You're assuming all you can do is prompt it. Surely you could also constrain its output to tokens that genuinely contain no e’s (or make only max 4 letters per word). LLMs actually output a probability distribution of next tokens; ChatGPT just always picks the top one, but you could totally just always filter that list by any constraint you want.
But the problem is that the tokens are subwords, which means that if you simply disallowed tokens with es, you'd make it hard to complete a word given a prefix.
For example, it may start like this "This is a way to solv-", or "This is th-"
If I understand it correctly, that's a valid concern but the way structured generation library like outlines[1] work is that they can generate multiple variants of the inference (which they call beam search).
One beam could be "This is a way to solv-". With no obvious "good" next token.
Another beam could be "This way is solv-". With "ing" as the obvious next token.
Yes, that would probably work quite well, given enough training data. However, I interpreted the question/claim as a task that LLMs excell at, meaning that writing text while avoiding a certain character is a task for a general purpose LLM.
I tried something like that some time ago. The problem with that strategy is the lack of backtracking.
Let's say I prompt my LLM to exclusively use the letters 'aefghilmnoprst' and the LLM generates "that's one small step for a man, one giant leap for man-"[1]. Since the next token with the highest probability ("-kind") isn't allowed, it may very well be that the next appropriate word is something really generic or, if your grammar is really restrictive, straight up nonsense because nothing fits. And then there's pathological stuff like "... one giant leap for man, one small step for a man, one giant leap for man- ...".
[1] Toy example - I'm sure these specific rules are not super restrictive and "management" is right there.
What I will add is that constrained generation is supported by the major inference engine like llama.cpp, vllm and the likes, so what you are describing is actually trivial on locally hosted models, you just have to provide a regex that prevent them to use the letter 'e' in the output.
That is not a counter point! The output has a probability distribution so you can assing zero to any e-containing token and scale everything else up accordingly.
