before it answers? 模型是否在回答前
就知道答案？ 02 Can representations
move between models? 表征可以在不同模型
之间移植吗？ 03 How much behavior can
one neuron carry? 一个神经元可以承载
多少行为？ 04 Can human attention
teach vision models? 人类注意力能否
指导视觉模型？
Independent AI research · amrmatrix 独立人工智能研究 · amrmatrix
Understand models
from the inside.
从内部理解模型。
Mechanistic interpretability, representation editing, and human visual attention — experiments that ask what models know and how their behavior can change. 机制可解释性、表征编辑与人类视觉注意——追问模型知道什么，以及它们的行为如何被改变。
Latest releases 近期发布
Research notes 研究笔记
- Loading research index… 正在读取研究索引…
Working surfaces 工作界面
Code, experiments, interfaces. 代码、实验与界面。
amr_wtf
GHOST — extracting, comparing, and editing behavior vectors inside language models. GHOST——抽取、比较并编辑语言模型内部的行为向量。
ScanThink
Eye-tracking guided vision models and human attention signals. 由眼动数据与人类注意信号引导的视觉模型。
Model interface 模型界面
A restrained interface for direct model interaction and evaluation. 用于直接模型交互与评估的克制界面。
About 关于
Small, inspectable experiments over grand claims. 用小而可检验的实验，替代宏大的断言。
amrmatrix is an independent research practice by chenmoacr. The work focuses on locating internal signals, testing causal interventions, and documenting failures as carefully as successes. amrmatrix 是 chenmoacr 的独立研究实践，关注内部信号定位、因果干预测试，并像记录成功一样认真记录失败。