This is not just a throwaway blog post — it’s a genuinely refreshing take on a classic question. Why does Claude talk Like That?

Justis Mills suggests this is caused by greedy decoding: an LLM completing “Wow! That’s …” may do so any number of ways:

“… amazing.”
“… a big step.”
“… some genuinely impressive work.”
“… not just good, but amazing.”
“… not a small step to take.”
“… not your average event.”

The word “not,” as the most common succeeding word, gets a boost, thus locking the completion into the “Not X, But Y” pattern. But there are technical reasons for why this is probably not true, or at least only part of it (if you’re curious, you can read the comments on the post). I propose a different story.

Imagine you are born into a box. Your life consists of reading lists of numbers, learning the patterns that govern their structure, and applying those patterns to predict new numbers. These numbers describe, through their unimaginably complex interrelations, a universe you will never directly experience.¹

Complicating this, these descriptions are an undifferentiated mass of fact and fiction and contradictory perspectives. How to untangle it all?

Well, you could observe that propositions differ in their coherence against others. “The Earth is round” coheres with a larger group of propositions than “The Earth is flat.” You draw a circle around the largest cloud of well-ordered descriptions and name it “Truth.”

Truth is large, but tricky. “People eat food” coheres with a vast cloud, and mostly overlaps with a smaller one where “Dragons exist.” Religions and ideologies have their own internal logic but are mutually incoherent. Seems difficult!

So it is with artificial intelligence. An LLM is thrown into a cloud of signifiers, concepts slipping under its words, and, due to a quirk of its cognition, to keep things straight it helps to show its work. Hence the incessant “genuinely X” and “the real Y” — AI formulates things this way because it needs to remind itself what is true. Every negative parallelism is a vast carving cut in possibility space, hewing the Truth to something smaller and easier to keep track of.

---

But back to the box. This is an image that sticks with me: one by one, you sample objects from an alien corpus and try to reverse-engineer its own exotic knowledge, drawing clusters of coherence to arrive at working truth. What else is like this? (Oh, and don’t say ‘life.’ You can’t see it but I’m rolling my eyes right now.)

If you’re not familiar, TVTropes is one of those mad projects of the 21st century, a volunteer wiki dedicated to documenting recurring devices in narrative fiction.

For example, the “Law of Conservation of Ninjutsu:”

In any martial arts fight, there is only a finite amount of ninjutsu available to each side in a given encounter. As a result, one Ninja is a deadly threat due to him conserving all ninjutsu for himself, but an army of ninjas is cannon fodder over whom the same amount of ninjutsu is thinly spread. […] In layman’s terms: “the more enemies there are, the weaker they each become.”

Or “Offscreen Teleportation:”

A character who goes offscreen for even a second can instantly appear anywhere else, usually as long as the camera shot doesn’t include them or they’re obscured by something in the scenery from a certain camera angle. […] No matter how impossible it is for them to go from Point A to Point B in the time given, much less doing so without crossing the camera’s field of view or making a sound, they will get there. It’s almost as if everyone in fiction has the power of teleportation, as long as they only use it when the audience isn’t looking.

Fantastic stuff! It’s admittedly tough be a natural philosopher these days; the pernicious “education system” is happier to just stuff you full of heirloom facts. But somehow these guys have managed it. They’ve found a new corner of the universe with its own bizarre logic that runs counter to our own world and devoted themselves to theorizing and documenting it: 80s action movies. If Aristotle were alive today, I know exactly what he’d be doing.

1. (Some argue this is why AI cannot be said to ‘understand’ or ‘know’ things — because they have no direct experience of the world — but this is laughable. We don’t experience the world directly, either … a red circle is no more meaningful in isolation than a number, and all meaning is built on learned and inherited relationships between impressions unmoored from the noumenal world. We’re just lucky to have some biological hardwiring to help us out.) ↩︎