Frozen 72B model's 'lived experience' stored as internal state allegedly changes later decisions
Nearby_Indication474 · reddit · 2026-11-12
The AkbasCore project's TEST878 experiment claims a frozen Qwen2.5-72B writes a lived action-outcome experience to disk as its own internal tensor state—not text—and reloading that file changes the model's decision in a state that never lived the experience. Two runs on a single A100 on different days produced identical numbers.
- The "cognitive cartridge" (Experience-CUT3) is pitched as a third memory path beyond RAG text memory and fine-tuning: it stores layer-3 output plus KV tensors of layers 0-3, each a 12.4MB SHA-256-sealed file
- No weight changes, no LoRA, no optimizer; the frozen model recomputes layers 4-79 on reload
- A 33-page technical paper, code, and GPU logs are on Zenodo
Caveat: self-published, non-peer-reviewed personal project—treat conclusions with skepticism.
More from Research
- Teknium on Unify's memory research: Hermes shows both memorization and ~15% gain over finetuning — Teknium · 2026-11-12
- Robotics take: get intelligence first, then distill to 27B sparse MoE — ChongZzZhang · 2026-11-12
- Meta releases MetaEncoder: one bi-encoder for retrieval and decision-making via natural language — pmttyji · 2026-11-12
- Physics-informed learning by IITM professor slashes AI simulation compute costs — ravi_iitm · 2026-11-12
- InvisibleMentor: AI That Watches Your Screen Recording to Suggest Better Workflows — burkov · 2026-11-12
- SteerDistill: closed-loop teacher steering recovers quantized LLM agent performance — Genming Cui · 2026-11-12