A Language Model is an artificial intelligence system designed to analyze an input expressed in natural language and, by recognizing the statistical patterns of language through training on large quantities of text, generate an output consistent with the input it has received, producing new text in natural language. Without a doubt, it is correct to say that LLMs are a subcategory of LMs. A Large Language Model, as the name suggests, is a large-scale Language Model; the term usually refers to systems that operate using several billion internal parameters and that, despite requiring extremely costly training, deliver performance levels beyond the reach of a simple LM. By “natural language,” we mean any language used by humans for communication, such as Italian, English, or Japanese. Natural language is characterized by nuances and ambiguities because, unlike other types of language, such as mathematical language, it is much less rigid, and its meaning can depend heavily on the context of the communication.

Among the most important stages in training these models are Masked Language Modeling, in which the system is required to fill in incomplete sentences with appropriate words; Autoregressive Prediction, in which the task is to predict a word based on the preceding words; and Instruction Tuning, a stage in which the algorithm acquires the ability to identify instructions within an input and follow them faithfully. This final stage of training is crucial for all AI chatbots that have now become part of our daily lives, which are based on Large Language Models such as GPT, Gemini, and Claude.

Compared with a standard LM, a high-quality LLM has a significantly greater ability to understand the input text, particularly when it comes to recognizing context. This subsequently enables the model to provide a response that is highly consistent with the input. Indeed, the responses generated by Large Language Models are characterized by remarkable fluency and a strong stylistic resemblance to texts produced by humans in general or, when requested, by experts in a specific field. Coherence is another fundamental characteristic of LLMs: while a small Language Modelmay be able to write a grammatically correct but poorly coherent text, a large one can maintain its coherence even when writing long paragraphs or throughout an entire conversation. For these two reasons, the quality of the outputs generated by Large Language Models is clearly and unquestionably superior to that achievable with a simple Language Model. However, robust as they may be, the feature that perhaps most distinguishes these algorithms is their excellent ability to follow the instructions they receive, which, as already mentioned, is acquired through the training stage known as Instruction Tuning.

Traditional LMs are designed to generate natural language and, in general, are not intended for interaction with users. LLMs, on the other hand, have an excellent ability to adapt to the input they receive. Consequently, for example, they are not only able to answer a question correctly, but can also develop their response according to the user’s instructions. Indeed, when asked to provide a concise answer, a Large Language Model will focus on the main key points of the topic, whereas the same model, if asked for a more in-depth explanation, will opt for a longer and more detailed output.

LLMs are tools that can be extremely useful, but their power, if used with malicious intentions, can pose a serious threat to cybersecurity. Attacks on software, operating systems, or other IT infrastructures through the exploitation of Large Language Models are highly relevant and concerning issues today. However, the topic most widely debated nowadays is certainly the use of these algorithms to attack users directly.

LLMs are capable of generating extremely realistic and convincing content that can, of course, also be used for fraudulent purposes and disinformation. A very common example concerns the field of social engineering, which refers to the set of techniques used to manipulate an individual and persuade them to disclose personal information or voluntarily carry out transactions. The involvement of a Large Language Model in this practice can be decisive because the system can provide the scammer with information about a particular individual and subsequently incorporate that information into a personalized and highly credible message, making it more difficult for the victim to recognize the deception.

Giorgio Fumagalli

(translated by MIRAI)