4 ms·
Wow this sounds like they trained it on ChatGPT outputs. However I could not reproduce this output.
by ucha 3y ago
Wow this sounds like they trained it on ChatGPT outputs.
However I could not reproduce this output.
- agucova 3y agoIt 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.
- agucova 3y agoAlso, training over ChatGPT outputs is terribly expensive and Anthropic already has a working competing paradigm for fine-tuning (Claude worked before ChatGPT was released).
- throwaway9274 3y agoYeah, 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/ https://projectdecember.net/