It sounds (towards the end) that this is saying that accepting a approximation of correct is often good enough, and that can often be made polynomial time.

Which is of course correct. There are lots of algorithms where there exist theoretical minimum complexities, but good approximations can do far better: lossy compression for example vastly beats the Shannon limit, google maps doesn’t take decades to plot a route from San Francisco to New York, etc