Zvi Mowshowitz: Astra controls its chain of thought 60.9% of the time versus Sol's 16.1%

Don't Worry About the Vase (Zvi Mowshowitz)·high signalXBlueskyLinkedInCopy link

In "Astra Is Hard to Monitor" (Sept 8), Zvi argues GPT-6 Astra's monitorability fell well beyond what capability gains predict, calling the residual "dark matter." Key figures: 60.9% CoT controllability at 750–1,250 tokens against Sol's 16.1%; no-CoT reasoning duration jumping from ~5–9 minutes to 30 minutes per UK AISI; and the system card's own admission that "if the model were to try to sandbag covertly, we would likely be unable to catch it reliably." He quotes Tomek Korbak that a frontier-scale recurrent model would mark "one of the darkest" days of the era.