A large network can contain millions of adjustable parameters. One iteration sends a batch forward, measures the error, carries a correction backward, and updates those parameters once. Training repeats that loop across many batches and many passes through the data.

This creates an enormous amount of arithmetic. A GPU makes suitable parts practical by dividing similar calculations among many parallel workers.

But fast training still needs something worth learning from. It needs many examples, human-provided answers, and a shared test that reveals whether the resulting model actually works.

The next chapter follows the people who assembled that material and the experiment that made the result hard to ignore.