Limits of AI: Using Kant’s Transcendental Idealism to Find the Limitations of Transformer Models for Policy Making, Computer Science, and Epistemology

AI models are exponentially improving while policy makers lag behind, leading to societal concerns. We analyze the core of one such model, Transformer-architecture models, to find out what its capabilities and limitations are regardless of compute through the lens of Kantian Transcendental Idealism. We find that such models are limited vis-à-vis human cognition due to their inability to perceive the world through categories, leading them to always act on input as given in experience only. This implies that such models cannot reach true Artificial General Intelligence, and that alternative architectures should be investigated towards that end.