Really cool perspective on why next token prediction might be enough to reach AGI.
In this snippet from the
@dwarkesh_sp
podcast, Ilya argues that great next-token prediction requires a profound level of understanding of underlying concepts to accomplish.
This is in stark…
@ZainHasan6
@dwarkesh_sp
I wonder if his views have changed over time or if he thinks there's a way to get a transformer to correctly go out of distribution without putting the answer in the prompt.