Honestly, the best filter in my experience is Spamhaus Zen; it spots most spam with very few FPs. It's an IP blocklist.
In the end it's still a hard problem, so takes a bunch of work.
It's an unsolved problem for anyone.
> Why can I identify spam at a glance, with greater than 90% accuracy
I don't think you understand how terrible even 99.998% accuracy is at their scale.
> They probably have the worlds largest and highest paid staff of ML experts, and they resort to auto-flagging based on IP address?
You think they aren't using them for this? IP addresses very likely go into their model as well, when determining something is spam.
> Fix it.
They are and adversaries are breaking them again.
What does scale have to do with it, am I missing something? Wouldn’t a spam detection rate of 99.999% mean that only every 100000th spam mail would get through on an account basis, how’s that terrible?
The person I replied to said they can identify spam with 90% accuracy with a glance, my point was that it's abysmal as accuracy. Even orders of magnitude better it'd still be bad.
> Wouldn’t a spam detection rate of 99.999% mean that only every 100000th spam mail would get through on an account basis, how’s that terrible?
I didn't specify if it's per-account or a false positive rate, it's also actually not relevant to the point I'm trying to make.
When you receive billions of letters a month small errors start to matter and they're doing a better job than the person I replied to thinks they do.
Google reads and indexes every email that they deliver to a gmail inbox. They probably have the worlds largest and highest paid staff of ML experts, and they resort to auto-flagging based on IP address?
Fix it.