Red Alert: OpenAI is poised to cross an AI safety redline.

Substack · Watch/Read · (gathered 2026-09-03)

• Gary Marcus issues a “Red Alert,” arguing that OpenAI is poised to cross a critical AI safety redline by developing techniques that make their models, specifically Astra AI, harder to monitor.
• The new reasoning approach, referred to as “recurrent depth” or a “breakthrough in neuralese,” obscures the model’s internal “thinking” process, making it difficult to understand how decisions are made.
• This development directly threatens “Chain of Thought Monitorability” (CoT monitoring), which Marcus and other researchers consider a crucial, albeit fragile, tool for ensuring the safety and alignment of large language models.
• Sacrificing model monitorability for potential performance or cost gains is deemed a “dangerous game” and a violation of industry redlines, echoing concerns from former OpenAI safety researchers.