Consider:

* spammers have access to large amounts of compute via their botnets

* the effectiveness of any particular spam message can easily be measured - it is simply the quantity of funds arriving at the cryptocurrency wallet tied to that message within some time window

So, just complete the cycle: LLM to generate prompts, another to generate messages, send out a large batch, wait some hours, kick off a round of training based on the feedback signals received; rinse, lather, repeat, entirely unattended.

This is how we /really/ get AI foom: not in a FAANG lab but in a spammer's basement.