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

Eight things to know about large language models [pdf]

cims.nyu.edu

Post reply on HN
Print view View original

61–70 of 114 posts

Re: Eight things to know about large language models [pdf]

#61

Post by bigcat12345678 »

Kiro wrote: post #58
somenameforme wrote: post #57

Earlier quoted context omitted.

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

I don't get it. What is the correct answer?

In 2 comments we are showing gpts are not worse than human

Re: Eight things to know about large language models [pdf]

#62

Post by PopePompus »

somenameforme wrote: post #57
PopePompus wrote: post #45

I have a little weather station mounted outside my house, and there is a small linux machine (Raspberry Pi) logging the weather data into an SQLite database. I have almost two years of data now, so I asked GPT-4 to write a Python program that would query the database, calculate the average temperature for each day, fit a sine wave to those data points, find the first minimum of the sine wave and use that to calculate…

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

> It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words.

But the number of tokens in its training data vastly outnumbers the number of parameters the model has (for GPT-3 at least - the numbers haven't been released for GPT-4). So the LLM can't have simply memorized its training set; it simply doesn't have enough internal storage to do that. So it is not just regurgitating portions of its training set.

Re: Eight things to know about large language models [pdf]

#63

Post by mpliax »

Kiro wrote: post #58
somenameforme wrote: post #57

Earlier quoted context omitted.

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

I don't get it. What is the correct answer?

6, heaven,8, fine? Every other number is replaced with a similarly-sounding word.

Re: Eight things to know about large language models [pdf]

#64

Post by jayd16 »

Kiro wrote: post #58
somenameforme wrote: post #57

Earlier quoted context omitted.

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

I don't get it. What is the correct answer?

Odd numbers are replaced with sound alikes. One/done, Three/tree, five/hive

Re: Eight things to know about large language models [pdf]

#65

Post by jay_kyburz »

i_cannot_hack wrote: post #59
Kiro wrote: post #58

Earlier quoted context omitted.

I don't get it. What is the correct answer?

Every even item is the corresponding digit, every odd item is a word that rhymes with the corresponding digit. A continuation could be "6, heaven, 8"

I would be very interested to know if you could explain the pattern to GTP4 and then ask it again.

given a sequence of numbers 0 to 10, replace every second number with a word the rhymes with the number.

Re: Eight things to know about large language models [pdf]

#66

Post by amoss »

PopePompus wrote: post #62
somenameforme wrote: post #57

Earlier quoted context omitted.

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

> It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. But the number of tokens in its training data vastly outnumbers the number of parameters the model has (for GPT-3 at least - the numbers…

The number of tokens in the training set is not a good measure of the information content of the training set - if it was then the training would not work as the training set would be statistically random.

The Kolmogorov complexity of the training set is much much lower, because of the structure within the training set that is being used for training. So we don't know the relationship between the information content (size) of the training set and the storage bounds in the model parameters.

Re: Eight things to know about large language models [pdf]

#67

Post by OmegaPoint »

somenameforme wrote: post #57
PopePompus wrote: post #45

I have a little weather station mounted outside my house, and there is a small linux machine (Raspberry Pi) logging the weather data into an SQLite database. I have almost two years of data now, so I asked GPT-4 to write a Python program that would query the database, calculate the average temperature for each day, fit a sine wave to those data points, find the first minimum of the sine wave and use that to calculate…

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

I asked my coworkers in the office but none was able to answer that. Not sure because they were are ESL (me included) or because they were GPT in disguise.

Re: Eight things to know about large language models [pdf]

#68

Post by somenameforme »

PopePompus wrote: post #62
somenameforme wrote: post #57

Earlier quoted context omitted.

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

> It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. But the number of tokens in its training data vastly outnumbers the number of parameters the model has (for GPT-3 at least - the numbers…

It's not about directly regurgitating the training set, but about creating associations that allow it to indirectly regurgitate, and more critically - mix, its training set. It's easier to understand this in a different domain. Imagine we want to make a letter parsing algorithm.

The first step is to simplify the input and create a consistent format. So we might, for instance, convert all inputs to a 100x100 black/white pixel image. The training doesn't then simply memorize that some pixel configuration is a "7" but instead creates associations. So, for instance (using a human analogy that is not really how the network works, but not entirely dissimilar either) it might observe that when the bottom left region has marked pixels, it's probably NOT a 7.

Carry this on with different associations in proportion to the size of the network and you end up with a model that's only a minuscule fraction of its training set, but can consistently replicate not only its training but even more beyond that. Now change those 10,000 input pixels to words instead of pixels, with the target output being a word instead of a letter, and you get ChatGPT. It's the whole magic of neural networks, and really neat. But I think also paired with not really justifiable hyperbole.

Re: Eight things to know about large language models [pdf]

#69

Post by thunderbong »

DonaldPShimoda wrote: post #3

This is a personal correspondence typeset via LaTeX — it is not an academic paper, and it was not peer-reviewed. (The document does not claim otherwise, but I think it's common for people to assume that documents that have been typeset in such a format are more rigorous than this is.) Leaving that aside, I really take issue with the style used by the author. For example, section 3 begins: > There is increasingly subs…

It's been a long time since anyone reminded me of Richard Bach's 'illusions'

Thank you for a pleasant break!

https://www.patheos.com/blogs/wakeupcall/2020/05/10-thought-...

Re: Eight things to know about large language models [pdf]

#70

Post by SanderNL »

somenameforme wrote: post #57
PopePompus wrote: post #45

I have a little weather station mounted outside my house, and there is a small linux machine (Raspberry Pi) logging the weather data into an SQLite database. I have almost two years of data now, so I asked GPT-4 to write a Python program that would query the database, calculate the average temperature for each day, fit a sine wave to those data points, find the first minimum of the sine wave and use that to calculate…

It's a word prediction algorithm. Literally any collection of words (sometimes also known as a fact) it was trained on, regardless of how esoteric or domain specific, should generally be able to be regurgitated and, to a lesser degree, associated with similar collections of words. If you want to see it fail, don't try to "outsmart" it, but simply consider how it's programmed. --- Me: "continue the sequence: 0, done,…

I apologize for being enormously stupid, but the word is a rhyme with the underlying number right?

Edit: I want to add I even get GPT3 getting it right without explanation to be honest. Even that one is outsmarting me. 4 is using terms like homophone. I'm not sure you proved what you thought you proved.

Post reply on HN