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

I asked GPT-NeoX-20B a hundred arithmetic questions

twitter.com

Post reply on HN
Print view View original

131–140 of 235 posts

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#131

Post by gamegoblin »

throwanem wrote: post #97
schleck8 wrote: post #93

Earlier quoted context omitted.

I tried this with the new open source GPT NeoX 20B model on GooseAI playground and it suceeded too: Q: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? A: Yes. Q: What was the word that I put in between? A: The word was "here." It's capable of other languages as well, although not quite as fluent. Q: Kannst du auch Deutsch sprechen?…

This makes me wonder about a variant of the Turing test that I don't think I've seen explored. How would a conversational AI prove beyond reasonable doubt that it is an AI?

Sentences that embed even simple numerical things such as

“What was the 3rd letter of the 46th word in this conversation?”

Would take a human many seconds at minimum, but could be answered instantly by an AI.

A more interesting question may be to constrain the test to something akin to postal correspondence, where there is a significant delay between messages.

I think it could still be solved, though. Arbitrarily complicated numerical tasks can be conceived.

For instance, the AI sends correspondence:

“I have demonstrated proof of my identity by rewriting the children’s story ‘Green Eggs and Ham’ such that it still rhymes and retains the original plot, but every sentence has an md5 digest that ends with the byte 0x42”

Composing such a text would take much, much longer for a human, whereas an AI could just brute force through all the possibilities until it finds one that works.

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#132

Post by iggldiggl »

schleck8 wrote: post #93
tehsauce wrote: post #84

The latest version of gpt-3 spooked me the other day with: Me: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? Gpt: Yes, I can understand what you are asking if you put words in between. Me: What was the word that I put in between? Gpt: The word that you put in between is "here."

I tried this with the new open source GPT NeoX 20B model on GooseAI playground and it suceeded too: Q: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? A: Yes. Q: What was the word that I put in between? A: The word was "here." It's capable of other languages as well, although not quite as fluent. Q: Kannst du auch Deutsch sprechen?…

> It's capable of other languages as well, although not quite as fluent.

Haven't had a chance to play around with this one yet, but with the smaller GPT-J model, there's a clearly noticeable difference:

In English it'll happily generate reams of text that are – at least internally – quite coherent. Any absurdity and humour mostly only comes in because the text as a whole might only have a loose connection with reality as we know it.

In German on the other hand, it comparatively much more often produces gibberish at the individual sentence level, makes up nonsense words (although they are at least indeed German-sounding), etc. Somewhat interestingly it doesn't do too bad in terms of grammar and especially orthography, though, it just often fails to turn it into a fully coherent sentence.

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#133

Post by syntaxing »

schleck8 wrote: post #93
tehsauce wrote: post #84

The latest version of gpt-3 spooked me the other day with: Me: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? Gpt: Yes, I can understand what you are asking if you put words in between. Me: What was the word that I put in between? Gpt: The word that you put in between is "here."

I tried this with the new open source GPT NeoX 20B model on GooseAI playground and it suceeded too: Q: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? A: Yes. Q: What was the word that I put in between? A: The word was "here." It's capable of other languages as well, although not quite as fluent. Q: Kannst du auch Deutsch sprechen?…

Did you try this on GooseAI? I was not able to replicate this

Edit: oops just noticed you mentioned GooseAI, what settings did you use?

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#135

Post by Samin100 »

rexreed wrote: post #125

Amazing how a very costly to train system using billions of neural nodes on millions of dollars of compute performs more poorly than an 8-bit 1970s pocket calculator. Not sure why people are expecting some sort of "intelligence" to emerge from a text generator model trained on Internet corpus data. GPT-3 doesn't calculate, it pattern matches. I do get why people might be surprised, on the other hand, that it actually…

According to Kahneman, when a chess pro makes an “intuitive” and unexplainable move it’s just pattern recognition happening subconsciously.

If language models like GPT-3 are “just” pattern recognizers, wouldn’t that makes them capable of intuition?

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#136

Post by MauranKilom »

mordymoop wrote: post #65

When you toss “2241 + 19873 =” into an applet that shows you the default tokenization scheme GPT-3 uses, you get this: (224)(1)( +)( 198)(73)( =) I’ve heard it remarked before that, while tokenization is obviously an unavoidable part of a model with an architecture like GPT, this is a very silly way of tokenizing number strings for the purposes of learning or doing arithmetic. Indeed, I think a lot of GPT-3’s puzzlin…

Gwern noted that adding things like thousands separators and $ signs to the input makes GPT significantly better at math.

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#137

Post by MaxMoney »

Samin100 wrote: post #135
rexreed wrote: post #125

Amazing how a very costly to train system using billions of neural nodes on millions of dollars of compute performs more poorly than an 8-bit 1970s pocket calculator. Not sure why people are expecting some sort of "intelligence" to emerge from a text generator model trained on Internet corpus data. GPT-3 doesn't calculate, it pattern matches. I do get why people might be surprised, on the other hand, that it actually…

According to Kahneman, when a chess pro makes an “intuitive” and unexplainable move it’s just pattern recognition happening subconsciously. If language models like GPT-3 are “just” pattern recognizers, wouldn’t that makes them capable of intuition?

No post body was provided.

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#138

Post by baalimago »

mordymoop wrote: post #65

When you toss “2241 + 19873 =” into an applet that shows you the default tokenization scheme GPT-3 uses, you get this: (224)(1)( +)( 198)(73)( =) I’ve heard it remarked before that, while tokenization is obviously an unavoidable part of a model with an architecture like GPT, this is a very silly way of tokenizing number strings for the purposes of learning or doing arithmetic. Indeed, I think a lot of GPT-3’s puzzlin…

Many words have several semantic definitions depending on definition. This is why the word "is" is a very good token to have in a vocabulary (as an example), since it can mean so much depending on what tokens came before and after it.

Numbers have very limited semantic value. "123816" only means that number, and it's used very rarely in comparison to basically any other word (and the higher the number, the less chance of use, statistically peaking).

So the question becomes; to what extent do you expand the vocabulary using only numbers? "1", "2", "3", ... "1000000" would probably be a huge waste of words in an AI vocabulary (1MB input nodes), yet still not very impressive arithmetically even with 100% calculationrate. In comparison, a hand calculator from 30 years ago could do this with ease. It's not a question of being able to cleverly tokenize.

Calculations like this is an inherent flaw of vocabulary based AI until the semantic meaning of number sequences are somehow taught to it. Basically it needs to understand that "12" and "1" + "2" has the same contextular meaning, something which very rarely is explained in anything but 7 year old's schoolbooks. The problem is the dataset.

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#139

Post by andreyk »

tehsauce wrote: post #84

The latest version of gpt-3 spooked me the other day with: Me: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between? Gpt: Yes, I can understand what you are asking if you put words in between. Me: What was the word that I put in between? Gpt: The word that you put in between is "here."

Might actually not be a huge feat given the transformer architecture and it's attention component - here was not that related to the rest of the sentence. Still, quite cool...

Re: I asked GPT-NeoX-20B a hundred arithmetic questions

#140

Post by andreyk »

rexreed wrote: post #125

Amazing how a very costly to train system using billions of neural nodes on millions of dollars of compute performs more poorly than an 8-bit 1970s pocket calculator. Not sure why people are expecting some sort of "intelligence" to emerge from a text generator model trained on Internet corpus data. GPT-3 doesn't calculate, it pattern matches. I do get why people might be surprised, on the other hand, that it actually…

What's your definition of 'intelligence'? Many of the things GPT-3 does clearly exhibit intelligence (just not human level intelligence).
Post reply on HN