interesting. has chatgpt/copilot been trained on nim to the point where they creates code as good they do with python?
I ask because I've always treated coding as a necessary evil. When I worked as a programmer, I never gave 2 shits about the language I was required to use. I'd write a specification, code to the specification and tried to update the spec if I need to deviate from it.
Because of my damaged hands, now, I do not have enough typing capacity to learn a new language the old trial and error way. chatgpt is my accessibility tool for creating code.
I've learned how to give chatgpt a spec, skip the coding step and goto validation. I've trained myself and chatgpt/copilot to generate python code like I had created it, making desk and unit testing easier and faster.
Learning a new language would start with a specification of the problem as above, then seeing what chatgpt/copilot generates. However, it is easier to learn if the LLM model understands the language. It is sometimes hard to tell the difference between hallucinations in the code vrs hallucination in the specification.
It ultimately depends upon what you mean by "as good". There is no clear single metric. Once you have >1 metric it becomes a subjective "who's priorities?" game of weighting them / projecting them into a single metric.
The best answer to a question like yours is: give it a try on some easy problems and see what you think yourself. No one else can really know the kinds of problems / answers / code you most work with (and sometimes the future is pretty murky even to oneself, even in these vague categories).
Disclaimers issued, some things can be said which might help. Since Nim is primarily a highly ergonomic static ahead-of-time compiled language with code running as fast as C often does, errors may be caught more conveniently. Because Python is popular, especially for teaching programming, training coverage will always be better, but Nim has some core features & keywords "kinda similar" to Python which may help on the other side.
Not sure about 4.0, but ChatGPT-3.5 does poorly on basic Nim things without Python equivalents. To give just one concrete example (out of many), `a,b = b,a` is a common way to manifest swapping in Python while in Nim one uses `swap a,b`.
So, if you are willing to do more "compiler-assisted clean-up" or have a/develop a knack at steering the random sampling toward code which compiles, Nim could be about as effective as Python used this way.
In terms of code entry work for your specific hands problem, parentheses can often be left off in Nim code and in general it seems to have much less punctuation / syntactic noise. Of course, keys can be rebound & such & maybe you do that, too. Nim definitely has more powerful abstraction mechanisms like user-defined operators, templates, and syntax macros.
Your suggestion at the end is a form of "speaking the keyboard" problem that has plagued speech-driven programming for years. It is so wonderful that I can dictate the specs and then verbally cut and paste them into the LLM. If I had the energy, I would build an LLM interface that used speech recognition-friendly text areas that one could use speech for editing and revising results.[1]
Your response, however, touches on the meta-problem of adding additional information to a training set. For example, as people learn and generate more nim code, the community can expand the LLM capabilities without going through open AI or whoever. I know training requires a lot of GPU time, so training over distributed GPUs is necessary. For example, I would buy one or two GPUs and pay for the electricity to contribute to community LLM training efforts.
[1] For anyone interested in helping, it would be a simple two-panel design: top for editing, lower for LLM results, and two buttons, first for "copy to clipboard," second for "submit to LLM."
I ask because I've always treated coding as a necessary evil. When I worked as a programmer, I never gave 2 shits about the language I was required to use. I'd write a specification, code to the specification and tried to update the spec if I need to deviate from it.
Because of my damaged hands, now, I do not have enough typing capacity to learn a new language the old trial and error way. chatgpt is my accessibility tool for creating code.
I've learned how to give chatgpt a spec, skip the coding step and goto validation. I've trained myself and chatgpt/copilot to generate python code like I had created it, making desk and unit testing easier and faster.
Learning a new language would start with a specification of the problem as above, then seeing what chatgpt/copilot generates. However, it is easier to learn if the LLM model understands the language. It is sometimes hard to tell the difference between hallucinations in the code vrs hallucination in the specification.