Let’s start with the name. OpenAI is the company that developed ChatGPT. Chat stands for “chatbot”, which is software capable of simulating a human conversation. Generative means that this system generates something, in this case, texts. Pretrained indicates that ChatGPT is pre-trained (on texts). The term transformer refers to the type of computer model used, and finally, 3.5 is the version we are considering. ChatGPT 3.5 is currently free. Its big brother, 4.0, is instead paid, and we will talk about it later.

ChatGPT is part of the so-called large language models (LLM). They are systems trained on enormous amounts of texts to recognize the relationships between words, even distant ones, and work by trying to predict the next word in a text based on what has been written previously by the user or the model itself.

An LLM uses its vast information on language structures to complete and enrich the text based on user requests. These models are founded on so-called neural networks, which can be imagined as networks of small processing units (nodes) connected to each other. You can find here or here a series of accessible videos on these networks, on LLMs, and other interesting topics.

ChatGPT interacts with the user in natural language, for example English, can choose different words to vary the contents of the responses, can generate coherent texts with different degrees of complexity, structure, and tone, and remembers to a certain extent the contents of a conversation.

The training of ChatGPT 3.5 ended in January 2022, so the system does not contain any reference to facts subsequent to that date (for example, Argentina’s victory in December at the FIFA 2023 football championship).

The understanding of the meaning of texts is only simulated, and the information provided (such as facts or bibliographic citations) might be inaccurate or false. This becomes a critical problem only if the model is used improperly (we will talk about this in other posts). ChatGPT is not a search engine, but it can be connected to the web, and it is not very suitable for performing complex numerical calculations or operations that involve them. At any time, the maximum workable unit is about an A4 sheet (but it depends on the cases). Finally, conversations are stored for a certain period and can be analyzed by human operators; information covered by privacy or subject to other dissemination constraints should not be entered.

To begin using ChatGPT you can start here and follow the instructions. The web is full of materials on the use of this model, but in future posts, in addition to telling a little more, we will also focus on perhaps less discussed aspects. See you soon.