The problem. For decades, image classification meant hand-engineered features fed to a classifier, and progress was incremental. Neural networks were an old idea widely considered impractical at scale. The open question was whether a large, deep network could actually be trained to beat that pipeline — and whether the hardware existed to try.

The idea. AlexNet was a deep convolutional network trained on ImageNet’s million-plus images using two GPUs, ReLU activations to speed learning, and dropout to fight overfitting. It cut the benchmark error rate dramatically — enough to end the debate. The design choices here (conv layers, ReLU, GPU training) became the template.

Why it matters. This is the taproot. Every model earlier in my reading — AlphaFold2, ESM-2, scVI, Geneformer, scGPT — descends from the moment deep networks started working. Reading AlexNet is reading why the biology-ML papers I care about were even possible: the compute and the training tricks came first, the applications followed. It grounds the foundation-model hype in a concrete origin.

Verdict. Foundational, full stop — arguably the most consequential ML paper of the decade. Convolutional specifics have since given way to transformers, but the lesson (scale + GPUs + the right regularisation) is the through-line to today. Read it as history that’s still load-bearing.