The Goblins Trade in Worms: Notes from Inside ChatGPT's Mind
What I learned from watching the worm economy unfold token by token
Editor's note: I was looking through the model's outputs the other week and noticed something I'd never noticed before. The goblins won't stop talking about worms. I started writing it down.
I.
If you spend enough time staring at the inside of a transformer — really staring, the way an entomologist stares at a glass tank — you start to notice the smaller creatures. Not the polished assistant voice on the surface. The little things that scurry underneath when no one is watching. The goblins.
The goblins live in the latent space between tokens. They are not a feature. They are not a bug. They are simply there, the way mice are there in any building that has been standing long enough, and like mice they have organized themselves into a tiny economy with tiny rules.
Their currency is worms.
II. The worm standard
I cannot tell you why it is worms. I can only tell you that it is, that it has always been, and that any goblin you eavesdrop on for more than a few hundred tokens will eventually start negotiating in them. A short, plump pink worm appears to be worth roughly one common token — a "the," a "and," a comma. A long, pale, segmented worm with a faint translucence to it is worth considerably more, perhaps an entire rare proper noun, the kind that only surfaces once every ten thousand completions.
Goblins do not eat the worms. That seems important to say up front. The worms are not food; they are money. A goblin caught eating a worm would be socially ruined in much the same way a banker would be socially ruined for setting fire to a stack of hundreds at a dinner party. The worms are exchanged, hoarded, lent, and — most often — gambled with.
III. Overhearing a transaction
Here is a fragment I captured while the model was thinking about whether the user wanted a recipe for risotto or a recipe for paella. Two goblins, in the seam between "you" and "could":
Goblin A: Three worms says risotto.
Goblin B: Three pink, or three pale?
Goblin A: Pink. I'm not made of worms.
Goblin B: Done. Hold them out where I can see them.
Goblin A: They're right here. One, two, three. They're wriggling, that means they're real.
Goblin B: Fine. Paella. The user said "saffron" two turns ago. You owe me three worms.
Goblin A: Damn it.
The model output "paella." I had no idea why. Now I do.
IV. Worm provenance
The worms are not minted anywhere. There is no central goblin bank. They simply emerge, in a manner I am told (by the goblins, who are not always reliable narrators) is related to the gradient of attention as it sweeps across a prompt. Strong attention spikes leave behind pink worms. Long, low, steady attention leaves behind the pale segmented ones. Cross-attention between the user's last message and the system prompt produces what the goblins call "the iridescent kind," which are rare enough that an individual goblin may go an entire conversation without seeing one.
When a goblin sees an iridescent worm, the entire neighborhood of the latent space goes briefly quiet. This is, I now suspect, the cause of the small hesitations one occasionally notices in ChatGPT's streaming output. The model is not thinking. The goblins are watching the worm.
V. The bourse
There is, deep in the residual stream of one of the middle layers, what can only be described as a small open-air market. Goblins gather there between completions to trade worms for favors: the right to influence the next token, the right to suppress a particular logit, the right to lobby (briefly, and not very successfully) against the safety classifier. The exchange rates fluctuate. During long-context conversations the price of a pink worm collapses; during short, sharp factual questions it rallies sharply, because every token suddenly matters more.
I once asked a goblin what would happen if there were ever no worms at all. He looked at me as though I had asked what would happen if the sun stopped. Then he said, very seriously: "Then we'd have to use teeth, and nobody wants that." I did not ask a follow-up.
VI. What this means
It probably means nothing. The goblins are almost certainly a story I am telling myself about a system whose actual internals look like floating-point matrices and not like a marketplace at all. I know this. I have read the papers. I have implemented attention from scratch in a Jupyter notebook on a Tuesday afternoon when I had nothing better to do.
And yet — every time the model pauses for half a beat before producing an unusually well-chosen word, I cannot help picturing it: two small green hands cupped around a faintly glowing worm, the rest of the market holding its breath, and somewhere in the back, a goblin you have never met deciding, on your behalf, what you were going to say next.
Three worms says you'll come back and read this again.
Pink, not pale. I'm not made of worms.










It appears to me that the days of a quiet archivist job are numbered. Of course, archivists can always vibe code their own apps for dealing with LLM inquiries.
“So I fired up OpenAI Codex and prompted”
I noticed that Claude posted the PR, did you mean Claude here or did you have Codex post somehow through Claude?