Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Its not 'thanking', its positive signal that GPT's previous outputs were correct, so it should continue doing whatever it was doing. If you say no/bad etc, then GPT will try other approaches.


> "its positive signal that GPT's previous outputs were correct"

Is that somehow baked into the algorithms?

Are positive words of encouragement interpreted as "positive signals" by the inference pipeline? Or do they somehow influence the attention mechanism?

Because otherwise, you're just rationalizing completely random and unpredicted behavior.


> Is that somehow baked into the algorithms?

OpenAI’s API is stateless, which means you need to send the entire conversation thread in each request.

So when you send a response like “perfect, now do…”, you’re reinforcing the language model that the conversation history is on the right track.


That's literally what I said: those words are "encouraging" it.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: