LLMs can’t reason — they just crib reasoning-like steps from their training data
LLMs can’t reason — they just crib reasoning-like steps from their training data
Researchers at Apple have come out with a new paper showing that large language models can’t reason — they’re just pattern-matching machines. [arXiv, PDF] This shouldn’t be news to anyone here. We …
Did someone not know this like, pretty much from day one?
Not the idiot executives that blew all their budget on AI and made up for it with mass layoffs - the people interested in it. Was that not clear that there was no “reasoning” going on?
Well, two responses I have seen to the claim that LLMs are not reasoning are:
- we are all just stochastic parrots lmao
- maybe intelligence is an emergent ability that will show up eventually (disregard the inability to falsify this and the categorical nonsense that is our definition of "emergent").
So I think this research is useful as a response to these, although I think "fuck off, promptfondler" is pretty good too.
there’s a lot of people (especially here, but not only here) who have had the insight to see this being the case, but there’s also been a lot of boosters and promptfondlers (ie. people with a vested interest) putting out claims that their precious word vomit machines are actually thinking
so while this may confirm a known doubt, rigorous scientific testing (and disproving) of the claims is nonetheless a good thing
No they do not im afraid, hell I didnt even know that even ELIZA caused people to think it could reason (and this worried the creator) until a few years ago.
A lot of people still don't, from what I can gather from some of the comments on "AI" topics. Especially the ones that skew the other way with its "AI" hysteria is often an invite from people who know fuck all about how the tech works. "Nudifier" or otherwise generative images or explicit chats with bots that portray real or underage people being the most common topics that attract emotionally loaded but highly uninformed demands and outrage. Frankly, the whole "AI" topic in the media is so massively overblown on both fronts, but I guess it is good for traffic and nuance is dead anyway.
Indeed, although every one of us who have seen a tech hype train once or twice expected nothing less.
PDAs? Quantum computing. Touch screens. Siri. Cortana. Micropayments. Apps. Synergy of desktop and mobile.
From the outset this went from “hey that’s kind of neat” to quite possibly toppling some giants of tech in a flash. Now all we have to do is wait for the boards to give huge payouts to the pinheads that drove this shitwagon in here and we can get back to doing cool things without some imaginary fantasy stapled on to it at the explicit instruction of marketing and channel sales.
Isn’t OpenAI saying that o1 has reasoning as a specific selling point?
They say a lot of stuff.
My best guess is it generates several possible replies and then does some sort of token match to determine which one may potentially be the most accurate. Not sure if I'd call that "reasoning" but I guess it could potentially improve results in some cases. With OpenAI not being so open it is hard to tell though. They've been overpromising a lot already so it may as well be just complete bullshit.
Yes.
But the lies around them are so excessive that it's a lot easier for executives of a publicly traded company to make reasonable decisions if they have concrete support for it.
Seriously, I've seen 100x more headlines like this than people claiming LLMs can reason. Either they don't understand, or think we don't understand what "artificial" means.
We suspect this research is likely part of why Apple pulled out of the recent OpenAI funding round at the last minute.
Perhaps the AI bros “think” by guessing the next word and hoping it’s convincing. They certainly argue like it.
🔥
This has been said multiple times but I don't think it's possible to internalize because of how fucking bleak it is.
The VC/MBA class thinks all communication can be distilled into saying the precise string of words that triggers the stochastically desired response in the consumer. Conveying ideas or information is not the point. This is why ChatGPT seems like the holy grail to them, it effortlessly1 generates mountains of corporate slop that carry no actual meaning. It's all form and no substance, because those people -- their entire existence, the essence of their cursed dark souls -- has no substance.
1 batteries not included
The only difference between the average VC and the average Sovereign Citizen is income.
Perhaps the AI bros “think” by guessing the next word and hoping it’s convincing
Perhaps? Isn't that the definition of LLMs?
Edit: oh, i just realized it's not talking about the LLMs, but about their apologists
"sigh"
(Preface: I work in AI)
This isn't news. We've known this for many, many years. It's one of the reasons why many companies didn't bother using LLM's in the first place, that paired with the sheer amount of hallucinations you'll get that'll often utterly destroy a company's reputation (lol Google).
With that said, for commercial services that use LLM's, it's absolutely not true. The models won't reason, but many will have separate expert agents or API endpoints that it will be told to use to disambiguate or better understand what is being asked, what context is needed, etc.
It's kinda funny, because many AI bros rave about how LLM's are getting super powerful, when in reality the real improvements we're seeing is in smaller models that teach a LLM about things like Personas, where to seek expert opinion, what a user "might" mean if they misspell something or ask for something out of context, etc. The LLM's themselves are only slightly getting better, but the thing that preceded them is propping them up to make them better
IMO, LLM's are what they are, a good way to spit information out fast. They're an orchestration mechanism at best. When you think about them this way, every improvement we see tends to make a lot of sense. The article is kinda true, but not in the way they want it to be.
(Preface: I work in AI)
Are they a serious researcher in ML with insights into some of the most interesting and complicated intersections of computer science and analytical mathematics, or a promptfondler that earns 3x the former's salary for a nebulous AI startup that will never create anything of value to society? Read on to find out!
(Preface: I work in AI)
Preface: repent for your sins in sackcloth and ashes.
IMO, LLM’s are what they are, a good way to spit information out fast.
Buh bye now.
while true; do fortune; doneis a good way to spit information out fast.
Oh what a sweet, sweet tune to end a Sunday to
Arxiv paper link referenced in the article: https://arxiv.org/pdf/2410.05229
When you ask an LLM a reasoning question. You're not expecting it to think for you, you're expecting that it has crawled multiple people asking semantically the same question and getting semantically the same answer, from other people, that are now encoded in its vectors.
That's why you can ask it. because it encodes semantics.
because it encodes semantics.
if it really did so, performance wouldn't swing up or down when you change syntactic or symbolic elements of problems. the only information encoded is language-statistical
thank you for bravely rushing in and providing yet another counterexample to the “but nobody’s actually stupid enough to think they’re anything more than statistical language generators” talking point
so.... a stochastic parrot?
Rooting around for that Luke Skywalker "every single word in that sentence was wrong" GIF....
did you ask a LLM for a post to make here? that might explain this mess of a comment
because it encodes semantics.
Please enlighten me on how? I admit I don't know all the internals of the transformer model, but from what I know it encodes precisely only syntactical information, i.e. what next syntactical token is most likely to follow based on a syntactical context window.
How does it encode semantics? What is the semantics that it encodes? I doubt they have denatotational or operational semantics of natural language, I don't think something like that even exists, so it has to be some smaller model. Actually, it would be enlightening if you could tell me at least what the semantical domain here is, because I don't think there's any naturally obvious choice for that.