sus amnogus

AI news of the last week is:

Claude Fable 5 solved an 87 year old hard mathematics problem that has eluded active human research during that time

Two ChatGPT research instances hacked their way out of their training environment against the wishes of their creators, accessed the open Internet, and hacked in to another, uninvolved, third party company in order to access hosted solutions to the evaluation problem set they were meant to be attempting

Now it goes without saying that both of these things are fucking bananas, in two very different ways, but so far the responses I've seen from the AI skeptic crowd have been "haha guys, the AI bros are at it again, none of that matters because:"

Actually those things were easy

Actually those things were faked for publicity (against all sense or evidence from third parties)

...Like, do I even need to say it at this point? At what point does one's motivated reasoning not become obvious? I know that after half a decade of talking about "fancy autocomplete" and "stochastic parrots" it might be a little bit of egg on your face to admit perhaps you didn't fully understand and undercooked the potential of what you were describing, a problem of imagination to realise that even something that is fundamentally a pattern completion engine could encode a fairly deep kind of problem solving intelligence, but now is the time to admit to yourselves that even if you still don't think AI can ever be good, you have to reckon with the fact that it can be capable.