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

Understanding ChatGPT

atmosera.com

Post reply on HN
Print view View original

11–20 of 241 posts

Re: Understanding ChatGPT

#11

Post by Vespasian »

rain1 wrote: post #4
Bishonen88 wrote: post #2

This articles describes much of what many youtubers explained in their videos in the recent few weeks. While I understand the core concept of 'just' picking the next word based on statistics, it doesn't really explain how chatGPT can pull off the stuff it does. E.g. when one asks it to return a poem where each word starts with one letter/next alphabet letter/the ending of the last word, it obviously doesn't 'just' pi…

I think this talk goes into really good clear detail about how it all works https://www.youtube.com/watch?v=-QH8fRhqFHM but actually the best content that goes into a little bit more technical depth that I've found is this series by Hedu AI: https://www.youtube.com/watch?v=mMa2PmYJlCo&list=PL86uXYUJ79...

Andrej Karpathy made a video implementing and training a simple transformer and together with his other 4 videos on the topic it clicked for me.

https://youtu.be/kCc8FmEb1nY

Re: Understanding ChatGPT

#12

Post by carrolldunham »

Bishonen88 wrote: post #2

This articles describes much of what many youtubers explained in their videos in the recent few weeks. While I understand the core concept of 'just' picking the next word based on statistics, it doesn't really explain how chatGPT can pull off the stuff it does. E.g. when one asks it to return a poem where each word starts with one letter/next alphabet letter/the ending of the last word, it obviously doesn't 'just' pi…

I think you need to consider conditional statistics. "What are high probability options for the next word, given that the text i'm working on starts with the words please rhyme, and that the text 10 words ago was 'sun' and the text 20 words ago was 'fun'?" How it knows which parts of the text are relevant to condition on is the attention mechanism which is like "what is the probability this word is important to how to finish this sentence?". Both of these can be extracted from large enough example data

Re: Understanding ChatGPT

#13

Post by calf »

ChatGPT is a glorified word predictor. It isn’t sentient. It doesn’t know what it’s saying, and yes, you can coax it into admitting that it wants to take over the world or saying hurtful things (although it was specially conditioned during training to try to suppress such output). It’s simply stringing words together using an expansive statistical model built from billions of sentences.

Is this true though? The public debate albeit poorly explained by many, is whether the emergent behaviors users are seeing are caused by emergent algorithms and structures arising in the neural network. So for example some scientists claim that they can find fragments of syntax trees or grammars that the neural network emergently constructs. That would point to higher-level phenomena going on inside ChatGPT and its ilk, than merely statistics and predictions.

I'm curious as to the answer but it's not implausible to me that there's stuff happening on two levels of abstraction at the same time. Analogous to hardware/software abstraction, nobody says a Mac running Safari is a glorified Boolean circuit. I don't know the answer but it's not implausible, or maybe I don't know enough about machine learning to understand the author's quote above.

Re: Understanding ChatGPT

#15

Post by detrites »

Bishonen88 wrote: post #9
raincole wrote: post #7

Earlier quoted context omitted.

> it obviously doesn't Why?

simply because I think that it's rather statistically unlikely, that just because my first word started with "A", the next word should start with "B", "C" ...

My understanding is, if you have 175 billion parameters of 16-bit values that all effectively transact (eg, multiply) together, the realm of possibility is 175b^65536; really rather a large number of encodable potentials.

The length and number of probability chains that can be discovered in such a space is therefore sufficient for the level of complexity being analysed and effectively "encoded" from the source text data. Which is why it works.

Obviously, as the weights become fixed on particular values by the end of training, not all of those possibilities are required. But they are all in some sense "available" during training, and required and so utilised in that sense.

Think of it as expanding the corpus as water molecules into a large cloud of possible complexity, analysing to find the channels of condensation that will form drops, then compress it by encoding only the final droplet locations.

Re: Understanding ChatGPT

#16

Post by scotty79 »

Bishonen88 wrote: post #2

This articles describes much of what many youtubers explained in their videos in the recent few weeks. While I understand the core concept of 'just' picking the next word based on statistics, it doesn't really explain how chatGPT can pull off the stuff it does. E.g. when one asks it to return a poem where each word starts with one letter/next alphabet letter/the ending of the last word, it obviously doesn't 'just' pi…

Try this one:

https://www.youtube.com/watch?v=yGTUuEx3GkA

This series of video explains how the core mechanism works. There are few details omitted like how to get good initial token embedding or how exactly positional encoding works.

High level overview is that main insight of transformers is just figuring out how to partition huge basic neural network and hardcode some intuitively beneficial operations into the structure of the network iteself and draw some connections between (not very) distant layers so that gradient doesn't get eaten up too soon during backpropagation.

It all makes the whole thing parallelizable so you can train it on the huge amount of data despite it having enough neurons altogether to infer pretty complex associations.

Re: Understanding ChatGPT

#17

Post by aqme28 »

> ChatGPT is a glorified word predictor. It isn’t sentient. It doesn’t know what it’s saying, and yes, you can coax it into admitting that it wants to take over the world or saying hurtful things (although it was specially conditioned during training to try to suppress such output). It’s simply stringing words together using an expansive statistical model built from billions of sentences.

How do you differentiate it from the human mind? Do we understand ourselves well enough to say that we aren’t also just self-reflective reinforcement learners doing statistical inference on a library of all our “training data”?

Re: Understanding ChatGPT

#18

Post by missingdays »

calf wrote: post #13

ChatGPT is a glorified word predictor. It isn’t sentient. It doesn’t know what it’s saying, and yes, you can coax it into admitting that it wants to take over the world or saying hurtful things (although it was specially conditioned during training to try to suppress such output). It’s simply stringing words together using an expansive statistical model built from billions of sentences. Is this true though? The publi…

Of course Mac running Safari is a glorified boolean circuit. It's programmed in a way to show the pixels on the screen in a way that you understand, and take inputs from your mouse and keyboard in a way you understand

Re: Understanding ChatGPT

#19

Post by scotty79 »

calf wrote: post #13

ChatGPT is a glorified word predictor. It isn’t sentient. It doesn’t know what it’s saying, and yes, you can coax it into admitting that it wants to take over the world or saying hurtful things (although it was specially conditioned during training to try to suppress such output). It’s simply stringing words together using an expansive statistical model built from billions of sentences. Is this true though? The publi…

It's pretty much the truth. What the ChatGPT is good at is "keeping in mind" various associations between words that occurred in the session so far. To keep those associations some internal structure bound to get conjured. It doesn't mean the transformer understands anything or can do any kind of reasoning, despite the fact that it can mimic a bit how reasoning output looks like and even get it right sometimes if the context is fairly close to something it seen in the training data.

Re: Understanding ChatGPT

#20

Post by colinmorelli »

aqme28 wrote: post #17

> ChatGPT is a glorified word predictor. It isn’t sentient. It doesn’t know what it’s saying, and yes, you can coax it into admitting that it wants to take over the world or saying hurtful things (although it was specially conditioned during training to try to suppress such output). It’s simply stringing words together using an expansive statistical model built from billions of sentences. How do you differentiate it…

Somewhat related to this:

We seem to operate on the assumption that sentience is "better," but I'm not sure that's something we can demonstrate anyway.

At some point, given sufficient training data, it's entirely possible that a model which "doesn't know what it's saying" and is "stringing words together using an expansive statistical model" will outperform a human at the vast, vast majority of tasks we need. AI that is better at 95% of the work done today, but struggles at the 5% that perhaps does truly require "sentience" is still a terrifying new reality.

In fact, it's approximately how humans use animals today. We're really great at a lot of things, but dogs can certainly smell better than we can. Turns out, we don't need to have the best nose on the planet to be the dominant species here.

Post reply on HN