I have discovered that ChatGPT, when asked to translate from natural language to Morse code, generates code that it cannot correctly convert back into dots and dashes to produce the correct text. Even the generated Morse code itself is incorrect in one direction. This is quite interesting for a language model, as it essentially involves only a font change from ASCII to Morse.
Yeah, it seems to use patterns based on the data it was on, so it does certain forms of encoding poorly, even GPT-4. For example, it’s blind to things like large ASCII art and it can have trouble trying to reverse unusual words and such.
Most probably down to tokens not representing letters but word parts.
A thing to test would be to insert - between letters or spaces, turning the m-e-s-s-a-g-e- -i-n- -t-o- - t-h-i-s if you do that then each letter will be it’s own token and you should get correct operation.
Hi, Just tried that and it works
I tried in diferent languages for few hours yesterday and checked it with gpt itself and algorhytmic aproach of:
So when I passed a fraze in something else than english gpt genaretad dots and lines, but when I reset sessions and give gpt this series to translate back it were gibberish. Better - every time it was a bit different gibberish.
Probably. But this task is algorhytmic sign to sign and it looks like model lose the point trying to overcomplicate solution.
However gpt wrote at first glance a program to translate letters to morse.
I think it is just intresting, because morse is just a “font”, and gpt is a language model.