Interesting article. Thanks for posting.

> Assuming they could solve the problem of the headers, the spam of the future will probably look something like this: > > Hey there. Thought you should check out the following: > http://www.27meg.com/foo

Funny. 20 years later, that’s indeed how many spam messages look like.

> what’s the difference?

The key difference here is personalization.

Traditionally, if a message was personalized it fell under 'cold outreach' and users were more likely to interact and play along. Just like what happened with the author (the same applies for everyone).

It's like the difference between receiving a flyer vs being contacted by a sales representative. Even if it's they advertise the same product, the perception is different, the results are different.

If you're mean the difference from a pure technical spam detection perspective, I'm not familiar, but would love to read more about the subject and the state of the art techniques if anyone has some resources to recommend.

Do you read/answer cold outreaches then? Why?

Unless you're specifically looking for unsolicited offers, in which case you probably have a process for them, they seem like a waste of time.

> Do you read/answer cold outreaches then? Why?

Do you only read emails from recognized addresses? No new communication whatsoever unless it's initiated by you?

Not if they're trying to sell me something...

> Not if they're trying to sell me something...

How do you know they're trying to sell you something without even reading the email?

Your question was "Do you read/answer cold outreaches then? Why?" which doesn't make much sense. For me, and I imagine the same applies for most people:

1. You read until you find a clue that its content is not of interestt. Usually the email subject doesn't say much.

2. You only reply if you need.

Cold outreach are genuine emails that covers colleagues, new clients, job opportunities, someone reaching out to collaborate, etc. How you deal with it depends on your profile and who you've given your email address to. Personally, I have many email addresses, for some I don't even check my inbox.

> You read until you find a clue that its content is not of interestt. Usually the email subject doesn't say much.

You confusing "read" with "quickly skim"? :)

a) If someone manages to generate a letter that I actually find useful and interesting then I’m not sure I would mind if it was unsolicited. I don’t believe that the likeliness for that is super-high, though. And if a crappy message would get past the spam filter I would just flag it.

b) If you want to read more, feel free to check the link I posted. Paul Graham has thought/written a lot about this. I think one reason people has forgotten about those articles is that today, a huge number of us use Gmail, so we don’t actually need to think so much about how spam filtering is implemented.

> If someone manages to generate a letter that I actually find useful and interesting

But that's inconsistent with the example you put forward. For the email to be interesting a human would need to research and approach every prospect independently, how many emails a day they can do? 5, 10, 20, 100?

It's simply not possible for a human to generate 100,000 personalized email by hand. That's the difference.

Using a language model, one can craft an individually targeted email for each of those 100000 recipients. How do you "handle" this without doing anything current spam filters don't? Can you prevent an individual from sending 100000 emails a week? Can you make it cost them money?

Using an LLM to generate 100000 letters is hardly free, is it?

And AFAIK, Bayesian filtering (by the recipient) doesn’t require any knowledge of what other people has received.

> Using an LLM to generate 100000 letters is hardly free, is it?

No, but with further advances it might easily get cheap enough that spammers think it's worth it.

> Bayesian filtering (by the recipient) doesn’t require any knowledge of what other people has received.

Agreed. However, assuming people don't individually configure those filters -- which they currently do not and scaling this up would be something quite novel --, this seems quite gamable

John Doe is probably very good a generating sales leads! By definition, most sales leads are generated from unsolicited communications -- email, phone, etc. I expect the very best sales people will be using a combination of ChatGPT and genuine personalisation for unsolicited communications.