• 0 Posts
  • 28 Comments
Joined 25 days ago
cake
Cake day: February 27th, 2025

help-circle
rss








  • That’s something people really have to get into their heads: an “answer” by an LLM ist just a series of high probability tokens. It’s only us humans who interpret reason and value into it. From the system’s standpoint it’s just numbers without any meaning whatsoever. And no amount of massaging will change that. LLMs are about as “intelligent” as a fancy database query.



  • Every couple of years I try out the latest “assistant”. I ask very simple queries, stuff that you might expect it to be able to handle. It always disappoints hilariously.

    Questions where previous assistants consistently and utterly failed: “What’s the weather going to be like today?” “I want to go to [address], using public transport, can you tell me which routes I have to take?” “What are my appointments today?”

    You’d think these are pretty obvious use-cases, but every iteration of assistants was completely incapable of getting even close to a satisfying answer. I don’t expect LLMs to fare better this time around.