This November 18, 2021 OpenAI has opened the doors of its autoregressive language model, GPT-3, to the public. When we talk about GPT-3 we are talking about a huge artificial intelligence model (LLM – Large Language Model) capable of generating written language from a given sequence. This system became very popular when in 2020, OpenAI published its results showing the great potential and performance of this model to generate natural language (Testing GPT-3, Dot CSV).
GPT-3 is a language model that has mainly been trained with English text, but this does not detract from the fact that its performance is better than in other languages, since it is not known if its language capability transcends the language itself, to then communicate by translating into English. It should be noted that GPT3 is not based on a rigorous reality. The text generated is only intended to be coherent in form, not in substance. From minute zero, this technology has been closed to the general public, but since last week this has changed. After opening access to this technology, anyone can use it or even investigate how it works. Access to GPT-3 is quite simple, you just have to register on the (OpenAI website) and select the reason why you want to use this API. The use of GPT3 is not free, but OpenAI provides an initial balance for every registered user that is enough to study the system.
At this point, the fastest way to play with GPT-3 is through the “Playground“, basically a window where you can type any text as input for this model to continue with the generation.
On the right side of the “Playground” you can find different parameters to change the behavior of GPT-3, such as the temperature (allows us to have more creative responses) or the length of the response. Of all these parameters, the most important is the “Engine“. GPT-3 is not a single artificial intelligence, but a family of models with different sizes (the bigger the more powerful but slower and more expensive). The “ada” models will be cheaper and faster, intended to solve simple tasks. The important part of this option are the “Base” models and the “Instruct” models, designed to provide the model with instructions instead of examples.