AI by Hand: Technical Education via Math and Algorithm Visualization

AI by Hand provides a first-principles approach to understanding AI

AI by Hand is the research publication of By Hand Research, founded by Prof. Tom Yeh. The platform focuses on model interpretability and explainability by breaking down complex AI architectures, algorithms, and mathematics into "by hand" exercises. This method aims to move beyond the use of black-box frameworks, allowing learners to understand the underlying mechanics of frontier AI through manual calculation and visualization.

Educational Framework and Content Pillars

The platform organizes its technical content into two primary tracks: Foundation and Frontier.

Foundation: Core AI Mathematics

The Foundation track focuses on the essential building blocks of deep learning. Key resources include:

  • Interactive Workbooks: Digital versions of the Deep Learning Math Workbook, including specific modules for Matrix Multiplication (Matmul) and Linear Layers.
  • Attention Mechanisms: A comprehensive series of 11 interactive diagrams covering Single vs. Multi-Head Attention and Self vs. Cross Attention.
  • Activation Functions: A dedicated chapter featuring 12 interactive lessons on how neurons are activated within a network.
  • Visualization Tools: Specialized tools like "Matmul Flow" and "Self-Attention Flow" designed to visualize the movement and transformation of data.

Frontier: Advanced Architectures and Algorithms

The Frontier track analyzes the latest developments in large-scale AI models and hardware. Notable topics include:

  • Model Analysis: Technical seminars and recordings on models such as Qwen 3.6 and Gemma 4.
  • Optimization Algorithms: Deep dives into reinforcement learning and alignment techniques, specifically the progression from PPO (Proximal Policy Optimization) to DPO (Direct Preference Optimization), GRPO, and the use of Rubrics.
  • Hardware Architecture: Analysis of the Google Ironwood TPU, tracing the data path from bits to High Bandwidth Memory (HBM).
  • Specialized Components: Technical breakdowns of the SwiGLU activation function and the PPO Clipped Policy Loss algorithm.

Community Perspectives and Alternative Resources

While some users praise the "by hand" approach for its rigor, others in the technical community have highlighted challenges regarding the platform's accessibility and user experience.

Critiques of Accessibility

Some users noted that the platform's UX can be a barrier, with several comments mentioning that a significant portion of the content is locked behind a subscription wall or requires navigating a subscription page before accessing links.

Complementary "From Scratch" Resources

Community members have suggested several alternative or complementary resources for those seeking to build LLMs from the ground up:

  • Code-Based Implementations: Projects such as llm-from-scratch and ml-by-hand (inspired by micrograd) emphasize the connection between calculus and code.
  • Structured Courses: Sebastian Raschka's Build a Large Language Model (From Scratch) is cited as a valuable pairing with explainer videos.
  • Academic Primers: The book Deep Learning: A Visual Approach by No Starch Press is recommended as a foundational primer.
  • Theoretical Papers: The paper Foundations of Large Language Models (arXiv:2501.09223v2) is suggested for theoretical grounding.

Sources

Related