02 / THE FOUNDATIONS
AI Foundations
From first principles to deep understanding.
Explore how AI works, why its methods were designed this way, and how the ideas connect. Build intuition without losing the mathematics, implementation details, or trade-offs.
- Model architecture
- Training & reasoning
- Inference & efficiency
- Multimodal AI & agents
KNOWLEDGE ARCHIVE
Online Softmax：为什么多算几次，反而能少读写显存？
From stable normalization to blockwise attention: equations, worked examples, original paper figures, memory traffic and engineering limits.
Read the article ↗Four ways to explore
A connected learning path
- First principles. The problem, the intuition, and a concrete worked example.
- Deep dives. Equations, tensor shapes, implementation choices, and practical limits.
- Evolution of ideas. How a method developed and which bottlenecks drove each change.
- Research surveys. A map of the literature, competing approaches, and open questions.