Thread by @nptacek

anyone else mix and match platforms while making #aiia?

i’m currently consulting with chatGPT to help build a post-processing script for art i’m generating with stable diffusion and upscaling with yet another ai platform

this has supercharged my entire workflow 🤖

debugging chatGPT zsh output

ah, a logic bug, and my fault at that.

my program specification was incorrect, basically

this is actually a fascinating look at how small errors in communication can make a big difference in outcome

the first zsh output from chatgpt threw a bunch of errors but also seemed to partially work, so i just copied a truncated version of the error output, pasted that back into chatgpt, and described the parts that did appear to be working correctly.

this was enough for chatgpt to diagnose and fix its own bug, with just a slight bit of prompting on my part

imagine a future version of chatgpt hooked up to your QA pipeline with some form of automated unit testing (a la @github actions) in the mix

this wasn’t anything super complex, but it was something where i would have had to look up stuff up on google

i knew *approximately* what i needed to do, but not *exactly*

and this is precisely where platforms like chatGPT shine

chatGPT cut what would have been maybe an hour’s worth of research, coding, and debugging on my own down to about 7 minutes total, including the time crafting the prompts.

this is incredibly powerful stuff

think of what it looks like at scale

in this case, it was doing double duty at augmenting my time and ability:

the specific shell script i had chatGPT craft is going to mimic the output of video editing apps that i don’t have the time or inclination to learn (and which are themselves abstractions for CLI tools)

since i could describe what kind of effect i wanted to have on the output visually, chatGPT was able to point me at a command line tool that did what i wanted.

from there, it was simply a matter of giving chatGPT a brief code snippet as an example, and testing the results

obviously you would want to be very careful with running code generated in this way

think of it as unsanitized input, and defend accordingly:

at minimum, i would pass the output to another ML model specifically set to perform a sanity check on the code.

this might look something like “does this code do anything unpleasant on a system that runs it?” or “is this code safe to run?”

would also recommend virtualization, let your emulator take the hit if the code is bad, rather than your main system

whether running code in an emulator or on a production system, keeping good backups is incredibly important, and only will become more so going forward.

waiting for homebrew to do its thing with ffmpeg install is somehow taking longer than everything else combined up until now 😴

following my own advice and backing up ~3.5 GB of data i generated with the first chatGPT-supplied zsh script before running the ffmpeg command

(just in case, i didn’t bother checking if it overwrites files by default or not because yolo)

huh.

the first ffmpeg command supplied by chatGPT didn’t work but the second one appears to be.

it’s certainly giving my cpu a workout

i have absolutely no idea how long this process is going to take, this might be one to let run all night while i get some sleep 🥱

𝘸𝘰𝘳𝘥𝘴 𝘤𝘳𝘦𝘢𝘵𝘦 𝘸𝘰𝘳𝘭𝘥𝘴