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

41–50 of 114 posts

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

#41

Post by olalonde »

judge2020 wrote: post #14
canjobear wrote: post #10

Earlier quoted context omitted.

> Gaining capabilities unpredictably is a bit exaggerated - it's more that a lot of apparent capabilities are embedded in language and these models approximate the truly vast amounts of text they've digested (imo). The fact is that they acquired abilities no one expected them to acquire. In hindsight you can say it was embedded in language and maybe you could have seen it coming, but it is an empirical fact that this…

It's still fairly exaggerated, because the LLM's purpose is basically "knowing human communication" and it still only exhibits that, but it knows how humans would respond if they obediently followed directions given to it exactly as instructed. And it was only unexpected to those that weren't following the news the GPT 3 paper was published in 2020 and detailed this wild advancement in its capabilities[0,1]. 0: https…

> It's still fairly exaggerated, because the LLM's purpose is basically "knowing human communication" and it still only exhibits that

That's wrong. It's been demonstrated that GPT4 has a model of physics, it knows how to play chess, arithmetics, it can interpret and run code, etc. (not perfectly, but to some degree). It goes way beyond basically "knowing human communication".

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

#42

Post by sanxiyn »

cjbprime wrote: post #24
DonaldPShimoda wrote: post #15

Earlier quoted context omitted.

> Seriously just watch. He's not actually going to be able to coherently define his "reasoning" in a way that can be tested. Google gives the following definition of the verb "reason": > think, understand, and form judgments by a process of logic. LLMs do not think, they do not understand, and they do not form judgments. They do not come to their own conclusions. They do not have the physical capability. They are sta…

New bar for people claiming LLMs can't reason: invent a specific, testable problem, representable in text, that many humans can solve and LLMs can't, and tell us what it is.

You know, François Chollet literally did this, and people don't listen. People should listen to Chollet more.

https://twitter.com/fchollet/status/1638643323748618240

https://arxiv.org/abs/1911.01547

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

#43

Post by Sharlin »

jawiggins wrote: post #35

> 6. Human performance on a task isn’t an upper bound on LLM performance Is that true? Because LLMs are trained on the work of human: a LLM that learned all of the content of the articles would at best yield output equal to that of the human writing. The reasons the author gives seem unfounded to me: > First, they are trained on far more data than any human sees If the human writer has access to google (everyone in m…

> If the human writer has access to google (everyone in modern society) this point is moot.

Does not follow at all. There’s a vast difference between being able to search for and internalize a tiny, tiny sample of all the available information, and literally having sampled and synthetized all of it.

> If it isn't, refer to #1.

Meaning what? LLMs use adversarial learning, the same thing that allowed AlphaGo to reach superhuman levels in Go.

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

#44

Post by RubyRidgeRandy »

jawiggins wrote: post #35

> 6. Human performance on a task isn’t an upper bound on LLM performance Is that true? Because LLMs are trained on the work of human: a LLM that learned all of the content of the articles would at best yield output equal to that of the human writing. The reasons the author gives seem unfounded to me: > First, they are trained on far more data than any human sees If the human writer has access to google (everyone in m…

I don't agree with you. Human knowledge itself is not bound only to the previous knowledge before it, so why wouldn't AI be any different? This is more of where AGI will shine, but I still think LLMs can produce novel innovation.

There's also a difference between having access to and being able to 'comprehend' all previous works in a given genre for generating music, that a human simply won't have the ability to process and recall when creating their own music and the AI could surpass. Humans simply having access to information is different than using information.

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

#45

Post by PopePompus »

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 what day of the year should be coldest on average, then plot everything. It did that easily. But the big surprise was this: just for fun, I told GPT-4 that one of the quantities it had calculated could be used to estimate the orbital period of the Earth, and I asked it which quantity it was. It correctly answered that the reciprocal of the frequency of the fitted sine wave is an estimate of the Earth's orbital period. All I had told it to do was pull out the temperature data for each day, and fit a sine wave. It knew that the reciprocal of the frequency was the Earth's orbital period, without me mentioning anything at all about the Earth when I was prompting it to write the program. It had calculated dozens of variable values on its way to producing the plots, and it knew the right one to pick, and it knew the answer was its reciprocal.

I thought that perhaps the program just recognized that the only value that could be used to calculate any period was that frequency, and perhaps the model didn't really know anything about the Earth motion through its orbit affecting the average temperature, so I asked:

"Can any of the values be used to estimate the Moon's orbital period?"

and it replied:

"No, the code and the data used in this program are not suitable for estimating the Moon's orbital period. The temperature data being analyzed is influenced by the Earth's rotation around the Sun and other factors like weather patterns, but it doesn't directly correlate with the Moon's orbital period."

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

#46

Post by olivermuty »

olalonde wrote: post #41
judge2020 wrote: post #14

Earlier quoted context omitted.

It's still fairly exaggerated, because the LLM's purpose is basically "knowing human communication" and it still only exhibits that, but it knows how humans would respond if they obediently followed directions given to it exactly as instructed. And it was only unexpected to those that weren't following the news the GPT 3 paper was published in 2020 and detailed this wild advancement in its capabilities[0,1]. 0: https…

> It's still fairly exaggerated, because the LLM's purpose is basically "knowing human communication" and it still only exhibits that That's wrong. It's been demonstrated that GPT4 has a model of physics, it knows how to play chess, arithmetics, it can interpret and run code, etc. (not perfectly, but to some degree). It goes way beyond basically "knowing human communication".

Have you tried gpt4? Because I have and this sounds like wild cherrypicking of lucky results.

«How to to play chess» -> devolve into illegal moves quickly?

So does that mean that its model of physics will let it levitate as well then?

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

#47

Post by sanxiyn »

kqr wrote: post #20
DonaldPShimoda wrote: post #15

Earlier quoted context omitted.

> Seriously just watch. He's not actually going to be able to coherently define his "reasoning" in a way that can be tested. Google gives the following definition of the verb "reason": > think, understand, and form judgments by a process of logic. LLMs do not think, they do not understand, and they do not form judgments. They do not come to their own conclusions. They do not have the physical capability. They are sta…

This is the problem with non-operational definitions, because now we need to know how you define "think" and "understand" and "form judgments", to move on. Instead, could you operationally define "reason" in a way that a human is, say, 90 % likely to pass the test and GPT is 10 % likely to do?

Yes, François Chollet released ARC(Abstraction and Reasoning Corpus) benchmark for this in 2019, and the benchmark can be scored automatically. Humans solve 100% of tests and GPTs solve 0% of tests and GPTs made exactly zero progress from 2019 to 2022.

https://twitter.com/fchollet/status/1631699463524986880

https://github.com/fchollet/ARC

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

#48

Post by cjbprime »

sanxiyn wrote: post #42
cjbprime wrote: post #24

Earlier quoted context omitted.

New bar for people claiming LLMs can't reason: invent a specific, testable problem, representable in text, that many humans can solve and LLMs can't, and tell us what it is.

You know, François Chollet literally did this, and people don't listen. People should listen to Chollet more. https://twitter.com/fchollet/status/1638643323748618240 https://arxiv.org/abs/1911.01547

Is ARC a benchmark that GPT-4 can be tested against today? I would be curious to see its results.

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

#49

Post by graypegg »

cjbprime wrote: post #36
graypegg wrote: post #29

Earlier quoted context omitted.

So I get that the evidence is that you can do insanely complex things with consistency, like playing a text adventure. You pick up a thing and it stays in your inventory etc, ergo, it’s modelled a character with an inventory in a text adventure. vvv But, coming at this as *respectfully and curiously* as I can here: ^^^ I guess I’m a little more skeptical about attributing magic to something that I know is only workin…

> something that I know is only working from words as a source. Your intuition about the architecture might be causing the distrust. Yes, the ends of the model are words in and words out (although GPT-4 is now multi-modal and can accept images as input too), but that doesn't mean the calculations in the middle are dealing in words. The "language" in "large language model" is misleading. It's a large generalized model…

That is totally possible, the explosion of stuff around this space has kind of blindsided me, so I don’t have much intuition for precisely how these models are built.

It all still feels rather “attribute magic to complexity” to say that there’s some human element skimmed off the patterns of our language, hidden in the coefficients somewhere.

I’m also not a magician, so really who am I to say :)

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

#50

Post by sanxiyn »

cjbprime wrote: post #48
sanxiyn wrote: post #42

Earlier quoted context omitted.

You know, François Chollet literally did this, and people don't listen. People should listen to Chollet more. https://twitter.com/fchollet/status/1638643323748618240 https://arxiv.org/abs/1911.01547

Is ARC a benchmark that GPT-4 can be tested against today? I would be curious to see its results.

Yes it is, and it was tested. GPT-4 can't solve any of the tests.
Post reply on HN