#LLMs make poor #teachers because they were never #studemts.
#story time.
I'm part of a #team of #volunteer co-teachers teaching #Python to #HighSchool students remotely over Zoom. But that's only the #process. The goal is to grow other-subject teachers into #ComputerScience teachers by teaching them along with their students.
This week, our students are learning how to use the input() function to get information from the user. The project is to write a program that introduces itself as a #genie and asks the user for three #wishes, then prints out the wishes. Our classroom teacher was thinking how he could do it in a loop. Because he learned that when previous classes had studied #loops.
We discussed that this was one of the reasons that volunteers were expected to do the work we are assigning to students using only what the students had been taught. So we could help them better when they encountered problems.
And this right here is a big problem with LLMs as teachers. These models are trained, but they don't learn like we do. They were given all the #answers as part of their #training #data. When they cannot find an answer, they don't try to determine the answer by replicating the steps that students are taught — they make a #statistical #guess based on their training data.
Worse yet, they are incapable of actual learning from their #errors.
And "errors" is the correct term. "#Hallucinations" are a subset of errors that involve the #perception of nonexistent stimuli. #Techbros use this term because it sounds nice to them. Like, "I was high, and I had this awesome idea," instead of, "I'm just making shit up so I sound like I know what I'm doing." The latter is how they engage with the world, because they prioritize #confidence over #correctness. Which is a very bad look for a teacher.
There are legitimate uses for LLMs in #education, but none of them are as personal tutors.
