← Back to context

Comment by esseph

7 days ago

Everybody always thinks everything is AI. AI learned from consuming writing.

This is a ouroboros that will continue.

(Not saying this is or isn't, simply that these claims are rampant on a huge number of posts and seem to be growing.)

This is strictly true but not correct. LLMs were trained on human-written text, but they were post-trained to generate text in a particular style. And that style does have some common patterns.

  • So are you saying all LLMs were post-trained in that style then?

    Because, well, there's a huge number of models. Are they all, as they say, "in cahoots"? (working together, clandestinely)

    • Examples of LLM-style text: short & punchy sentences, negative parallelism ("not just X, it's Y"), bullet points especially with emojis and bolded text. Overuse of em-dash.

      This is a good list: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

      It's one thing to observe "LLM-generated writing all looks the same". Whether the LLMs were all post-trained the same way is a different question.

      I don't agree "everyone says everything is AI". Do you have examples where a consensus of people are accusing something of being AI generated, where it lacks those indicators?

      7 replies →

    • There was a time everyone trained their models with ChatGPT output. You can still find open source models that tell you they're ChatGPT if you ask.

    • Seems like many train on the output of other models for post-training and catch some kind of cooties.

    • if the people who develop and release these models were all optimizing for the same goals, they could converge on strategies or behaviors, without coordinating.

  • I'm one of the unlucky ones who has coincidentally trained myself over the past fifteen years to write in the style that is now largely recognized to be the ChatGPT style— bolded lists, clear section breakdowns with intro and concluding sentences, correct and liberal use of semicolons and em-dashes. The only parts of it I don't do are litter my text with random emojis or directly address the reader with simpering praise.