Next big section is on residuals which has devolved to something extremely extremely cursed.
They first compare HyperConnections (learnable write, read, mix operators) against a simplied variant where its just a linear combination of various branches and the output is written to a single branch in a roundrobin fashion.
They compare both of these with their own variant GatedResidual:
1) RMSNorm each branch
2) Apply projection via a low rank bottleneck
3) Layer is the input coefficients of the different branches
4) Write uses a single branch level scalar