What is "Training" in AI?
Training an AI refers to creating or tuning an AI model, which is a configuration file used by some AI-powered program. The training process requires a dataset containing the data that the program is expected to reproduce is required, as well as a separate training program with a specialized algorithm to update the model based on the output of the main program.
In a common method, the training program uses the main program, the model, and the dataset to generate some results, and then compares those results to what it expected to receive. The difference between what it expected to get and what it got is "how wrong" the AI model is. This is technically called the loss or cost. Based on this loss factor, the training program adjusts the AI model and tries again. This process can be repeated indefinitely. The loss never becomes zero as it's implausible that the training program will ever find a perfect combination of configurations that always generate perfect results. Therefore, when the loss becomes low enough, the training is simply stopped, as it has become reliable in generating a specific set of results.
An alternative method would be to have some measurable goal produce some "score" based on the results, and then tweak the model to achieve higher scores in that measurement. This is often said to lead to the AI "cheating" the test by seeking to maximize the score even if it does so in a way that will never success at the task the researcher expected to lead the AI toward when they created the measurement.
It's worth noting that training an AI model often makes that model unable to achieve other results than what it was trained to do. That's because the AI model normally has a fixed number of parameters to configure. New parameters aren't added during training, existing parameters are adjusted. Thus, an AI model configured to recognize birds in images, for example, has its parameters tuned for that specific purpose, and if we used this model and trained it to recognize dogs, it most likely would stop being able to recognize birds, since it's the same parameters for all possible tasks. That said, it may be possible to train the model to recognize both birds AND dogs if the dataset includes observed samples of both animals, but in this case it may be possible that the model won't be as good at either task compared to models trained specifically for those tasks.