Spam has two levels of cost - the victimization of the people who receive the spam, and the bandwidth and processing costs borne by ISPs and network operators. I know from firsthand experience working at a large ISP that anti-spam alone was a multi-million dollar cost center and accounted for around 85% of our bandwidth costs.
Based on what I've seen, improving filter efficiency at the backbone/ISP level has the paradoxical effect of increasing spam traffic - if enough of their messages aren't getting through, the spammers will just keep throwing more spam at you until enough leaks through for them to get a sucker. If you improve your filter efficiency by an order of magnitude, they'll just throw an order of magnitude more spam at you, and since they are typically using stolen resources to do their spamming, it's not costing them much (if any) more to do so. The cost remains near-zero and even a single successful scam can potentially net hundreds if not thousands of dollars.
The core problem is that SMTP is fundamentally broken with regards to spam (among many other problems). It's a problem the authors just didn't (and probably couldn't) envision. The problem won't go away as long as sending millions of emails is effectively free (in terms of time, computation, and money)
The rate at which a bot or rogue server can send email is entirely network-bound. Email has to be redesigned from the top down so that it is inherently rate-limited by some irreducible factor. Increasing the computational cost to send email using a memory-bound function seems to be the best way to do this (the rationale being that memory bandwidth has historically only increased at a fraction of CPU speed).
The tl;dr version is that crime pays... and unfortunately that will never change. What can do is change is the economics of sending email by increasing the computational cost by several orders of magnitude. This may not solve the problem of individuals getting scammed, but it will dramatically reduce network congestion and infrastructure costs.