[felipeerias]

I gave Claude Fable $25 in Pangram API credits and, after hundreds of attempts, it was unable to produce a single readable original piece of writing that was not immediately identified as AI.

This seems to be a hard problem for LLMs, as passing would probably require good self-perception ("oh no, I am writing like an AI!") and fine-grained control over its own output ("let's write like a human instead!").

[inigyou]

I think the actual way you'd do that is to run RLHF fine-tuning with Pangram as the judge.

[ijk]

I wonder if it partially because "write like a human" is kind of a vacuous request. Like, it's the objective everyone including me has been saying that we want, but there's no one way to write like a human and and humans don't even have a good definition past "I know it when I see it."

There's a lot of work in the humanities about different aspects of good writing, but that's not quite the same thing. And anyway they tend to assume a pre-existing level of writing ability. Students are supposed to learn good writing through practice; there are rules and exercises but they're incomplete.

[dofm]

I think as much it is that people write by grappling for the right phrase to represent some inner feeling or concept, writing in part for themselves, whereas LLMs write always and only for an audience.

It’s much easier to understand this once you think about other generative forms. MidJourney never just sits down and draws for fun, so fun never informs its art (only the outward appearance of others’ fun, separate from the fun itself). Suno doesn’t waste hours trying to find riffs on a guitar, so its output is never informed by the direct joy of getting it right. Its music is never optimised for playability on a particular guitar with a scratchy seventh fret and a too-high action. Neither Midjourney nor Suno have evolved their styles due to short-sightedness or carpal tunnel.

If you had a human writer who over a long career only ever wrote articles from an outline given to them by someone else, and you had all the outlines and all the resulting articles from those outlines, and you could train an LLM to generate an article from an outline, it still would not be kicking itself frustrated by an inelegant phrase in a prior article, it would not avoid certain phrases out of a passive aggressive reaction to some editor’s note, it would not ever just rush an article because everyone is gathering at the pub, and it would not choose an analogy just to rub the author of a bitchy critical letter to the editor the wrong way. An LLM could not “subtweet”. It could not write a series of articles hoping one important person will spot that they are auditioning for a job.

Creators have unseen, undocumented influences and motivations that inform their work over a long period. I don’t mean to say that these individual influences can be reliably detected in individual pieces of work. I do mean to say that I think their broad absence tends to be felt in LLM writing. As readers we develop an affinity for writers as much as for their writing, and we do this in part because we deduce things about them.

[dingdongditchme]

Very elegantly put! So much so in fact, that I finally put down my phone to grab my laptop in order to write this. I have had a similar much cruder version of this thought: LLMs are not human, they don't know what it's like to stub their toe, despite having read probably millions descriptions of it. Even amongst humans there are experiences which are impossible to share, despite us having evolved to communicate quite effectively. Writing is just a small part of it, and cannot replace the actual experience of being human.

[felipeerias]

Each person writes in a different personal way, so writing “like a human” would actually require a model being able to purposefully make the specific choices that an individual human writer does.

However, general purpose LLMs like Fable have been trained on huge amounts of all kinds of data, and therefore find it exceedingly hard to break out of the grooves carved by that data. They can’t avoid defaulting to centroids and averages, even when they are trying not to. This makes it possible for classifiers like Pangram to discriminate their writing.

A plausible way to work around this limitation would be to train a LLM on a limited and cohesive subset of writing materials, so it would absorb their specific writing style.

One example might be Talkie, a LLM trained on pre-1930’s English text. Talkie is a far smaller and less powerful model than Fable.

And yet, Talkie’s writing is so distinctive that it is often classified as human by Pangram.

[mr_mitm]

Is pangram using knowledge about specific models? Could you throw it off by having several models generate a paragraph each or so?