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

Ask HN: How does ChatGPT work?

news.ycombinator.com

Post reply on HN
Print view View original

11–20 of 185 posts

Re: Ask HN: How does ChatGPT work?

#11

Post by dtagames »

nullish_signal wrote: post #6
dtagames wrote: post #3

ChatGPT is really very simple. Imagine you could analyze a million books and identify all the words within them -- not the meanings of the words, just the actual letters they contain. Now, when someone asks you about the history of France (or why the sky is blue), you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! It's like a kid in the…

OP question mentions specific network types, while reply uses a vague "million-book library" metaphor, but at least mentions LLM? Also arbitrarily sorts human { teacher, journalist, writer } as "people who do custom work with each client" which makes a bit of sense for a Teacher in a small-class environment. But Writers and Journalists write at SCALE, and while each may speak with gusto to individuals just as well as…

I wouldn't deign to have ChatGPT write my comments. They're all me!

Obviously, an author's audience can include more than one person. But every good piece of human writing is tailored for some audience. That was my point. You've heard the expression, "Know your audience?" ChatGPT cannot.

Re: Ask HN: How does ChatGPT work?

#12

Post by phillipcarter »

dtagames wrote: post #8
phillipcarter wrote: post #4

Earlier quoted context omitted.

I don't think this is an accurate depiction and I think there's a lot more going on there than you're saying there is. The models also involve Codex, which is able to effectively do pattern matching with more context you give it so that it can emit novel code. I've seen this happen numerous times with Copilot, where it works very well the more open documents you have, especially if those documents are related. True,…

Yes, I use Copilot, too for that very reason. But we need to be very careful about words like "semantic" and "understanding" as the method is neither. We'll get the best use out of these technologies if we don't ascribe to them magical qualities they don't have. Poke around. You'll find out it just statistical math with tokens (letters and punctuation). No meaning is ascribed to anything.

I...did say that it doesn't have a full semantic understanding of your code. Copilot uses tree-sitter under the covers to be able to figure out several things that you can get by analyzing syntax alone, which is actually quite far. It lets you identify what kinds of declarations (e.g., a type) and what kind of expressions (e.g., pattern matching on that type) exist, provided there's a grammar for it, and then lets that influence the suggested code. This isn't perfect because indeed, you need a full language service to actually understand things, especially in resolving ambiguities in symbols (like shadowing in some languages), but, like I said...it's good enough.

Re: Ask HN: How does ChatGPT work?

#13

Post by dtagames »

krackers wrote: post #5
dtagames wrote: post #3

ChatGPT is really very simple. Imagine you could analyze a million books and identify all the words within them -- not the meanings of the words, just the actual letters they contain. Now, when someone asks you about the history of France (or why the sky is blue), you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! It's like a kid in the…

> you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! This is not sufficient to explain how LLMs are able to synthesize novel, coherent poems or song lyrics. What you're describing seems closer to a markov model. So far I have yet to see a good explanatin of _why_ transformer models seem to have this emergent behavior as you scale it up.…

Ah, but in fact it does. Code and poetry seem different to people but tokens are tokens to the computer, and it knows not one from the other.

The reason you get a poetry-style answer when you ask for one is that the mix of words and styles can be taken from different parts of the corpus. This is how you can have it write (bad) poetry about something for which only prose was scanned.

Re: Ask HN: How does ChatGPT work?

#15

Post by dtagames »

phillipcarter wrote: post #12
dtagames wrote: post #8

Earlier quoted context omitted.

Yes, I use Copilot, too for that very reason. But we need to be very careful about words like "semantic" and "understanding" as the method is neither. We'll get the best use out of these technologies if we don't ascribe to them magical qualities they don't have. Poke around. You'll find out it just statistical math with tokens (letters and punctuation). No meaning is ascribed to anything.

I...did say that it doesn't have a full semantic understanding of your code. Copilot uses tree-sitter under the covers to be able to figure out several things that you can get by analyzing syntax alone, which is actually quite far. It lets you identify what kinds of declarations (e.g., a type) and what kind of expressions (e.g., pattern matching on that type) exist, provided there's a grammar for it, and then lets th…

Yes, Copilot has the advantage that programming language syntax is highly constrained and regularized, allowing it to slot-in your own variable names and functions into new code as you write it. This is the best feature of Copilot. It's not part of ChatGPT because that's not meant for writing code. And I wouldn't call it "understanding." But, it is useful.

Re: Ask HN: How does ChatGPT work?

#16

Post by nullish_signal »

dtagames wrote: post #13
krackers wrote: post #5

Earlier quoted context omitted.

> you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! This is not sufficient to explain how LLMs are able to synthesize novel, coherent poems or song lyrics. What you're describing seems closer to a markov model. So far I have yet to see a good explanatin of _why_ transformer models seem to have this emergent behavior as you scale it up.…

Ah, but in fact it does. Code and poetry seem different to people but tokens are tokens to the computer, and it knows not one from the other. The reason you get a poetry-style answer when you ask for one is that the mix of words and styles can be taken from different parts of the corpus. This is how you can have it write (bad) poetry about something for which only prose was scanned.

I think that the levels of Associativity between Linguistic Terms goes beyond a mere "mix of words and styles". When pulling on one part of the Prompt causes portions of the Output to shift or vanish, or change genre or word choice or composition or focal point or focal distance - - -

then, that's the very opposite of "mix" -That's "Stable Diffusion" in text2img, or the "Reversal of Entropy" in Physics terms..

I'm still blown away that computers have become inage "Imaginers", decades after Humans first started doing 3D through Computers. It's really all quite unbelievable.

Re: Ask HN: How does ChatGPT work?

#17

Post by subroutine »

How does it know when to stop when asked for a description or summary? Sometimes it outputs a few sentences, sometimes a few paragraphs.

Does it know how much output it has already provided when deciding on the next token? How does it decide to start a new sentence or paragraph, or if it's 'satisfied' with its current response?

Re: Ask HN: How does ChatGPT work?

#18

Post by akelly »

The way they went from GPT-3 to ChatGPT is really quite genius. My understanding is that it's something like this:

1. Start with GPT-3, which predicts the next word in some text and is trained on all the text on the internet

2. Take thousands of prompts, generate several responses for each of them, and have human reviewers rank the responses for each prompt from best to worst

3. The GPT model needs a massive amount of training data, it would be cost prohibitive to get enough human feedback to fine tune GPT manually. So you train another model, called the reward model, to predict how the humans will rate each response. Then you train the GPT model against the reward model millions of times

5. Feed a small percentage of the output from that training process back to the human reviewers to continue training the reward model, based on heuristics like reward model uncertainty which predict how helpful the human feedback will be towards improving the reward model

6. Release ChatGPT to the public, and use user feedback like response upvotes/downvotes to further optimize the reward model, while continuing to train ChatGPT against the reward model

https://openai.com/blog/chatgpt/

https://openai.com/blog/deep-reinforcement-learning-from-hum...

Re: Ask HN: How does ChatGPT work?

#19

Post by maxbond »

dtagames wrote: post #13
krackers wrote: post #5

Earlier quoted context omitted.

> you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! This is not sufficient to explain how LLMs are able to synthesize novel, coherent poems or song lyrics. What you're describing seems closer to a markov model. So far I have yet to see a good explanatin of _why_ transformer models seem to have this emergent behavior as you scale it up.…

Ah, but in fact it does. Code and poetry seem different to people but tokens are tokens to the computer, and it knows not one from the other. The reason you get a poetry-style answer when you ask for one is that the mix of words and styles can be taken from different parts of the corpus. This is how you can have it write (bad) poetry about something for which only prose was scanned.

I'd propose that you come up with a prompt to test your hypothesis and try it out. I think seeing is believing here, I don't know how ChatGPT works but after experimenting with it I am sure it's not simply looking for runs of similar words to the prompt in it's training set.

Here's an example to consider: https://news.ycombinator.com/item?id=33874110

Re: Ask HN: How does ChatGPT work?

#20

Post by throwaway1851 »

krackers wrote: post #5
dtagames wrote: post #3

ChatGPT is really very simple. Imagine you could analyze a million books and identify all the words within them -- not the meanings of the words, just the actual letters they contain. Now, when someone asks you about the history of France (or why the sky is blue), you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! It's like a kid in the…

> you could simply pluck out of your library the most common strings of word that seem to follow the words that were in your question! This is not sufficient to explain how LLMs are able to synthesize novel, coherent poems or song lyrics. What you're describing seems closer to a markov model. So far I have yet to see a good explanatin of _why_ transformer models seem to have this emergent behavior as you scale it up.…

I don't really see how it's different from, say, a large convolutional neural network learning progressively higher-order features as you progress through the layers of the network. At the lowest layers it's learning simple edge filters, which get combined into shapes, which get combined into filters that activate on faces, which get combined in ways that can be recognized as "family portrait", and so on. Of course, transformers have some unique advantages in terms of having very large context windows, being very parallelizable, etc.

When ChatGPT or any generative model produces output from a prompt, it's sampling from the (frozen) statistical structure it has learned. It makes sense that as you increase model capacity and the volume of training data, you can capture statistical patterns that occur at a very high level of abstraction. So, instead of just predicting the next token based on token-to-token patterns, it can predict the next token based on something resembling concept-to-concept patterns. (This somewhat demystifies the poetry generation / style transfer stuff that it can do. If you ask for a breach of contract complaint written as a sonnet, it can sample from both the patterns it's learned from legal documents and the patterns it's learned from poetry.)

What I wonder is how far this track of scaling up can take us. At the end of the day, interacting with ChatGPT is not that different from "interacting" with y=3x+7 by plugging in a value for x. It's just a much, much larger function.

Post reply on HN