GPT-3 works by means of Machine Learning algorithms that were fed with thousands of texts available on the internet. This is the shortest explanation, but it is worth going deeper to improve the understanding of this model.Like all Artificial Intelligence, the GPT-3 had to go through a learning process. This means that the development team entered a lot of information to "teach" the algorithm how it should behave with each input.
This is where the machine learning starts to happen, as the system learns from the data passed to it and is able to create complete texts for different contexts. In practice, what happens is as follows:
The program receives a text as input and, based on all the information it has previously processed, it calculates which is the most plausible answer for the context that has been presented.
However, this does not mean that GPT-3 understands the meaning of the texts it writes. It is important to remember that it is an NLP (natural language processing) based system "trained" to learn, recognize and reproduce linguistic patterns. In other words, despite its incredible capacity, GPT-3 is not able to reason abstractly like a person.