Resolution:
Success in this case means at least one student sits down and learns their friggen basic parts of speech by the end of the year. This will be assessed via a diagnostic (pre-lesson quiz) and a cumulative assessment (post-lesson quiz). If they can reliably (>80% accuracy) differentiate basic parts of speech (noun, verb, adjective, adverb) whereas they couldn't before, this market will resolve as Yes.
Background:
I'm a ninth grade English teacher, and I was just as horrified as you're about to be that around 40% of my students had no idea what a noun is at the beginning of the year. With extensive teaching over a month or two, I managed to drive that number down to around 20%... but with the Pareto principle in mind and standardized tests looming, I decided to move on.
Concurrently, I am studying instruction design in grad school, and have been messing around with LLMs since AI dungeon was using GPT-2. I see a connection, since any model worth mentioning can explain what a noun is and tell you if a word is a noun. The problem is that isn't teaching, just giving information. Teaching needs to be tailored to the student: adapted for their existing knowledge and designed to build on it.
The Project
I'll be a bit vague on the details, but the plan is to script the AI to implement a scaffolded (scaffold=teacher jargon for iterating on existing knowledge) lesson. As mentioned, I'll be using Anthropic's API (Claude 3). I'm engineering prompts to have it assess student responses for understanding, and having Claude decide on when to move through the lesson.
My Prospects
As usual, I'm not betting in this market. My gut says I have a 75% chance of success. I'm building this in a colab notebook for now, and have found Anthropic's pre-built materials quite helpful/intuitive.
My brain, however, knows what optimism bias is, and it says a 15-25% chance is more realistic. My coding experience is anything but formal. While it might be feasible to get an AI to guide users through a lesson, giving it a workable UI and fixing the various problems I'm not anticipating will be much more challenging. Hofstadter's Law would suggest even this 15-25% chance is too bullish given there's basically two months left in the year. Regardless, I've already committed to using this for a project due at the end of the month, so I'll have to have something to show soon. Risking failure is the price of ambition I guess.