
AI by Hand: Prof. Tom Yeh's Newsletter Teaches Machine Learning on Paper
Prof. Tom Yeh's AI by Hand ✍️ newsletter has passed 77,000 subscribers and walks readers through the math, algorithms and architectures of modern AI by hand rather than through black-box APIs.
Prof. Tom Yeh's newsletter "AI by Hand ✍️" has passed 77,000 subscribers with a promise that runs against the grain of most AI coverage: readers work through the mathematics of modern machine learning themselves — on paper, in spreadsheets and through interactive diagrams — instead of treating models as sealed black boxes. Its tagline is short: "Math, Algorithms, Architectures, by hand".
Working through the mechanics, not around them
The archive is built around that principle. Individual lessons dissect attention with entries such as "Single vs Multi-Head Attention" and "Self Attention vs Cross Attention", while a dedicated attention series assembles 11 interactive diagrams. An activation chapter was completed with 12 interactive lessons. The Deep Learning Math Workbook has been released in an interactive edition, followed by interactive workbooks for matrix multiplication and the linear layer, plus two visualization tools — Matmul Flow and Self-Attention Flow — that let readers trace how numbers move through a network step by step.
Seminar recordings on frontier systems
Beyond written lessons, AI by Hand publishes seminar recordings covering current models and the algorithms behind them: Kimi 3 and Qwen 3.6, Gemma 4, the training pipeline from PPO to DPO, GRPO and rubrics, the PPO clipped policy loss, Google's Ironwood TPU explained "from bits to HBM", and a comparison of Meta Superintelligence Labs with Facebook AI Research. Agent work is not left out — OpenClaw seminars and a companion study guide on debouncing in OpenClaw sit in the same library. A separate series, "Counting AI by Hand", hosted Mohsena Ashraf on measuring what AI costs, in Excel.
Why the approach matters
As frontier models are increasingly reached through chat windows and APIs, the arithmetic underneath them becomes invisible, and intuition about scaling, precision and training trade-offs gets harder to build. By keeping the calculations small enough to fit on a page, the newsletter gives engineers and students a way to check their understanding with a pencil. The 77,000-subscriber figure suggests that demand for that kind of grounding is not a niche concern.
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.