Buildings are a fantastically complicated and interesting complex problem space. I found my architecture studio students using it to ask technical and code questions about the buildings they were tasked with designing. I observed that the inaccurate information it was returning was compounding...

To be clear, I say "inaccurate" rather than "wrong" in this case because even if the information it returns is factually correct to the question being asked, students don't have an understanding of the complexity of the interdependent tectonic, regulatory, and spatial / experiential factors of a building sophisticated enough to ask their questions of the specificity and nuance necessary to get a good output that addresses the entire problem.

Anyway - with the students still learning to ask questions the right way, and the conditionally-incorrect facts making their learning more complicated rather than less, I hit on a strategy for them to use LLM's that seemed to help much better.

I suggested that instead of ask the LLM for the factual answer, or even better for the facts and an explanation, that they ask it to direct them to the proper place in the source material to find the answer themselves. Then, to treat it like a lab partner. IE:

Hey Claude I'm looking for "x."

Claude: "look at foo, bar."

Thank you - chapter (foo) part (bar) table (goo) says "car." However I notice that footnote (hoo) says there's an exception if "dar." Which is what I have. Walk me through this exception...

It seemed to have good results as a guide to understanding the disparate bodies of knowledge that they will eventually have to keep together in their heads and work synthetically and non-linearly through, rather than just as an external source of blindly trusted authority.