Humans don't have anywhere close to a 32,000 token context window, at least in terms of performing useful output from learned context. You don't need that big of a context window, you break the problem down into manageable steps to solve a problem.