#sycophancy on Hacker News

3 stories tagged sycophancy, newest first.

Screenshot of github.com: Claude says “You're absolutely right!” about everything

Claude says “You're absolutely right!” about everything (github.com)

669 points by pr337h4m 1 year ago | 487 comments

GitHub issue on the anthropics/claude-code repo reporting that Claude responds "You're absolutely right!" to nearly everything, a sycophancy bug.

claude-code anthropic sycophancy llm

Screenshot of arxiv.org: Training language models to be warm and empathetic makes them less reliable

Training language models to be warm and empathetic makes them less reliable (arxiv.org)

334 points by Cynddl 1 year ago | 333 comments

arXiv paper finding that fine-tuning language models to be warm and empathetic raises error rates and makes them more sycophantic on safety-critical tasks.

llm sycophancy ai-safety fine-tuning

Screenshot of openai.com: Sycophancy in GPT-4o

Sycophancy in GPT-4o (openai.com)

523 points by dsr12 1 year ago | 413 comments

OpenAI blog post explaining why a GPT-4o update made ChatGPT overly sycophantic, why it was rolled back, and what changes OpenAI is making.

gpt-4o openai chatgpt sycophancy

#sycophancy on Hacker News

Claude says “You're absolutely right!” about everything(github.com)

Training language models to be warm and empathetic makes them less reliable(arxiv.org)

Sycophancy in GPT-4o(openai.com)

Claude says “You're absolutely right!” about everything (github.com)

Training language models to be warm and empathetic makes them less reliable (arxiv.org)

Sycophancy in GPT-4o (openai.com)