native API behavior exactly.