Computer Use vs. Coding Agents: Which Is the Future?
Computer-use agents are improving quickly, but code and APIs may be the more efficient long-term interface for functional work.
Read articleNotes & working ideas
LLM Handyman
A practitioner's notes on training, inference, and LLM systems.
Archive
All articles
Trie Packing — Forward, Loss & Backward
A practical guide to sharing prefix computation across training sequences, preserving the loss, and accumulating gradients through ordinary backpropagation.
Read articleWe're Not at Yahoo Yet. We're Waiting for the Browser.
Zeng Ming's three-stage framework for AI industrialization—and why the native application era has not arrived yet.
Read articleQwen3-VL Video GRPO Training Pipeline
An interactive visualization of the Qwen3-VL GRPO training pipeline.
Read articleQwen3-VL Video SFT Training Pipeline
An interactive visualization of the Qwen3-VL supervised fine-tuning pipeline.
Read articleActivation Functions
A practical reference for activation functions and gated feed-forward networks used in modern language models.
Read articleBasic Math Used for Loss Functions
A practical introduction to entropy, cross-entropy, KL divergence, and softmax in language-model training.
Read articleHow autograd works
Explore how the forward pass builds a computation graph and the backward pass applies the chain rule to calculate gradients.
Read articleMatrix Multiplication Gradients
A visual guide to backpropagation through matrix multiplication, with worked examples, the scalar chain rule, and Jacobians.
Read articleTransformer Prefill & Decode
An interactive two-layer transformer walkthrough covering tensor shapes, prefill, decoding, and layer-specific KV cache updates.
Read article
No articles in this topic yet.