Comment by nextos

5 months ago

You have a point but current LLM architectures in particular are very fragile to data poisoning [1,2].

[1] https://www.anthropic.com/research/small-samples-poison

[2] https://arxiv.org/abs/2510.07192

No idea why you're being downvoted. We can't yet even demonstrate that LLMs will withstand training on their own output as they pollute the Internet.