Random thought: why isn't LLM-based spam filtering ubiquitous yet? Some of the obvious spam that lands into my inbox would be caught easily even by the tiniest, cheapest models.
High-effort spam is quickly becoming the norm and comes from spammers who take the time and expense to warm up their domain, correctly configure their DNS records, use TLS certificates and ports, are careful not to send too many messages to one provider at a time, address and format their emails to make them look like newsletters, and so on. Traditional spam mitigation system are not very effective against these techniques because it was always believed that spammers could never justify the expense of high-effort spam. That may have been true at one point, but it was silly to think it would always be true.
An LLM that can read an email and quickly perform sentiment/classification analysis on the content itself is the only thing that I can see which will fill that gap. (Aside from just choosing to live with it, of course.) I'm just getting into this myself but my understanding is that very small models can do this surprisingly well.
As to why it's not ubiquitous yet, I suspect it's only a matter of time. My Outlook (for work) has been automatically summarizing all of my emails for months now (something I can't opt out of, by the way). Hooking that up to an action that sends spam messages into the Junk Email folder is just a few lines of code on Microsoft's part.