AI Resonance · 阅读最新日报 · AI 入门推荐

2026-10-10 · America/Los_Angeles · 社区动态 · #17

AkbasCore MAM-BÇ: I saved a frozen LLM's memory to a 4.7 MB file, reloaded it bit for bit, and the model answered from the file alone, without seeing the text again (Qwen2.5-7B, open code, DOI)

Two days ago I shared AkbasCore MAM on Mistral-7B. Across three platforms it passed 50,000 views, which I honestly did not expect. Thank you. Today's release is the step I promised: the memory now leaves the model as a file and comes back. The short version. A frozen Qwen2.5-7B-Instruct, with not a single weight changed, answers a question from numerical memory. That memory was selected from a bank of 96 facts, written to disk as a 4.7 MB file, read back, checked bit for bit, and plugged into the model again. The source sentence is never given to the model when it answers. Everything is open…

AkbasCore MAM-BÇ: I saved a frozen LLM's memory to a 4.7 MB file, reloaded it bit for bit, and the model answered from the file alone, without seeing the text again (Qwen2.5-7B, open code, DOI)

热度 44.0 / 100;排名与评分保留该期记录。