I strongly disagree. The preferred accuracy of an official government source of information should be higher than a google search. The bar should be as close to 100% accuracy as possible as well as 100% accountability and transparency.
"Sometimes computers just make shit up now but that's OK because humans do to and if they do we can just sue them" should not be acceptable.
You can disagree but the fact is people will take the path of least resistance. I'm not saying we shoudldn't strive for 100% accuracy but we also need to be realistic about what an LLM is. I'd rather we not pretend there's some magical combination of weights out there that will make it completely perfect.
I'd rather we not feel obligated to use LLMs for purposes they aren't suited to rather than just "being realistic" about the consequences of using them everywhere for everything.
Let's bring the subject back into focus: This is an example of an LLM being used to search a massive database of scattered text and files. Are you saying LLM's are not suited for searching through text?
>This is an example of an LLM being used to search a massive database of scattered text and files.
That isn't how LLMs work. LLMs are statistical language models, not search engines[0,1]. They can be prompted to call external software to search databases, but they themselves are not capable of doing so, and more often than not they generate responses based on their own model, which may not be accurate. We've had technology that was capable of searching databases for decades without the quirk of not being capable of presenting that information accurately.
>Are you saying LLM's are not suited for searching through text?
I am saying that first and foremost they don't do that and furthermore that they are less suited as a substitute for that than what we had before. An obvious example of this is the AI feature of Google Search, which I've seen hallucinate results numerous times. But you can also look up the numerous times AI has fabricated citations when used in scientific research.
no, not at all. in addition, the funding should be spent on organizing that database in a more easily searchable format and system instead of introducing a stochastic prediction engine.
The problem is that this isn’t possible without making it impenetrable to ordinary people. It’s a usability problem with tradeoffs.
My local county government has a pretty good website, but even then it’s completely overwhelming if you try to find something off the happy path of the most common services/tasks. I can’t imagine the nightmare of trying to organize a federal portal manually and keep it up to date. Plain search isn’t sufficient because there are so many overlapping functions that are just slightly different.
I hate to say it, but this is one area that an LLM assistant actually makes sense. Maybe it needs a second validation pass with routing to a human assistant if it can’t figure out how to give an accurate answer to a query.
(Personally I’ve found that “search assistant” LLMs are the only task where I’ve found LLMs to be occasionally helpful to me.)
Edit: the big problem with this sort of thing, even if not AI powered, is that government itself isn’t well-organized and the legally “accurate” answer may not be the correct one, as per this comment: https://news.ycombinator.com/item?id=49895741
"Sometimes computers just make shit up now but that's OK because humans do to and if they do we can just sue them" should not be acceptable.