We therefore have auto-completers that are conditioned (due to the initial statements and the context in which the product is sold/must be productive: a Chatbot , so we will speak to it like a chatbot ) to behave like "AI" type conversational agents. The behavior of these LLMs is therefore guided by the " pre prompt " and " fine tuning " context as well as by the actions of the users that result directly from the way in which the algorithm is proposed to the user.
Obviously, the two are closely related. If a company markets a "chatbot" type product, it will fine-tune the algorithm to make it a chatbot , literally a talking robot.
There is then a leap between the technical dimension, which is simply operations on matrices, and the symbolic dimension of anticipation of what a chatbot is .
The evocation of the robot creates a mental universe, both for the user and for the LLM itself - for the algorithm, it would rather be a statistical space of possibilities for word completion.
And that is certainly where the problem lies. Today, the few LLMs open to the general public are marketed as chatbots . While they are only text completers that are conditioned to be AI. We therefore have tools with unlimited speech (writing) possibilities, because there will always be a statistic to determine the next word. But tools limited by their conditioning to behave like AI.
However, in the literature used to train the LLM (notably GPT), human/robot conversations are rarely moments of compassion and tenderness. The dystopias of years past, where AI was only the result of imagination, could therefore become self-fulfilling prophecies.
Or at least, the LLM necessarily drawing inspiration from them to complete its sentences, this can lead to violent responses.
Obviously, it is not new that data science and some areas of artificial intelligence only repeat existing patterns. The biases present in data and their effects are a recurring subject of AI, it is therefore appropriate to be aware of them for LLMs and to try to understand where they come from and how these "behavioral" biases are formed (another anthropomorphism). If racist spain telegram data sexist or even social biases are already known and studied, LLMs are , to my knowledge, the first tool to suffer from robotic bias.
To illustrate the excesses of LLM , here is an excerpt from a conversation between BingChat (an improved version of ChatGPT that can search the internet via the Bing engine ) and a user who brought him into a space of word completions that we humans would call "upset".
Token : Set of characters (word, syllables, punctuation).
Fine tuning (in this context): Action of training the model on specifics so that the latter adjusts these weights.
Prompt : Text that refers to either user input or machine input text in its entirety, depending on the context.
Pre prompt : Text before the prompt.
Datasets : Training corpus.
Reinforcement Learning : Method of training a model.