My fantasy with ChatGPT is to be able to describe in natural language some explosion or magic spell shader effect and for ChatGPT to output the code (maybe even auto convert it to gif or webm afterwards, but not necessary). Important: this is different to asking Stable Diffusion "draw a picture of an explosion" with a text prompt etc, I want the output as text code not just the picture. So ideally something like these one line twitter opengl code from twigl.app eg
mountains with volumetric light rays, abstract acid fantasy landscape
https://nitter.net/zozuar/status/1558237871189925888dungeon cavern with pillars
https://nitter.net/kamoshika_vrc/status/1613535884594524165mist mountains with clouds
https://nitter.net/i/status/1443012484189888515If you had enough of these shader demos concatenated in virtual text worlds, but more importantly responsive to NATURAL LANGUAGE text input this would make for a convincing generated virtual realm.
I wonder if ChatGPT can handle this highly condensed demoscene type code, high Kolmogorov complexity where there is not a lot of redundancy / compressability compared to say spoken English text / literature where many words are meaningless syntax fillers and so amenable to all that model parameterisation and token chaining.
The mapping from natural language to dense code to end picture result shader world effect is not obvious.
The analogy is like if you asked ChatGPT to come up with pi, it might just recall
3.14159265358979323846... etc or various mathematical series expansions from online sources but you want it to come up with
#include <stdio.h>
int a[52514], b, c=52514, d, e, f=1e4, g, h;
main(){
for(;b=c-=14;h=printf("%04d",e+d/f))
for(e=d%=f;g=--b*2;d/=g)
d=d*b+f*(h?a[b:lit]:f/5),a[b:lit]=d%--g;
getchar();
}
instead (of course, the above spigot algorithm already exists and is searchable online so it would be probably be part of any AI training corpus of data). You can execute and verify this code at any online c++ compiler, eg.
https://www.onlinegdb.com/online_c++_compilerAfter all most people just regurgitate and resynthesise information and data they read/remember/learnt, ChatGPT can match and exceed and automate this with vast data to a greater extent.
Perhaps if you ingest enough text and code eventually you can just derive the breakthrough insight to produce these sorts of compact worlds?
Or maybe ChatGPT forever remains at the level of algorithmic mimicry / recombinant autocomplete. What probably then results is some sort of horrifying information bloating, where overuse of AI automation just results in superfluous stochastic parrot recitation versus focused human problem solving.
