I'm not forgetting that. It's just that any specific random file will have (with high probability), patterns that can be exploited. It's not guaranteed to work on all files of course. The theoretical limit (per bijection argument) is that at most half the strings of length n can be losslessly represented in strings of length n-1, because there are half as many strings of length n-1 than of length n.

So if, for example, 2% of files can actually be reduced in size (I don't know what the actual number is, and if it is even computable) that's still positive EV in a $100 vs $5000 bet.