The comparison nobody selling a course wants to publish
fast.ai is free and it is excellent. If our neural-networks module cannot survive being placed next to fastbook Chapter 4 in public, you should not pay us. So we did the comparison (2026-04-25) and here is the honest score.
Where fast.ai wins
Mechanistic depth. Chapter 4 builds SGD, the loss function, and gradient computation from scratch before ever touching a library. Our NN module is library-first: tf.layers.dense, model.compile. A learner who finishes Chapter 4 can debug a training loop; a learner who finishes only our NN module can ship one but not yet take it apart. (Our training-loop module, one module later, closes part of that gap: you pay the cost of nudge-by-hand gradient estimation before backprop is named, and the chain rule is derived in an optional drawer. But at equal points in each course, fast.ai is deeper.)
Volume. ~18.7k words and 137 runnable cells against our ~3.5k words and 5 exercises for the equivalent concepts.
Where we win
Time-to-first-prediction. ~130 minutes to a working browser model versus 3-5 engaged hours in Colab.
Bridges to what you already know. fastbook assumes a general programmer. We assume a frontend developer and anchor every concept to an FE primitive, with the analogy's breaking point named on the page.
The browser. Every exercise runs and is graded in real TFJS in your tab. No Colab, no Python, no environment setup.
Narrative. A continuous story does real retention work across 56 lessons. fastbook has none. Whether that is worth money is your call; the mechanism is real.
The verdict we ship
They optimize for explanatory depth. We optimize for learner momentum in a JS-native path. If you want the math from first principles, go do fast.ai Chapter 4, sincerely. If you want to ship browser ML as an FE dev and actually finish the course you start, that is the product we sell.