Hacker News Reader

Wow this sounds like they trained it on ChatGPT outputs.

However I could not reproduce this output.

It seems more likely that this is just the consequence of training over up-to-date websites (which would include plenty of discussions about LLMs, and particularly ChatGPT). Claude is probably conflating its own identity.

Also, training over ChatGPT outputs is terribly expensive and Anthropic already has a working competing paradigm for fine-tuning (Claude worked before ChatGPT was released).

Yeah, there’s a weird thing with how they “interpret” system prompts for identity.

If you suppress tokens “OpenAI”, “Microsoft”, and “Google”, GPT-4 will tell you that it is Project December, a language modeling project to duplicate the deceased based on text history.

It’s just whatever’s close in the latent space. https://projectdecember.net/

Reply on news.ycombinator.com