Hacker News new | ask | show | jobs
by Al-Khwarizmi 1175 days ago
> Regarding the AI, it’s literally just a wrapper of OpenAI completions API (turbo-3.5 model) and a chat interface, so it should give the same answers as the free version of ChatGPT.

Have you tested this? My impression after a couple of tests is that its answers are worse than free ChatGPT (no idea why).

Maybe I just got unlucky with the RNG, though.

2 comments

The free user interface does something behind the scenes to provide context to chatgpt but I haven't figured out what it is
> My name is George

Nice to meet you, George. How can I assist you today?

> What is my name

As an AI language model, I don't have access to your personal information such as your name. Can you please tell me your name?

You need to pass the whole chat history in your request:

    [{:role "assistant", :content "Ciao, sono PizzaGPT come posso aiutarti?"}
     {:role "user",      :content "My name is George"}
     {:role "assistant", :content "Nice to meet you, George. How can I assist you today?"}
     {:role "user",      :content "What is my name"}
     {:role "assistant", :content "Your name is George."}]
On the ChatCompletion API endpoint
I think it is doing something a bit sophisticated in choosing what to pass in the context when the conversation grows beyond the token limit
Yes, I know, but it was a deliberate choice not to send the entire history, to keep the token consumption and costs low
Not really had much time to test the results. Yeah it should be possible to tweak the settings (mostly temperature and prompt) to get better results. Also I think I’m not sending the full history right now, to avoid consuming too many tokens
>Also I think I’m not sending the full history right now, to avoid consuming too many tokens

You can ask it to summarize the dialog, thus compressing the dialog history into a few sentences. That will give you some basic dialog context.