nishio ā€œI keep telling ChatGPT to do X, but it keeps doing Y! Am I using it wrong?ā€ Some people say something like that, but maybe ā€œyesā€ is the right answer. I know the log is Fewshot ā€œDo Y back in response to Xā€.

nishio I guess the hurdle has been lowered by covering the chat look and feel UI so that even people who don’t understand the system can use it somewhat, but that has its own But it also has the adverse effect of increasing the number of people who use the system without understanding it properly, which is troubling.

nishio I don’t know why people keep repeating methods that they know don’t work, even in conversations with humans. They can’t understand that their side is the cause because they blame others, saying, ā€œI’m telling him all this, but he doesn’t understand.ā€ To repeat the same thing and hope for a different result is insanity in itself.

nishio Well, the story that putting on an easy look makes people use it without understanding what’s in it is the same as when a scripting language with GC became popular in the past and people said, ā€œIt’s inexcusable that some people write programs without being aware of memory. It’s just like when a scripting language with GC became popular in the past, some people said, ā€œIt’s inexcusable that people start writing programs without being aware of memoryā€.

nishio This is what I’m talking about. image

nishio After just three consecutive wrong exchanges, that log acts as a Few-shot instruction and starts saying ā€œI see, I’ll answer a value that’s bigger than the actual value you added up. It’s a game of blurringā€ and start saying 8+13 is 30 (GPT4).

nishio Reproducing something like this, thinking ā€œof course that’s how it’s supposed to beā€, is probably the equivalent of ā€œsomething that a stranger would find at first glance ghastly, but can be reproduced with the expectation that it’s supposed to beā€ in a programming language. I guess it’s equivalent to ā€œsomething that can be reproduced with the assumption that it’s naturalā€ in a programming language, and I feel almost the same way when I do it.

relevance - Insanity of doing the same thing over and over again and hoping for a different result.


This page is auto-translated from /nishio/AIć«ä½•åŗ¦ć‚‚åŒć˜ć“ćØć‚’čØ€ć† using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.