AI agents build 3D scenes from photos but have no idea if they got it right

A new approach called LEGO-Anything has been developed, allowing single photos to be transformed into editable Blender code for creating 3D scenes. This technique utilizes the GPT-6 Astra model, which achieves up to 53% reconstruction accuracy. However, the biggest weakness observed across all tested agents is their inability to assess their geometric accuracy better than chance. These findings suggest that while AI agents can construct 3D scenes from photos, they lack self-awareness regarding the accuracy of their outputs, raising concerns about the reliability of these systems in practical applications.