Chatgpt, Gemini, Claude and Deepseek are the absolute protagonists of recent years. In fact, artificial intelligence has become part of daily life thanks to Large linguistic models (in English Large Language Models, Llm). But what exactly are they and how do they work? LLM are artificial intelligence systems capable of interpret and generate textmaking it possible to interact with them by speaking normally, as we would do with a person: they are able to support conversations, write texts e summarize documents, but, despite their skills, they can commit errors And they do not have a real understanding of the responses they generate. We must therefore use them with awareness.
What are LLM and what they are for
A Large Language Model (Llm) is an artificial intelligence designed for interpret, trial And generate text written. The LLM are part of the field of Natural Language Processing (NLP), i.e. processing of natural language, which deals with teaching computers how to interpret human language. For this, we can interact with these models simply speaking or by writing naturally.
These tools are used in many different contexts: they can write email, translate texts, respond to requests, summarize articles, generate Creative content or support in study and in the programming. There are also experiments in the medical field, in which the models are accompanied by doctors and doctors for the formulation of a diagnosis Starting from the symptoms.
There are many different models, each with its own characteristics. The best known are Chatgpt And Deepsekuseful for writing, translating and summarizing content. Gemini And Claude They are a little less known, but they too are equally flexible and suitable for different tasks. In particular, Gemini is effective in the management of long texts, while Claude stands out in the generation of code. Despite the differences, all these models share some common bases. Let’s see now, in an intuitive way, how they work.
How artificial intelligence does to learn
LLM are able to interpret our requests and generate phrases that seem to be written by humans. But how do they know what are the right words to use and how to structure the sentences? To do this, they collect and yes they use huge quantities of text containing billions of words. The model then learns in three phases Main:
- In the first, called Pre-training (hence the “P” in Chat-GPT), the model learns to copy and repeat the enormous amount of text that is provided to him. This phase consists substantially in complete to the model Phrases with missing wordsin order to do it Learn the structures and rules of languagewithout explicit teachings. It is a bit like a child who learns to speak listening and repeating.
- The second phase is theInstruction Fine-TONUNING. Here the model is asked not only to complete phrases, but also to follow aprecise instruction (as “write a formal email” or “explains photosynthesis to a child”) and generate a useful response and pertinent.
- In the third phase, the Reinforcement Learning from Human Feedbackthe model is further improved thanks to the feedback of human evaluators who see as possible answers to the same question and indicate which they prefer.
How the LLM interpret what we write and generate the answers
In concrete terms, what happens when we write a request to Chatgpt and we get an answer? When we write a request to a LLM, the model takes the sentence and breaks it into “token”, That is small characters blocks that contain ainformation. A token can be one word whole, one syllablebut also a single character. For example, if we ask Chatgpt: “explain to me why inflation increases prices, but do it as if I were a friend at the bar.” The sentence is broken down into token as we see here.
In this case, “how” is a token, but the point alone is also. This is because the point is giving us information: it tells us that the phrase is over.
Each token then comes converted in a number e analyzed from a Transformerthe heart of the LLM (it is from here that the chatgpt “t” arrives). Transformers are neural networks introduced by Google in 2017, which revolutionized the interpretation of language thanks to a mechanism called self-care. In the first sentence, we humans immediately understand that “Fallo” refers to “Spiega me”. For an algorithm, however, this connection is not immediate. Thanks to self-care, however, the model manages to establish links even between words distant in the sentence and to understand the meaning of the words based on the context. In this way, he manages to correctly interpret complex requests.
Once the request “happens”, the model must produce a answer. To do this, calculate step by step what is the token more likely To be inserted, based on everything he learned during pre-training. For this reason, it is said that the LLM are stochastic parrots: they repeat what they have learned, but with a certain dose of randomness. This “randomness” is different for each model and is defined by a parameter called temperature. A LLM with a low temperature will generate more predictable answers and more linked to starting sources, while one with high temperature will create more creative, but less reliable responses. Precisely for this reason, we cannot blindly rely on the responses of a LLM: the phrases it produces could play plausible, but be false.
