Back to index
OtherSelf paced exercises and workbooks, plus occasional live cohorts, work through at your own speed·Much of the content is free, with paid workbooks and cohort offerings

AI by Hand (Tom Yeh)

4.4

A refreshing, genuinely original antidote to black box learning. Tom Yeh has you work the mathematics of neural networks and transformers by hand, on paper and in spreadsheets, and the understanding that produces is remarkably durable. It is a supplement rather than a full curriculum, but as a way to make the maths finally click, few things do it better.

What We Liked

  • A brilliant, memorable way to build real intuition for the mechanics behind modern AI
  • Much of it is free and accessible, with a friendly, low intimidation approach to hard maths
  • Working attention or backprop by hand exposes exactly what the equations are doing
  • Trending upward for good reason, and increasingly popular with educators and self learners alike

What Could Be Better

  • It is a supplement, not a complete course, with no coding curriculum or path to a job
  • The hands on, do the arithmetic approach is deliberately slow and will frustrate the impatient
  • No certificate and no formal structure to hold you accountable
  • You still need a coding focused resource alongside it to turn intuition into working models

Detailed review

AI by Hand, the project of Professor Tom Yeh, is built on an idea so simple it is almost subversive in an era of one click model training. If you want to actually understand how a neural network or a transformer works, do the maths yourself, by hand, on paper or in a spreadsheet, one multiplication and addition at a time. The exercises walk you through computing a forward pass, running backpropagation, and working the attention mechanism manually, and the effect is genuinely different from watching an explainer or reading a derivation. When you have personally filled in the cells that turn a query and a key into an attention score, the abstraction stops being a black box and becomes something you can reason about, and that understanding tends to stay put in a way that passive learning rarely manages.

It has been spreading steadily, particularly among educators and self learners who have hit the wall where the equations are technically legible but never quite intuitive, and the popularity is deserved. Part of the appeal is how unintimidating it is, the maths is broken into small, concrete steps that feel more like a puzzle than a lecture, and a lot of the material is freely available, with paid workbooks and occasional live cohorts for those who want more structure. It is a rare thing in this space, a resource whose whole purpose is intuition rather than coverage or credentials. The honest framing is that this is a supplement and was never meant to be anything else.

There is no coding curriculum here, no path from these exercises to a deployable model or a job, and no certificate at the end, so it needs to sit alongside a practical, code focused course rather than replace one. The deliberate slowness, the whole point of which is that you do the arithmetic yourself, will irritate anyone in a hurry, and there is no external structure holding you to it. None of that undermines the value, it just defines the role. My honest view is that if you have ever felt that you can follow the transformer maths on the page but do not truly feel it, AI by Hand is close to the best fix I have come across, and I would happily recommend it to anyone as the intuition building half of a pairing, with a coding course supplying the other half.

[ final ]

The verdict.

A superb companion for anyone who has stared at the transformer equations and never quite felt them land, and who is willing to slow down and actually work the numbers. It is not a standalone route into AI, so pair it with a proper coding course, but as the thing that finally makes the maths make sense, it is one of the more genuinely useful ideas in AI education right now.