> Even if we're just talking about the level of what GPT 3.5 is capable of, that still corresponds to, let's say, a college freshman level of writing. Indeed -

> Even if we're just talking about the level of what GPT 3.5 is capable of, that still corresponds to, let's say, a college freshman level of writing. Indeed - the output of ChatGPT to a very quickly thrown together prompt and essay question, was able to easily beat the average of a cohort of freshmen in terms of writing skills. I imagine this is partly due to a decline in writing skills generally (?), but also partly due to an LLM having a significant advantage in being able to roll a die and produce a completely grammatically valid sentence on any topic each time. I suspect that we will see LMMs reach a point where they can pose a threat to weaker non-straight-A students - at least at freshman level, you can already get GPT to write an essay that's more factually accurate than the weaker students in a cohort. It won't be factually perfect, but it will be more factually accurate than the students' own attempts. Fundamentally though, I do wonder - if the marginal cost of prose reduces to near-zero (through bullshit generation via AI), what will we move to? Lengthy prose that you know someone won't read is a perfect situation for LMM generation. 5x short, sharp, action-focused bullet points that get to the crux of a situation and how to resolve it are far harder to bullshit, as ultimately people are focused on the substance, rather than style and presentation of your point. This would (I hypothesize) disadvantage both human and AI bullshitters equally. I tend to see the best students feel less of a need to pad and write more, as they know they have said what needs to be said, and are finished. Those with a need to bullshit (perhaps through not having as robust an understanding of the subject matter) will pad, flounder, circumlocute, and eventually get to something resembling a point, eventually.