Hot take on GPT-6 Astra

Substack · Watch/Read · (gathered 2026-09-04)

• Gary Marcus offers a “hot take” on OpenAI’s GPT-6 Astra, acknowledging its impressive SOTA performance on ARC-AGI-3 (63% raw, 99% with adapter harness) and its explicit use of symbolic world models, which he views as a vindication of his long-standing advocacy.
• He challenges claims of Astra being AGI, questioning the robustness of its symbolic modeling capabilities and predicting problems with open-ended real-world tasks, despite its success in verifiable domains.
• Marcus expresses concern over the lack of transparency regarding the system’s inner workings, noting it appears “less monitorable” than prior systems, which he deems problematic for safety when combined with increased capability.
• He warns that initial enthusiasm for such systems often gets tempered over time, suggesting that the limited access given to skeptics is a marketing strategy that can be misleading.