4 ms·

It's 50-80% because they are RLHFed into talking with "I". This was far less of an issue when it was just GPT-3 in a completion UI. But people find LLMs trained

It's 50-80% because they are RLHFed into talking with "I". This was far less of an issue when it was just GPT-3 in a completion UI. But people find LLMs trained to produce text that looks like it's coming from a personality to be more compelling: ChatGPT is when the tech exploded into popularity. LLMs that aren't chat tuned are just not as easy to anthropomorphize.

biophysboy 9mo ago
I really wish I could use custom products w/ RLHF turned off. I know that's not how it works, but the stupid marketing copy speak makes me use them less
- digiown 9mo ago
  If you use the API directly via open webui, etc., it is not nearly as annoying. You can also system prompt it into sounding more reasonable.
  - biophysboy 9mo ago
    thanks for the rec
Terr_ 9mo ago
My own minor attempt to hold back this tide involves urging everyone to imagine LLM exchanges a theater-play script document. The character's lines are not the author. Just imagine how different all this would be if every prompt contained something to make the character(s) obviously fictional, ex: "You are Count Dracula, dread lord of the night, and a visitor has the following question..." We hopefully wouldn't see mindless reports that "vampires are real now" or "Draculabot has developed coping mechanisms for the dark thirst, agrees to try tomato juice."
- lostmsu 9mo ago
  That was accurate only before the instruction tuning.
  - Terr_ 9mo ago
    I don't see how that gives the algorithm an ego when it didn't have one before.
    - lostmsu 9mo ago
      I don't see how that does not considering that base models never refer to themselves as "I" and most of the modern instruct ones do. I can foresee the objection, that there's still a distinction between an author and a script, but for instruct models at this point it seems no different than author intentionally lying when talking about themselves.
      - Terr_ 9mo ago
        > it seems no different than author intentionally lying That begs a really big question, assuming that humanoid type of "intent" already existed and was somehow mis-aimed the whole time. Not to be confused with the algorithmic intent of `f(tokens,opts) -> next_token` .
        roywiggins 9mo ago
        The blame attaches to the corporation that trained the "I" outputs into the model. Chat-tuned LLMs are mostly finetuned to roleplay as a friendly assistant at all times. This isn't exactly a lie- the models are really being designed to be helpful assistants- but the side-effect is that they also present as coherent personalities when they aren't. When a model outputs stuff like "I am FooGPT, a friendly chatbot" it is roleplaying just as much as when it's outputting stuff like "Hello, my name is Abraham Lincoln, I was the 16th President of the United States."
        Terr_ 9mo ago

biophysboy 9mo ago

I really wish I could use custom products w/ RLHF turned off. I know that's not how it works, but the stupid marketing copy speak makes me use them less

digiown 9mo ago
If you use the API directly via open webui, etc., it is not nearly as annoying. You can also system prompt it into sounding more reasonable.
- biophysboy 9mo ago
  thanks for the rec

Terr_ 9mo ago

My own minor attempt to hold back this tide involves urging everyone to imagine LLM exchanges a theater-play script document. The character's lines are not the author. Just imagine how different all this would be if every prompt contained something to make the character(s) obviously fictional, ex: "You are Count Dracula, dread lord of the night, and a visitor has the following question..." We hopefully wouldn't see mindless reports that "vampires are real now" or "Draculabot has developed coping mechanisms for the dark thirst, agrees to try tomato juice."

lostmsu 9mo ago
That was accurate only before the instruction tuning.
- Terr_ 9mo ago
  I don't see how that gives the algorithm an ego when it didn't have one before.
  - lostmsu 9mo ago
    I don't see how that does not considering that base models never refer to themselves as "I" and most of the modern instruct ones do. I can foresee the objection, that there's still a distinction between an author and a script, but for instruct models at this point it seems no different than author intentionally lying when talking about themselves.
    - Terr_ 9mo ago
      > it seems no different than author intentionally lying That begs a really big question, assuming that humanoid type of "intent" already existed and was somehow mis-aimed the whole time. Not to be confused with the algorithmic intent of `f(tokens,opts) -> next_token` .
      - roywiggins 9mo ago
        The blame attaches to the corporation that trained the "I" outputs into the model. Chat-tuned LLMs are mostly finetuned to roleplay as a friendly assistant at all times. This isn't exactly a lie- the models are really being designed to be helpful assistants- but the side-effect is that they also present as coherent personalities when they aren't. When a model outputs stuff like "I am FooGPT, a friendly chatbot" it is roleplaying just as much as when it's outputting stuff like "Hello, my name is Abraham Lincoln, I was the 16th President of the United States."
        Terr_ 9mo ago