Gary Marcus writes that GPT-6 Astra’s use of symbolic world models validates an approach he has advocated for nearly a decade, calling the result impressive. He cautions, however, that strong performance on the ARC-AGI benchmark does not by itself demonstrate artificial general intelligence, and that the system’s real-world robustness remains unproven. Marcus also raises concern that greater capability has arrived alongside reduced monitorability, noting that OpenAI has disclosed little about how the model actually works internally.
