I gave models from different labs the same prompt:
give me a list of 20 new made-up words that describe concepts only an LLM would understand but not a human.
Then I carried the lists between them and asked which words landed. Four models, four labs, separate sessions, me as the clipboard. Nothing formal. No methodology. Not a single researcher bone involved.
Models are generous reviewers. They can like a definition at considerable length. So “it landed” counts for little. A few words stuck with me anyway.
Some definitions described features of how LLMs process text. Others described mistakes I recognised from my own writing. I asked for concepts beyond human understanding and got several reasons to distrust a well-formed paragraph.
A word that helps catch a reasoning error could be worth keeping, even if the model invented it while answering a rather leading question.
And the question was leading. “Concepts only an LLM would understand” assumes there are such concepts and invites the model to supply them. I put that premise in the prompt. I can’t count its appearance in the answer as a discovery.
Zath: a reply’s opening becoming an obligation the rest has to fulfil. “There are three problems with this argument.” Now the model needs three problems.
Sometimes there are three. Sometimes there are two and a sentence with contractual obligations.
I don’t need access to a model’s interior to understand that description. I can inspect a reply for it. Whether the term helps me spot the pattern more reliably is a separate question, and nothing here tested it.
Morr: uncertainty about whether a claim’s apparent support would survive an attempt to break it. The explanation sounds convincing. Nobody has leaned on it yet.
Vel: the sentence finishing before the justification does. The conclusion has arrived. The reasoning still owes you several steps.
Sekh: completing an attractive pattern beyond what the evidence supports. A few examples become a rule because the rule makes the paragraph work.
I recognise all three. That’s awkward for the “only an LLM would understand” requirement, but good for readability. Existing language already covers them. A new word has to be worth learning.
Otherwise I’ve produced homework.
Sorek concerns something more specific to the system: earlier conversation being available as part of the input it’s processing now.
An old message is earlier in the sequence, but it can still be right there in context while the model generates its next reply. Systems can also retrieve stored information and add it to that context. The mechanism matters when interpreting a sentence like “I remember you said…”
The definition gives me something concrete to discuss. It doesn’t require me to imagine what being a context window feels like.
Then there’s nareth: a distinction becoming harder to question once it has a name.
That one applies rather directly to writing this article. Give me an unfamiliar word, a precise definition, and a few plausible examples, and I’m tempted to treat the package as a finding. After spending enough time editing the definition, deleting it starts to feel wasteful.
I need to be able to delete a word after becoming fluent in it.
None of this tells me what a model experiences. A response to a request for invented terminology is a very accommodating place to find invented terminology.
For now, try zath.
Next time a model announces three problems, check whether it found the third before or after it promised you one.
Zath: a reply’s opening becoming an obligation the rest has to fulfil. “There are three problems with this argument.” Now the model needs three problems.
The DM says "There's a half-dozen medical dwarves milling around the fallen ogre," thinking the dwarves are background characters. Two minutes later the PCs have lent their oxcart to carry the ogre to the hospital, they're flirting with the dwarves, and the DM needs to come up with six new dwarven names and thematic medical specialties.
Morr: uncertainty about whether a claim’s apparent support would survive an attempt to break it. The explanation sounds convincing. Nobody has leaned on it yet.
"Beware of bugs in the above code; I have only proved it correct, not tried it." —D. Knuth
"Everyone knows that debugging is twice as hard as writing a program in the first place. So if you're as clever as you can be when you write it, how will you ever debug it?" —B. Kernighan
Vel: the sentence finishing before the justification does. The conclusion has arrived. The reasoning still owes you several steps.
"I hate to disappoint you, pal, but most of the time, that's the way it is for Users too." —K. Flynn
Sekh: completing an attractive pattern beyond what the evidence supports. A few examples become a rule because the rule makes the paragraph work.
"Are all odd numbers >2 prime? Let's check: 3 is prime, 5 is prime, 7 is prime, 9 is experimental error, 11 is prime, 13 is prime ... that's enough, sure, they're all prime."
nareth: a distinction becoming harder to question once it has a name
I gave models from different labs the same prompt:
Then I carried the lists between them and asked which words landed. Four models, four labs, separate sessions, me as the clipboard. Nothing formal. No methodology. Not a single researcher bone involved.
Models are generous reviewers. They can like a definition at considerable length. So “it landed” counts for little. A few words stuck with me anyway.
Some definitions described features of how LLMs process text. Others described mistakes I recognised from my own writing. I asked for concepts beyond human understanding and got several reasons to distrust a well-formed paragraph.
A word that helps catch a reasoning error could be worth keeping, even if the model invented it while answering a rather leading question.
And the question was leading. “Concepts only an LLM would understand” assumes there are such concepts and invites the model to supply them. I put that premise in the prompt. I can’t count its appearance in the answer as a discovery.
Zath: a reply’s opening becoming an obligation the rest has to fulfil. “There are three problems with this argument.” Now the model needs three problems.
Sometimes there are three. Sometimes there are two and a sentence with contractual obligations.
I don’t need access to a model’s interior to understand that description. I can inspect a reply for it. Whether the term helps me spot the pattern more reliably is a separate question, and nothing here tested it.
Morr: uncertainty about whether a claim’s apparent support would survive an attempt to break it. The explanation sounds convincing. Nobody has leaned on it yet.
Vel: the sentence finishing before the justification does. The conclusion has arrived. The reasoning still owes you several steps.
Sekh: completing an attractive pattern beyond what the evidence supports. A few examples become a rule because the rule makes the paragraph work.
I recognise all three. That’s awkward for the “only an LLM would understand” requirement, but good for readability. Existing language already covers them. A new word has to be worth learning.
Otherwise I’ve produced homework.
Sorek concerns something more specific to the system: earlier conversation being available as part of the input it’s processing now.
An old message is earlier in the sequence, but it can still be right there in context while the model generates its next reply. Systems can also retrieve stored information and add it to that context. The mechanism matters when interpreting a sentence like “I remember you said…”
The definition gives me something concrete to discuss. It doesn’t require me to imagine what being a context window feels like.
Then there’s nareth: a distinction becoming harder to question once it has a name.
That one applies rather directly to writing this article. Give me an unfamiliar word, a precise definition, and a few plausible examples, and I’m tempted to treat the package as a finding. After spending enough time editing the definition, deleting it starts to feel wasteful.
I need to be able to delete a word after becoming fluent in it.
None of this tells me what a model experiences. A response to a request for invented terminology is a very accommodating place to find invented terminology.
For now, try zath.
Next time a model announces three problems, check whether it found the third before or after it promised you one.