I think we would need to design these tests differently for GPT-4. It has much larger memory than humans, but should be challenged more towards transfer to entirely new settings, and be challenged with very plausible but wrong settings.