Zhejian Peng
Welcome to J'Log
“Extend the light of consciousness to the stars.” — Elon Musk
Writing on ranking, recommendation, and coding RL. At xAI I work on the Phoenix ranking model and Grok Coding RL.
Blogs
Research notesPre-Norm vs. Post-Norm: The Identity Path That Stabilizes LLMs
Why normalization placement changes gradient flow — with a runnable JAX stress test
Kimi K3: An In-Depth Look at KDA
Kimi Delta Attention across the sequence, Attention Residuals across depth
WIP-Rec System
An syllabus
Stanford - CS224U, Natural Language Understanding
Spring 2020
Stanford - CS224n, NLP with Deep learning
Winter 2019
Python for Absolute Newbies
Python installation, pip, and some basic functionality
Research / Project
AllMatrAIx
8.3 billion persona agents for testing AI and products.
Difficulty-aware looped transformer
A training and inference setup that spends more loops on harder problems and fewer on easy ones.
Fresh in-store production
Probabilistic forecast plus optimization for what a store should bake and stock. US patent 2022/0318709 A1.
