LLMs reason exclusively in natural language—a convention rarely questioned. We argue this constitutes a fundamental bottleneck: natural language is an inefficient medium for structured thought. Mathematicians do not write 'the sum of all integers from one to n'—they write Σi. We show that .ava notation achieves 2.65x compression, and hypothesize that compressing the medium of thought could enable proportionally more reasoning steps within a fixed token budget. Also available in French: Le langage naturel est-il le bon médium pour la pensée machine ?
Cros et al. (Sun,) studied this question.