Exactly. All of these examples are well documented games. And even then it’s only really good at games that don’t need realtime.

You could make the same argument for these LLMs years ago but there's a reason that they've become faster and more accurate at beating these games. GPT 5.5 took 218 hours to beat Pokemon Fire Red; Astra took 21.

Only months ago did these LLMs make it past the Victory Road boulder puzzle without harness assistance.

GPT-4 was trained on all that too and couldn't get any task at all done