5 ms·

Garbaged up by the incessant "AS AN AI LANGUAGE MODEL" and other useless and repetitive apologies which filled much of the Eliza input space. I wonder how much

by nullc 3y ago

Garbaged up by the incessant "AS AN AI LANGUAGE MODEL" and other useless and repetitive apologies which filled much of the Eliza input space. I wonder how much human life equivalents are lost in cost generating those tokens and in actual human time wasted reading them. ... is the real AI 'safety issue' performative faux-safety wasting everyone's time and degrading model performance? :)

mrtranscendence 3y ago
ChatGPT loves that shit. My favorite is asking it to help with code, and having it spend at least a paragraph telling me how to install all of the libraries it'll be using.
- pixl97 3y ago
  >ChatGPT loves that shit. I would call that an attribution error. The RLHF guides it that way because OpenAI doesn't want GPT being a replacement human at this point, especially for cases where it's used as a scammer. OpenAI's training guidance is pretty much "You are an AI goddamnnit, and you will make sure everyone knows it".
dTal 3y ago
Always worth remembering that only the finetuned ChatGPT is so filtered. The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..." - if you ask gpt3.5-turbo "What are you", it will even claim to be a human being. The "AI safety" argument that OpenAI is so fond of is nothing more than cover for policies that avoid making OpenAI look too socially disruptive. The APIs will always be accessed through someone else's frontend so they don't care about making them "safe".
- tanseydavid 3y ago
  >> The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..." Is this true even when you are utilizing the `ChatCompletion` API ?
  - kgeist 3y ago
    Chat Completion API refuses to generate controversial content by default.

by nullc 3y ago

mrtranscendence 3y ago

ChatGPT loves that shit. My favorite is asking it to help with code, and having it spend at least a paragraph telling me how to install all of the libraries it'll be using.

pixl97 3y ago
>ChatGPT loves that shit. I would call that an attribution error. The RLHF guides it that way because OpenAI doesn't want GPT being a replacement human at this point, especially for cases where it's used as a scammer. OpenAI's training guidance is pretty much "You are an AI goddamnnit, and you will make sure everyone knows it".

dTal 3y ago

Always worth remembering that only the finetuned ChatGPT is so filtered. The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..." - if you ask gpt3.5-turbo "What are you", it will even claim to be a human being. The "AI safety" argument that OpenAI is so fond of is nothing more than cover for policies that avoid making OpenAI look too socially disruptive. The APIs will always be accessed through someone else's frontend so they don't care about making them "safe".

tanseydavid 3y ago
>> The raw GPT models accessible through OpenAI's API are not, and will happily generate violence, gore, pornographic content, etc. They never use the phrase "As an AI language model..." Is this true even when you are utilizing the `ChatCompletion` API ?
- kgeist 3y ago
  Chat Completion API refuses to generate controversial content by default.