4 ms·

What do you call something humans built (artificial) that can solve all topics of AP tests (general) with most being 5/5, better than almost all humans (intelli

What do you call something humans built (artificial) that can solve all topics of AP tests (general) with most being 5/5, better than almost all humans (intelligence)?

roarcher 3y ago
I call it an LLM, not AGI. Consider this scenario: - Ask ChatGPT to produce some output meeting certain criteria - ChatGPT responds, but the output does not meet the criteria in one or more obvious ways - Point this out and ask for a corrected version - ChatGPT says "I apologize, here's the corrected version:" and spits out a nearly identical answer with the same obvious mistake. - Repeat steps 3 and 4 ad nauseam. I experience this frequently when I ask it to do anything somewhat outside the norms of its training. It cannot correct itself or even see its own mistake. It cannot even conclude that it doesn't have the ability to do what I've asked and tell me so. Whatever sort of intelligence it may have, I would categorize it as not "general" enough to fall under AGI. Being able to answer AP test questions cannot be a sufficient measure of generality if a model can do it with flying colors but still fail at far more basic tasks.
- hyperadvanced 3y ago
  I’d say that LLM’s inability to handle DSL’s in particular is an interesting question on the topic of knowledge and language. Should a generally intelligent entity be able to quickly figure out a new dialect with little background? If I were to air-drop you into an extremely foreign nation, how long would it take you to organically decode the local language and be able to meaningfully craft expressions in it? I’m not sure if humans are “generally intelligent” but a common bar for AGI is for it to be able to beat or meet average humans at normal tasks. I’m not sure if something like writing perfect k8s yaml should be a requirement, though I do agree that LLM’s inability to do word puzzles or relatively straightforward math should disqualify them.
  - roarcher 3y ago
    I agree with your points in general, though I wasn't talking about DSL generation or any specific task. I was talking about ChatGPT's general tendency to cheerfully apologize for mistakes, explain exactly what the mistake was, and then present the same mistake while claiming that it's been corrected. You can ask ChatGPT what it means to be asked to correct a mistake, and it will give you a perfectly thorough and eloquent answer. But it is often unable to apply this concept to its own behavior. If asked, it can explain back to you exactly what correction you want it to make. You can even make it pledge to correct the mistake in exactly the manner that it just described. And then it will completely fail to do it. It reminds me of that "repeat after me" meme from Friends [1]. This makes me lean in the direction of "stochastic parrot" when I think about what LLMs are. As impressive as it is, ChatGPT demonstrably lacks a sense of self. It talks as if it understands that it's an agent in control of its behavior, but then fails to control or even recognize its own actions. [1] https://en.meming.world/images/en/c/c2/Friends-_Repeat_After_Me_origin_2.jpg https://en.meming.world/images/en/c/c2/Friends-_Repeat_After...
kmnc 3y ago
An advanced calculator.
giantrobot 3y ago
We don't give AP tests to humans to determine whether or not they are in fact an intelligent being. We give them to humans to show that over the course of about a decade and a half that human went from not realizing it had hands to ingesting a bunch of knowledge and had the capability to recall it answer questions. ChatGPT was trained on hundreds of terabytes of information. I would hope that with all that effort it could answer AP test questions. That doesn't make it intelligent.
- fragmede 3y ago
  what's your definition of intelligence that you're using here. there are various takes on what it means, so I'd like to know your take on it. That it's able to spit out essays makes it something, intelligence is too loosely defined to describe it for everybody, so we need to develop new words and meanings and new language for what it does. It doesn't think, but it does dot products to tensors to get create its output.
  - serf 3y ago
    intelligence in animal species, roughly, is how well the animal deals with novel situations. Corvid tool use is a common example of animal intelligence. It is interesting to consider AIs that way given that almost every AI will fail catastrophically when confronted with an area of expertise that is outside of their training corpus, although I guess Corvid tool use isn't exactly out of the animals' range of expertise.
otabdeveloper4 3y ago
Google search can solve AP tests too. (And before that library indexes did the same.) This isn't some new technology.
- ben_w 3y ago
  Google (and library indexes) can solve exactly and only the questions in their databases. All AI, even narrow ones, necessarily have the ability to interpret novel questions in some fashion — how novel and how well they interpret being the "G" and the "I" in AGI.
IncreasePosts 3y ago
Fairly boring - "The answer is known and I want to see if you know the answer" is good for judging students abilities, but no real world problem exists like that. I will be much more hopeful for AGI when AI starts producing new and interesting knowledge, rather than regurgitating already known facts.

What do you call something humans built (artificial) that can solve all topics of AP tests (general) with most being 5/5, better than almost all humans (intelligence)?

roarcher 3y ago

I call it an LLM, not AGI. Consider this scenario: - Ask ChatGPT to produce some output meeting certain criteria - ChatGPT responds, but the output does not meet the criteria in one or more obvious ways - Point this out and ask for a corrected version - ChatGPT says "I apologize, here's the corrected version:" and spits out a nearly identical answer with the same obvious mistake. - Repeat steps 3 and 4 ad nauseam. I experience this frequently when I ask it to do anything somewhat outside the norms of its training. It cannot correct itself or even see its own mistake. It cannot even conclude that it doesn't have the ability to do what I've asked and tell me so. Whatever sort of intelligence it may have, I would categorize it as not "general" enough to fall under AGI. Being able to answer AP test questions cannot be a sufficient measure of generality if a model can do it with flying colors but still fail at far more basic tasks.

hyperadvanced 3y ago
I’d say that LLM’s inability to handle DSL’s in particular is an interesting question on the topic of knowledge and language. Should a generally intelligent entity be able to quickly figure out a new dialect with little background? If I were to air-drop you into an extremely foreign nation, how long would it take you to organically decode the local language and be able to meaningfully craft expressions in it? I’m not sure if humans are “generally intelligent” but a common bar for AGI is for it to be able to beat or meet average humans at normal tasks. I’m not sure if something like writing perfect k8s yaml should be a requirement, though I do agree that LLM’s inability to do word puzzles or relatively straightforward math should disqualify them.
- roarcher 3y ago
  I agree with your points in general, though I wasn't talking about DSL generation or any specific task. I was talking about ChatGPT's general tendency to cheerfully apologize for mistakes, explain exactly what the mistake was, and then present the same mistake while claiming that it's been corrected. You can ask ChatGPT what it means to be asked to correct a mistake, and it will give you a perfectly thorough and eloquent answer. But it is often unable to apply this concept to its own behavior. If asked, it can explain back to you exactly what correction you want it to make. You can even make it pledge to correct the mistake in exactly the manner that it just described. And then it will completely fail to do it. It reminds me of that "repeat after me" meme from Friends [1]. This makes me lean in the direction of "stochastic parrot" when I think about what LLMs are. As impressive as it is, ChatGPT demonstrably lacks a sense of self. It talks as if it understands that it's an agent in control of its behavior, but then fails to control or even recognize its own actions. [1] https://en.meming.world/images/en/c/c2/Friends-_Repeat_After_Me_origin_2.jpg https://en.meming.world/images/en/c/c2/Friends-_Repeat_After...

kmnc 3y ago

An advanced calculator.

giantrobot 3y ago

We don't give AP tests to humans to determine whether or not they are in fact an intelligent being. We give them to humans to show that over the course of about a decade and a half that human went from not realizing it had hands to ingesting a bunch of knowledge and had the capability to recall it answer questions. ChatGPT was trained on hundreds of terabytes of information. I would hope that with all that effort it could answer AP test questions. That doesn't make it intelligent.

fragmede 3y ago
what's your definition of intelligence that you're using here. there are various takes on what it means, so I'd like to know your take on it. That it's able to spit out essays makes it something, intelligence is too loosely defined to describe it for everybody, so we need to develop new words and meanings and new language for what it does. It doesn't think, but it does dot products to tensors to get create its output.
- serf 3y ago
  intelligence in animal species, roughly, is how well the animal deals with novel situations. Corvid tool use is a common example of animal intelligence. It is interesting to consider AIs that way given that almost every AI will fail catastrophically when confronted with an area of expertise that is outside of their training corpus, although I guess Corvid tool use isn't exactly out of the animals' range of expertise.

otabdeveloper4 3y ago

Google search can solve AP tests too. (And before that library indexes did the same.) This isn't some new technology.

ben_w 3y ago
Google (and library indexes) can solve exactly and only the questions in their databases. All AI, even narrow ones, necessarily have the ability to interpret novel questions in some fashion — how novel and how well they interpret being the "G" and the "I" in AGI.

IncreasePosts 3y ago

Fairly boring - "The answer is known and I want to see if you know the answer" is good for judging students abilities, but no real world problem exists like that. I will be much more hopeful for AGI when AI starts producing new and interesting knowledge, rather than regurgitating already known facts.