[ogogmad]

> They can't go beyond the boundaries of their training set.

TFA says they just did. That's what the ARC-AGI benchmark was supposed to test.