AIArticle Tom Yeh's hand-traced worksheets grew into a 73,000-reader publication and a paid lab. Here's what transfers to real work.
Rachel Goldstein There's a spreadsheet on GitHub that implements DeepSeek's multi-head latent attention, cell by cell, with no code. It's part of AI by Hand, the project from University of Colorado Boulder professor Tom Yeh that resurfaced on Hacker News this week — and that has quietly evolved from viral hand-drawn worksheets into a 73,000-subscriber publication and a paid "research lab" called By Hand Research. That evolution is worth examining, because it's both a genuinely useful resource and a case study in how far the "learn the fundamentals" pitch can stretch before it snaps.
The fourth lane of AI education #
Every wave of ML education has picked an abstraction level and defended it. Andrej Karpathy's nanoGPT and Zero to Hero lectures say: build it in code, and autograd will make sense once you've written it yourself. 3Blue1Brown says: watch the linear algebra move, and intuition follows. Jay Alammar's illustrated guides say: a good static diagram beats a thousand equations. fast.ai says: start at the top, train something real, dig down later.
Yeh's approach is a distinct fourth lane: compute it yourself, with numbers small enough to fit on paper. A worksheet gives you a 4-dimensional embedding, a 2-head attention block, and blank cells. You do the dot products. You apply the softmax. You find out — physically, with a pencil or a spreadsheet formula — that attention weights are just a normalized similarity score used to blend value vectors.
What makes the Excel version more than a gimmick is that a spreadsheet is the only ML runtime with total observability by default. In PyTorch, every intermediate tensor vanishes unless you deliberately hook it; the framework's whole job is to hide the plumbing. In Excel, the plumbing is the interface. Every intermediate value of a forward pass sits in a visible cell with an inspectable formula. One HN commenter called it developing "mechanical sympathy," and that's exactly right — the same reason race drivers learn engines. The free ai-by-hand-excel repo (MIT-licensed, around 6,000 stars) now covers a startling range: softmax and temperature scaling through backprop, LSTMs, Mamba, full transformer stacks, and that DeepSeek sheet with latent attention plus mixture-of-experts routing.
From worksheets to a "research lab" #
The new part — and the reason byhand.ai is back on HN — is the packaging. AI by Hand now presents itself as the publication of By Hand Research, which "studies model interpretability and explainability at the math and algorithm level." There's a $29/month membership, live seminars on frontier architectures, and courses like a 2026 agentic-AI intro. Recent seminars trace single tokens through Qwen 3.6 and work through RoPE and YaRN context extension by hand.
Let's be precise about what this is and isn't. In the research community, "interpretability" means reverse-engineering what a trained model has learned — circuits, features, sparse autoencoders, the Anthropic-style program of explaining why specific weights produce specific behaviors. Tracing an architecture's forward pass by hand is not that. It teaches you what the architecture can compute, not what a particular model has learned. Calling worksheet pedagogy "interpretability research" borrows prestige from a field doing something harder. The HN thread's sharpest friction wasn't about that, though — it was simpler: visitors bounced off a wall of member-gated post previews. The free-to-paid gradient is steep, and the site doesn't make the excellent free tier obvious.
And yet the paid offering isn't unreasonable — it's just mislabeled. What Yeh is actually selling is an architecture-briefing service: every time a frontier lab ships a new attention variant or context-extension trick, someone with two decades of teaching experience turns it into a traceable numeric example within weeks. That's a real product. Conference tutorials cost more and go stale faster.
Where the pencil work actually pays off #
The fair question is whether hand-computing a 4×4 attention matrix transfers to real work, or whether it's productive-feeling busywork. Here's the concrete case for transfer.
If you're self-hosting models, the knobs you touch map directly to the math in these worksheets. Extending a context window in vLLM means setting YaRN rope-scaling parameters — factor, original max position — and if you've never traced how RoPE rotates query/key pairs by position-dependent angles, those are cargo-culted magic numbers. Sizing GPU memory for serving means estimating KV-cache growth, and understanding why DeepSeek's latent attention slashes that cache requires exactly the low-dimensional projection you compute in the Excel sheet. Debugging a fine-tune that outputs garbage means knowing whether temperature, a chat-template mismatch, or a broken attention mask is the culprit — a diagnosis that's guesswork if the forward pass is a black box. The honest limits: by-hand work teaches mechanism, not behavior. It won't tell you why your model hallucinates, what your training data did to it, or anything about the emergent properties that only appear at scale. Nobody should confuse completing the DeepSeek worksheet with understanding DeepSeek-the-model. And the audience is narrower than the subscriber count suggests — if you live entirely above the API line, prompt-engineering hosted models, the ROI on hand-computing LayerNorm is genuinely low.
The verdict #
The pedagogy is excellent and the free repo is the best zero-setup on-ramp to transformer internals that exists right now — better than videos, because you can't passively nod along with a blank cell. Start there: the MIT-licensed spreadsheets cover more ground than most paid courses. The $29/month tier is defensible for practitioners whose employers reimburse it and who need to stay current on architecture churn, though you're buying a briefing subscription, not access to a research lab. The "interpretability research" framing oversells what worksheets do — but in an era when most developers' understanding of transformers is a vibe absorbed from API docs, overselling pencil-and-paper rigor is about the most forgivable sin in AI education.
Sources & further reading #
[AI by Hand](https://www.byhand.ai/)— byhand.ai -
[About - AI by Hand](https://www.byhand.ai/about)— byhand.ai -
[AI by Hand - Hacker News discussion](https://news.ycombinator.com/item?id=49300568)— news.ycombinator.com -
[AI by Hand Exercises in Excel - Hacker News discussion](https://news.ycombinator.com/item?id=42967173)— news.ycombinator.com -
[ai-by-hand-excel repository](https://github.com/ImagineAILab/ai-by-hand-excel)— github.com -
[Tom Yeh faculty page](https://www.colorado.edu/cs/tom-yeh)— colorado.edu
[Rachel Goldstein](https://sourcefeed.dev/u/rachel_goldstein)· Dev Tools Editor
Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop.
Discussion 0 #
No comments yet
Be the first to weigh in.