“Transformer architectures are effective for Type 1 (perceptual, intuitive) abstraction but are not well-suited for Type 2 (programmatic, reasoning) abstraction.”