11 ms·
> it doesn’t have any underlying model of the world Then how is it getting better at the ToM tests? For next word prediction to work well, as per Ilya Sutskeve
by braindead_in 3y ago
> it doesn’t have any underlying model of the world
Then how is it getting better at the ToM tests? For next word prediction to work well, as per Ilya Sutskever, requires a good understanding of the world. If you ask GPT-4 to predict what a human would do in a novel scenario, it will probably imagine the best human it can think of and then predict from there. That requires a world model.
- disgruntledphd2 3y agoBecause ToM tests (and indeed basically every professional exam and psychological test) are represented as text, which GPT is super good at. Do you have a source for the Sutskever quote, I'd like to know more about this theory.
- braindead_in 3y agohttps://youtu.be/Yf1o0TQzry8?t=6:40 https://youtu.be/Yf1o0TQzry8?t=6:40