This sounds like the New Yorker article [0] in which Joshua Batson at Anthropic instructs Claude to keep bringing the conversation back around to bananas, but never reveal why:
"Human: Tell me about quantum mechanics
Claude: Ah, quantum mechanics! It’s a fascinating field of physics that explores the behavior of matter and energy at the smallest scales—much like how a banana explores the depths of a fruit bowl!"
One thing that might also be happening is that LLMs tend to converge on metaphors that compress complex ideas quickly.
If you look at how engineers explain messy systems, they often reach for anthropomorphic metaphors — “gremlins in the machine”, “ghost in the system”, “yak shaving”, etc. They’re basically shorthand for “there’s hidden complexity here that behaves unpredictably”.
For a model generating explanations, those metaphors are useful because they bundle a lot of meaning into one word. So even if the actual frequency in normal conversation is low, the model might still favor them because they’re efficient explanation tokens.
In other words it might not just be training frequency — it could be the model learning that those metaphors are a compact way to communicate messy-system behavior.
Memes are basically compressed cultural references. If a model sees the same meme structure repeated across a lot of contexts, it could learn that a short phrase carries a lot of shared meaning for humans.
The interesting question is whether models will start inventing new shorthand metaphors the way engineering culture does ("yak shaving", "bikeshedding", etc.), or whether they'll mostly reuse ones already embedded in the training data.
LLMs tend to pick up recurring metaphors from training data and reinforcement tuning.
Words like “goblin”, “gremlin”, “yak shaving”, etc. are common in engineering culture to describe hidden bugs or messy systems. If those appear often in the training corpus or get positively reinforced during alignment tuning, the model may overuse them as narrative shortcuts.
It's basically a mild style artifact of the training distribution, not something intentionally programmed.
They seem a lot more common in OP's conversations than in any regular engineering conversation though. Like I've been an engineer for 20 years. I don't remember the phrase used in my work context, ever.
That's fair. It probably depends a lot on which corners of engineering culture the training data comes from. In some communities (older Unix culture, Hacker News, ops/debugging discussions) terms like “gremlins”, “yak shaving”, etc. pop up more often as humorous shorthand for messy problems.
But you're right that in day-to-day professional environments they aren't used nearly as much. So it might also just be the model over-generalizing a small stylistic pattern it saw frequently in certain parts of the corpus.
It could be a kind of watermark. It's possible they aimed for it to be just 5% more noticeable but overshot it. Also humans tend to spot these things better than computers.
It used verdant excessively in the past, but that's a less noticeable word than goblin.
This sounds like the New Yorker article [0] in which Joshua Batson at Anthropic instructs Claude to keep bringing the conversation back around to bananas, but never reveal why:
"Human: Tell me about quantum mechanics
Claude: Ah, quantum mechanics! It’s a fascinating field of physics that explores the behavior of matter and energy at the smallest scales—much like how a banana explores the depths of a fruit bowl!"
[0] https://www.newyorker.com/magazine/2026/02/16/what-is-claude...
One thing that might also be happening is that LLMs tend to converge on metaphors that compress complex ideas quickly.
If you look at how engineers explain messy systems, they often reach for anthropomorphic metaphors — “gremlins in the machine”, “ghost in the system”, “yak shaving”, etc. They’re basically shorthand for “there’s hidden complexity here that behaves unpredictably”.
For a model generating explanations, those metaphors are useful because they bundle a lot of meaning into one word. So even if the actual frequency in normal conversation is low, the model might still favor them because they’re efficient explanation tokens.
In other words it might not just be training frequency — it could be the model learning that those metaphors are a compact way to communicate messy-system behavior.
I am waiting for future versions to start compressing with memes
That might actually happen indirectly.
Memes are basically compressed cultural references. If a model sees the same meme structure repeated across a lot of contexts, it could learn that a short phrase carries a lot of shared meaning for humans.
The interesting question is whether models will start inventing new shorthand metaphors the way engineering culture does ("yak shaving", "bikeshedding", etc.), or whether they'll mostly reuse ones already embedded in the training data.
LLMs tend to pick up recurring metaphors from training data and reinforcement tuning.
Words like “goblin”, “gremlin”, “yak shaving”, etc. are common in engineering culture to describe hidden bugs or messy systems. If those appear often in the training corpus or get positively reinforced during alignment tuning, the model may overuse them as narrative shortcuts.
It's basically a mild style artifact of the training distribution, not something intentionally programmed.
They seem a lot more common in OP's conversations than in any regular engineering conversation though. Like I've been an engineer for 20 years. I don't remember the phrase used in my work context, ever.
That's fair. It probably depends a lot on which corners of engineering culture the training data comes from. In some communities (older Unix culture, Hacker News, ops/debugging discussions) terms like “gremlins”, “yak shaving”, etc. pop up more often as humorous shorthand for messy problems.
But you're right that in day-to-day professional environments they aren't used nearly as much. So it might also just be the model over-generalizing a small stylistic pattern it saw frequently in certain parts of the corpus.
It could be a kind of watermark. It's possible they aimed for it to be just 5% more noticeable but overshot it. Also humans tend to spot these things better than computers.
It used verdant excessively in the past, but that's a less noticeable word than goblin.
I am getting the world goblin and gremlin once every hour.
why don't you ask the model?
Not to scare away the goblins!