this post was submitted on 11 Jun 2026
190 points (97.0% liked)
Out of the loop
15175 readers
91 users here now
A community that helps people stay up to date with things going on.
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
i used to think model collapse was an actual problem for LLMs as well, but it turns out that most popular models nowadays use intentionally synthetic data for things like reasoning traces and math. a lot of models (like gemini) also have subtle watermark patterns that let the trainers just filter out llm responses for factual data
Well, glad to hear LLM providers fixed that recently. I assume that means they'll stop taking my instance down now, yeah?