|
|
|
|
|
|
by TuringTest
20 days ago
|
|
|
> Which an unreliable answer is not. > Hallucinations are not a matter of some "details" being off. They are a matter of plausible, confident-sounding claims that are just plain wrong. This is no worse than Wikipedia, or the original encyclopedia for that matter. Those contain dubious claims that you'll need to verify on your own too. LLMs help because they have a gigantic amount of compressed knowledge, and they are able to find relevant information and present it incredibly fast. You wouldn't trust the ten first results of a Google search either, but you wouldn't say that having a search engine is totally useless and in no way an improvement over your local library, would you? > the poor person who's asking can't tell is wrong, because it sounds plausible and is stated with such confidence. True, but having to learn how to use a tool properly doesn't make the tool useless, even if it can hurt those who use it carelessly. |
|
|
A search engine that just indexes the web and gives me results is of course a great improvement over my local library.
But LLMs are not the same as a search engine. LLMs don't give me links. (Well, sometimes they do, and sometimes the links don't even exist, or if they exist, they don't actually say what the LLM said they say. At least with a search engine it's just the link, with no claims about what I'll find if I click on it.) They give me authoritative-sounding text. That's what they're for. And no, that text is not coming from "compressed knowledge". Text is not knowledge. Knowledge requires connections with the real world. LLMs don't have that. All they have is text.
A comparison I've used before is between LLMs and Wolfram Alpha. If I ask an LLM what's the distance from New York to Tokyo, the LLM has no idea that New York and Tokyo are places on the Earth, that the distance is a physical distance that can be measured, and that there is a right answer to the question. LLMs just generate text based on what their algorithm spits out as the most likely text to follow my prompt. The LLM doesn't even have any concept of what's happening if it gives me a wrong answer and I tell it the answer is wrong. It will spit out text saying, oh, yes, you're right, that's a wrong answer...and then spit out more text that might contain a different wrong answer, or even the same wrong answer. It literally has no concept that I am trying to extract meaning from its text.
Wolfram Alpha, on the other hand, if I ask it what's the distance from New York to Tokyo, figures out that I'm asking for a geographic distance, looks it up in its database of geographic distances (which has been curated by humans using actual geographic data from actual measurements on the actual Earth), and formats the answer in readable text. That's still a very simple connection to the real world outside of text, but at least it's some connection. LLMs have none. And that is what makes them useless as tools for trying to learn actual knowledge.