HISE as a 3D renderer

[scottmire]

@Oli-Ullmann
That is amazing! If you don't mind me asking....could you share the prompt you used in ChatGPT?

[Oli Ullmann]

@scottmire
Of course! :-)

Since I'm from Germany, the original prompt was in German. However, the translation is quite simple:

"Please use the HISE scripting language to draw me a 3D grid of a rotating cube."

But don't be fooled—chatGPT is in no way reliable when it comes to generating code! Today, I used the same prompt to generate a cube that doesn't change in size, and chatGPT failed completely. It didn't even manage to reproduce the same result. Creating a working script with chatGPT is pure luck.

[d.healey]

@Oli-Ullmann It seemed good for little snippets like this

[Oli Ullmann]

@d-healey
Yes, sometimes it works well. But sometimes it doesn't. There are always little mistakes. For example, “var” variables are created in functions instead of “local” variables every time.

But chatGPT often helps you get on the right track.
For example, chatGPT has often helped me calculate mathematical formulas for LAF functions.

It's quite an impressive little piece of technology! :-)

[Chazrox]

@Oli-Ullmann facts. For now its only good for redundant tasks and heavy mathing. In those aspects I abuse gptty. lol.

[scottmire]

@Oli-Ullmann
My understanding is that Claude (https://claude.ai/) is supposed to be much better than ChatGPT at writing usable code.

[Oli Ullmann]

@scottmire
I have never used Claude, but I have heard that it is supposed to be better in terms of coding. However, I am uncertain whether this also applies to the new GPT5 version.

[dannytaurus]

@Oli-Ullmann gpt-5 was terrible for me initially. I switched back to claude-sonnet after a few days. Then I heard gpt-5-fast was decent, so I enabled that and it's been fine.

Overall, claude-4-sonnet has been consistently good for me for coding for the last several months.

I think they rushed gpt-5 out too soon. Seen some horror stories about its performance.

[d.healey]

Isn't GPT5 basically the same as the previous version just that it auto selects the best model instead of you having to do it manually?

[rglides]

@Oli-Ullmann yeah claude is still better than gpt5, at least for HISE, although I tested it with some C++ and Claude is better at that too imo. Both still depend largely on the user's patience, prompt and knowledge of course. This cube is amazing by the way! Would be interesting to see if it could do some 3d XY pad stuff with knobs. I guess it would be called an XYZ pad

[dannytaurus]

@d-healey That wasn't my experience at all. Search the internet for "GPT5 fails" and you'll see all the negative reactions.

In my case it was very slow, showed way too much 'thinking' text, refused to execute terminal commands and had a much higher hallucination rate than previous GPT versions or Claude.

Keep in mind, this wasn't with HISE. This was mainly in a Ruby on Rails app with ~30k LOC and a bunch of other much smaller projects. I haven't used AI for HISE at all yet, mainly because I don't think there's enough HISE code out there for it to give really useful answers.

For casual stuff it might be fine, but for coding - and in particular in the Cursor IDE, it was horrible.