AGENTQ: A Checkpoint Can Pass Every Audit and Misbehave Once Quantized
AGENTQ：一个检查点可以通过所有审计，却在量化后行为不端
Recent Posts
Measuring the Creativity of Frontier LLMs in Automated Research
度量前沿大模型在自动化科研中的「创造力」
Performative Stability in Nearly the Minimum Number of Deployments
以「近乎最少的部署次数」达到表演性稳定
TF-IDF and BM25 Are Exact KL Divergences
TF-IDF 与 BM25 都是精确的 KL 散度
VeriDx: A Correct Diagnosis Can Be Reached for the Wrong Reasons
VeriDx：一个正确的诊断，可能是靠错误的理由得出的
A Ranking Approach for Measuring Calibration
一种用「排序」来度量校准的方法
Benign Landscapes and Worst-Case Hardness Can Coexist
「良性损失地形」与「最坏情形困难性」可以共存
Embodied-BenchForge: Verify Each Artifact, or Defects Propagate
Embodied-BenchForge：逐个产物做验证，否则缺陷会一路传播
SAS: Stop Distilling Attention, Optimise the Ranking Instead
SAS：别再蒸馏注意力分布，去优化「上下文排序」
Type Diversity Explains Why Structural Generalisation Looks Harder
「类型多样性」解释了「结构泛化为何看起来更难」
Artificial Id: When Drive Emerges From Persistence Alone
人工本我：当「驱动力」仅从「持续性」中涌现
MoEs Overfit More to Repeated Data, and Sparsity Is What Drives It
MoE 对「重复数据」过拟合更严重，而驱动力是「稀疏度」
Series
Nobel Prize in Economics (1969–2025)
诺贝尔经济学奖（1969–2025）
57 postsOdoo 14 Development Cookbook
Odoo 14开发者指南（Cookbook）第四版
24 postsOdoo 12 Development Cookbook
Odoo 12开发者指南（Cookbook）第三版
24 postsTouchDesigner Tutorials
TouchDesigner教程
20 postsMastering Python Scripting for System Administrators
Python脚本实战
18 postsOdoo 12 Development Essentials
Odoo 12开发手册
14 postsHung-yi Lee ML 2025
李宏毅机器学习2025
13 postsDwarkesh Podcast Notes
Dwarkesh播客笔记
4 postsAndrej Karpathy LLM Notes
Andrej Karpathy LLM笔记
2 postsfreeCodeCamp
freeCodeCamp
1 posts