this post was submitted on 21 Oct 2024
88 points (100.0% liked)
Technology
37725 readers
709 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
This is an awesome use of an LLM. Talk about the cost savings of automation, especially when the alternative was the reviews just not getting done.
Specialized LLMs trained for specific tasks can be immensely beneficial! I'm glad to see some of that happening instead of "Company XYZ is now needlessly adding AI to it's products because buzzwords!"
Given the error rate of LLMs, it seems more like they wasted $258 and a week that could have been spent on a human review.
LLMs are bad for the uses they've been recently pushed for, yes. But this is legitimately a very good use of them. This is natural language processing, within a narrow scope with a specific intention. This is exactly what it can be good at. Even if does have a high false negative rate, that's still thousands and thousands of true positive cases that were addressed quickly and cheaply, and that a human auditor no longer needs to touch.
What do you believe would make this particular use prone to errors?
The use of LLMs instead of someone that can actually understand context.
I think you may have misunderstood the purpose of this tool.
It doesn't read the deeds, make a decision, and submit them for termination all on its own. It reads them, identifies racial covenants based on patterns of language (which is exactly what LLMs are very good at), and then flags them for a human to review.
This tool is not replacing jobs, because the whole point is that these reviews were never going to get the budget and manpower to be done manually, and instead would have simply remained on the books.
I get being disdainful or even angry about LLMs in our unregulated-capitalism anti-worker hellhole because of the way that most companies are using them, but tools aren't themselves good or bad, they're just tools. And using a tool to identify racial covenants in legal documents that otherwise would go un-remediated, seems like a pretty good use to me.
So, what? They're going to pay a human to OK the output and the whole lot of them never even gets seen?
Say 12 minutes per covenant, that's 1 million work hours that humans could get paid for. Pay them $50 an hour and it's $50 million. That's nothing, less than 36 hours worth of the $12.5 Billion in weapons shipments we've sent to Israel in the last year. We could pay for projects like this with the rounding errors on the budget for blowing up foreign kids, and the people we pay to do it could afford to put their kids through college.
Instead, we get a project to train a robotic bigotry filter for real estate legalese and 50 more cruise missiles from the savings.
I think you are confused about the delineation between local and federal governments. It's not all one giant pool of tax money. None of Santa Clara County's budget goes to missiles.
Also, this feels like you are too capitalism-pilled, and rather than just spending the $240 to do this work, and using the remaining $49,999,760 to just fund free college or UBI programs, you're like, "how about we pay these people to do the most mind-numbingly, soul-crushingly boring work there is, reading old legal documents?"
You know what would actually happen if you did that? People would seriously read through them for 1 day, and then they'd be like, "clear", "clear", "clear" without looking at half of them. It's not like you're gonna find and fund another group to review the first group's work, after all. So you'd still be where we are now, but you also wasted x* peoples' time that they could have been enjoying doing literally anything else.
I am not, I simply don't believe the delineation is relevant since taxpayers fund both the state and federal budgets.
This is me being "reasonable" and working within the constraints of the system. If we aren't going to have free universal college et al then we can at least trade some of the bloated military budget for a public works program.
Sounds to me like a 50% improvement over zero human eyes.
Why not? We could hire three teams to do it simultaneously in every state in the country and the cost would still be a tiny fraction of how much was wasted on the F-35 program.
It certainly would be. Thankfully, there's many more than zero human eyes involved in this.
Then what's the LLM for?
Quickly filtering out a subset of them to prioritize so that we get the most value possible out of the time that humans spend on it.
One of LLMs main strengths over traditional text analysis tools is the ability to "understand" context.
They are bad at generating factual responses. They are amazing at analysing text.
LLMs neither understand nor analyze text. They are statistical models of the text they were trained on. A map of language.
And, like any map, they should not be confused for the territory they represent.
If you admit that they have issues with facts, why would you assume that the randomly generated facts their "analysis" produces must be accurate?
I mean they literally do analyze text. They're great at it. Give it some text and it will analyze it really well. I do it with code at work all the time.
Because they are two completely different tasks. Asking them to recall information from their training is a very bad use. Asking them to analyze information passed into them is what they are great at.
Give it a sample of code and it will very accurately analyse and explain it. Ask it to generate code and the results are wildly varied in accuracy.
I'm not assuming anything you can literally go and use one right now and see.
The person you're replying to is correct though. They do not understand, they do not analyse. They generate (roughly) the most statistically likely answer to your prompt, which may very well end up being text representing an accurate analysis. They might even be incredibly reliable at doing so. But this person is just pushing back against the idea of these models actually understanding or analysing. Its slightly pedantic, sure, but its important to distinguish in the world of machine intelligence.
I literally quoted the word for that exact reason. It just gets really tiring when you talk about AIs and someone always has to make this point. We all know they don't think or understand in the same way we do. No one gains anything by it being pointed out constantly.
You said "they literally do analyze text" when that is not, literally, what they do.
And no, we don't "all know" that. Lay persons have no way of knowing whether AI products currently in use have any capacity for genuine understanding and reasoning, other than the fact that the promotional material uses words like "understanding", "reasoning", "thought process", and people talking about it use the same words. The language we choose to use is important!
No it's not. It's pedantic and arguing semantics. It is essentially useless and a waste of everyone's time.
It applies a statistical model and returns an analysis.
I've never heard anyone argue when you say they used a computer to analyse it.
It's just the same AI bad bullshit and it's tiring in every single thread about them.
I never made any "AI bad" arguments (in fact, I said that they may be incredibly well suited to this) I just argued for the correct use of words and you hallucinated.
LLMs arent "bad" (ignoring, of course, the massive content theft necessary to train them), but they are being wildly misused.
"Analysis" is precisely one of those misuses. Grand Theft Autocomplete can't even count, ask it how many 'e's are in "elephant" and you'll get an answer anywhere from 1 to 3.
This is because they do not read or understand, they produce strings of tokens based on a statistical likelihood of what comes next. If prompted for an analysis they'll output something that looks like an analysis, but to determine whether it is accurate or not a human has to do the work.
LLMs cannot:
LLMs can
Semantics aside, they're very different skills that require different setups to accomplish. Just because counting is an easier task than analysing text for humans, doesn't mean it's the same it's the same for a LLM. You can't use that as evidence for its inability to do the "harder" tasks.
You forgot to put caveats on all the things you claim LLMs can do, but only one of them doesn't need them.
Why would you think that LLMs can do sentiment analysis when they have no concept of context or euphemism and are wholly incapable of distinguishing sarcasm from genuine sentiment?
Why would you think that their translations are of any use given the above?
https://www.inc.com/kit-eaton/mother-of-teen-who-died-by-suicide-sues-ai-startup/90994040
The human capacity for reason is greatly overrated. The overwhelming majority of conversation is regurgitated thought, which is exactly what LLMs are designed to do.
I don't really dispute that but at least we are able to apply formal analytical methods with repeatable outcomes. LLMs might (and do) achieve a similar result but they do so without any formal approach that can be reviewed, which has its drawbacks.
Did you see something that said it was an LLM?
Edit: Indeed it's an LLM. They published the model here: https://huggingface.co/reglab-rrc/mistral-rrc
Considering that it's a language task, LLMs exist, and the cost, it's a reasonable assumption. It'd be pretty silly to analyse a bag of words when you have tools you can use with minimal work with much better results. Even sillier to spend over $200 for something that can be run on a decade old machine in a few hours.