> if that string can be a number, that's obviously what was intended
Maybe that would be a reasonable assumption in a world where programmers were omniscient and never made mistakes. In practice these "do what I mean" rules just tend to make it easier for poorly-specified programs to produce nonsense without signaling that anything is wrong. If your program intends to add two numbers and one of the inputs turns out to be a string then someone gave you bad input; the right solution to that is to report an error, not continue with bad data. Perhaps it's only an accident that the string happens to look like a number this time. Maybe it wasn't intended to be a number at all, and some parameters got swapped or an input file was corrupted.
Also, that rule is too broad. The string "5" could be treated as a number, sure, but so could "five" or "⑨" or "¾". So why is "five" + "⑨¾" an error and not 14.75 (or even "14¾")? If you want to accept number-like strings as numbers, fine, but there should be an explicit process where the programmer defines exactly what form the input can take and how to perform the conversion to a number. Implicit conversions skip that step and expose the internal defaults of the language to the end user.
If you could do relative comparisons like "5" < 5 and get a boolean answer, that would be surprising. I don't see anything controversial, however, about unequivocally reporting that objects of different types are unequal. In Lua anything can be tested for (in)equality but asking whether a string is greater than or less than a number is an error, as is performing an ordered comparison involving any other type(s).