How can you possibly believe this? If this were true, why do open weight models think in real English tokens and not arbitrary characters?