CLRS includes two dynamic programming algorithms for solving the same problem: all-pairs shortest paths.
The kernels of these algorithms (side-by-side) look almost identical, except that they seem to exploit two different dynamic programming relationships about this problem.
What's the intuition behind the difference between these two dynamic programming relationships?
To be clear, the code for EXTEND-SHORTEST-PATHS is actually used by the routine below, with a total runtime of $O(n^4)$ which is slower than Floyd-Warshall above, which runs in $O(n^3)$.
