> Bayesian filters worked very well for email spam.
As someone flagged me, I'd like to know how well current filters catch LLM-generated spam? By hypothesis, not well, especially given that you can always run a LLM output through f(g(h(...))).
It doesn't matter if human/LLM or its simply preset text is the source. Spam has to be consumed and "bought" by a human, so it is easy to classify as humans are predictable in what scams we fall for.[1]
Most spam is advertisement for something to make you click, In the 90s and 00's in the early days of email, it was things like Viagra etc, today it is "coin drops" and crypto related things.
Naive Bayesian filtering is a just a matter of training on the probability of such words in regular issue/PR/discussion comment threads and assigning a probability for the post and flagging it when crossing a threshold
In the case of Github, they would probably refine and improve this by adjusting the weights for different topics.
There is a good chance they already do this, and it just that the sensitivity for crypto scam words is set too low in crypto related projects as they would be probably used more by real people as well, and that is why OP noticed this as issue and rest of us rarely see much spam in Github.
You could add reputation for the user globally and with respect to the project (akin to ESP reputation) and many other refinements in addition to Bayesian filtering.
[1] Nigerian prince emails are written in poor English for a reason for example. You could bypass a filter yes, but people are far less likely fall for such evolved language defeating the purpose of sending spam in the first place.
Google ( and most others) has been able to keep spam classifed long before transformers were even introduced with mostly Bayesian filters.
Not saying it can or can’t be solved with transformers, just that could do it with lower cost(computational) older methods just as easily .