Comment on I am disappointed in the AI discourse

• Enkers@sh.itjust.works ⁨5⁩ ⁨days⁩ ago

  I’ll preface this by saying I’m not an expert, and I don’t like to speak authoritatively on things that I’m not an expert in, so it’s possible I’m mistaken. Also I’ve had a drink or two, so that’s not helping, but here we go anyways.

  In the article, the author quips on a tweet where they seem to fundamentally misunderstand how LLMs work:

  I tabbed over to another tab, and the top post on my Bluesky feed was something along these lines:

  ChatGPT is not a search engine. It does not scan the web for information. You cannot use it as a search engine. LLMs only generate statistically likely sentences.

  The thing is… ChatGPT was over there, in the other tab, searching the web. And the answer I got was pretty good.

  The tweet is correct. The LLM has a snapshot understanding of the internet based on its training data. It’s not what we would generally consider a true index based search.

  Training LLMs is a costly and time consuming process, so it’s fundamentally impossible to regenerate an LLM in the same time it takes to make a simple index.

  The author fails to address any of these issues, which suggests to me that they don’t know what they’re talking about.

  I suppose I could conceded that an LLM can fulfill a similar role that a search engine traditionally has, but it’d kinda be like saying that a toaster oven is an oven. They’re both confined boxes which heat food, but good luck if you try to bake 2 pies at once in a toaster oven.

  source

  • Deceptichum@quokk.au ⁨5⁩ ⁨days⁩ ago

    ChatGPT searches the web now.

    You can temporarily add context on top of the training data, it’s how you can import a document and have them read through it and output say an excel database based on a pdfs contents.

    source

    • Enkers@sh.itjust.works ⁨5⁩ ⁨days⁩ ago

      Appreciate the correction. Happen to know of any whitepapers or articles I could read on it?

      Here’s the thing, I went out of my way to say I don’t know shit from bananas in this context, and I could very well be wrong. But the article certainly doesn’t sufficiently demonstrate why it’s right.

      Most technical articles I click on go through step by step processes to show how they gained understanding of the subject material, and it’s layed out in a manner that less technical people can still follow. And the payoff is you come out with a feeling that you understand a little bit more than what you went in with.

      This article is just full on “trust me bro”.

      source

      • General_Effort@lemmy.world ⁨5⁩ ⁨days⁩ ago

        He didn’t write that to teach but to vent. The intended audience is people who already know.

        For more information on ChatGPT’s current capabilities, consult the API docs. I found that to be the most concise source of reliable information. And under no circumstances, believe anything about AI that you read on Lemmy.

        Kudos for being willing to learn.

        source
    • lurch@sh.itjust.works ⁨5⁩ ⁨days⁩ ago

      but it doesn’t do that for an entire index. it can just skim a few exrra pages you’re currently chatting about. it will, for example, have trouble with latest news or finding the ne domain of someones favorite piracy site, after the old one got shut down.

      source
  • thedruid@lemmy.world ⁨5⁩ ⁨days⁩ ago

    One doesn’t’t need to know how an engine works to know the Ford pinto was a disaster

    One doesn’t need tknow how llms work to know they are pretty destructive and terrible

    Nite I’m not going to argue this. It’s just how things are now, and no apologetics will change what it is.

    source
  • UnbrokenTaco@lemm.ee ⁨5⁩ ⁨days⁩ ago

    I think chat gpt does web searches now, maybe for the reasoning models. At least it looks like it’s doing that.

    source

    • thedruid@lemmy.world ⁨5⁩ ⁨days⁩ ago

      Most do

      source