A short and non-technical praise of AI slop. A writeup on the Muon Optimizer is coming shortly. I have a few ideas to test about it.
I came across a really nice blog post from Stephen D. Turner where, among other things, he argues that the fact that AI brought down to zero the marginal cost for text/idea generation will make it harder to tell what’s important or not.
He also thinks that institutions will gain power because they will be the only ones with enough bandwidth, skill and power to define the quality of new work.
I don’t think it’s true.
Slop is much easier to spot now than before: it has now much lower variance and thanks to this fact, pangram is enough to filter out 99% of it.
I scan every single substack post for AI text, and what remains is always high quality. Much higher quality than the average pre-2022 SEO-optimized blog post.
AI is a blessing in that sense: all attention-hacking attempts now look alike. It’s incredibly powerful because now we all have a reliable measure of:
whether the guy who’s writing to you is taking you for an unsophisticated moron (only then would he not fear you pangram his shit)
the time he took to write
his own lack of taste and originality
I’m probably going to look like a fucking snob, but you can immediately tell whether someone comes from an aristocratic background (or could be part of that world): aristocrats dont gaf about 99% of social conventions, and only care about a handful that really have discriminative potential. Basically everyone who wants to belong in this milieu just cannot, because they have no idea what’s important, and signal heavily on what’s not. They actively signal they don’t belong.
AI slop is similar. People who haven’t been used to communicating interesting/original ideas by writing things themselves have no idea how easy it is to spot AI slop. So they think: “God, this AI-generated text looks serious, no one will be able to tell it’s bullshit!!”, just like bourgeois people think “I sound so aristocratic with this high-pitch voice and disdainful attitude! No one will tell”. And so they actively signal “my shit can safely be ignored”.
I’m definitely a fucking snob.
Things were harder before AI. Slop signals were much more diverse and unreliable:
typos (some genuinely brilliant people suck at grammar, but statistically it worked rather well)
fancy emails, or any email starting with “we’re introducing blah blah”. But even then, sometimes, people would present interesting things. Speaking of which, I believe that email marketing was disastrously ineffective in the past, compared to what it could be. I believe I have more than a 40% response rate to my cold emails (even now), mostly because I always make sure I have a good reason to reach out to someone and just go fuckign straight to the point.
bad ideas. But you’ve got to understand the ideas telling knowing they’re bad. So you were halfway through an email or a blog post before figuring out you’d been wasting your time.
A lot of domain-specific stuff.
Before AI, a guy who ran a construction business and was pitched a new software had to dig deep for signals that the software guy was competent and trustworthy. Now, he can just check whether his pitch is AI-generated, and he will have a good idea about it.
If every crook and incompetent person had “don’t trust me” written on their forehead, life would be easier (except for the blinds). It’s now the case, at least for a little time!
.