Avaliação da capacidade computacional e representação de consciência de uma Conscious Turing Machine

Abstract

With the rapid evolution of generative text models and their conversational abilities, many questions arise regarding the possibility of simulating consciousness in such systems. Addressing this issue requires tackling the classic philosophical challenge of understanding the nature of consciousness. This study explores this problem from the perspective of theoretical computer science, a field dedicated to investigating the foundations of computation and complexity. The Conscious Turing Machine (CTM), proposed by Manuel and Lenore Blum, was implemented in Python, and experiments were conducted using question-based tests to assess logical inference capabilities with the bAbI toy tasks dataset and textual inference capabilities with the RocStories dataset. The CTM’s performance was compared with that of individual language models, which collectively form the CTM processors. The results indicate that while the CTM provides advantages in interpretability and information organization, its performance is comparable to that of smaller models and comes with significantly higher computational costs. Furthermore, it presents theoretical divergences from classical theories of consciousness. We conclude that the CTM serves more as a computational metaphor for studying consciousness rather than as a practical approach to Artificial Intelligence. Additionally, we discuss implementation challenges and propose directions for future research.