What AI do you use to solve CG software puzzles ?
sublime tool
I use $20 ChatGPT subscription and Google Gemini . Much to my surprise Gemini sometimes knows better when I need to recollect some condition checkbox I need to check in for something to work.
Wonder what you guys think? GPT seems lazy to look through actual documentation lately , I need to ask it to do it specifically. Am I imagining it ?
ps. Gemini sometimes replay me on a sort of general question : "we humans believe" that sounds a bit creepy.
Replies
All make up shit, all are pretty good at some things. In my experience they make similar errors and come to same conclusions, smooth talk you into believing what you decide it the best decision ever.
But ultimately also very helpful at times, making tools, explaining workflows, walking through nodesetups. IF THEY DONT MAKE UP STUFF
All models i tested use to shit talk about zscript, that was a bit funny. Also much more painful than doing simple things for blender or maya. Max is also a bit more painful at times
the models are much more inclined to read the docs and reason from them rather than checking stackoverflow / reddit for a premade solution and making shit up if they can't find it
I think the free tier is still pretty generous but even if not, the $20 monthly is pretty cheap and you can pause payment easily
What I know and what might come after are derivatives of what has been learned and can be figured out. There aren't really any "puzzles" anymore.
Short answer: I don't use AI to solve puzzles and give me answers/write my code/make my nodes. But AI is a really good "search" tool since it scraped the net anyways, now I can find that obscure bit of information.
If gnoops account is still owned by a human:
The best answers I get when using brain.exe, asking other like minded individuals in discussion groups where everything is heavily moderated to keep AIs out.
Docs, Wikis and toying with the tools you have helps too.
I use both GPT and Claude. Whenever I run into a problem and one of them starts losing the thread or hallucinating a bit, I make them challenge each other. I'll tell GPT, "Claude said this..." and then I'll do the same in Claude, passing the responses back and forth until they converge on a solution. Surprisingly, this works really well, and I feel like I get the best of both models.
I've found this especially useful when solving complex technical problems that can easily fall into loopholes or circular reasoning. I run a small game studio with only two developers, and we're building a fairly complex 3D multiplayer game. Because of that, I work across every discipline, from programming and networking to animation, art, and game design. This has become a regular part of my workflow.
The last time I used this approach, I developed a custom node for Unreal Engine's Animation Graph that transforms the final animation output into a stop-motion effect inspired by Ray Harryhausen's films. The system involves frame-rate control, storing previous and future animation frames, frame interpolation, and controlled jitter. The result captures the feel of traditional stop-motion surprisingly well while still running in real time. I'm very happy with how it turned out.