Deep Learning (Goodfellow)

The MIT Press textbook that became the field’s standard reference, written before the transformer and before scaling was the organising story. It builds from the mathematical prerequisites — linear algebra, probability, numerical computation — through the machine-learning basics to deep feedforward networks, regularisation, optimisation, convolutional and recurrent architectures, and then into the research frontier of the mid-2010s: autoencoders, representation learning, structured probabilistic models and generative models, the last of which Goodfellow had a hand in inventing. The full text has always been readable free online, which is part of why it spread as fast as it did.

Why it’s on the list: The textbook the current generation of AI researchers actually learned from — and, read in 2026, a snapshot of what the field believed was hard just before scaling reorganised the whole question.

Where I saw it: @tylercosgrove’s October 15, 2025 list of 25 books he offered to send to Jared Kushner.

Connections

• Deep Learning — the concept page; this is the textbook that named it for most practitioners.
• Machine Learning — the broader field the book builds up from.
• Transformers — the architecture that postdates this book and displaced much of its second half.
• AI — the field.