>GPT-5 Normal completes Pokémon Gen I in 6100 moves
>10700 moves less than o3
>With significantly fewer screen-to-action mismatches
It reminds me of early sandbox puzzle reinforcement opimization. The fun part would be if GPT-5 started exploring the glitched side of Pokémon in order to achieve certain conditions while exploring the hex. Such a system would be far more adept than human trial and error.