2 ms·
> As the recently released Apple study demonstrates The Apple study that did Towers of Hanoi and concluded that giving up when the answers would have been too
by ben_w 1y ago
> As the recently released Apple study demonstrates
The Apple study that did Towers of Hanoi and concluded that giving up when the answers would have been too long to fit in the output window was a sign of "not reasoning"?
https://xcancel.com/scaling01/status/1931783050511126954 https://xcancel.com/scaling01/status/1931783050511126954
I mean, on that basis, anyone who ever went "TL;DR" is also demonstrating that humans don't reason.
> That's not "intelligence" however you define it because they can only solve things that are already within their training set.
This is proven untrue by, amongst other things, looking at them playing chess. They can and do play moves not found in the training data: https://www.lesswrong.com/posts/yzGDwpRBx6TEcdeA5/a-chess-gpt-linear-emergent-world-representation https://www.lesswrong.com/posts/yzGDwpRBx6TEcdeA5/a-chess-gp...