> It’s as if the world now requires you to be more precise about everything.
This is something I've been thinking a lot about. We have trended from subjective language to objective language. Why?
Computing. Software is written with objective language. Everything is clearly unambiguously defined. Blue is no longer a category, it's #0000FF. Logic must always reduce to a binary truth value. Most of what we have to talk about is somehow relative to software. Software even structures most of what we write! We don't just talk to each other, we tweet, email, message, post, search, etc. These structures each imply a specific set of phrase structures that can make sense.
Lately, it's hard to go even a day without reading some complaint that such and such was written by "AI" (an LLM). Why is this so obvious? Well, the core advantage that LLMs provide is that they don't compute. Inside an LLM, there is no arithmetic, no logical branches, no truth values. Phrases aren't generated to define or to resolve. They are generated to continue. Sure, we can direct the story to follow the steps of logical deduction, but that isn't anything like calculation. An LLM simply isn't invested in logic, precision, correctness, etc. the way we expect modern writers to be. It's not the em-dashes or the word choice that illustrates this, it's the fundamental perspective of the system.
We are sorely missing subjectivity. Natural language never was, and never will be, computable. You can't reduce a natural story to binary truth values without choosing an arbitrary perspective that resolves its ambiguity. The more precisely abstract our language gets, the more detached from reality our stories become. The more objective our assertions about reality are, the less relevant they can be.
My answer to this is to make the arbitrary choice of perspective a first-class feature. If we can explicitly decide what meaning is relevant, we should be able to weakly solve natural language processing. It seems like a pretty simple and obvious idea, but so far is easier said than done.
I don't agree with the assumed equivalence between precision and objectivity. Language can be precise when describing a subjective perspective, and it can be imprecise or ambiguous when describing objective facts. The tendency toward increasing precision is a good thing, in my opinion. If anything, precision in language allows writers to tell better, more nuanced stories about subjectivity.
Increasing precision is good to the end that is disambiguation. The problem is that that end is not always what language is good for. Metaphor is an incredibly useful tool, because it allows us to convey meaning with imprecise definition, but precise implications.
The meaning of an expression is its implication, not its definition. The implication must be decided by the reader. This is either done by strict definition, or by arbitrary inference. Inference is a powerful feature of language, because it expands the set implications that can be expressed without abstracting the language. Without inference and metaphor, we must instead create a tower of well-defined abstractions that rule over us; having already determined what implications are relevant: objective truth. With inference and metaphor, the writer/speaker can articulate their implications with familiar language; relying on the reader/listener to infer from the expected relevant perspective: subjective truth.
Essentially the tradeoff isn't whether or not you get precision, but whether that precision can be shared by the writer and the reader, or be managed entirely by the writer. Defined abstractions allow the writer to be unambiguous to anyone (objective), while inference allows the writer to offload that responsibility to the context and perspective of the reader (subjective).
Does that make sense?
Eh, this seems somewhat right, but somewhat not right either.
Language has to do with our Monkeysphere, that humans over long periods of time in the past were limited to a very small subset of people they interacted with socially and closely. A few institutions likely had a large effect on your language like the church depending on where you were. After that it was the people you interacted with to stay alive. Because almost everything was in person or person to person transfer of information a lot of socially encoded clues were involved which lessened the need for well defined words.
Books were the first stage of homogenizing language as they could be shared over long distances and to many people, but more sequentially than latter forms of communication. After that radio and TV had a huge effect, for example the 'General American' used in broadcasts that was based heavily on a midwestern accent.
As we encroached on the 70s and 80s the previous technological advances and things like high speed interstates and trucking shrank America to something you could drive across in less than a week, and you could reach anywhere by voice nearly instantly. Suddenly people in California, Texas and New York all could be in the same meetings and local colloquialisms would need explained, so people would trend to a shared vocabulary.
It's also odd to me to say an LLM isn't subjective. Each LLM has it's own behavior, it's that there are like 20 or 30 big LLMs in all, and people are using them millions to billions of time so we're getting that one LLMs language everywhere. And that's why I disagree and will say natural language is computable, but it's also lossy and probabilistic. And for the most part it's single prompt and being ran by the user for the cheapest price possible.
> It's also odd to me to say an LLM isn't subjective.
I was saying the opposite. The core feature of an LLM is that it is subjective. The implications of a written expression (prompt) are not well-defined objective truth, but instead a vague probability. We can compute the probability, but that doesn't ever intersect with logical reduction or arithmetic; so the implications we get are just vague guesses on the progression of the story. With enough examples and training, the LLM can guess correct arithmetic answers, but critically, it does not actually perform the arithmetic that verifies them.