This is a special post for quick takes by axel_sdq. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Has anyone had the experience of trying to explain their idea to an LLM, but it fails to grasp the basic concept?
Asking because I don't feel like this has happened to me (from my limited usage). When it can't connect the dots, it's because I haven't provided enough dots.
(Edit: examples against much appreciated if any come to mind)
More like the opposite, LLMs are great for the "tip of my tongue" type of questions, where I can describe something or give an example but I don't know the official keyword for that.
I'm not sure if this is the same thing, but I frequently talk to Claude about research ideas, and if the idea is close enough to a different idea that it knows about, it repeatedly collapses back into talking about the idea it's familiar with.
One I remember from this week:
I'm looking into ways to make intermediate values more visible in the logit lens, and Claude really wants to talk about the tuned lens, which does the opposite of what I want[1]. Even if Claude itself has explained why this doesn't make any sense, it will repeatedly suggest trying the tuned lens.
I feel like I had another case where it took forever to get it to grasp what I was even talking about, but I don't remember the details unfortunately.
1. ^
Specifically, the tuned lens makes the next token's representation more clear and actively erases anything else.
1axel_sdq
Thank you for the example, this definitely counts in my mind.