dm1ar
commented
1 week ago Certainly LLM code generators are far from perfect, as I have written. And JSMAG will need some more iterations to run smoothly. Wait and see.
There are certainly professional HAM software developers around, but the majority of OM's and YL's has other priorities and limited time and SW maturity levels. There are worlds between running a "Hello World" Python script and developing a reliable and robust C++ real time application with a GUI. Open Source HAM software ranges from apps with notoriously weak GUI's to professionally designed packages (e.g. SDRangel by F4EXB or the seasoned FLDIGI by W1HKJ).
LLM code generators will improve further and can currently help on each maturity level, but will not produce perfect code. Actual, there is no perfect code. But it is a great new tool for beginners and professionals alike.
Vy 73, Axel
Dear OM Axel, browsing though CQDL 12/2024, I come across this small AI generated project. It got my attention as I tried to generate morse code generating program a year ago(??) when ChatGTP was relative new.
The results I got had been of poor quality. But at least I can report that at least something had been generated even then. Now I had a review of your code and didn't find a significant improvement. HI HI My resume: ChatGTP is a fantastic liar! It generates something not too bad on a first glance, but I falls completely apparat when you investigate deeper on the output/code. This might improve very rapidly, but as of now the code output is in a shape that you need a software developer to fix the bugs, or at least as long time for prompt engineering that just program directly.
As result: ChatGTP might be useful at the moment only for software developers that can fix the output or extreme simple cases.
Here two example of bugs I could find in very brief source review of less than a minute: 1) the source does not handle prosigns correct. If you example enter the German word "KNAST" (EN: jail) it's recognize prosign KN. But it's not even playing KN, but complete mess the output: skipps N and plays AS as prosign. 2) The calculated time of transmission is not correct. Example: O and S is both calculated 0.,18s. The reason in source is quick to find: dit and dah are multiplied with same factor, so output complete nonsense.
The quality of the code is poor: it generates a Oscillator for each did or dah. As ham radio operator we instantly can recognize no LPF or other key envelope shaping applied by AI so output has respective harmonics (Key clicks). Beside I would not trust the calculated "Shannon-Entropy", I'm really wonder (why?) it even tries?
Here just an old small/simple example of WebAudio Morse Code generation on how to correctly play morse code: https://fritzsche.github.io/morse_player/
Don't get me wrong. Even I consider the output of poor quality, it's still an interesting experiment. BEST 73, Thomas