← sources

LLMs are really bad at exposition

source: reddit (reddit:singularity)published 2026-10-10 16:06ingested 2026-10-11 02:39

retained text

The LLMs were always bad at free form explanations, and this became especially visible with the released proofs by openAI.

Just like with code where the code itself works, but its mostly incomprehensible, the proofs are equally regarded incomprehensible by working mathematicians. You have to use LLM and decipher line by line, symbol by symbol.

People here note from time to time that somehow LLMs are super awesome especially at "explaining and exposition", so heres a challenge for you. Write a lecture on LLMs.

My version is generated with Astra, Pro tier, 200 chat subscription. [https://drive.google.com/file/d/127UmCOxHDi3Qn-1fXJBP0hXoxRt22Y\_8/view?usp=sharing](https://drive.google.com/file/d/127UmCOxHDi3Qn-1fXJBP0hXoxRt22Y_8/view?usp=sharing)

Its a generic prompt, "give me a lecture about LLMs". Since its a generic model, it should sort of orient itself to create a reasonable lecture. And at a glance the lecture seems reasonable. But the contents are not reasonable, its generic AI slop, as it always is, if you ask it to free style educational content.

By the way, there is no amount of steering or iterations that make it reasonable, the models simply ignore the instructions, since they are RLed to hell to provide correct answers and lose this ability.

Slop includes completely mathematically trivial boxed example about "loss is an average of logarithms" but then just a couple of pages later it completely goes off the deep end with Bradley–Terry reward models and just generic formulas and name dropping. There is no real structure, no idea progression, things appear out of nowhere etc.

Skill issue? :)

document fetched 2026-10-11 00:30 · canonical https://reddit.com/r/singularity/comments/1x2izir/llms_are_really_bad_at_exposition