Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What is ChatGPT doing and why does it work?

writings.stephenwolfram.com

Post reply on HN
Print view View original

441–450 of 518 posts

Re: What is ChatGPT doing and why does it work?

#441

Post by kilgnad »

kqr wrote: post #401
kilgnad wrote: post #397

Earlier quoted context omitted.

Pfft more useless text written by a giant pile of atoms. Pointless. Kidding aside, in the current context of human communication etiquette, to even refer to something as atoms outside of a demonstration or example is a deliberate dismissive reduction.

I see the current context as something narrower, namely a discussion on the technical underpinnings of LLMs.

The output of LLMs implies a higher level structure. You can't solely discuss the lower level underpinning unless you deliberately do not acknowledge that the output of LLMs are implying something else.

Re: What is ChatGPT doing and why does it work?

#442

Post by sebzim4500 »

SilverBirch wrote: post #337
spion wrote: post #15

The answer to this is: "we don't really know as its a very complex function automatically discovered by means of slow gradient descent, and we're still finding out" Here are some of the fun things we've found out so far: - GPT style language models try to build a model of the world: https://arxiv.org/abs/2210.13382 - GPT style language models end up internally implementing a mini "neural network training algorithm" (…

Such a hacker news comment. The title isn't a question Stephen Wolfram is asking you a question, it's the title of an article he's written that answers the question.

What exactly do you think the point of this site is, if not to discuss the question raised in the article?

Re: What is ChatGPT doing and why does it work?

#443

Post by eru »

hinkley wrote: post #426
eru wrote: post #343

Earlier quoted context omitted.

I wonder how you see global warming having an impact at all here? Yes, Dennard scaling seems to be over, but Moore's law is still alive and kicking.

It really isn’t. I’m curious which pundits you’ve been listening to that are claiming Moore’s Law didn’t cap out back around 2015. We can only solve some of our problems with core count, and core count cares a great deal about Dennard’s Law, as well as Gustafson’s Law if not Amdahl’s. Data center energy usage is becoming a category of its own with regard to carbon footprint. And of course the power dissipation of a d…

Moore's law talks about the number of transistors in the chip that's cheapest per transistor. It's not talking about CPUs specifically.

So GPUs or even more 'exotic' beasts like TPUs count for Moore's law.

Moore's law doesn't say anything about how useful those transistors are. Nor does increasing core count somehow fall afoul of Moore's law.

Re: What is ChatGPT doing and why does it work?

#444

Post by sebzim4500 »

Salgat wrote: post #258
space_fountain wrote: post #242

Earlier quoted context omitted.

There are fascinating studies from people who have been blind through childhood and have their vision restored late enough that we can talk to them. For example ( https://pubmed.ncbi.nlm.nih.gov/28533387/ ). In particular it takes several months for these previously blind children to learn to distinguish faces from non faces. I recall a pop science article which I can't find the source for now that explained that peo…

I don't think you understand the scope of training data required for these models. We're talking thousands of lifetimes worth of reading for ChatGPT (GPT-3 for example is trained on 45TB of textual data).

I'm certain that humans are trained on far more than 45 TB of data, the vast majority of it is 'video' though.

Re: What is ChatGPT doing and why does it work?

#445

Post by TorKlingberg »

GuB-42 wrote: post #360
rcfox wrote: post #317

Earlier quoted context omitted.

What does "temperature" mean here though? Are you sure you didn't just ask it to generate two different responses?

That's the amount of randomness in the answers. Low temperature means it will take the most common path every time, at the risk of paraphrasing its sources. The "zero temperature" answer may very well been copied verbatim from a mainstream website. High temperatures means the system will get fed a lot of noise to create something original, at the risk of getting off rails or simply wrong.

Does asking it in the prompt to use different temperatures actually work? I would assume the temperature needs to be set with a side channel? I suspect this is just ChatGPT interpreting "temperature" as a regular word.

Re: What is ChatGPT doing and why does it work?

#446

Post by adammarples »

bottlepalm wrote: post #415
adammarples wrote: post #354

Earlier quoted context omitted.

Language is how people demonstrate that they have a good theory of mind. Solving these tasks is a proxy for people's internal theories. Language is a good way to demonstrate that ChatGPT has a good statistical model for what token comes next. Humans are not good at solving TOM tasks because they're good at guessing what token comes next, ChatGPT is not good at solving them because is has a good TOM. They're different…

What ‘statistical model’? ChatGPT isn’t a big Markov chain, it’s a neural net. It was trained on ‘full input’ transformed to ‘full output’. Generating the ‘next word’ is implementation details as it already knows the bigger idea it is trying to explain.

By a statistical model I mean a very large, highly dimensional data manifold on which classification is done to predict the next token. The correct token can be predicted very often because somehow the common patterns in the written training data are being efficiently compressed in this highly dimensional space. Humans are also able to efficiently compress the data they are trained on, but the structures inside a human and a transformer are completely different, what happens at inference time is completely different. They both have language in common as an output and can both answer TOM questions but this doesn't mean they both have a theory of mind.

Re: What is ChatGPT doing and why does it work?

#447

Post by int_19h »

janalsncm wrote: post #270
int_19h wrote: post #228

Earlier quoted context omitted.

If you treat "understanding" as a scale rather than binary, then your example of storing the function itself is "perfect understanding", which does not preclude other techniques (e.g. storing a much simpler function that approximates the original one) that could also be reasonably called "understanding", just imperfect. Directly mapping likely inputs and outputs is still a form of that, albeit obviously way too primi…

Actually understanding a function means that it can be computed to arbitrary precision. Memorizing inputs and outputs only allows you to interpolate between values you’ve already seen. That breaks down quickly for anything somewhat chaotic. You can try the following function in ChatGPT yourself: f(x) = x * (sin(sin(x)))^2 Ask it to give you the values of f for integers of x between -10 and 10. I tried it 5 times and…

It's not exactly a secret that GPT is bad at processing math in inputs, which seems to have something to do with the token representation being poorly suited for it, for starters. But I thought we were talking more generally about functions mapping inputs to outputs - the "perfect understanding" function is the one that captures all the relationships between the entities in the input and thus can always give a perfect answer to the extent that training data contains it, while an approximated "partial understanding" function internally has a much simpler model of those entities and their relationships (which can still turn out to be still good enough). Mapping inputs to outputs is an especially simple model, but it's still a model and an approximation.

And yes, in your example, of course it's simpler to just store the function itself - provided that you know in advance what it is, which, to remind, GPT does not. But when we're dealing with f(question)=answer of chatbots, the function that GPT ends up approximating is decidedly not simple.

Re: What is ChatGPT doing and why does it work?

#448

Post by bottlepalm »

adammarples wrote: post #446
bottlepalm wrote: post #415

Earlier quoted context omitted.

What ‘statistical model’? ChatGPT isn’t a big Markov chain, it’s a neural net. It was trained on ‘full input’ transformed to ‘full output’. Generating the ‘next word’ is implementation details as it already knows the bigger idea it is trying to explain.

By a statistical model I mean a very large, highly dimensional data manifold on which classification is done to predict the next token. The correct token can be predicted very often because somehow the common patterns in the written training data are being efficiently compressed in this highly dimensional space. Humans are also able to efficiently compress the data they are trained on, but the structures inside a hum…

Why do you think the wiring of a biological and artificial neural net are completely different? The results from both are remarkably similar.

Re: What is ChatGPT doing and why does it work?

#449

Post by famouswaffles »

britzkopf wrote: post #132
codeulike wrote: post #32

Wow this is 19,000 words. I like his summary at the end: At some level it’s a great example of the fundamental scientific fact that large numbers of simple computational elements can do remarkable and unexpected things. And this: ... But it’s amazing how human-like the results are. And as I’ve discussed, this suggests something that’s at least scientifically very important: that human language (and the patterns of th…

The thing I'm sort of confused about, but maybe someone can explain why I shouldn't be, is, why does there seem to be no implication for language translation? Or is there but coverage is overwhelmed by the fascination with chatGPT? In short, is machine language translation now a fully solved problem? A couple years ago when I tested Google translate in a non-esoteric conversation with my Russian speaking girlfriend a…

Bilingual LLMs are human level translators. I don't know why this frankly mindblowing fact isn't discussed or researched more but they are.

https://github.com/ogkalu2/Human-parity-on-machine-translati...

Re: What is ChatGPT doing and why does it work?

#450

Post by famouswaffles »

tkgally wrote: post #202
britzkopf wrote: post #132

Earlier quoted context omitted.

The thing I'm sort of confused about, but maybe someone can explain why I shouldn't be, is, why does there seem to be no implication for language translation? Or is there but coverage is overwhelmed by the fascination with chatGPT? In short, is machine language translation now a fully solved problem? A couple years ago when I tested Google translate in a non-esoteric conversation with my Russian speaking girlfriend a…

I worked as a translator for many years and have been following developments in machine translation closely. In my opinion, ChatGPT does represent a significant advance for machine translation. If you have the time to watch it, I made a video about the topic last week: https://youtu.be/najKN2bXqCo

Hey you might this. Bilingual LLMs really are human level translators. I don't know why this frankly mindblowing fact isn't discussed or researched more but they are.

https://github.com/ogkalu2/Human-parity-on-machine-translati...

Post reply on HN