I really don't think this is happening. The problems with LLMs is when they use recursion on their own output they propagate errors and degenerate over time.
It's certainly true for image generating models. They need humans to tell them "this is that thing" in their training. And even after going through a ridiculous amount of images they will make mistakes.
But for programming as well. Unless the improvement metrics are VERY specific, an LLM modifying its own source code in order to improve would just add a lot of errors and mess itself up. And its standards of correctness are only as good as the previous LLM can 'reason'. If an LLM modifies source code to make the model more accurate in answering a question how can it determine this? It only has the best answer it can come up with. Human intervention is needed to say "this is better quality writing" or a more accurate answer.
I really don't think this is happening. The problems with LLMs is when they use recursion on their own output they propagate errors and degenerate over time.
It's certainly true for image generating models. They need humans to tell them "this is that thing" in their training. And even after going through a ridiculous amount of images they will make mistakes.
But for programming as well. Unless the improvement metrics are VERY specific, an LLM modifying its own source code in order to improve would just add a lot of errors and mess itself up. And its standards of correctness are only as good as the previous LLM can 'reason'. If an LLM modifies source code to make the model more accurate in answering a question how can it determine this? It only has the best answer it can come up with. Human intervention is needed to say "this is better quality writing" or a more accurate answer.