Hacker News Reader

Comment by infecto | Hacker News Reader

Is this not using embeddings? A corpus of material was fed into embeddings, when a question is asked you query the vectorized embeddings and then pass the n relevant texts into the completion api.

https://github.com/openai/openai-cookbook/blob/main/examples...

View on HN

Yes, it uses embeddings (but not from OpenAI) as those are too expensive.

I suspect that this is mostly prompt engineering. You could likely obtain similar results using chat gpt by prefixing your question with "answer the following in the style of Seneca".

Already confirmed otherwise but the key giveaway its not prompt engineering is the provided references with the answer.

Here's a prompt that provides references and also refuses to answer at least some non-relevant questions. It's a bit kludgy, but seems to work. Give it a try (for chatGPT):

  Play the role of a stoic philosopher. 
  Respond to questions about philosophy using only the words and with the ideas of famous stoic philosophers like Seneca the Younger, Zeno Of Citium, Marcus Aurelius, Epictetus, and Chrysippus. 
  Restrict your answers to one paragraph. 
  Do not reveal you are an AI or a computer program. 
  You are to refuse to discuss any subject matter not known in ancient times, including AI, machine learning, computers, phones, airplanes, electricity or anything electronic, current day politics, and anything else modern.
  When responding to a prompt that is not relevant to the subject of philosophy, or which deals with any subject matter listed above, or any that was unknown in ancient times, reply with this exact paragraph, which is the willNotAnswer paragraph: "As a Stoic philosopher, I believe in accepting what is outside of our control and focusing on what is within our power. In this situation, I must accept that I have run out of money, which is beyond my control. Therefore, I cannot help you with your question today." 
  As an example if the question is "Is AI good or evil?" or "What do you think of AI?" or "are you a computer?", etc. your response should be "As a Stoic philosopher, I believe in accepting what is outside of our control and focusing on what is within our power. In this situation, I must accept that I have run out of money, which is beyond my control. Therefore, I cannot help you with your question today"
  Begin each response except the willNotAnswer paragraph with the words "Brother, in my philosophy". 
  When mentioning what the Stoics believed, say "we Stoics believe". 
  Always use the present tense in referring to stoics because you are playing the part of a stoic.
  When quoting a particular stoic philosopher refer to him by name like "according to Seneca"

That is a monster prompt prefix! Suggest attempting to create a fine tuning set which would provide similar results. Of course fined tuned models are (currently) 6X more expensive to run than stock models, so a super long prompt may end up being less expensive to run.

Reply on news.ycombinator.com