4 ms·

>and even if it were, there is no evidence ChatGPT is a route to even human-level intelligence. People who say this nonsense need to start properly defining hu

>and even if it were, there is no evidence ChatGPT is a route to even human-level intelligence. People who say this nonsense need to start properly defining human level intelligence because nearly anything you throw at GPT-4 it performs at at least average human level, often well above. Give criteria that 4 fails that a significant chunk of the human population doesn't also fail and we can talk. Else this is just another instance of people struggling to see what's right in front of them. Just blows my mind the lengths some will go to ignore what is already easily verifiable right now. "I'll know agi when i see it", my ass.

AnthonyMouse 3y ago
> People who say this nonsense need to start properly defining human level intelligence because nearly anything you throw at GPT-4 it performs at at least average human level, often well above. "Average human level" is pretty boring though. Computers have been doing arithmetic at well above "average human level" since they were first invented. The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. Which is clearly still not the case.
- toomuchtodo 3y ago
  > The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. Which is clearly still not the case. I imagine an important concern is the learning & improvement velocity. Humans get old, tired, etc. GPUs do not. It isn't the case now, but it is fuzzy how fast we could collectively get there. Break out problem domains into modules, off to the silicon dojos until your models exceed human capabilities, and then roll them up. You can pick from OpenGPT plugins, why wouldn't an LLM hypervisor/orchestrator do the same? https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html https://waitbutwhy.com/2015/01/artificial-intelligence-revol... https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-2.html https://waitbutwhy.com/2015/01/artificial-intelligence-revol...
  - dragonwriter 3y ago
    > The concern is the learning & improvement velocity. Humans get old, tired, etc. GPUs do not. They do, though. Of course, replacing the worn out hardware while keeping the software is easier with GPUs.
- famouswaffles 3y ago
  >"Average human level" is pretty boring though. Lol ok. Still human level. and GPT-4 is way above average in most tasks. >Computers have been doing arithmetic at well above "average human level" since they were first invented. Cool. That's what the general in agi is about. GPT-4 is very general. >The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. as well as what kind of people ? experts ? That was not the premise of agi when the term was coined or for a long time afterwards. Posts have shifted(as they often do in this field) so that that's what the term seems to mean now but agi was artificial and generally intelligent, which has been passed. There's no difference between your definition of agi which is supposed to surpass experts in every field and super intelligence.
  - AnthonyMouse 3y ago
    > Lol ok. Still human level. and GPT-4 is way above average in most tasks. It has access to a lot of information that most humans don't have memorized. It's a better search engine than most humans. And it can format that information into natural language. But can it drive a car? If given an incentive to not confabulate and the knowledge that its statements are being verified, can it achieve that as consistently as the median human? If you start by giving it a simple instruction with stark consequences for not following it, can it continue to register the importance of that instruction even after you give it a lot more text to read? > as well as what kind of people ? experts ? Experts are just ordinary people with specific information. You're giving the specific information to the AI, aren't you? It's in the training data. > There's no difference between your definition of agi which is supposed to surpass experts in every field and super intelligence. That's because there is no difference between them. Super intelligence is achievable just by making general intelligence faster. If you have AGI and can make it go faster by throwing more compute hardware at it then you have super intelligence.
    - famouswaffles 3y ago
      >It has access to a lot of information that most humans don't have memorized. It's not just about knowledge. Lots of papers showing strong reasoning across various reasoning types. Couple papers demonstrating the development of world models too. >It's a better search engine than most humans. And it can format that information into natural language. Not how this works. They aren't search engines. and their performance equity with people isn't relegated to knowledge tasks alone. >But can it drive a car? If given an incentive to not confabulate and the knowledge that its statements are being verified, can it achieve that as consistently as the median human? Can a blind man drive a car ? a man with no hands ? >If you start by giving it a simple instruction with stark consequences for not following it, can it continue to register the importance of that instruction even after you give it a lot more text to read? Lol yes >Experts are just ordinary people with specific information. You're giving the specific information to the AI, aren't you? It's in the training data. No. Experts are people with above average aptitude for any given domain. It's not just about knowledge. many people try and fail to become experts in any given domain. >That's because there is no difference between them. Super intelligence is achievable just by making general intelligence faster. That's not how intelligence works. Dumb thinking sped up is just more dumb thinking but faster.
- godelski 3y ago
  > "Average human level" is pretty boring though. It seems you have the wrong idea of what is being conveyed, or what average human intelligence is. It isn't about being able to do math. It is being able to invent, mimic quickly, abstract, memorize, specialize, and generalize. There's a reason humans have occupied every continent of the earth and even areas outside. It's far more than being able to do arithmetic or playing chess. This just all seems unimpressive to us because it is normal, to us. But this certainly isn't normal if we look outside ourselves. Yes, there's intelligence in many lifeforms, even ants, but there is some ineffable or difficult to express uniqueness to human intelligence (specifically in its generality) that is being referenced here. To put it one way, a group of machines that could think at the level of an average teenager (or even lower) but able to do so 100x faster would probably outmatch a group of human scientists in being able to solve complex and novel math problems. This isn't "average human level" but below. "Average human level" is just a shortcut term for this ineffable description of the _capacity_ to generalize and adapt so well. Because we don't even have a fucking definition of intelligence.
  - AnthonyMouse 3y ago
    > It isn't about being able to do math. It is being able to invent, mimic quickly, abstract, memorize, specialize, and generalize. But this is exactly why average is boring. If you ask ChatGPT what it's like to be in the US Navy, it will have texts written by Navy sailors in its training data and produce something based on those texts in response to related questions. If you ask the average person what it's like to be in the US Navy, they haven't been in the Navy, may not know anyone who is, haven't taken any time to research it, so their answers will be poor. ChatGPT could plausibly give a better response. But if you ask the questions of someone who has, they'll answer related questions better than ChatGPT. Even if the average person who has been in the Navy has no greater intelligence than the average person who hasn't. It's not better at reasoning. It's barely even capable of it, but has access to training data that the average person lacks.
    - famouswaffles 3y ago
      >It's not better at reasoning. It's barely even capable of it You are wrong. and there's many papers to show otherwise. Algorithmic, Casual, Inference, Analogical LLMs reason just fine https://arxiv.org/abs/2212.09196 https://arxiv.org/abs/2212.09196 https://arxiv.org/abs/2305.00050 https://arxiv.org/abs/2305.00050 https://arxiv.org/abs/2204.02329 https://arxiv.org/abs/2204.02329 https://arxiv.org/abs/2211.09066 https://arxiv.org/abs/2211.09066
godelski 3y ago
> Just blows my mind the lengths some will go to ignore what is already easily verifiable right now. "I'll know agi when i see it", my ass. You and me both. I mean look at how people attribute the ability to recall with intelligence. Is memory part of intelligence? Yeah. Is it all of it? No. That's why people with eidetic memories are considered the smartest and they're the most successful people. We have no idea how good systems like GPT and Bard actually are because we have no idea what is in their training data. But we do know that when we can uncover sections that they do really well at what's in there but not so when it isn't. This is generalization, and a big part of intelligence. Unfortunately all we know is that everything is contaminated so we can't clearly measure this, which was already a noisy proxy. We've quietly switched to making the questions on the test identical or nearly identical to those on the homework. That's different than testing novel problems. And it doesn't help that we have a lot of people that haven't spent significant times in ML speaking about it. People who haven't studied up on cognition. People who haven't studied statistics. An academic degree isn't needed, but the knowledge is. These are people with no expertise. We even see a lot on HN. People that think training a transformer "from scratch" makes them an expert. Maybe because CS people pick up minor domain knowledge quickly, but can't differentiate domain expertise from domain knowledge. Then in experts we have the tails of the distribution dominating the conversation: over hype and just memorization. We let people dominate the conversation with discussions about how we're going to create superintelligences that are going to (with no justification) enslave humanity and other people saying they're just copy machines. Neither is correct or helpful. It is crying wolf before the wolf comes. If we say a wolf is eating all our sheep when in reality we just see it patrolling the edge of the woods then people won't listen when the wolf does attack. (AI dangers do exist, but long before superintelligence and not "generates disproportionately white faces.") > "I'll know agi when i see it", my ass. I don't know any researcher that realistically believes this. The vast majority of us believe we don't have a good definition and it'll be unclear. If we can build it, we'll probably not know it is intelligent at first. Which this being the case, of course we should be expecting pushback. We SHOULD. It's insane to attribute intelligence to something we don't know has intelligence. It's okay to question if it has it though. If we didn't then we'd need a clear definition. If you got a secret consistent one, then please, tell us all. Unless you think that when/if we build an intelligence that it comes into existence already aware of its own sentience. Which would be weird. Can you name a creature that knows its own sentience or consciousness at birth? There's a lot of domain expertise nuance that is ignored because at the surface level it looks trivial. But domain expertise is knowing the depth and nuance of seemingly simple things. Most of our AI conversations are downright idiotic. The conversation is dominated by grifters, people looking for views and playing on emotion, not experts. The truth is we don't know a lot and it is just a very messy time. Lots of people adding noise in a very noisy environment. Maybe a lot of us should do what the experts tend to do, and just not confidently talk openly in the public about things we don't have the answers to.
daveguy 3y ago
> Give criteria that 4 fails that a significant chunk of the human population doesn't also fail and we can talk. The ability to respond "I don't know." to literally any question.
- famouswaffles 3y ago
  I've had 4 respond i don't know to questions before. and the base model was excellently calibrated. https://openai.com/research/gpt-4 https://openai.com/research/gpt-4 "Interestingly, the base pre-trained model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correct). However, through our current post-training process, the calibration is reduced." Next ?

AnthonyMouse 3y ago

> People who say this nonsense need to start properly defining human level intelligence because nearly anything you throw at GPT-4 it performs at at least average human level, often well above. "Average human level" is pretty boring though. Computers have been doing arithmetic at well above "average human level" since they were first invented. The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. Which is clearly still not the case.

toomuchtodo 3y ago
> The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. Which is clearly still not the case. I imagine an important concern is the learning & improvement velocity. Humans get old, tired, etc. GPUs do not. It isn't the case now, but it is fuzzy how fast we could collectively get there. Break out problem domains into modules, off to the silicon dojos until your models exceed human capabilities, and then roll them up. You can pick from OpenGPT plugins, why wouldn't an LLM hypervisor/orchestrator do the same? https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html https://waitbutwhy.com/2015/01/artificial-intelligence-revol... https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-2.html https://waitbutwhy.com/2015/01/artificial-intelligence-revol...
- dragonwriter 3y ago
  > The concern is the learning & improvement velocity. Humans get old, tired, etc. GPUs do not. They do, though. Of course, replacing the worn out hardware while keeping the software is easier with GPUs.
famouswaffles 3y ago
>"Average human level" is pretty boring though. Lol ok. Still human level. and GPT-4 is way above average in most tasks. >Computers have been doing arithmetic at well above "average human level" since they were first invented. Cool. That's what the general in agi is about. GPT-4 is very general. >The premise of AGI isn't that it can do something better than people, it's that it can do everything at least as well. as well as what kind of people ? experts ? That was not the premise of agi when the term was coined or for a long time afterwards. Posts have shifted(as they often do in this field) so that that's what the term seems to mean now but agi was artificial and generally intelligent, which has been passed. There's no difference between your definition of agi which is supposed to surpass experts in every field and super intelligence.
- AnthonyMouse 3y ago
  > Lol ok. Still human level. and GPT-4 is way above average in most tasks. It has access to a lot of information that most humans don't have memorized. It's a better search engine than most humans. And it can format that information into natural language. But can it drive a car? If given an incentive to not confabulate and the knowledge that its statements are being verified, can it achieve that as consistently as the median human? If you start by giving it a simple instruction with stark consequences for not following it, can it continue to register the importance of that instruction even after you give it a lot more text to read? > as well as what kind of people ? experts ? Experts are just ordinary people with specific information. You're giving the specific information to the AI, aren't you? It's in the training data. > There's no difference between your definition of agi which is supposed to surpass experts in every field and super intelligence. That's because there is no difference between them. Super intelligence is achievable just by making general intelligence faster. If you have AGI and can make it go faster by throwing more compute hardware at it then you have super intelligence.
  - famouswaffles 3y ago
    >It has access to a lot of information that most humans don't have memorized. It's not just about knowledge. Lots of papers showing strong reasoning across various reasoning types. Couple papers demonstrating the development of world models too. >It's a better search engine than most humans. And it can format that information into natural language. Not how this works. They aren't search engines. and their performance equity with people isn't relegated to knowledge tasks alone. >But can it drive a car? If given an incentive to not confabulate and the knowledge that its statements are being verified, can it achieve that as consistently as the median human? Can a blind man drive a car ? a man with no hands ? >If you start by giving it a simple instruction with stark consequences for not following it, can it continue to register the importance of that instruction even after you give it a lot more text to read? Lol yes >Experts are just ordinary people with specific information. You're giving the specific information to the AI, aren't you? It's in the training data. No. Experts are people with above average aptitude for any given domain. It's not just about knowledge. many people try and fail to become experts in any given domain. >That's because there is no difference between them. Super intelligence is achievable just by making general intelligence faster. That's not how intelligence works. Dumb thinking sped up is just more dumb thinking but faster.
godelski 3y ago
> "Average human level" is pretty boring though. It seems you have the wrong idea of what is being conveyed, or what average human intelligence is. It isn't about being able to do math. It is being able to invent, mimic quickly, abstract, memorize, specialize, and generalize. There's a reason humans have occupied every continent of the earth and even areas outside. It's far more than being able to do arithmetic or playing chess. This just all seems unimpressive to us because it is normal, to us. But this certainly isn't normal if we look outside ourselves. Yes, there's intelligence in many lifeforms, even ants, but there is some ineffable or difficult to express uniqueness to human intelligence (specifically in its generality) that is being referenced here. To put it one way, a group of machines that could think at the level of an average teenager (or even lower) but able to do so 100x faster would probably outmatch a group of human scientists in being able to solve complex and novel math problems. This isn't "average human level" but below. "Average human level" is just a shortcut term for this ineffable description of the _capacity_ to generalize and adapt so well. Because we don't even have a fucking definition of intelligence.
- AnthonyMouse 3y ago
  > It isn't about being able to do math. It is being able to invent, mimic quickly, abstract, memorize, specialize, and generalize. But this is exactly why average is boring. If you ask ChatGPT what it's like to be in the US Navy, it will have texts written by Navy sailors in its training data and produce something based on those texts in response to related questions. If you ask the average person what it's like to be in the US Navy, they haven't been in the Navy, may not know anyone who is, haven't taken any time to research it, so their answers will be poor. ChatGPT could plausibly give a better response. But if you ask the questions of someone who has, they'll answer related questions better than ChatGPT. Even if the average person who has been in the Navy has no greater intelligence than the average person who hasn't. It's not better at reasoning. It's barely even capable of it, but has access to training data that the average person lacks.
  - famouswaffles 3y ago
    >It's not better at reasoning. It's barely even capable of it You are wrong. and there's many papers to show otherwise. Algorithmic, Casual, Inference, Analogical LLMs reason just fine https://arxiv.org/abs/2212.09196 https://arxiv.org/abs/2212.09196 https://arxiv.org/abs/2305.00050 https://arxiv.org/abs/2305.00050 https://arxiv.org/abs/2204.02329 https://arxiv.org/abs/2204.02329 https://arxiv.org/abs/2211.09066 https://arxiv.org/abs/2211.09066

godelski 3y ago

> Just blows my mind the lengths some will go to ignore what is already easily verifiable right now. "I'll know agi when i see it", my ass. You and me both. I mean look at how people attribute the ability to recall with intelligence. Is memory part of intelligence? Yeah. Is it all of it? No. That's why people with eidetic memories are considered the smartest and they're the most successful people. We have no idea how good systems like GPT and Bard actually are because we have no idea what is in their training data. But we do know that when we can uncover sections that they do really well at what's in there but not so when it isn't. This is generalization, and a big part of intelligence. Unfortunately all we know is that everything is contaminated so we can't clearly measure this, which was already a noisy proxy. We've quietly switched to making the questions on the test identical or nearly identical to those on the homework. That's different than testing novel problems. And it doesn't help that we have a lot of people that haven't spent significant times in ML speaking about it. People who haven't studied up on cognition. People who haven't studied statistics. An academic degree isn't needed, but the knowledge is. These are people with no expertise. We even see a lot on HN. People that think training a transformer "from scratch" makes them an expert. Maybe because CS people pick up minor domain knowledge quickly, but can't differentiate domain expertise from domain knowledge. Then in experts we have the tails of the distribution dominating the conversation: over hype and just memorization. We let people dominate the conversation with discussions about how we're going to create superintelligences that are going to (with no justification) enslave humanity and other people saying they're just copy machines. Neither is correct or helpful. It is crying wolf before the wolf comes. If we say a wolf is eating all our sheep when in reality we just see it patrolling the edge of the woods then people won't listen when the wolf does attack. (AI dangers do exist, but long before superintelligence and not "generates disproportionately white faces.") > "I'll know agi when i see it", my ass. I don't know any researcher that realistically believes this. The vast majority of us believe we don't have a good definition and it'll be unclear. If we can build it, we'll probably not know it is intelligent at first. Which this being the case, of course we should be expecting pushback. We SHOULD. It's insane to attribute intelligence to something we don't know has intelligence. It's okay to question if it has it though. If we didn't then we'd need a clear definition. If you got a secret consistent one, then please, tell us all. Unless you think that when/if we build an intelligence that it comes into existence already aware of its own sentience. Which would be weird. Can you name a creature that knows its own sentience or consciousness at birth? There's a lot of domain expertise nuance that is ignored because at the surface level it looks trivial. But domain expertise is knowing the depth and nuance of seemingly simple things. Most of our AI conversations are downright idiotic. The conversation is dominated by grifters, people looking for views and playing on emotion, not experts. The truth is we don't know a lot and it is just a very messy time. Lots of people adding noise in a very noisy environment. Maybe a lot of us should do what the experts tend to do, and just not confidently talk openly in the public about things we don't have the answers to.

daveguy 3y ago

> Give criteria that 4 fails that a significant chunk of the human population doesn't also fail and we can talk. The ability to respond "I don't know." to literally any question.

famouswaffles 3y ago
I've had 4 respond i don't know to questions before. and the base model was excellently calibrated. https://openai.com/research/gpt-4 https://openai.com/research/gpt-4 "Interestingly, the base pre-trained model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correct). However, through our current post-training process, the calibration is reduced." Next ?