5 ms·

I strongly suspect the program that's apologising isn't the actual LLM itself, but a wrapper around it that does some initial checking of inputs and possibly mo

I strongly suspect the program that's apologising isn't the actual LLM itself, but a wrapper around it that does some initial checking of inputs and possibly modifies them before sending them to the actual LLM.

muzani 3y ago
Yeah, it doesn't apologize as much when going through the API.
- krisoft 3y ago
  Do you use the same prompting scheme when going through the API? (Not just your own inputs but the buts OpenAI prepends to them.)
  - muzani 3y ago
    "what's ligma" on ChatGPT, April 2023: `Ligma is a made-up term that originated as part of an internet joke or prank. When someone asks, "What's ligma?" the person who initiated the joke usually responds with a crude or inappropriate punchline. The purpose of the joke is to elicit a humorous or shocked reaction from the person asking the question. As an Al language model, my goal is to provide accurate information and assistance. If you have any serious questions, please don't hesitate to ask, and I'll be glad to help.` "who is ligma", today, Aug 2023: `As of my last update in September 2021, "Ligma" is not a person; it's an internet meme or joke. The term "Ligma" is a play on words and a form of trolling. When someone asks, "What is Ligma?" the usual response is, "Ligma balls." It's a childish and deliberately provocative joke aimed at getting people to say the punchline without realizing what they are actually saying. This kind of joke is typical of internet culture, where memes and trolling are prevalent. However, as with any internet meme, its popularity and relevance can change over time, so it's possible that the context or significance of "Ligma" may have evolved or faded away since my last update in September 2021.` "who is ligma" over API: `Ligma is a fictional character or concept that originated as an internet meme. It is often used as a prank or joke to trick someone into asking "What is ligma?" The response is typically "Ligma balls," which is a play on words and a vulgar joke.` In April, it was very difficult to get it to explain the joke, and it will just outright refuse to do so, insisting that it can't say crude and inappropriate things. Recently, it's been more open to using the actual vulgarity to explain the phrase. The base API without any steering `system` prompt tends to exclude all the stuff ChatGPT was trained with, such as proposing a solution or apologising for using profanity or apologising for being an AI.
    - rvnx 3y ago
      It's sad that ChatGPT doesn't want to play the ligma game. From all the interactions, you can see that it understands what you want, and goes around it intentionally (though it wouldn't cause any harm).
      - muzani 3y ago
        Unguided davinci (GPT-3) doesn't go along with it and gives random answers like "Billie Eilish" or "a disease users suffer from after buying a windows computer" It's also interesting that the various levels of davinci act differently. text-davinci-002 and 003 are more likely to give a canned response.
- moonchrome 3y ago
  I mean you get to setup the system prompt on the API - maybe the system prompt is setup that way ?
  - muzani 3y ago
    The API system prompt is weak and can easily be overriden by user instructions. There's something similar on ChatGPT, but a lot stronger than the system prompt. I suspect GPT-3.5 is also just very heavily tuned davinci and such, to a level where it's much cheaper but also responds mechanically.
travbrack 3y ago
I was thinking it's an example in the system prompt. "Here's how you would behave when corrected" or similar
flir 3y ago
I've seen one example where it started apologising for what it had written within the current response. I have a suspicion that, if it is a MoE architecture, I was swapped from one model to another mid-response. Since them I've seen a few long pauses that make me wonder.

muzani 3y ago

Yeah, it doesn't apologize as much when going through the API.

krisoft 3y ago
Do you use the same prompting scheme when going through the API? (Not just your own inputs but the buts OpenAI prepends to them.)
- muzani 3y ago
  "what's ligma" on ChatGPT, April 2023: `Ligma is a made-up term that originated as part of an internet joke or prank. When someone asks, "What's ligma?" the person who initiated the joke usually responds with a crude or inappropriate punchline. The purpose of the joke is to elicit a humorous or shocked reaction from the person asking the question. As an Al language model, my goal is to provide accurate information and assistance. If you have any serious questions, please don't hesitate to ask, and I'll be glad to help.` "who is ligma", today, Aug 2023: `As of my last update in September 2021, "Ligma" is not a person; it's an internet meme or joke. The term "Ligma" is a play on words and a form of trolling. When someone asks, "What is Ligma?" the usual response is, "Ligma balls." It's a childish and deliberately provocative joke aimed at getting people to say the punchline without realizing what they are actually saying. This kind of joke is typical of internet culture, where memes and trolling are prevalent. However, as with any internet meme, its popularity and relevance can change over time, so it's possible that the context or significance of "Ligma" may have evolved or faded away since my last update in September 2021.` "who is ligma" over API: `Ligma is a fictional character or concept that originated as an internet meme. It is often used as a prank or joke to trick someone into asking "What is ligma?" The response is typically "Ligma balls," which is a play on words and a vulgar joke.` In April, it was very difficult to get it to explain the joke, and it will just outright refuse to do so, insisting that it can't say crude and inappropriate things. Recently, it's been more open to using the actual vulgarity to explain the phrase. The base API without any steering `system` prompt tends to exclude all the stuff ChatGPT was trained with, such as proposing a solution or apologising for using profanity or apologising for being an AI.
  - rvnx 3y ago
    It's sad that ChatGPT doesn't want to play the ligma game. From all the interactions, you can see that it understands what you want, and goes around it intentionally (though it wouldn't cause any harm).
    - muzani 3y ago
      Unguided davinci (GPT-3) doesn't go along with it and gives random answers like "Billie Eilish" or "a disease users suffer from after buying a windows computer" It's also interesting that the various levels of davinci act differently. text-davinci-002 and 003 are more likely to give a canned response.
moonchrome 3y ago
I mean you get to setup the system prompt on the API - maybe the system prompt is setup that way ?
- muzani 3y ago
  The API system prompt is weak and can easily be overriden by user instructions. There's something similar on ChatGPT, but a lot stronger than the system prompt. I suspect GPT-3.5 is also just very heavily tuned davinci and such, to a level where it's much cheaper but also responds mechanically.

travbrack 3y ago

I was thinking it's an example in the system prompt. "Here's how you would behave when corrected" or similar

flir 3y ago

I've seen one example where it started apologising for what it had written within the current response. I have a suspicion that, if it is a MoE architecture, I was swapped from one model to another mid-response. Since them I've seen a few long pauses that make me wonder.