Hacker News Reader

My experience is that ChatGPT is dreadful at everything but the simplest ffmpeg invocations, and will often produce command lines with subtle quirks (such as "only works if the input is an even number of pixels wide").

But then again, my experience is that ChatGPT is dreadful at everything but the simplest anything.

View on HN

My experience is similar, but I find it can still be useful if you go step by step, check its work and explain errors and corrections as you go. Admittedly when I describe it like that it doesn't seem very useful, but if I'm figuring things out by myself then most of that work is a given anyway and the bot helps that process along.

Sometimes... but most often, after I explain in detail an error it did, it will just say "I apologize for the mistake, you are correct, <re-iteration of my explanation of the error>. Here is a version with the error fixed:", followed by either the exact same output or an equally wrong alternate output.

The fact that these glorified Markov chains manage to fool people into thinking they posses some kind of actual intelligence or ability to reason baffles me.

> my experience is that ChatGPT is dreadful at everything but the simplest anything.

Have you tried Claude 3 Opus (or even just Claude 3 Sonnet)?

https://www.anthropic.com/news/claude-3-family

No

Reply on news.ycombinator.com