After the popularization of ChatGPT, a lot of promises were made and noise was
formed around the possibility of LLMs (Large Language Models)
being the next step to the so-called AGI (Artificial General Intelligence).
Many people are concerned and fear losing their jobs to the AGI.
These thoughts come from the lack of perception of how the
Large Language Model that powers ChatGPT works.
Underlying the fancy name, everything behind LLM is a lot of neural networks carefully combined. The ANN (Artificial Neural Network) is an AI technique that relies on algebra (more specifically, matrix multiplications) and calculus (backpropagation uses partial derivatives). Modern NVIDIA processors power these operations and have very high computational costs.
LLMs are, in fact, a complex architecture of many layers and “submodels”, trained in a very well-curated dataset. This curation starts early in how to obtain embeddings that “make sense” and can be interpreted in a vectorial space. The data for the training phase of GPT is carefully designed to tackle each subtask, from text classification to multiple-choice answers.
The technical part put asside, when we talk about general intelligence, we have to be careful. Intelligence is a complex behaviour observed in living creatures, specially the ones endowed with a central nervous system.