Covert Systems — On the Intelligence of Large Language Models

In light of recent developments in Artificial Intelligence, particularly the development of Large Language Models (LLMs), the question if machines can be said to think is examined. The problem was already addressed by Alan Turing in the 1950s, laying the groundwork for our argumentation. Our text refutes four common objections. Firstly, it challenges Ada Lovelace’s claim that computers are incapable of creating truly original content, demonstrating that neural networks, such as Transformer models, independently generate novel structures. Secondly, the assertion that LLMs merely statistically predict the next word is discussed. They are shown to be capable of constructing compressed internal representations from their input data through nonlinear operations, distinguishing them from statistical models. A third argument concerns AI’s lack of sensory access to the world. However, as Turing has demonstrated, the development of mental capabilities requires only communication between a teacher and a student, and sensory experience can be technically replicated at any time. The Chinese Room thought experiment, which suggests that algorithms merely simulate intelligence through formal symbol manipulation, is already anticipated by the Turing Test, which posits that any machine successfully feigning mental functions should be deemed to possess them. The text explores how LLMs generate a perfect compression of their input data, resulting in an accurate model of the world, a convincing simulation of human intelligence, and mastery of language. It discusses the phenomenon of «emergence», the appearance of higher-order cognitive capabilities with increasing amounts of training data. Although the successes in implementing intelligence are notable, the neural network methodology employed results in the enigmaticity of the underlying basis, failing to illuminate the true nature of the mind as originally hoped. This obscurity was already evident in the foundational works of Kurt Gödel and Alan Turing, highlighting the complexities inherent in understanding human and machine intelligence alike.