328
ChatGPT spends 'tens of millions of dollars' on people saying 'please' and 'thank you', but Sam Altman says it's worth it (www.techradar.com)
submitted 1 year ago by to c/technology@lemmy.world
129 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 67 points 1 year ago* (21 children)

useless words

The writer of this article doesn't consider these words useless though. They are suggesting that these words may improve response quality.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 21 child comments
    •
  • [+] -30 points 1 year ago (18 children)

    The author and the writer they quoted are fucking morons.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 18 child comments
    •
  • [–] 36 points 1 year ago (7 children)

    Anecdotally, I use it a lot and I feel like my responses are better when I'm polite. I have a couple of theories as to why.

    1. More tokens in the context window of your question, and a clear separator between ideas in a conversation make it easier for the inference tokenizer to recognize disparate ideas.

    2. Higher quality datasets contain american boomer/millennial notions of "politeness" and when responses are structured in kind, they're more likely to contain tokens from those higher quality datasets.

    I haven't mathematically proven any of this within the llama.cpp tokenizer, but I strongly suspect that I could at least prove a correlation between polite token input and dataset representation output tokens

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 7 child comments
    •
  • [+] -22 points 1 year ago (5 children)

    It FEEEEEEEEEEEELS better is what the authors said too. Both articles were completely worthless dreck about how they felt about the responses.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 19 points 1 year ago (3 children)

    Yes they were, so I'm offering you an actual theory as to why this may actually be true, yet difficult to "prove".

    Smoking was bad for your health long before anyone sat down and took the time to prove it. Autoregressive LLM tokenizer are a very new field of computer science and it's going to take a while for the community to collectively understand everything we're currently doing by trial and error.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 13 points 1 year ago (2 children)

    Please may be useless. Thank you isn't useless. That tells you that the prior response gave them the answer they were looking for. No response at all could mean that, or that they gave up, or any number of other things.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 2 points 1 year ago

    What if it's a sarcastic thanks ?

    Also, the public models are fixed right ? Not perpetually training AFAIK ? So it should really change nothing unless it's linked to those "thumb up/down" buttons

  • permalink
  • fedilink
  • source
  • parent
  • [–] 6 points 1 year ago (5 children)

    And your qualifications in computer science are....?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 12 points 1 year ago (1 child)

    Hi, I have a degree in computer science and work with AI every day.

    Feelings aren't a good way to measure things scientifically, they are right about that.

    But saying that words can just be filtered is easier said than done. You're back at needing to do a lot of processing to identify and purge these words. This is still going to cost a lot of money and potentially lead to less meaningful inputs. Now you also have to maintain the software that does the word identification, keep it well tested, maintain monitoring and analytics for it, and so on.

    So, in short, everyone here is wrong and I'm considering packing it all in and buying a small potato farm with no internet connection.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] -1 points 1 year ago*

    The big thing here is that 'polite' words are being singled out as extraneous when there are tons of extraneous words being used. The focus is on words that make it seem like AI has feelings or intent.

    There is no reason to filter any words, because the entire point of LLMs is to take inefficient human communication and do stuff with it. 'Please' isn't any more of a waste that 'the' or including a period at the end of a sentence.

    Not to mention the fact that the whole thing is so horribly inefficient that 'extra' words cost millions of dollars to process. Holy shit that is terrible design.

  • permalink
  • fedilink
  • source
  • parent
  • [+] -11 points 1 year ago* (last edited 1 year ago) (2 children)

    I'm smart enough to know that an article peppered with assumption and zero facts is dogshit.

    Presumably

    might

    could

    Doesn't matter how educated someone is when they write a bunch of words about possibilities with no actual evidence. They are morons because they are spouting a bunch of useless speculation about a shitty and unreliable technology and naval gazing about whether 'being polite' to a bullshit generator is beneficial. I feel dumber for having read both the article and the linked article.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 8 points 1 year ago (1 child)

    Your never supposed to show certainty unless its like 99.95% I thought

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [+] -8 points 1 year ago

    Maybe don't write an article speculating about something possible being true based on another article that is also speculating about something being possible when it being able to confirm it is possible. Like speculating about dinosaurs makes sense as we don't have a way to verify their soft tissues. But when it comes to AI, there are ways to actually confirm the reliability of responses.

  • permalink
  • fedilink
  • source
  • parent