One of my co-workers uses her chatgpt as the go-to for everything.
She’s doing a masters in psych, all of her assignments are written by the thing, and she was just telling me that there’s a 3 second submit-to-recieve-grade window. Meaning… nobody even reads the papers, composed by one dumb robot, fed into a second dumb robot that pukes out a grade.
To point out the obvious, LLMs are objectively terrible graders, no matter how big or smart they are, because:
They have zero context outside the prompt, no memory of previous papers it “graded”
Wild variance in output thanks to high temperature sampling (which everyone uses because it’s default).
Tokenization artifacts making non-binary decisions like this.
It’s… incredible they’re using it like that.
This is the other part of LLM usage that gets me. It’s used in the laziest, most broken ways possible.
The grader, theoretically, could have at least waited to feed a long context model a whole classes papers at once, with a scheme to rank them. Or something like that? But nope, lazily feed them one at a time, don’t even check if it’s right.
ChatGPT’s UX and training is designed to encourage this, I think, kind of like how social media has all sorts of dark/addictive patterns. IMO this should be highlighted more.
Yeah I know. The true point of my comment was a few layers of inference deep.
There’s no learning happening. She knows little-to-nothing about the subject matter and the school doesn’t give two shits.
And she asks me primitive questions (relating to the field we work in) that would be the equivalent of a Michelin Star chef asking me how to cook potatoes.
e: She is just finishing her masters in psych, and she’s about to become a therapist. She knows fuck about shit. Asked me what “theory of mind” was the other day ffs
One of my co-workers uses her chatgpt as the go-to for everything.
She’s doing a masters in psych, all of her assignments are written by the thing, and she was just telling me that there’s a 3 second submit-to-recieve-grade window. Meaning… nobody even reads the papers, composed by one dumb robot, fed into a second dumb robot that pukes out a grade.
Yeah.
To point out the obvious, LLMs are objectively terrible graders, no matter how big or smart they are, because:
They have zero context outside the prompt, no memory of previous papers it “graded”
Wild variance in output thanks to high temperature sampling (which everyone uses because it’s default).
Tokenization artifacts making non-binary decisions like this.
It’s… incredible they’re using it like that.
This is the other part of LLM usage that gets me. It’s used in the laziest, most broken ways possible.
The grader, theoretically, could have at least waited to feed a long context model a whole classes papers at once, with a scheme to rank them. Or something like that? But nope, lazily feed them one at a time, don’t even check if it’s right.
ChatGPT’s UX and training is designed to encourage this, I think, kind of like how social media has all sorts of dark/addictive patterns. IMO this should be highlighted more.
Yeah I know. The true point of my comment was a few layers of inference deep.
There’s no learning happening. She knows little-to-nothing about the subject matter and the school doesn’t give two shits.
And she asks me primitive questions (relating to the field we work in) that would be the equivalent of a Michelin Star chef asking me how to cook potatoes.
e: She is just finishing her masters in psych, and she’s about to become a therapist. She knows fuck about shit. Asked me what “theory of mind” was the other day ffs