Aligning with the natural language programming by LLMs

As generative AI grows increasingly capable, I feel that, where natural language is concerned, we have already entered the age of natural-language programming. By transcribing speech, translating it, and refining the resulting text, we can now process and produce writing across linguistic boundaries far more rapidly than ever before.

Large language models can generate an infinite amount of text. Yet I occasionally encounter people whose sense of ownership seems to have become confused. However ingeniously you may devise the prompt that produces a piece of writing, it is not truly your writing. It contains none of the hip-hop joy of expressing yourself. I believe this kind of misunderstanding can be potentially fatal.

For my part, I have made it a routine to speak my thoughts aloud, have generative AI turn them into prose, refine the form, and translate the result. This particular text was originally written in Japanese and published on my Japanese X account @kenichiromogi Because the original concept and central intention come from me, I feel that the resulting text is my own—just as I would regard as mine a piece written by a professional writer on the basis of what I had said. There is, however, an important criterion to be considered.

Even when the Japanese or English text produced by generative AI is based on my own speech, I still revise it according to my sense of rhythm, continuity, and the overall feel—the qualia—of the prose. In that process of revision, the only thing I can rely on is my own linguistic sensibility. AI is of no help whatsoever there.

In this age of generative AI, we can program natural language through prompts, speech-to-text conversion, translation, and stylistic refinement. Yet it is crucial to remember that, ultimately, the person who judges the quality of the writing and gives it its final form is a living human being, like you and me. This is precisely why we must read a great deal of excellent writing and continually sharpen our sensibilities of language.

In my case, this particular principle applies not only to Japanese, which is my mother tongue, but increasingly to English. In natural-language programming in English, what matters most is my own ability to judge the quality of English prose derived from my speech, whether originally delivered in Japanese or English—to experience its qualia and make the final adjustments. In other words, what is required is an overwhelming command of language, which I always strive to acquire somehow.

With the aim in mind that I must refine my sense of language still further, I am now reading James Joyce’s Ulysses and enjoying every sentence of it.