Comment

Comment on An LLM would probably run the USA better

Slippery_Snake874@sopuli.xyz ⁨1⁩ ⁨week⁩ ago

We don’t know why LLMs return the responses they return.>>

While I agree with most of your points, this is a strange thing to say. Sure, you and I don’t know why LLMs return the responses they do, but the people who actually make them definitely know how they work.

source

Sort:hotnew top

BertramDitore@lemm.ee ⁨1⁩ ⁨week⁩ ago
It isn’t just you and me. Not even the people who designed them fully understand why they give the responses they give. It’s a well-known problem. Our understanding is definitely improving over time, but we still don’t fully know how they do it.

Here’s the latest exploration of this topic I could find.

LLMs continue to be one of the least understood mass-market technologies ever

Tracing even a single response takes hours and there’s still a lot of figuring out left to do.

source
- Slippery_Snake874@sopuli.xyz ⁨1⁩ ⁨week⁩ ago
  Hmm. That is interesting, and I admit it does seem like the company that made it is also still researching their own model, but some parts of the article seem a bit dramatic (not sure if there is a better word).
  
  Like when it says the model doesn’t “admit” to how it solved the math problem when asked. Of course it doesn’t, it is made for humans to interact with so it is not going to tell a human how a computer does math, it makes more sense for it to explain the “human” method.
  
  Interesting stuff though, thanks for the article!
  
  source
  - rikudou@lemmings.world ⁨1⁩ ⁨week⁩ ago
    The wording is very sensationalist. It’s currently hard to trace any single output specifically, but that doesn’t mean we don’t know how it works. The algorithm and theory behind it are very well understood. It’s just that the algorithm is very complex (and uses very large amounts of data), which is what makes tracing a single response token by token so hard.
    
    When you read something about how we don’t understand how AI works, you can safely skip that article, they intentionally use dishonest language to make it sound like something mystical is going on.
    
    source
    Slippery_Snake874@sopuli.xyz ⁨1⁩ ⁨week⁩ ago
    Yeah that makes a lot more sense to me. I can definitely understand that the algorithm might be very large and complex for one person to understand, but it isn’t magic.
    
    source