The individual human in your comparison is arguably cheating, as they are probably vocalising (or otherwise expressing, for people without inner monologue) seve

The individual human in your comparison is arguably cheating, as they are probably vocalising (or otherwise expressing, for people without inner monologue) several thoughts inside of their head before actually deciding to say anything. Conversely, if you allow a transformer to have an "inner monologue" (token stream it can build up without it counting as an output) too, I don't see why it shouldn't be able to solve these types of problem. A human connected to one of those EEG devices that can predict your rock-paper-scissors hand before you consciously know it and subjected to the constraint that they must not output (as detected by the device) anything but some result would appear rather limited, too.