Six growth classes, one per row, and the input size n doubling along the columns. Each cell will hold the operation count that class predicts. Big O is about how the count changes when n doubles, not about the count itself.
Growth table for six classes and sizes 1 to 32.
n = 1. At n = 1 every class costs about the same: 1, 0, 1, 0, 1, 2. Constants hide the differences.
n equals 1: counts 1, 0, 1, 0, 1, 2.
n = 2. Doubling n added 1 to log n (0 to 1), doubled n, roughly doubled n log n (0 to 2), quadrupled n squared (1 to 4) and squared 2 to the n (2 to 4).
n equals 2: counts 1, 1, 2, 2, 4, 4.
n = 4. Doubling n added 1 to log n (1 to 2), doubled n, roughly doubled n log n (2 to 8), quadrupled n squared (4 to 16) and squared 2 to the n (4 to 16).
n equals 4: counts 1, 2, 4, 8, 16, 16.
n = 8. Doubling n added 1 to log n (2 to 3), doubled n, roughly doubled n log n (8 to 24), quadrupled n squared (16 to 64) and squared 2 to the n (16 to 256).
n equals 8: counts 1, 3, 8, 24, 64, 256.
n = 16. Doubling n added 1 to log n (3 to 4), doubled n, roughly doubled n log n (24 to 64), quadrupled n squared (64 to 256) and squared 2 to the n (256 to 6e4).
n equals 16: counts 1, 4, 16, 64, 256, 6e4.
n = 32. Doubling n added 1 to log n (4 to 5), doubled n, roughly doubled n log n (64 to 160), quadrupled n squared (256 to 1024) and squared 2 to the n (6e4 to 4e9).
n equals 32: counts 1, 5, 32, 160, 1024, 4e9.
At n = 32 the rows read 1, 5, 32, 160, 1024 and about 4 billion. That spread is the point: O(2^n) is hopeless beyond tiny inputs, O(n^2) is fine for thousands, O(n log n) for millions, and O(log n) barely notices n at all.
The classes diverge by a factor of billions by n equals 32.