Re:ARC-AGI without pretraining
by tripplyons on Tuesday March 04, 2025 @09:32PM
(#43261917)
I haven't seen a convincing argument that it is more sample efficient than a neural network that has seen an equivalent amount of lifetime information. Yann LeCun gave an interesting explanation of how even in the first few years of a child's life, they have seen much more information than the largest pretrained models have.
[ Parent
| on HN ]
There are no comments on this comment yet.