02 / BLOG02 / 技术文章

Technical Notes技术笔记

Concepts, methods, derivations, and implementation notes.核心概念、技术方法、公式推导与工程笔记。

CONCEPTS / METHODS / DERIVATIONS / IMPLEMENTATIONS核心概念 / 技术方法 / 公式推导 / 工程实现06 ENTRIES篇文章

Writing & working notes文章归档与工作笔记

6 ENTRIES篇
LANGUAGE /语言筛选 /
2026.10Embodied AI中文

两万字学习 π0.7:从多模态输入到实时动作生成

结合推理时间线与训练前向数据流,拆解 π0.7 如何将多视角图像、视觉子目标、文本提示与本体感知编码为连续与离散联合动作控制。

2026.10Mechanistic InterpretabilityEN

Neural Networks Are Already Clean

The cause of polysemanticity in early model organisms wasn’t superposition, but several ordinary, mostly fixable issues. A working hypothesis from Goodfire on neurons, generalization, and circuits.

2026.04Embodied AI中文

VLA 真的通用吗?—— 机器人具身策略的泛化性与记忆

结合斯坦福 MSL 实验室 Mac Schwager 团队多项实证研究,剖析当前 VLA 与扩散策略的“锁定效应”、动作片段查表记忆化、稀疏自编码器诊断与元认知破局路径。

2026.03Representation LearningEN

Constructive Self-Supervised Learning (Part 1): Designing Generalisable Deep Self-Supervision

Designing generalisable deep self-supervision by predicting across an abstraction hierarchy rather than dispersing lower-level details—introducing cI-JEPA (+6.02% linear-probe top-1 over I-JEPA on ImageNet-100).

2025.11Embodied AI中文

VLA 架构解析:告别“黑箱”与“空谈”,探寻自动驾驶的“大脑”进化史

从 VA 端到端黑箱、快慢双核 VLM+VA 过渡形态到全可微 VLA 统一大脑,系统拆解视觉(SigLIP+DINOv2 / 3DGS)、语言(LLaMA / MindGPT)、动作解码(Diffusion Transformer)与四阶段演进路线。

2025.01Representation LearningEN

Thinking About a Bias for Better Semantic Coherence and Reuse

Why current generative next-token prediction and subtractive latent SSL struggle to bias neural networks toward unified, reusable semantic invariances across levels of abstraction.