How My Dad Out-Generates a Language Model At The Dinner Table

 

Post adapted from "Reading Between the Lines" Issue #2

Temese Szalai, Principal & CEO, Subtextive


Many of us either make up words and expressions or know someone who does. Maybe both. As humans, we do this a lot. It turns out that invented expressions teach us a lot about context and shared understanding.

“Hey, kids, settle down and stop being so armbunctious!”

My entire life, there have been some words and phrases that are unique to my dad. A native speaker of Hungarian, he learned English mostly through immersion when he came to the United States at 20 years old. He celebrated his 90th birthday in September 2026, which makes for seven decades of speaking English.

Overall, his command of English is excellent. He speaks fluently, reads complex content, has authored books, and has an uncanny ability to appreciate humor and make jokes.

The expressions I’m thinking of go well beyond what we think of as a person’s idiolect. (No, that is not a dialect for idiots. An idiolect is the distinctive way that a particular person uses language.)

These are words and phrases that he has made up using building blocks that legitimately occur in English, i.e., patterns that he has observed. He uses them fluently. And, they are understandable from context.

“Armbunctious” is one example. Another is “swallowed up” for “swollen”: “Her ankle was all swallowed up because she twisted it.” He even once referred to the character “Large Marge” in “Pee-wee’s Big Adventure” as “Fat Martha.” So close!

He’s prolific with the expressions he makes up. As far as I can tell, he’s never let not knowing a word stand in the way of communicating. He just makes one up.

He’s also been known to productively riff on other people’s made-up words. My sister made up the endearing term “nu-nu” to refer to one of our dogs. Once, when we were out somewhere and trying to figure out which direction to go, my dad said, “Let’s not stand around like a bunch of nu-nus.” Previously, this term had only ever been applied to the dog: “Oh, you’re such a nu-nu.”

Are made-up words inventions or mistakes?

Each of these expressions starts from a word or concept he clearly understands and is built from real English sound, meaning, or sentence patterns. He isn’t anglicizing Hungarian. He’s producing something new that follows rules he’s absorbed. It conveys exactly what he means and how we’re supposed to react.

It just so happens that the constructions he’s using aren’t standard English. Are they really inventions? Or, are they mistakes? I lean toward “inventions” because that’s how human language works.

We invent expressions all the time.

Whether it’s for world-building in movies like “Wicked” with made-up words like “encouragerize” and “disrespectation,” creating catchy brand names like “Shopify” or “Google,” or simply adding dramatic flair with f-bomb infixes, we make up new words constantly, and with supporting context, others understand them on first exposure.

It’s all part of our innate generative capacity for language. It’s our ability to apply known patterns and rules to create and interpret novel utterances or expressions. This works when there’s shared context and shared norms and standards.

Usually when we encounter this from people, rather than machines, we don’t blink an eye. Sometimes, we giggle. At other times, the new expression catches on and other people start to use it.

Large language models (LLMs) invent things, too.

We know LLMs make things up. It’s not surprising. Their job is, literally, to create content!

When we think about LLMs inventing things, though, AI hallucinations and slop come to mind. Hallucinations are dangerous because they sound like coherent facts but aren’t true. Slop is different. Slop is content, often produced in bulk, made of sentences that are grammatically sound but have little or no meaning behind them.

For example, not long ago, as part of some editorial workflows I was testing out, one of the major chatbots gave me this explanation regarding one of its editing suggestions: “I just renamed the section header slightly less than I considered.”

What??? I felt like I’d fallen down the rabbit hole. Grammatically, it appeared fine. It just wasn’t anchored to meaning.

The LLM was being a nu-nu. It had fallen back on simply generating content for the sake of generating content.

Why do we think of many LLM inventions as mistakes?

We think of many of the things that LLMs make up as errors. Their made-up expressions, unsubstantiated factual statements, and semantically void sentences aren’t endearing to us at all. We’re counting on these systems to be reliable. Don’t they get it?

Well, they are reliable in a certain way. I mean, these systems have been built to make things up. And, actually, it turns out that they do that quite reliably!

But, that’s not really what we have in mind when we say “reliable,” is it?

So, we get annoyed when their inventions don’t meet our standards, which are often implicit, contradictory, or situational.

In the case of my dad’s invented words and expressions, I’ve never seen anyone get annoyed. He uses them in situations where there is enough supporting context and shared norms that they’re interpretable, even if someone hasn’t had years of dinner-table conversations with him. Importantly, they’ve been anchored to meaning, reality, and intent to communicate, all of which are implicit.

And, this is why I think that the things LLMs make up are so aggravating. They often aren’t anchored to any of these things because the implicit hasn’t been made explicit. We expect them to know better. But, they don’t, and they won’t unless we tell them.

AI needs the context we usually leave unsaid: know the way and show the way.

We usually have the context we need to make communication work, and when we don’t, we can fill in the blanks because we share understanding and purpose.

That’s why my dad out-generates a large language model. He’s in it to communicate, not just to generate content. It’s the quality of what he invents that outperforms LLMs.

AI systems don’t come equipped with human understanding, purpose, and standards. Because they need explicit context that establishes those things, we have to supply it.

That means spelling out who will use their outputs and be affected by their actions, the situation and assumptions they’re working with, the guardrails they operate within, and the quality standards that matter. (For a real-world example of what happens when nobody spells out what a word like “insight” should mean, see When Everything Is an Insight, Nothing Is.)

Context won’t make them reason like us, and it won’t stop every bit of slop or every hallucination. But, without it, they don’t stand a chance.

We have to know the way and show the way so that they can create human-centered, human-interpretable experiences that we can trust.

Fortunately, we can keep being as armbunctious as we like as we do that :)

Thanks for reading!

If you enjoyed reading this post, consider signing up for our newsletter “Reading Between The Lines”. It’s distributed up to twice monthly, some time around the middle of the month and the end of the month/early the next month. All newsletter issues get posted on our blog, too.

Reach us at hello@subtextive.com.

Temese Szalai