alt Hacker NewsThe quote has absolutely nothing to do with my point.
The scaling to an interval in the quote is about formal mathematical reasons,in particular that polynomials do not approximate continuous functions globally. This is totally unrelated to numerics.
The issue is that in particular the interval 0 to 1 has to be chosen, as otherwise the numerics totally fall apart. The message of the article is that high degree polynomials pose no danger, but that is wrong. All the examples in the article only work because of a specific choice of interval. All the major numerical issues are totally ignored, which would immediately invalid the core thesis of the article. If you calculate 10^100 in 64 bit floating point you will run into trouble. The article pretends that will not be the case.
However, if you normalize your data to [0,1], you'll never have to compute 10^100 and thus never face any numerical issues. "Never" assumes no distribution shift.
Indeed, the examples work thanks to this choice of the interval, but this comes with the choice of the basis. Of course Bernstein basis functions explode outside [0,1], but I think the point is that high-degree polynomials pose no danger if you scale the data *according to the polynomial* (use [0,1] for Bernstein and [-1,1] for Chebyshev, for example). So the "magic combo" is polynomial + scaling to its interval. Otherwise all bets are off, of course.