Update README.md

README.md

@@ -16,3 +16,5 @@ This variation was a bit hit or miss, but showed high promise and fair accuracy
 The outcomes and train results speak for themselves and the notebook is included.
 
 The encoder is a bit lackluster - the attention is hit or miss and causes disruptions to the high learn rate crystalization process. The outcomes show potential promise and the speed isn't quite up to the standards I was hoping for - so I began again from this point and shrank the encoder while advancing the pentachoron structure for the next notebook.
+
+The geometry for this one is fickle so high attention often causes overfitting early, which is why the latest version does not use multihead attention in the encoder.