Ah, got it. Yeah, I think they have a lot more utility in an environment where you can trivially validate the correctness of the output without risk than in one where you only learn if they were right or not after the damage is already done.
Comments
Log in with your Bluesky account to leave a comment
These are essentially search-like functions, right? Instead of going to Google and pulling some particular, well-regarded website, you get a statistical aggregate of what such sites might have said.
I think a huge aspect people seems to be (purposely) ignoring is the understanding that the LLM produced results, especially the summaries are often FUSIONS of data from pages. If some of the data is wrong or an opinion, it's a delusion result.
What's bad is people not VERIFYING info LLMs spit out.
Yeah. But when it takes longer to verify an LLM's output than it does to just do the work yourself, it seems like the utility of the LLM is pretty low.
So basically you're using the LLM to steal clicks from the websites that actually have the info you want, like Wikipedia. That in fact is the primary purpose of AI, to capture information from the people who created it and deal it out as if they created it instead.
But that's a skillset issue, not a generalised criticism of the platform. Analogous to how I type faster than I think, but my boss uses a dictaphone and edits the output.
I'm not making a generalized criticism of the platform. I'm saying that if it takes longer to validate the LLM's output than it does to generate my own -- which is p much always, because I need to perform my own original research in order to validate that output -- the LLM's utility is limited.
I spend more time doing what we did before LLMs as after, just with the added time sink of scrolling past LLM summaries to find answers and the psych damage of knowing it's helping burn (more) the planet too.
Something it's *very* good at is avoiding the very-specialized type search syntaxes that e.g. https://hoogle.haskell.org has.
You can ask questions like "foul machine, give unto me a standard library function to transmute the lead of a series of `Option`s and yield the gold of an `Option<(T, U, V)>`"
It’s more than that for what Chris is saying. It has deep knowledge on Python. I have very limited coding knowledge, I can’t google something that will get me what I need unless it’s the exact same issue getting posted on a message board. It’s not just googling.
I have several reports I have to drop the same data into for different clients in their portal and in their preferred format and word length and they’re good for saving me a few hours of mucking around.
This is it precisely I think. There are terrible downsides to LLMs (environment, copyright, etc) but some people in some use cases do get utility from them. I’m in tech and we do see productivity gains in some areas of coding. This is absolutely not a blanket endorsement
Yeah. In my field, using something generated by AI is the rough equivalent of getting a CD with an executable on it in the mail and blindly installing it to prod.
I haven't seen this characterization before, Kathryn. But unlike it. And those of us of a certain age can understand each of those words, individually and the "you did WHAT?!" impact of them collectively
Which copilot are we talking about? Microsoft or GitHub? They are entirely different products with the same name and I’ve found that this causes a lot of people to talk at cross purposes.
imo this is consistent w/ the idea the majority of use cases currently being promoted to the broader public have, for most general users, limited utility, no utility, &/or greatly overrated utility, especially contrasted with their social, enviro & econ costs. but there are many specialized uses
Comments
What's bad is people not VERIFYING info LLMs spit out.
While what you cite is an actual problem, that's on top of the fact that it will just make up stuff whole cloth just to have an answer
Obviously that's not very useful if you're just using an LM like Claude or copilot direct from their websites. Those are basically novelties.
And for those of us with less than active social lives, it's not that hard to download and setup your own LM.
I spend more time doing what we did before LLMs as after, just with the added time sink of scrolling past LLM summaries to find answers and the psych damage of knowing it's helping burn (more) the planet too.
You can ask questions like "foul machine, give unto me a standard library function to transmute the lead of a series of `Option
It’s a fairly trivial use case.
But instead of me doing that codegen over a few minutes, it spits it out in seconds. Which adds up!
I'm guessing the github version is more tailored but less encompassing.