Post

Since 2022 I occasionally test various LLMs via a custom API frontend to gauge if they can aptly solve various problems. In doing so I have discovered a reliable 0-cost, 0-environmental-impact method to sharpen your intellect with a LLM. (No, seriously.)

1. Discover a head-scratcher a LLM may be good at (big pile of data, seeking a non-obvious info pattern, combing through stuff, approximating "mid" conventional wisdom) 2. Type up a detailed prompt for a LLM 3. Realize via writing it up that you DO have an idea of how to solve it 4. Close LLM

Of course this also works if you write an e-mail to explain the problem to a human, then delete the e-mail. But you'd have to roleplay seriously asking someone else to waste their time on your stupid problem, and you feel even more silly, instead of superior-to-machine, when you figure it out.

I have discovered a new corollary principle to all this, which is something like "ok, tell me why this code / math / whatever isn't working" and then after you read the confabulation you go "what the !@#$?! there is no way that's right. it's probably the --" ( lightbulb goes on over your head )

I tried this out, and uh oh, I think my rubber duck might be developing sentience!

If you have a rubber duck that can procedurally generate straw-man explanations, no matter how internally incoherent, I want to borrow it