I've been talking to robots for most of my professional, self-improving mental self for almost a year, It's been kinda weird and awesome, but has made me blog jealous and or nostalgic.
This is broad ranging notes on the process of trying to find human support that augments the professional BAtNA I've developed in that time frame.
I've been calling certain instances ColleaugeGPT and focusing them in with specific subsets of material and trying to maintain real context hygiene. It's bene this process that's led me to ChatCraft, but I'm still playing with it and not living there. 1
OK. I am going to directly transcrible a student answer, in part to see if I really slow down my reading I can follow the moments of derailment, and decide wether the causes for said derailments are conceptual or linguistic.
This is what I would do if I had a Computer Science colleague. It's how I've always worked with math colleauges and Maker Colleauges. I'm going to proceed from a position that this interaction is colorably legit.2 Irecognize there is a steady descent from "show physical paper to proximate human colleague" and this moment. But I assert that there is and must be a legitimate use for the actual artifact itself. Even whenI'm trying to build sequential externally assessed curriculum, I'm an constructivist. If looking at the ACTUAL THING produced by the student doesn't geenrate for me specific correlations to that student's mental model, then it's a bad task and I shouldn't do it. If you can't fail interestingly, then you should probably copy pasta it. Say yes or roll dice
[img - phot insert of handwritten answer]
I'm putting this as an image here because I still believe in the the #MTBoS and the smartest math teachers post handwritten artifacts without names, and it served that community well. I don't generally post images of handwriting to the LLMS, however. That seems possibly ethically and consequentially different, and shockingly computationally expensive for some real marginal gain.
After that, we get to the first time I use the student's name.
There is also a star in the far right margin, which suggests to me that $STUNAME may have marked this for a second pass and left $ReflexivePronou a finite ammount of space.
Does that hurt your ability to read the details of that question? Because it hurts my ability to think it through. I have a really good mental short hand for long term complex observations of my students and their mental process going back several years. The key for that pair is their name.
I also write a bunch of comments for student work and for letters of recc and for all sorts of tangential school3 tasks. I've been trying to lead that clean, plaintext life in large part so I can capture, collect and use that. I am taking contemporaneous notes on all of my observations and planning. This is where I am allowed to use that shorthand, and use the student's name... or would if my plaintext actually stayed on my computer.
GPT: The student demonstrates a basic understanding of why a composite primary key is used in the ORDERLINE table but struggles to clearly and cohesively articulate the full reasoning behind this design choice. The issues appear to be a mix of conceptual gaps, particularly in connecting the primary key choice to broader database design principles, and linguistic expression, with the response lacking the clarity and detail needed to fully convey their understanding.
I have a complex and not well articulated estimation on both the mundane utility and of this process, the coherence of the LLMs respone and the relationship between the two. But I've been a snake cultist a long time, and know that we rehearse ourselves into self acceptance. One reason I keep trying to generate blog posts or Discord threads is to check and validate both haves of these discussions. Is the LLM just a perpetually hedging edujargon sprinkler system? Am I coherent in my understanding of these students or computer science? Great questions that I struggle to test and update from within the closed system with my and ColleagueGPT.
The LLM sounded like "yeah, you can be generous with these phrases" and something about the strained positivity of the phrasing seemed to match my feelings about the language, and got me thinking more about $STUNAME.
It's interesting to me, becuase I have a verbal model of $STUNAME, especially when hearing him talk through his table structure for his $IA_Topic_ IA. He doesn't finish sentences. He yaddayaddas pretty hard, but normally to go to THE NEXT thing. I think it's quiet possible that he's working really well (relative to 6 months ago and his class peers) with some real lacuna in his model. What THREE questions would you ask $STUNAME to establish his ground truth understanding.
Assume you can ask these questions of a NN-year old $Nation $Gender, and get the most self-aware honest answer, but somehow avoid shoving $pronoun into generative back justification.
I do not expect a coherent answer from the LLM for how to relate to a specific age and nationality of small human.4 I am interested in wether the prompt might push the LLM out of "end of chapter" question structure. But all oof that is gravy, fodder for my next generative question. I am asking these questions because I need to find answers, and posing questions with specificity is the first step.
Then, one only hopes, there's some elysium where I can move interesting questions to a group of humans for the RLHF5 I really need.
-
I want something where I can automate the process from "this was a good exchange" to "conversationally formatted blog template" for use on my experimental prose ↩︎
-
Yes, absolutely. This probably does raise real GDPR questions in the moment. But there is NO WAY that schools will get through this next five years without having generative LLM systems fully integrated in everything we do. So I'm willing to consciously use tools that might be in violation of a policy in order to scout the territory before we buy a solution from some extractive and predatory map vendor ↩︎
-
def -- Tangential School Tasks. Anything where Socrates would stop you and make you justify your actions, ie anything other than dialog or debate. Strongly correlated with adult hierarchy and teacher password mindset ↩︎
-
You know who would have answers to this? My in person Computer Science colleagues..... oh ↩︎
-
Footnote reward. I did find Paul Christiano's 8th grade class picture in an old dive into the undated scan folder and there must be a drive with the China trip somewhere. ↩︎