The Basic Reasoning Test That Separates Real Intelligence from AI

This study draws attention to a crucial contradiction in artificial intelligence: the theoretical expressiveness of Transformer-like models and their ability to reason in real-world scenarios. This paper shows that deep neural networks and Transformers are consistently unsuccessful in practice on basic, structured reasoning tasks, despite the fact that they are formally proven to be Turing-complete, which means that they can theoretically simulate any Turing machine and solve any computable problem with polynomial resources. A "height comparison" job serves as an example of this, in which the model must combine multiple provided facts to infer a relationship.