After reading a completely credulous article attributing the success of OpenAI's latest model on the ARC-AGI test to the spontaneous emergence of genuine AGI and not, as seems more likely, the fact that the benchmark has recently become fashionable and therefore they've been specifically aiming for it as they develop the model, I finally took a look at the test itself.
It's sort of a fun little challenge for the bots, but it's absolutely nothing compared to playing the game of Go as they managed to do in 2016. Meanwhile actual AGI as we traditionally think of it will be when an artificial intelligence can hear about the game of Go somewhere, having never heard of or trained on it, find it interesting, and teach itself to play.