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

And yet It Understands

borretti.me

Post reply on HN
Print view View original

171–180 of 231 posts

Re: And yet It Understands

#171

Post by YeGoblynQueenne »

PoignardAzur wrote: post #169
YeGoblynQueenne wrote: post #151

Guys guys! Stop talking about LLMs a minute and look at this! I gave my phone's calculator app this very hard multiplication problem and it got it right! Look! 2398794857945873 * 10298509348503 = 2.47040112696963e+28 My calculator can do arithmetic! But only humans can do arithmetic! Therefore, my calculator must understand arithmetic! And I bet it always gets it right, too! That means it must understand arithmetic b…

Wow, being sarcastic about AI! I bet you're the first person to come up with that idea! Man, I wish we could all be so smart and clear-seeing.

To be clear, I'm not being sarcastic about AI. I am pointing out that there are many examples of systems that can perform calculations that humans can also perform because we understand calculation, but (those systems) do not understand calculation, or what they are calculating, or that they are calculating anything at all.

Yet, nobody is surprised that a calculator can perform arithmetic operations without understanding arithmetic, but here we are, with a whole bunch of people surprised that a language model can calculate the probability of a string in a natural language without understanding language.

Re: And yet It Understands

#172

Post by PoignardAzur »

skybrian wrote: post #134

Arguing over whether it “understands” or not is bad philosophy. It’s like there’s a magic show and you’re arguing over whether it’s “real magic” or whether there’s “some trick to it.” There are always tricks, but until you know what they are, the mystery is still there and you haven’t solved it. If God told you “yes it understands” or “no it doesn’t,” what would you have learned? The mystery would still be there. It’…

Well, you can make falsifiable prediction about whether an AI "understands" something at a deep or shallow level, though both these concepts and the predictions themselves will be a bit fuzzy.

As a concrete example, take the "wolf, goat and cabbage cross a river" puzzle. you can make several experiments which distinguish at which level an AI "understands" it.

- Can it solve the problem at all?

- Can it solve the problem if you translate it in a different language?

- Can it solve the problem if you switch the names of the characters around but maintain the framing of "a boat crossing the river"?

- Can it solve an equivalent problem with completely different wording where the solution is still logically equivalent?

A model that can do 1 but not 2-3 is probably just pattern matching a sequence of words; it doesn't "understand" the problem. A model that can do 1 and 2 but not 3 or 4 is still pattern-matching the problem, but it's matching abstract concepts (like "the concept of a wolf" instead of just the token "wolf"). A model that can do 3 but not 4 is probably pattern-matching the general-shape of the problem, as in "mutually-incompatible characters being transported on a boat". A model that can do 4 is the real deal.

(I think ChatGPT currently sits between 2 and 3)

Re: And yet It Understands

#173

Post by actuallyalys »

awfulneutral wrote: post #152
PoignardAzur wrote: post #64

Earlier quoted context omitted.

> Sydney is not supposed to give coherent 3-part messages using them, right? Right, that's the "what the actual fuck" part. This raises some very interesting questions about how Sydney generates its output and the input suggestion. Presumably the LLM is given a prompt like "First generate an answer to the previous text, then generate three input suggestions for the user"; also, the fact that Sydney "hides" the messag…

I'd be careful of anthropomorphizing this too much though. Yesterday I was experimenting with a ChatGPT (3.5) Twitch streamer that played a text adventure, that was supposed to return JSON like this: { "speak_out_loud": "Hey chat, what's up, etc", "game_cmd": "go north" } And it occasionally would put the "speak" part into the "game" part so the game would get long sentences that were supposed to be spoken out loud.…

Yeah, this seems like it could be the model not understanding that suggestions are meant to be separate or a bug in how suggestions returned from the model are parsed. I haven't used Bing's AI features so I don't know how reliable the suggestions normally are.

Re: And yet It Understands

#174

Post by Veedrac »

Minor correction to an otherwise valid article: AI does not pass the Turing Test, and what LaMDA did was not a Turing Test. Reading the original article by Turing is illustrative. This is not to say AI is not impressive in a measure that the Turing Test is meant to take a measure of.

Re: And yet It Understands

#175

Post by akiselev »

glenstein wrote: post #16

>I was a deep learning skeptic. I doubted that you could get to intelligence by matrix multiplication for the same reason you can’t get to the Moon by piling up chairs I've always been fascinated by this example. I've also heard it referred to as climbing a tree won't get you to the Moon. Because, for some reason, people think that's an argument against the possibility of getting to the Moon when it's actually a prof…

You can't use chairs to get to the moon not because of distance but because of orbital mechanics. That's why chairs work and colors don't: you think you're getting closer by stacking chairs when in reality, all you're going to do is fall down.

Until you grasp orbital mechanics, the concept of orbital velocity, and delta-V you can't possibly get to the moon. The distance might as well be infinite because the gravity well will pull you down.

It's also a completely invisible problem until you really reach altitude unless you're got a lot of theoretical background which we don't have on artificial intelligence. We don't even really know what the AI rockets will look like when compared with AI chairs.

Re: And yet It Understands

#176

Post by famouswaffles »

PoignardAzur wrote: post #172
skybrian wrote: post #134

Arguing over whether it “understands” or not is bad philosophy. It’s like there’s a magic show and you’re arguing over whether it’s “real magic” or whether there’s “some trick to it.” There are always tricks, but until you know what they are, the mystery is still there and you haven’t solved it. If God told you “yes it understands” or “no it doesn’t,” what would you have learned? The mystery would still be there. It’…

Well, you can make falsifiable prediction about whether an AI "understands" something at a deep or shallow level, though both these concepts and the predictions themselves will be a bit fuzzy. As a concrete example, take the "wolf, goat and cabbage cross a river" puzzle. you can make several experiments which distinguish at which level an AI "understands" it. - Can it solve the problem at all? - Can it solve the prob…

Gpt-4 can do all four. Just keep in mind that it has human like failure modes. It can give you an answer that is just applying common but false assumption reasoning steps. However if you rewrite the question to avoid biasing common priors, it gets it. And( Or at least with Bing), if you tell it it's making a wrong assumption somehow (not necessarily what the wedding assumption is), it gets it.

Re: And yet It Understands

#177

Post by im3w1l »

glenstein wrote: post #166
im3w1l wrote: post #128

Earlier quoted context omitted.

The quote isn't talking about physical distance. It's talking about "close" in the sense of progress. You haven't made any progress towards walking on the moon even if the pile of chairs is as big as Mount Everest. There isn't just the question of toppling either, the chairs would start physically crushing each other and then they are no longer chairs.

I think this misunderstands my point across the board. Physical distance is progress, and not only that, it's the most fundamental kind of progress for this type of problem. Chairs run up against practical limitations but not in-principle limitations, and people invoke the piling-up-chairs argument because they are confused about which kinds of arguments are about practicality and which are about illustrating a conce…

The moon is 30 earth diameters away. There isnt enough gravel on earth to build a pile that high. It's just not a working approach. If you are standing on a giant pile and want to get to the moon your best bet is walk all the way down again and start building a rocket.

Re: And yet It Understands

#178

Post by tjr »

marcosdumay wrote: post #141
kypro wrote: post #76

Earlier quoted context omitted.

I mean, isn't this the whole point of large + deep NNs? To model complex relationships in data? It's odd so many people seem to deny this with GPT and try to trivialise what it does by saying, "it just predicts the next word". This idea that GPT only works at the level of words and develops no deeper understanding of the concepts in language seems silly given its behaviour. And at the very least it's not what we obse…

> This isn't knowledge regurgitation. What makes you say that? Why do you think it's "reasoning" an answer, instead of looking up that people being grounded makes them frustrated?

Right, in this scenario I think it's more that. Who Daisy is (or if Daisy even exists) is irrelevant toward formulating a response.

Which is still impressive!

Re: And yet It Understands

#179

Post by Swenrekcah »

glenstein wrote: post #65
raincole wrote: post #42

Earlier quoted context omitted.

I don't know. I've been using Copilot, ChatGPT, and Bing Chat intensively in the past month. So far I still think the metaphor "you can’t get to the Moon by piling up chairs" aged very well. > A criticism that would make more sense would be something along the lines of "piling up colors you won't get you any closer to the Moon", since colors aren't even the right kind of thing, and you can't aggregate them in a way t…

>means that even chairs and rockets share some attributes (helping you get higher), the difference between them is still qualitative, not quantitve. I don't think so, because the needed 'quality' is the ability to traverse space. So I don't think I agree that the qualitative piece is missing. Perhaps the moon example is helpful here because the real solution, a rocket ship, uses propulsion rather than sheer mass, and…

>I don't think so, because the needed 'quality' is the ability to traverse space. So I don't think I agree that the qualitative piece is missing.

It's not the distance that keeps us from the moon but gravity. The needed quality to get to the moon is the ability to escape the gravity well, and chairs don't provide that.

These models are impressive but they aren't general intelligence, and I'm not sure this type of model ever will be.

If we look at it the other way around. These models have consumed more textual and graphical information than any human could ever do in a million lifetimes, and yet for most questions they don't really answer any better than the average human could. (They're much faster, so for the specific cases that they're right that's very useful).

So there is some key thing missing. I would say it is the ability to understand the text they are consuming and generating.

Re: And yet It Understands

#180

Post by wizofaus »

rhn_mk1 wrote: post #113
wizofaus wrote: post #75

Earlier quoted context omitted.

The question is whether an LLM can ever be trained to be good at maths. Currently LLMs can string together tokens that roughly equate to words in order to form sentences that carry particular semantic value. But how much change to the underlying technology would be required to give them the ability to string together digits to represent numbers and then numbers and operators in order to represent equations with parti…

To some limited extent, they are already good at maths: https://ar5iv.labs.arxiv.org/html/2201.02177 This paper makes me believe it's less about the lack of ability to understand maths, but the power to learn more of it.

Is that paper discussing LLMs though? That some sort of neural network can be trained to perform maths doesn't seem surprising and I would expect eventually it'll be a standard expectation they can do so even without being fed masses of training data (ideally no more than a human needs). It's the path from ChatGPT-4 to such a network I'm curious about.
Post reply on HN