Add an omnigrok note

notebooks_jason/max_of_2_grokking.py

@@ -714,5 +714,7 @@ def compute_traces_and_frames(
 # However, because the test set includes so many more sequences, and the loss-reduction is carefully tuned to the train set, we see a sharp increase in test loss.
 #
 # But once the monotonicity violation is resolved, there's a sharp reversal in the generalization from the train set to the test set: now every adjustment helps the test set even more than the train set, so we see a sharp drop in loss.
-
+#
+#
+# Added note: [2210.01117: Omnigrok: Grokking Beyond Algorithmic Data](https://arxiv.org/abs/2210.01117) claims that "Grokking is caused by the mismatch between training and test loss landscapes."  This demo shows that this explanation isn't a complete picture, though, because there's still a phase transition in *training loss* even apart from the train-test loss landscape mismatch.
 # %%