LLMs don't understand any words.
yes. and you wouldn't believe¹ what's in the replies when you make this simple and obvious statement.
¹ who i am kidding. of course you know.
I both agree and disagree. I think of them as golems. They do understand how to respond, but that's as deep as it goes. It's simulated understanding, but a very very good simulation... Okay maybe I do agree.
I think that at best you could say that they understand the relationship between tokens. But even that requires a really generous definition of the word "understand".
There's a saying..."Knowledge is knowing a tomato is a fruit. Wisdom is knowing not to put it in fruit salad."
Meanwhile, LLMs are telling us to put glue on pizza so the cheese sticks. Even if the technology could eventually deliver on the promise, by the time we get there, nobody intelligent will trust it because the tech bros are, again, throwing half-baked garbage out into the world to try and be first to market.
I didn't trust it from the very moment of the announcement.
it's almost like this thing has no internal conceptual representation! I know this can't possibly be, millions of promptfans and prompfondlers have told me it can't be so, but it sure does look that way! wild!
It must have some internal models of some things, or else it wouldn't be possible to consistently make coherent and mostly reasonable statements. But the fact that it has a reasonable model of things like grammar and conversation doesn't imply that it has a good model of literally anything else, which is unlike a human for whom a basic set of cognitive skills is presumably transferable. Still, the success of LLMs in their actual language-modeling objective is a promising indication that it's feasible for a ML model to learn complex abstractions.
if I copy a coherent sentence into my clipboard, my clipboard becomes capable of consistently making coherent statements
Yes, but that's not how LLMs work. My statement depends heavily on the fact that a LLM like GPT is coaxed into coherence by unsupervised or semi-supervised training. That the training process works is the evidence of an internal model (of language/related concepts), not just the fact that something outputs coherent statements.
let me free up some of your time so you can go figure out how LLMs actually work
if I have a bot pick a random book and copy the first sentence into my clipboard, my clipboard becomes capable of consistently making coherent statements. unsupervised training 👍
this isn't necessarily true. patterns in data aren't by nature proof of an underlying system of logic. if you run the line-fitting machine on any kind of data, its going to output a line. considering just how much data is encoded into these transformers, i don't think we can conclusively say that it has a underlying conception of how language works, much less an understanding of the concepts that language represents. it could really just be using the vast quantities of data it has to output approximately correct statements. there's absolutely structure there, but it doesn't have to have the kind of structured understanding humans have about language to produce language, in the same way a less sophisticated machine learning model doesn't have to know what kind of data its fitting a line to to make a line.
it doesn't. that's why we're calling it “spicy autocompletion” .
It does, which is why it's autocompletion and not auto-gibberish.
It must have some internal models of some things, or else it wouldn’t be possible to consistently make coherent and mostly reasonable statements.
Talk about begging the question
Ha, I love the sauce on that headline.
It's not the headline used by the publication.
yes, this is the anti-HN
it seems like it's not the worst way to write text if I don't want to allow an ai to parse my messages...
not being not sure to fail to not write like this could become the opposite of interesting after a time that isn't long, though
Wow... It's not easy trying not to misunderstand sentences...
This article is over a year old and you all seem to be buying it as relevant to the current state of things. Can anyone reproduce the experiments/conversations where it fumbles with double negatives etc? I tried a couple examples with chatgpt and it seemed to handle them fine
we don’t care that your instance of a nondeterministic, unreliable system can’t replicate someone else’s results, and we don’t take marching orders from SSC readers.
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community