3 ms·

I remember when Claude identified as ChatGPT a long time ago. It proves nothing else than that there is a lot of training material on the internet with Claude a

I remember when Claude identified as ChatGPT a long time ago. It proves nothing else than that there is a lot of training material on the internet with Claude as the AI.

scotty79 3mo ago
I don't understand why labs don't include correct model name in the training process. Almost nobody seems to be willing to tell their models who they actually are.
- embedding-shape 3mo ago
  If the model works correctly, you don't need to include the model name in the training process. You just add in the system prompt "You are X, trained by Y" and the model will claim to be that. That also allows users to "white-label" the outputs sort to say. This approach is basically how all "models know who they are" (they don't actually typically "know" that at all), it's just a system prompt instruction in the platform you use.
  - brookst 3mo ago
    But I would expect they would at least use post-training to discourage identifying as the wrong company.
    - XenophileJKO 3mo ago
      That only makes sense if you don't sell access to the model for 3rd party product development. One of those products is chat personas/agents. The company integrating the model will want it's own identity.
    - dannyw 3mo ago
      A lot of API customers don't want their chatbot or agent to identify as Claude or GPT, they want it to identify as <customisable marketingname>.
      - scotty79 3mo ago
        That's probably the true reason.
  - scotty79 3mo ago
    You can put anything in the prompt. And if you have bare model that you don't give any prompt from the start and want to find out which model you are talking to, you are out of luck. I think Qwen teaches its models that they are Qwen. Most others don't bother.
- irthomasthomas 3mo ago
  Because they do not know the name of the model before they train it. There is also distillation, where multiple models will be trained from a larger one. E.G. Sonnet was promoted to Opus at one point after it surpassed expectations.

I remember when Claude identified as ChatGPT a long time ago. It proves nothing else than that there is a lot of training material on the internet with Claude as the AI.

scotty79 3mo ago

I don't understand why labs don't include correct model name in the training process. Almost nobody seems to be willing to tell their models who they actually are.

embedding-shape 3mo ago
If the model works correctly, you don't need to include the model name in the training process. You just add in the system prompt "You are X, trained by Y" and the model will claim to be that. That also allows users to "white-label" the outputs sort to say. This approach is basically how all "models know who they are" (they don't actually typically "know" that at all), it's just a system prompt instruction in the platform you use.
- brookst 3mo ago
  But I would expect they would at least use post-training to discourage identifying as the wrong company.
  - XenophileJKO 3mo ago
    That only makes sense if you don't sell access to the model for 3rd party product development. One of those products is chat personas/agents. The company integrating the model will want it's own identity.
  - dannyw 3mo ago
    A lot of API customers don't want their chatbot or agent to identify as Claude or GPT, they want it to identify as <customisable marketingname>.
    - scotty79 3mo ago
      That's probably the true reason.
- scotty79 3mo ago
  You can put anything in the prompt. And if you have bare model that you don't give any prompt from the start and want to find out which model you are talking to, you are out of luck. I think Qwen teaches its models that they are Qwen. Most others don't bother.
irthomasthomas 3mo ago
Because they do not know the name of the model before they train it. There is also distillation, where multiple models will be trained from a larger one. E.G. Sonnet was promoted to Opus at one point after it surpassed expectations.