Large-Language Failures

One is surprised such a thing is possible at all

Large-Language Failures

Large Language Failures

Join the Nerd Roundup Discord Server!
Check out the Nerd Roundup community on Discord - hang out with 81 other members and enjoy free voice and text chat.

I'm teaching myself to write long form fiction.

As a self-taught programmer and very analytical person, this has been quite an adventure! Thinking mathematically and creating meaningful narratives seem to me to be quite opposite endeavors.

In a big way, math is context-free. When I say x^2 + y^2 = z^2, I'm stating the relationship between the length of the sides of a right triangle. But it's no particular right triangle; it's every right triangle. It's up to the human to take some particular right triangle, measure some sides, and "plug in" the formula.

In fact, it's not about triangles at all. You could use the formula (or a version of it) to measure the distance between points in a multidimensional space. It works because you can also think of the relationship between Cartesian points as involving a right triangle.

The point is that there's no triangle at all, really. It's just how some symbols can be matched up, by a human, to some things they observe in the outside world. The "matching-up" part is a human one, not a mathematical one. Without humans taking various pieces of math and associating them with various things in the world, math just "is". It exists, completely outside our experience and our universe. The Pythagorean Theorem existed a long time before monkeys were drawing triangles in the dirt and will exist billions of years after we're gone. Math, if you think about it, is really not related to the universe at all. Some mathematical expressions may never be able to be mapped to anything we could observe in this universe but that might not be the case in other universes. (For instance, curvature of the universe would make the Pythagorean Theorem work different ways)

The human experience, however, is on the other side of the spectrum-of-absoluteness. Today I might see a butterfly and cry. Tomorrow I might see a butterfly and laugh. There may be no other stimulus an outside observer could see. There might not even be a stimulus that I consciously feel myself, yet still this relationship holds.

Math is "everything is absolute yet has no context". Existence is "everything is context and nothing is absolute" It's the intersection of the two that intelligent, sentient species play in.

One might think we could take every human, every butterfly, record every interaction, create some probabilistic tree that would closely approximate what might happen when they interact. One would be correct, life is extremely probabilistic. This is the basis of insurance companies.

In that way we could "fake out" about every part of the human experience, eventually being able to create artificial mathematical universes in which seemingly real people go about their daily lives. An outside observer would be unable to tell the difference between these simulated people and real ones. Sorry Turing.

But it would not capture the actual experience I have seeing that butterfly, no matter how deeply or precisely those people were created. That's because we're teaching the systems to learn as an outsider observing things. We actually become human from the inside-out, starting as a little baby and eventually learning to communicate with words and text. Once we start communicating we begin creating training data for LLMs to operate on. But that's a long ways down the road from conception.

It also presumes that human language, or any language, is a prerequisite for intelligence. It is not. Plenty of illiterate people, people who can't speak or listen, people with various disabilities, are extremely intelligent. As far as we know, humans were plenty intelligent for eons before they started writing things down. Those people were just as intelligent as any of us.

So when you're writing about humans or any other sentient creature experiencing a narrative, you're trying to take words, a latecomer to the game, to describe non-verbal things. It's like trying to eat blueberries with boxing gloves: it's possible, but to really do it you're working against the only tools you have.

A few months back in my fiction work I was introducing a character, a 19-year-old daughter, and her family. They obviously loved one another, but love is a funny thing. Perfect families are pretty boring and cookie-cutter from the outside, as Tolstoy observed.

So I had to describe the dynamics of the family, not what they did or felt. How does the emotional engine that is this family function? Well, let's give dad a quirk. Let's have dad love telling historical stories. I have this quirk myself so I should be able to write about it.

"Dad loved telling these stories at the oddest times. She hated it. She loved it. He could never know."

Last part of that: "She hated it. She loved it. He could never know.""

What the hell does that mean, anyway? I mean, I know what it means. I suspect you do too. It's words in sentences, so anybody or anything could parse it.

To me, that's fiction writing. It's when you've got a bunch of words, but they express something that words can't express. We are storytellers. We experience stories in a pre-verbal manner. In fact, it's those seemingly nonsensical, contradictory hunks of text that make stories so powerful. When I create that contradiction in the mind of the reader, each reader resolves the contradiction on their own.

Life consists of each of us resolving the contradictions on our own.

One day we laugh. One day we cry.

It's not the butterfly.

I'm out.