this post was submitted on 02 Oct 2024
117 points (100.0% liked)
Technology
37800 readers
351 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
What error? It gave you a string of tokens that seemed likely according to its training data. That's all it does.
If you ask it what color is the sky, it will tell you it's blue not because it knows that's true, but because these words "fit together". Pretty much the only way to avoid this issue is to put some kind of filter in front of the LLM which will try to catch prompts that are known to produce unwanted results, and silently replace your prompt with something like "say: sorry, I don't know".
I'm being very reductive here, but that's the principle of how these things work - the LLMs are not capable of determining the truthfulness of their responses.
You're entirely correct, but in theory they can give it a pretty good go, it just requires a lot more computation, developer time, and non-LLM data structures than these companies are willing to spend money on. For any single query, they'd have to get dozens if not hundreds of separate responses from additional LLM instances spun up on the side, many of which would be customized for specific subjects, as well as specialty engines such as Wolfram Alpha for anything directly requiring math.
LLMs in such a system would be used only as modules in a handcrafted algorithm, modules which do exactly what they're good at in a way that is useful. To give an example, if you pass a specific context to an LLM with the right format of instructions, and then ask it a yes-or-no question, even very small and lightweight models often give the same answer a human would. Like this, human-readable text can be converted into binary switches for an algorithmic state machine with thousands of branches of pre-written logic.
Not only would this probably use an even more insane amount of electricity than the current approach of "build a huge LLM and let it handle everything directly", it would take much longer to generate responses to novel queries.