glendive (part 4)

Show hard words

🔥 the room was toasty. toasty like bread.

yesterday was minimally productive. but some ideas have clarified.

----

education-related LLM-related maybe-testing-related ideas:

  • the goal is "have a 5 minute video, see how people perform". they will need to read an LLM-generated text for a target grade level. ⚔️ well, actually, we will pre-generate a few thousand texts. ⚙️ for now, we can just have them read some Harvard Sentences. 🔥 It's easy to tell the depth of a well.
  • a better "linear" metric. when your reading "increases by 6%", what does that mean? presumably that your test score went up by 6%. 💡 this relieves the concept of "Grade Level" from its overloaded meaning. 🔥 and the desire of "above grade level" from its paradoxical meaning.
  • interactive / non-interactive learning. there is something to be said for "useful videos tailored to a student's interests" 💡 but, also, easy paths towards them being over-optimized
  • the machine can calculate values like "confidence" 💡 maybe "over-confidence" for students who have rehearsed, "fluency", etc. 💡 the machine can also complete the list of metrics
  • determining what intervention can solve what condition ... is not something the machine can do currently.
  • finding "gaps" in knowledge / thoroughly tracking existing knowledge. 💡 the machine won't track this, but a database that is attached certainly can. 🔥 there has to be some math to adjust for possible forgetting.

----

other ideas / tasks / goals

  • the "series of questions about a text" project. and, then, exception handling. 💡 it's not clear if a linear flowchart is sufficient; or even a linear flowchart with GOTO. ⚔️ well, actually, with enough GOTO it is clearly possible; just possibly confusing
  • single-domain experts. right now, people are doing this for "simple" domains like "customer service".
  • "politics" of education. the white whale is Not Everybody Should Go To College. this, already, might be more unpopularity than I can chew.

----

and, the latest Atacama hopes and dreams, deliberately slow-walked to avoid spending too much time here:

  • a list of the syntax tags and what they look like, on some documentation page
  • fix the "multi-level channel" listing 💡 including that "restricted" channels shouldn't be named publicly
  • something with "articles". 💡 which turns into "how much do I want to dig into my past writings" and "how much do I want to amplify anything"