There are some in the research community that agree with your take: THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET

Basically the long and short of that paper is that LLMs are inherently biased towards likely responses. The more their training set is LLM generated, and thus contains that bias, the less the LLM will be able to produce unlikely responses, over time degrading the model quality throughout successive generations.

However, I tend to think this viewpoint is probably missing something important. Can you train a new LLM on today's internet? Probably not, at least without some heavy cleaning. Can you train a multimodal model on video, audio, the chat logs of people talking to it, and even other better LLMs? Yes, and you will get a much higher quality model and likely won't get the same model collapse implied by the paper.

This is more or less what OpenAI has done. All the conversations with 100M+ users are saved and used to further train the AI. Their latest GPT4 is also trained on video and image recognition, and they have also been exploring ways for LLMs to train new ones, especially to aid in alignment of these models.

Another recent example is Orca, a fine tune of the open source llama model, which is trained by GPT-3.5 and GPT-4 as teachers, and retains ~90% of GPT-3.5's performance though it uses a factor of 10 less parameters.