Why is AI generated language so hard to understand? (arxiv.org)

And pretty much universally, they were awful. It was like somebody had blindly taken an English writing textbook and applied every little helper rule to every clause of every sentence, paying no attention whatsoever to deeper concepts like pacing, drawing attention to key words, and moderating emotion.

And I think that's basically what's going on in all of the frontier models as well. When they are being fine tuned, some reward score is telling them to be concise, not to use filler words, to avoid run-on sentences, etc. And there's no counter-balance in the fine-tuning that looks at the writing and flow as a whole.

At least in the case of ChatGPT - it was (and mostly still is) objectively optimized for engagement. Since ChatGPT was leading the way, and much distilling has occurred, there are at least partial remnants of that problem everywhere.

The Chinese models may or may not really care about this at all, but they're doing a lot of distilling, so they're getting it whether they want it or not.

The American models for the most part care about engagement - some a lot more than others.

Maybe we'll get assistant models that will get optimized for human readable prose, but I think the labs have deluded themselves when it comes to utility/coding models and think the humam isn't going to be a variable soon.

I have a jira-esque ticketing system for agents/subagents and the tickets are basically unreadable to me. I hate it. But if I force them to write the tickets in a prose I prefer their preformance degrades. Its almost like this gibberish hard to read prose is designed for llms to communicate betweem themselves probably because the llms are doing most of the RLHF (RLAIF??) for training now.

Things that are easy for humans, like remembering a character is holding a gun during a conversation, gets forgotten, because there is no clear architectural/spacial understanding of how the words are describing an underlying reality. And if the LLM gets this right it might instead forget which character has a Southern drawl, or make an illogical metaphor.

This is all more clear with programming, but the problem is the same - a lack of interwoven semantic understanding and friction between the real world and the self-contained world of the LLM's context.

And then it is also trained on it's our claude code sessions, so the more people accept claudish the more it becomes present in training data.

Languages change over time, and one of the well-documented "laws" of language change (using that term very fuzzily) is that people like to take irregular features of a language and generalize them to make them more regular. And there are many English verbs (particularly verbs expressing a change of state) which can be used transitively or intransitively, with equal validity:

> "I opened the door" -- "the door was opened" -- "the door opened"

> "the heat melted the ice" -- "the ice was melted" -- "the ice melted"

> "the impact cracked the glass" -- "the glass was cracked" -- "the glass cracked"

This generalization has been common, especially in jargon-y contexts, long before the rise of LLMs. A pedant might insist on saying "wait for the compiler to compile the code", and Merriam-Webster thinks "compile" can only be used transitively. But I think most developers would say "wait for the code to compile [i.e. be compiled]" is fine too. (see e.g. this XKCD comic from 2007: https://xkcd.com/303/)

In my view, this usage is completely fine as an occasional thing, especially when the social process of language change has normalized it over time for a particular verb. The issue is that LLMs drastically overuse it and pepper it everywhere. Just like "it's not X, it's Y", which is fine occasionally in human writing, but has become a dreadful cliche through overuse.

In your other comment, “Talking also costs.”, it’s clarified through context that talking costs tokens.

I think it’s a side-effect of the AI companies trying to reduce verbosity and redundancy, which were my personal main complaints with LLM writing. Though I believe LLMs are still verbose, and I suspect it’s because they don’t know what to remove, and these are cheap gains where the LLM removes verbosity that is obvious but expected because everyone does it.

This is entirely speculative and highly disputed, even if you are doing the more sophisticated thing here and by Language you mean something like mentalese.

Much cognition is embodied and makes use of visual, tactile, auditory, kinaesthetic, and other sense modality representations. "Thought is mostly a Language" has been a wordcel kind of argument with little empirical or philosophical support for a couple of decades now, IMO. Language is how we communicate thought, not primarily how we think.

Not that this refutes your main point: LLMs are trained almost entirely on (written) language, and since language is only a single, limited aspect of real cognition, this makes them so annoying and alien to read.

The neat thing about LLMs is that in a way they proved that only the truth turns out to be grammatical (statistically), largely because there are so many more ways to be wrong than right that the mode is almost always right.

But people also lie and bullshit in consistent ways, with the intention of fooling people with well-known empty arguments. Part of being a "good writer" is to make the meaningless sound meaningful - and the techniques are ancient.

Train on a bunch of people trying to push their framework and get rich, whether in online garbage, literary garbage, or in ostensibly serious nonfiction or scientific works, and you'll end up with modal language that is trying to play deceptive games with reasoning, but has nothing in particular it's trying to convince you of (no motives.)

People would either fall for that language (if they weren't thinking much and the context was one that comforted them), or they would methodically try to put that language into a model (let's say a series of syllogisms for an argument about what to do or what to believe, but it could be any model), and see that it failed. And speaking again of the context, the context of humans interacting with human content implies motives that help the listener focus on particular parts of their model to look for rhetorical tricks. The LLM, again having no motive, just does this in random places.

The latest advances in AI have been in the tooling around AI, which I think largely reduces to recursively creating models and populating them in outputs short enough that hopefully the text generation won't go off the rails.

Symbolic AI is coming back, it's just that the mortar is random word association.

Call me crazy, but I think Georges Bataille saw all of this. In my interpretation, he simultaneously thought of random associations as the most important component of reason and also completely irrelevant in the reality of the world. The only way you could discover the structure of things is through random collisions; your view of the universe was entirely constructed by your prejudices, lusts, and fears; but whether that view was wrong or right was based on how it worked and the particular content of the id-garbage that got you there was ultimately irrelevant. The fabric of the world is completely id-less, but can only be digested by id-driven beings. LLMs (without the tooling around them) are all id.

I cannot imagine a typical suburban American consumer tolerating a chatbot as blunt and direct as Linus.

I think it's not the raw training corpus, but the samples collected specifically for reinforcement learning, direct preference optimization, or the other instruction-tuning steps that get a base model to turn into a chat/reasoning model. (We're working on testing this with the Allen AI models that have open training data.)

Which gives a relative incentive to hard to understand text.

Is it? I mean they get payed more if they return longer answers.

"Baselinesystemstatetransitionlatencyvalidation" is surprisingly enough less confusing to me because I know it's one word describing one thing, and my eyes registers each joined word with (years of) training, even though what my eyes look at might be the following chunks in order:

1) [ Baseline ] systemstatetransitionlatencyvalidation

2) Baseline [ systemsta ] tetransitionlatencyvalidation

3) Baselinesystems [ tatetransitio ] nlatencyvalidation

4) Baselinesystemstatetransiti [ onlatency ] validation

5) Baselinesystemstatetransitionlaten [ cyvalidation ]

My eyes essentially scan the word, and although it might be a tad bid slower than reading "baseline system state transition latency validation", reading and understanding is faster because I already know it's one thing before I know which thing it is.

Edit: The example would be easier to demonstrate if HN supported bold font, but I don't think it does (or know how).

Actually LLM generated text is almost impossible to skim. It's too flowery and the actual meaning is spread too thinly across the text, which means that you can't just read the "important" sentences of a paragraph.

AI language, it seems, is optimized to support reasoning from the interior monologue.

In contrast, most human communication and writing is intended for other humans, and we constantly hit failures here that we learn to iterate on and improve. Humans who do math do so with the constraints that their math needs to be reviewed by another human and in the context that all their writing needs to be understood by another human.

The process causes a warped and compressed version of language.

Of course if you optimize for natural-sounding language it will learn to sound more natural.

I mean, maybe that's all right for when it's producing code - code is not beautifully flowing prose in the first place.

I find Claude’s writing unbearable. That ChatGPT offers personalization options (Base style, warmth, enthusiasm) and Claude doesn’t suggests Anthropic doesn’t recognize this as a problem. Plus, people sharing ChatGPT writing won’t all have the same voice.