Groups
Grokking is when a model suddenly starts to generalize well long after it has already memorized the training set.
In underdetermined linear systems (more variables than equations), gradient descent started at zero converges to the minimum Euclidean norm solution without any explicit regularizer.