This article has a precise aim: to showcase the capabilities of the GPT-5 model when employed to elaborate scientific studies at the doctoral level. The text is not intended as an original contribution to academic research but rather as a demonstration of the model’s ability to handle complex subjects, weaving together mathematics, theoretical physics, and neural network theory. The goal is to test how far GPT-5 can go in producing a rigorous analysis, complete with advanced formalisms and conceptual references, while maintaining clarity of exposition. For this reason, the article first offers a popularized explanation and, in a second stage, provides the complete study in its full version for download.
In recent years, artificial intelligence models based on the transformer architecture have followed a trajectory that can be read as a steady evolution. It is a process that preserves a stable core while enriching it with new features at each generation. The study that introduces the idea of a “DNA” for generative models focuses precisely on this: identifying which elements remain unchanged when moving from one version to another in systems such as GPT-4, GPT-4o, and GPT-5, and how these elements allow us to trace a genuine lineage within the landscape of artificial intelligence.
At the heart of the analysis lies the identification of traits that define the deep identity of these models. The ability to predict the next step in a sequence is the foundation of their functioning, reducing uncertainty in the data by compressing regularities into structured patterns. Supporting this function is the multi-layer attention mechanism, which organizes computation and makes it possible to connect even distant pieces of information. Of particular significance is distributional semantics: meanings and relations are placed in latent geometric spaces, where distances and directions acquire precise value. Completing the picture is context plasticity, a property that allows information to be maintained and transformed in the short term, fostering flexible reasoning adaptable to different situations. This combination of principles is what remains stable across the innovations introduced.
The study does not confine itself to theoretical observation but also proposes a method for measuring the persistence of the core. For this purpose, it introduces the Core Heredity Score, an index built from three fundamental quantities: the similarity of internal activations across successive generations of models, the ability to adapt to contexts of varying length, and the measurable advantage when the model develops intermediate reasoning. These three aspects are treated as observable variables, capable of providing concrete and repeatable data through experimental protocols. The approach aims to distinguish what truly belongs to an inherited core from what instead results from contingent adjustments or implementation choices.
The expected findings highlight a key point: the strength of the models lies in the innovations introduced at each step, combined with their ability to preserve an underlying continuity. The expansion of context windows, for example, has extended the horizon of computation without replacing the core. The same applies to the integration of multimodality: images, sounds, and text are projected into the same vector space without altering the internal operating principles. The deliberative aspect, linked to the ability to perform intermediate reasoning, is further proof of the solidity of the core. When models are given the space to develop an intermediate reflective process, they show a measurable advantage in solving complex and compositional tasks.
The relevance of this study extends beyond the technical dimension. On one hand, it provides tools for rigorously understanding the genealogy of models, distinguishing between continuous evolution and changes that signal the emergence of distinct branches. On the other, it opens a broader reflection on the very nature of generative artificial intelligence. Instead of being considered as a sequence of disconnected products, it is described as a transforming family that preserves shared traits over time. To speak of a “DNA” is therefore to give shape to a notion that combines rigor and clarity: today’s systems, such as GPT-5, are the most advanced versions of a project that continues to evolve without abandoning its original framework. This perspective helps us understand how progress does not erase what came before, but instead grows upon a structure capable of ensuring continuity and identity.
