Inside a transformer, computation happens as geometric transformations across a high-dimensional continuous latent space. But under standard chain-of-thought, the model is forced at each cycle to collapse that rich latent state through a bottleneck: projecting it into discrete tokens—twenty-six glyphs invented by Phoenician turquoise miners in 1800 BC—written onto an English scratchpad, which must then be re-embedded back into latent space before the next layer can touch it. This bottleneck is a severe capability tax, but an accidental alignment godsend: humans can monitor intermediate thinking in plain English. The specter of 'neuralese recurrence' is what happens when you cut out the middleman and loop the computation directly within latent space—letting the model circulate raw activation vectors indefinitely without ever translating into human glyphs. It unbottlenecks thought, but leaves us completely blind, because interpretability has no dictionary for arbitrary thousand-dimensional latent trajectories. astralcodexten.com