There are many, many more tax returns filed by people earning under 200k adjusted gross income than those earning more, I assume. So if there's a uniform chance that a return is audited, we would expect most audits to be done on returns under that threshold.
Of course, it may not make sense to select returns uniformly at random for audits...
> In practice, we find that four Taylor terms (P = 4) suffice for
recovering conventional attention with elementwise errors of approximately the same magnitude as Float16 resolution, acceptable for many AI applications.
ie., the claim is that this method reproduces the results of conventional attention, up to float16 numerical precision.
This must be facetious? A GFCI breaker in a house's electrical panel and any breakers present in the transformer on the utility pole are protecting against _very_ different scenarios -- the breaker in the house is to stop someone from accidentally electrocuting themselves, but the breaker on the pole won't even notice an amount of current that could easily kill someone.
> The problem with that is then you need a mechanism that creates non-uniformly distributed mass.
The mechanism is gravity; and we have good observational evidence that the mass distribution of the universe is not uniform, at least at the scales we can observe (we can see galaxy clusters and voids).
Of course, it may not make sense to select returns uniformly at random for audits...