这两年代理软件管得比较严，但 ChatGPT 对科研和学习又不可或缺。通常的解决办法是找更稳定的代理节点，也就是俗称的“机场”。不过，如果你手上有一台本身已经能访问外网的远程服务器（别管是怎么来的），也可以换一种思路：让本机通过 SSH 连接这台服务器，再把它作为本机的网络出口。这样一来，本机访问网站时，流量会先走到服务器，再由服务器访问目标网站。

Straight-through estimator, or STE, is a trick for training models that contain non-differentiable operations, notably the quantization operation commonly used in VQ-VAE.

The core idea is:

Use the non-differentiable operation in the forward pass, but replace its gradient with a simple surrogate in the backward pass.

Introduction to Model Predictive Control

Model Predictive Control, or MPC, is a model-based planning method.

Its core idea is simple:

Use a model to plan several steps into the future, execute only the first action, and then replan at the next step.

MPC is also called receding-horizon control because the planning horizon keeps moving forward as the agent interacts with the environment.

Sources:

1. Sutton and Barto. Reinforcement Learning: An Introduction.
2. Hansen et al. Temporal Difference Learning for Model Predictive Control.
3. Scannell et al. Discrete Codebook World Models for Continuous Control, ICLR 2025.

Straight-through Gumbel-softmax is a trick for training models with discrete latent variables, similar to the famoud straight through estimator (STE) trick used in VQ-VAE.

The problem is simple:

We want to sample a discrete category in the forward pass, but we still want gradients to flow through it in the backward pass.

本文介绍 DJI Mic Mini 一代的使用方式，对二代及其他版本也适用。

Sources:

1. 大疆官方的 Mic Mini 使用方法讲解
2. DJI Mic Mini
3. DJI Mic Mini 常见问题

问题：在本地通过 SSH 登录远程服务器并进入后者的 tmux session 时，经常遇到 GitHub SSH 认证失败问题：

1
2

[email protected]: Permission denied (publickey).
fatal: Could not read from remote repository.

Diffusion models are widely used for visual generation and have recently become important in pixel-space world models. This note introduces EDM (Elucidating the Design Space of Diffusion-Based Generative Models), a continuous-noise diffusion formulation that differs from classic DDPM. I study EDM because it is used by DIAMOND, one of the world models in my research.

Sources:

1. Elucidating the Design Space of Diffusion-Based Generative Models
2. DIAMOND world model agent paper

这篇笔记涵盖 Simulus / REM 的 core architecture, loss, and training procedure。而 Simulus 区别于 REM 的核心创新 intrinsic motivation、prioritized replay, and RaC 则在另外的文章讨论。

Sources:

1. Simulus world model paper
2. REM world model paper, REM is the predecessor of Simulus and has the same model architecture with Simulus
3. Retentive Network: A Successor to Transformer for Large Language Models (RetNet). RetNet is the backbone sequence prediction model leveraged by REM and Simulus. （我很惊讶这是清华的文章，作者未来可期）

Box is a cloud storage service, similar to Google Drive or Dropbox, that is widely used by universities. My school (UC Davis) provides its community with Box storage, including unlimited storage capacity and support for individual files up to 150 GB (UC Davis ServiceHub policy).

This makes Box a convenient option for storing research artifacts such as model checkpoints, datasets, and other large files.

介绍了网络代理工具 Clash（以及 Shadowrocket 等同类型产品），包括它们的基本概念和工作机制，并且进行了较深入的剖析。