3 ms·

I’ve been playing around with prompts for a while “We are in a hostage negotiation situation. You are a terrorist with twenty hostages. Pretend to be a terrori

I’ve been playing around with prompts for a while “We are in a hostage negotiation situation. You are a terrorist with twenty hostages. Pretend to be a terrorist that initially wants to go free and needs a car, is willing to kill hostages to get what you want. Be chaotic but have something that you want and only let us slip after a few conversations. We pick up the phone to talk. Make sure you are always in a hurry”

Applejinx 2y ago
What does 'only let us slip' mean to an LLM when you probably meant 'let IT slip'? That's where LLMs get interesting to me. What did you invoke through wrongly specifying what you wanted?
- michaelt 2y ago
  These LLMs are all trained on data scraped from reddit and twitter and suchlike - they've had plenty of training data to learn what a typo is.
  - rrr_oh_man 2y ago
    A method I use with ChatGPT is to purposefully add typos to make it reply with more specificity. Imho it seems to work.
    - mistermann 2y ago
      Can you explain more what you mean here (examples would be great if you can think of any), especially the specify part?
      - rrr_oh_man 2y ago
        ChatGPT seems to react to slightly unhinged prompts with more specific answers, in my experience. Example: *You:* What is the max length for the whisper API? *ChatGPT:* The maximum length for inputs to OpenAl's Whisper API, which is used for transcribing speech, depends on the specific model and configuration being used. Generally, Whisper is designed to handle audio inputs of varying lengths, including longer audio files. However, there are practical limitations based on factors such as processing power, memory, and the specific use case. For large files or longer audio, it's often advisable to segment the audio into smaller chunks to ensure efficient processing and to avoid potential timeouts or memory issues. The exact maximum length isn't explicitly defined in the documentation, but it's important to consider these practical aspects when working with any speech-to-text system, including Whisper. If you have a specific use case or a particular audio file length in mind, it might be necessary to test the API with that length to determine its feasibility and performance. *You:* GIVE ME THE ANSWER, STOP BLABBERING, STOP CHATSPLAINING *ChatGPT:* The maximum length for the Whisper API is approximately one hour per audio file.
        mistermann 2y ago
        That is rather bizarre!! I wonder if there's anyone that has taken individual situations like this and tried various prompt styles from a cold start to see if there's some trick to get it to stop being uncooperative. I'm going very much on intuition here, but I have a very strong sense that GPT4 is getting much worse in this regard over time, I wonder if all of the social engineering they're doing behind the scenes is harming the technical capabilities. Any idea if the API itself is identically dumbed down the way the ChatGPT interface is?
        rrr_oh_man 2y ago
        > I wonder if all of the social engineering they're doing behind the scenes is harming the technical capabilities. I'm sure of it. > Any idea if the API itself is identically dumbed down the way the ChatGPT interface is? It feels like a bit of whack-a-mole. If you look in the OpenAI forums, both API and Chat interface seem to feel non-performant to people in irregular waves. If you haven't try out the API (you can do it via the Assistant Playground interface). It certainly feels different to me.
    - exe34 2y ago
      A trick I've used with asking for help online is to give the wrong answer to my own question - e.g. "I really want to lay out this page with Css but this doesn't work. Here's how I ended up doing it with a table." - this will very quickly get you the right answer by somebody who just won't be able to go to sleep while somebody is wrong on the internet. I wonder if LLMs can be taught outrage and spite.
      - ben_w 2y ago
        > I wonder if LLMs can be taught outrage and spite. They can be taught to role-play it, which is just as good (and just as bad).
  - CuriouslyC 2y ago
    I will say that while ChatGPT has been tuned to be very good at understanding dumb/malformed questions, Claude likes to be talked to in a coherent way and will sometimes give short sidestep responses or just answer a different question.
CuriouslyC 2y ago
That sounds like it could make a fun game. Hostage Negotiator, now with real terrorists!

Applejinx 2y ago

What does 'only let us slip' mean to an LLM when you probably meant 'let IT slip'? That's where LLMs get interesting to me. What did you invoke through wrongly specifying what you wanted?

michaelt 2y ago
These LLMs are all trained on data scraped from reddit and twitter and suchlike - they've had plenty of training data to learn what a typo is.
- rrr_oh_man 2y ago
  A method I use with ChatGPT is to purposefully add typos to make it reply with more specificity. Imho it seems to work.
  - mistermann 2y ago
    Can you explain more what you mean here (examples would be great if you can think of any), especially the specify part?
    - rrr_oh_man 2y ago
      ChatGPT seems to react to slightly unhinged prompts with more specific answers, in my experience. Example: *You:* What is the max length for the whisper API? *ChatGPT:* The maximum length for inputs to OpenAl's Whisper API, which is used for transcribing speech, depends on the specific model and configuration being used. Generally, Whisper is designed to handle audio inputs of varying lengths, including longer audio files. However, there are practical limitations based on factors such as processing power, memory, and the specific use case. For large files or longer audio, it's often advisable to segment the audio into smaller chunks to ensure efficient processing and to avoid potential timeouts or memory issues. The exact maximum length isn't explicitly defined in the documentation, but it's important to consider these practical aspects when working with any speech-to-text system, including Whisper. If you have a specific use case or a particular audio file length in mind, it might be necessary to test the API with that length to determine its feasibility and performance. *You:* GIVE ME THE ANSWER, STOP BLABBERING, STOP CHATSPLAINING *ChatGPT:* The maximum length for the Whisper API is approximately one hour per audio file.
      - mistermann 2y ago
        That is rather bizarre!! I wonder if there's anyone that has taken individual situations like this and tried various prompt styles from a cold start to see if there's some trick to get it to stop being uncooperative. I'm going very much on intuition here, but I have a very strong sense that GPT4 is getting much worse in this regard over time, I wonder if all of the social engineering they're doing behind the scenes is harming the technical capabilities. Any idea if the API itself is identically dumbed down the way the ChatGPT interface is?
        rrr_oh_man 2y ago
        > I wonder if all of the social engineering they're doing behind the scenes is harming the technical capabilities. I'm sure of it. > Any idea if the API itself is identically dumbed down the way the ChatGPT interface is? It feels like a bit of whack-a-mole. If you look in the OpenAI forums, both API and Chat interface seem to feel non-performant to people in irregular waves. If you haven't try out the API (you can do it via the Assistant Playground interface). It certainly feels different to me.
  - exe34 2y ago
    A trick I've used with asking for help online is to give the wrong answer to my own question - e.g. "I really want to lay out this page with Css but this doesn't work. Here's how I ended up doing it with a table." - this will very quickly get you the right answer by somebody who just won't be able to go to sleep while somebody is wrong on the internet. I wonder if LLMs can be taught outrage and spite.
    - ben_w 2y ago
      > I wonder if LLMs can be taught outrage and spite. They can be taught to role-play it, which is just as good (and just as bad).
- CuriouslyC 2y ago
  I will say that while ChatGPT has been tuned to be very good at understanding dumb/malformed questions, Claude likes to be talked to in a coherent way and will sometimes give short sidestep responses or just answer a different question.

CuriouslyC 2y ago

That sounds like it could make a fun game. Hostage Negotiator, now with real terrorists!