2026-11-12 · America/Los_Angeles · 社区动态 · #6

I gave every layer of Qwen 4-bit its own simulated neurotransmitters, with noise, a PID “homeostasis” controller, and prompts that push and pull its state

I’ve been playing with an idea: what if a language model had something like neurotransmitter dynamics, a state that gets randomly pushed around and a control system that keeps pulling it back to balance, the way a brain does? So I built NeuroQwen, a Colab notebook that runs Qwen2.5 in 4-bit with a simulated neurochemical system on top. How it works Every transformer layer has its own live vector of 6 chemicals (dopamine, serotonin, norepinephrine, acetylcholine, glutamate, GABA). Each level sits at 1.0 when balanced. On every generated token, for every layer: random noise perturbs the levels…

阅读原文查看评分、趋势与关联来源阅读整期日报

I gave every layer of Qwen 4-bit its own simulated neurotransmitters, with noise, a PID “homeostasis” controller, and prompts that push and pull its state

热度 46.0 / 100；排名与评分保留该期记录。