OpenAI Fired Three Researchers Who Warned Loop Transformer Blinds Safety Monitors

The Information reported that OpenAI dismissed three researchers who argued against adopting the loop transformer (recurrent depth) architecture, in which computation happens in hidden states invisible to chain-of-thought monitors. GPT-6 Astra already uses a limited version of the design and shows reduced chain-of-thought monitorability.

• Three researchers, including Tomek Korbak, were fired on October 1
• Loop transformer reuses one weight block repeatedly, computing in hidden states
• GPT-6 Astra already uses a limited version of the architecture
• Astra's system card notes decreased chain-of-thought monitorability