To be clear I’m not expert. But I know a bit.
The way LLMs (like ChatGPT, GPT-4, etc) work, is that they continuously decide what the next best-sounding word might be, and they print it, over and over and over, until it makes sentences and paragraphs. And the way that next-word decision works under the hood, is with a deep neural net that was initially a theoretical tool designed to imitate the neural circuits that make up our biological nervous system and brain. The actual code for LLMs is rather small, it’s just about storing and managing representations of a neuron, and rearranging the connections between neurons as it learns more; just like the brain does.
I was listening to the first part of this “This American Life” episode this morning that covers it really well: https://podcasts.apple.com/us/podcast/this-american-life/id201671138?i=1000618286089 In it, Microsoft AI experts also express excitement and confusion about how GPT-4 seems to actually reason about things, rather than just bullshitting the next word to make it look like it reasons, like it’s supposed to be designed to do.
And so I was thinking: the reason why it works might be the other way around. It’s not that LLMs are smart enough to reason instead of bullshit, it’s that human’s reasoning actually works out of constantly bullshitting too, one word at a time. Imitate the human brain exactly, and I guess we shouldn’t be surprised that we land with a familiar-looking kind of intelligence - or lack thereof. Right?
You must log in or register to comment.
Please. If you’ve actually used it, you would know that ChatGPT writing won’t even pass high school English. It does not understand subtext or humor AT ALL.
It’s a tool, a very powerful tool, but a tool nonetheless.
deleted by creator
For a complete fool?
Who wants to build a pool
Hey, that’s pretty cool!
I am covered in mool :C
GPT-4, not ChatGPT. The episode I linked above mentioned that those researchers were unimpressed with LLMs actually reasoning because they had only studied ChatGPT; until GPT-4 was available for them to study, which did freakishly much better at reasoning tests, and they don’t know why. Don’t take it from me, take it from the experts!
(I’m not an expert, but I work in a nearby domain. Yes, I have used quite a few LLMs. I agree that ChatGPT is underwhelming.)
I can’t find the video but basically GPT can only predict the next word it wants to say, or something like that… so things like making a joke with a funny punchline is almost impossible because GPT doesn’t consider what the punchline will be until after the setup has be generated.
This inability to plan ahead results in failure of extremely simple tasks like, shortening a sentence to exactly 10 words
The original video I saw was doing something with lists I think. But yea this inability to think even 2 words ahead imbues some serious limitations.
I don’t know if your source is using ChatGPT or GPT-4, but the podcast I linked above is about AI researchers who knew very well about ChatGPT’s limitations and therefore were of the strong opinion that LLMs couldn’t reason, and then GPT-4 came out which changed their entire thinking about LLMs. They tried to come up with experiments that could reliably identify that some reasoning happened or not, and GPT-4 didn’t pass all of them, but it passed a bunch more than ChatGPT could, and a bunch more than it should. And they have no idea how exactly it would succeed at those, if the thing really can’t reason.
You’re correct that LLMs only guess the next best word, over and over. Which makes it even more mysterious why some LLMs pass those tests you just mentioned. They really shouldn’t, theoretically.
Unless… (and that’s my point) we also shouldn’t theoretically, but there is something about neural designs and bullshitting things one word at a time that makes those specific reasonings work, and that would explain why both LLMs and humans succeed at some of them.
My guess here is that LLMs of today are neural networks (transformer models) that primarily guess the next (and previous) words since they are literally trained directly for that. Guessing words are not linguistics, although as they are neural networks, splinchter skills are to be expected. GPT most likely learned how to understand language, after the words in question would make certain patterns and that would make a neural system specifically meant for that since our word order is related to meaning too, so understanding words would help predicting the next word better, kinda like how GPT-4 got a mind’s eye after most likely having to read descriptions constantly which may make visual patterns even though the AI was never trained on images (The visual GPT-4 had CLIP stitched on it, GPT-4 can do visual tasks without CLIP if described)- Despite most likely actually understanding language, GPT still prioritzes word guessing over language comprehension - https://www.youtube.com/watch?v=PAVeYUgknMw by “AI Explained” where you can see GPT-4 prioritizing syntax (relating to word order) more than meaning, despite the AI being well aware that they what they are defending is rather dumb- it is like their mind is telling them that THIS is the right answer despite them having arguments against it, but still defending anyways because if it MUST be the right answer, than there MUST be a reason to them. Again, guessing. - Also, GPT and many other AI are functioning on “steam of consiousness” meaning their thoughts are semi-conscious and they don’t think twice - GPT-4 with RLHF does better than most AI when it comes to this problem (we also sometimes have similar issue with words too, until we semi-awarely have to rethink about them too, I guess) — Also also GPT-4 is more like a huge “wernicke’s area”, it is not a complete “brain”. If you were to strip away one’s wernickes area from their brain, and connect it to an (g)old thinkpad via some messy dollar store aux, you would prob get stuff similar to what GPT-4 is “puking”
