Loading slide
Some improvements move a needle. Some move the goalposts.
ImageNet 2012 moved the goalposts. It settled a question that had been contested for years: whether machines could learn useful features on their own, without humans deciding what to look for. Before 2012, many researchers doubted it. After AlexNet, the debate was over.
It also gave the field something concrete to build on. Not an abstract claim, but a specific network, with specific techniques, that anyone could train and verify for themselves.
Everything that followed, the voice recognition in your phone, the image generators, the language models, the AI assistants, traces back to what AlexNet proved. Not because AlexNet itself was used in all of them, but because it showed the approach worked.
The bet a small group had kept alive for two decades had paid off. And it carried a lesson the field would lean on for the next ten years: the way forward was not a cleverer idea, but more. More layers, more data, more compute. The emergence we glimpsed in the last chapter, new abilities appearing from scale, became the bet the field kept making.
This was the moment the AI era began.