Source: Hacker News
Why privacy breaks fairness at small scale—and how collaboration fixes both without sharing a single record The post I Evaluated Half a Million Credit Records w...
TL;DR LLMs train on stuff like documentation, GitHub repositories, StackOverflow, and Reddit. But as we keep using LLMs, their own output goes into these platf...
TL;DR I forced GPT‑2 to learn from its own output for 20 generations. By Generation 20 the model lost 66 % of its semantic volume and began hallucinating state...
Article URL: https://embd.cc/llm-problems-observed-in-humans Comments URL: https://news.ycombinator.com/item?id=46527581 Points: 24 Comments: 2...
