Thesis: ChatGPT is destroying education, students keep using it to cheat on homework and exams.
Antithesis: ChatGPT has no understanding of facts or semantics.
Synthesis: Homework and exams don't measure understanding of facts or semantics, and can be fooled by plausible-sounding bullshit.
It'd be nice if tech companies didn't take an accelerationist approach to exacerbating and widening every problem that already existed in society, but they have.
The best defense, in some ways, is to fix the shit that's always kinda-sorta-maybe worked. In this case, yeah, it sucks that ChatGPT is DDoSing existing problems with academic evaluation methods, but it is... fixing those *is* a form of resisting the further incursion of AI into educational settings.
Another way of putting it: ChatGPT is an absolutely amazing fuzzer for human systems. It just spams humans with so many plausible-sounding scraps of nonsense that a ChatGPT user *will* find places where meaning doesn't matter and exploit them.
If you think of LLMs as a spam generation technology, this is unsurprising, but it flies in the face of how tech companies tell us to think about them.
a fuzzer! yes!
Re the original point, climate disaster planning sometimes talks about the Theory of Anyway, all the stuff that would be worth doing even if the specific disaster doesn’t happen
@xgranade this always comes to mind when hearing about people using LLMs in the legal field. Like you want to take extremely careful and precisely intentioned language and pollute it with output from something that assembles sentences from a literally random grab bag of words!??
Both in programming and law there are multiple ways to define the behavior you want, but so many more to define not quite the behavior you were expecting, with costly consequences for failure.
@dotjayne At least with code, you can mitigate to some (truly insufficient degree) by testing and using memory-safe languages and by using static analysis tools. It's still a fucking bad idea, but you can limit the impact of that bad idea *slightly*.
Can't fucking unit test a judge, though.
@xgranade i'd love to put this on linkedin and watch llm-atrophied vibe coders try to argue their case
absolutely agree with this tho! in my schooltime pre-llm, i already got a hunch but that one good thing about LLM's is that they prove that shit right now. Only question is, is someone doing sth about it at some point, or are our reasoning skills going to continue atrophying?
Cassandra is only carbon now
clew
Jeremy Kahn
Uncle Slacky
Jayne
bhtooefr
Lucy (private alt) Ⓐ
Mastodon Migration
Christian Berger DECT 2763