The computational limitations of employed architectures constrain the problem-solving capabilities of artificial intelligence (AI) models. Whether such models are simple perceptrons or large language models (LLMs) does not change the fact that the underlying working principle behind these models is the same and limited in terms of computational time complexity. More specifically, such models use constant time to forward propagate the input to the final output layer. This fact makes us wonder how some problems requiring more irreducible instructions could be solved within the scope of current AI models. By putting the burden of solving a given problem on a fixed architecture and considering the size of the irreducible instruction set required to solve the problem, there is a hard limitation on what our AI models will and will not be capable of solving. We propose a potential solution to this problem by introducing a new framework called Column Extension that helps to dynamically allocate computational resources on the fly to make their computational complexity theoretically unbounded. We demonstrate how the Column Extension framework can be used in practice as a proof of concept and share experimental results. We conclude by discussing why our framework is a promising approach to solving some harder problems with AI.