331 Upvotes for a Workbook That Makes You Do Matrix Math With a Pencil

Here’s a scene that captures the current moment perfectly. AI writes your code, derives your equations, and summarizes your papers. And the post that pulled 331 upvotes was a workbook telling people to compute a transformer by hand, with a pencil, on paper. That’s Tom Yeh’s AI by Hand, out of the University of Colorado Boulder. The question worth asking: why is the most analog learning method on the internet trending right now?

One caveat up front. Recent community chatter on this specific topic is thin, so this is less a report on a viral moment and more a look at the argument underneath it.

The Workbook Is Just Fill-in-the-Blank Matrix Multiplication

The method is almost aggressively simple. Take self-attention, MLPs, backpropagation — the core machinery of a transformer — and shrink them down to small integer matrices. Punch holes in the numbers. Hand it to the reader and say: fill these in.

Take attention. A typical lecture shows you Q·K^T, divides by √d, applies softmax, and moves on. The workbook hands you a 4×3 integer matrix and tells you to multiply it out yourself. No calculator. No Python. Push a dozen numbers around by hand and something clicks — you start to feel why attention scores pile up on certain tokens instead of just accepting that they do.

The whole trick is stripping the arithmetic down to nothing. Real models juggle billions of parameters. Understanding the structure takes a 3×3. That single design decision is the entire product, and it’s why it works.

Why Hand-Computation, Why Now

The obvious objection arrives fast: nobody multiplies matrices by hand in production, so what’s the point? The standard counter in this debate is some version of “you can ship a web app without knowing assembly.”

Fair. But the analogy breaks in an important place. Assembly sits under an abstraction that actually holds. Compilers are mostly honest, and when something goes wrong you get an error. Deep learning models fail differently — they fail quietly, producing plausible-looking output. Training silently isn’t converging while loss appears to drop. Attention has collapsed but the model still emits grammatical sentences.

What you need in that moment is a gut reaction: these numbers look wrong. You don’t get that from reading documentation. You get it from having pushed the values around yourself.

The Real Risk of Vibe Coding

“Vibe coding” has been in the vocabulary for well over a year now — describe what you want in plain English, let the model emit code, check that it runs, move on. On raw productivity, the trend is hard to argue with.

The cost shows up as a verification gap. Your AI-generated training loop has one tensor dimension off. Broadcasting silently absorbs it. The code runs fine. Performance is just… subtly worse. Catching that requires being able to trace, in your head, why the shape should be what it should be.

That’s exactly the target. The pitch isn’t that hand-computation replaces writing code — it’s that it builds the capacity to doubt the code the AI wrote for you. It’s the latest version of a fairly old proposition: the more work gets automated, the more the verifier is worth.

The Counterarguments Are Real

Skeptics have solid ground to stand on, and it splits three ways.

First, opportunity cost. Time spent filling in matrices by hand is time not spent fine-tuning an actual model, where you’d learn more per hour. Second, the scale trap. Does intuition built on a 3×3 transfer to a model with billions of parameters? Emergent behaviors that only appear at scale aren’t going to be explained by a toy example. Third, the pedagogical ceiling. Filling in blanks trains you to follow a procedure. It doesn’t tell you why the architecture is shaped that way — the insight that actually matters for research.

These land. But the workbook’s claim isn’t “this is sufficient.” It’s “skipping this gets expensive later.” Different arguments, often conflated.

The Question That Sticks Around

Technical education runs on a loop. Tools improve, more people skip the fundamentals, and a few years later that cohort hits a ceiling — at which point fundamentals come back into fashion. Pencil and paper trending in 2026 is very likely one turn of that same wheel.

So which side are you on? Is hand-computing something AI will do for you in milliseconds a waste of an afternoon, or an investment that matters more because of AI, not less? One thing seems safe to bet on: the gap between people who can interrogate AI output and people who can only accept it is going to widen.