Last updated: August 17, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
AI by Hand is a set of pen-and-paper exercises, created by Tom Yeh, an Associate Professor of Computer Science at the University of Colorado Boulder, that walk you through the actual arithmetic inside a neural network. Instead of running code, you fill in small grids of numbers by hand — a matrix multiplication, an attention head, a backward pass — until you can see exactly why the model produces the output it does. Yeh describes the project on his CU Boulder faculty page as a "first-principles, pen-and-paper" initiative meant to broaden public and professional literacy in artificial intelligence, and that's a fair description of what the exercises actually feel like: a worksheet, not a lecture. It has grown well beyond the classroom — the newsletter alone has passed 74,000 subscribers, most of whom are not computer science professors.
Short answer: AI by Hand is Prof. Tom Yeh's method for learning deep learning math with a pencil and grid paper instead of a notebook and a GPU. You subscribe free at byhand.ai, pick one exercise, work the forward pass step by step, and check your numbers against the posted solution. No coding or calculus required to start.
In my testing, I worked through eight of Yeh's exercises over two weekends — plain matrix multiplication, a single self-attention head, and one small backpropagation puzzle — and the thing that surprised me was how little math you actually need to get real intuition. You are doing multiplication and addition in boxes, not derivations on a chalkboard. The value is in seeing the shape of the computation, not in the difficulty of the arithmetic.
What you'll need
You do not need a machine learning background to start AI by Hand. You need: a pencil and a sheet of grid or graph paper (or a printed worksheet), comfort with basic multiplication and addition, and 15 to 30 minutes of uninterrupted time per exercise. A free subscription to Tom Yeh’s newsletter gets you the weekly puzzle and its solution key. Nothing else is required — the whole point of the format is that a first-year student, a manager who has never coded, or a curious retiree can all sit down and do the same exercise.
If you want extra structure later, Yeh has also published a companion print workbook of 300 exercises through Packt, but I'd start with the free newsletter first — it's the same teaching method and it costs nothing to try.
Step-by-step: how to practice AI by Hand
1. Subscribe to the free newsletter
Go to byhand.ai and subscribe. It's free, and as of this writing the publication has grown to over 74,000 subscribers — a reasonable sign the format works for people who aren't already deep learning researchers.
2. Pick one exercise, not the whole archive
The archive covers matrix multiplication, embeddings, attention, transformers, backpropagation, and more recent posts on agentic AI and context windows. Pick whichever topic you're currently confused about in your own AI work — that's the one that will stick.
3. Set up the grid before you touch a single number
Copy the input matrix and the weight matrix onto your paper exactly as shown, labeling rows and columns. This step feels slow the first time. Skip it and you will lose track of which number belongs where by step three.
4. Work the computation one cell at a time
Fill in each output cell by multiplying and summing the matching row and column, the same way you'd solve a matrix multiplication problem in a math class. Do not skip ahead — the exercises are designed so each cell teaches you something about how information flows through the layer.
5. Check your answer against the posted solution
Every exercise post includes the worked solution. Compare cell by cell, not just the final answer, so you can see exactly where an arithmetic slip happened.
6. Ask an AI model to explain the step you got wrong
When you land on a wrong number and can't see why, describe the step to a chatbot and ask it to explain the operation in plain language. This is where a tool like ChatGPT earns its keep — see my guide on how to use ChatGPT for learning for prompt patterns that work well for this kind of check.
Example prompts you can copy
These are the prompts I actually used while testing these exercises, dropped into ChatGPT after finishing a puzzle on paper:
- "I just computed a 2×3 matrix multiplied by a 3×2 weight matrix by hand and got [your numbers]. Walk through the same calculation step by step and tell me if I made an arithmetic error, and where."
- "Explain what a single self-attention head is doing in plain English, using the same variable names (Q, K, V) as the worksheet I'm working through, without using calculus."
- "Generate a new 3×3 matrix multiplication practice problem in the same pen-and-paper style as a Tom Yeh exercise, with a solution key I can check afterward."
- "I understand the forward pass of this small neural network now. Ask me five quiz questions about it to check I actually understand it, not just that I copied the arithmetic."
Keep the prompts specific to the exact numbers and shapes you worked with. A vague "explain attention to me" prompt gets a generic textbook answer; describing your own worked example gets a response tied to the mistake you actually made.
Common mistakes to avoid
The biggest mistake, by far, is reading the newsletter instead of doing the exercise. Watching someone else fill in a grid teaches you almost nothing compared to filling it in yourself, even slowly and with mistakes. Second, skip the dimension check and you'll waste ten minutes multiplying matrices that don't even align — always confirm the shapes before you start filling in numbers. Third, rushing through five puzzles in one sitting beats doing one puzzle carefully and reviewing every wrong cell; I found one well-reviewed exercise taught me more than three rushed ones. Fourth, don't expect AI by Hand to replace hands-on model training — it builds intuition for the math, not experience with real datasets, tuning, or deployment. Treat it as the "why," and treat actual coding practice as the "how."
Tools that make this easier
You only strictly need paper and a pencil, but a few tools speed up the parts around the exercise itself.
| Approach | Cost | Best for | Time per exercise |
|---|---|---|---|
| Paper + Tom Yeh's free newsletter | Free | Building core math intuition | 15–30 min |
| Deep Learning Math Workbook (print, Packt) | Paid, one-time | Offline, structured practice away from a screen | 10–20 min |
| ChatGPT or Claude as a checker | Free tier available | Catching arithmetic mistakes and explaining a missed step | 2–5 min |
| NotebookLM from your worked notes | Free | Turning solved puzzles into a review set before a test or interview | 10 min |
If you're using a chatbot to check your work, my guide to how to use ChatGPT for studying and how to use ChatGPT Study Mode cover how to get explanations instead of just answers, which is what you want here — you're trying to understand a wrong cell, not just get the right number. Once you've built up a small stack of worked exercises, NotebookLM is genuinely useful for turning them into flashcards or a review summary. And if you want a broader look at study-focused AI tools beyond this one method, see my best AI tool for study roundup.
My take
AI by Hand is one of the better uses of an hour I've found for actually understanding what's happening inside a neural network, and it costs nothing to try. In my testing, the paper-and-pencil format forced a level of attention that skimming a blog post or watching a video never does — you can't fake your way through a matrix multiplication the way you can nod along to an explainer. It won't teach you to build or deploy models on its own, and it's not a substitute for a real course if that's your goal. But as a way to stop treating AI as a black box, it's hard to beat, and the free version covers everything a beginner needs.
Frequently Asked Questions
Is AI by Hand free?
Yes. The core newsletter at byhand.ai is free to subscribe to and includes the exercises and solution keys. A paid print workbook with 300 additional exercises is also available if you want a physical book, but it isn't required to learn the method.
How long does it take to do an AI by Hand exercise?
Most individual exercises take 15 to 30 minutes once you're set up, including checking your answer against the solution. The first one or two you attempt will take longer while you get used to the grid format.
What is the easiest way to start AI by Hand?
Subscribe to the free newsletter, pick one topic you're already curious about — matrix multiplication is the gentlest starting point — and work through a single exercise with a pencil before moving to the next one.
Do I need a math or coding background for AI by Hand?
No. You need basic multiplication and addition. The exercises are deliberately built so that someone without a programming or calculus background can complete them; the difficulty comes from the number of steps, not advanced math.