6 ms·

It is wild how seriously people are taking the development of these LLMs. There is talk of them leading to the extinction of life while neglecting to consider

It is wild how seriously people are taking the development of these LLMs. There is talk of them leading to the extinction of life while neglecting to consider that we can't even build self-driving cars, let alone have a robot clear the table after dinner. Also, many people seem to be mistaking "artificial intelligence" for "actual intelligence". You know that these LLMs are not actually intelligent, right? Right???

CamperBob2 3y ago
"Intelligence" is a word we made up. It doesn't mean much of anything at the end of the day. If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, or write a program that has never run before, it's intelligent enough to be called "intelligent." Conversely, if you can do these things in the absence of any training or outside influence, you're intelligent enough to be called a god. Assuming that this is not the case, don't hold machines to standards you're not willing to hold humans to.
- throw47474777j 3y ago
  “X is a word we made up. It doesn't mean much of anything at the end of the day.” This has to be a candidate for the most unhelpful statement possible. If you are sincere about it, then why not apply it to every word in your reply and see where that leaves you?
  - SpicyLemonZest 3y ago
    I can't speak for the original commenter, but I would definitely say that "is this actually X" is a pointless debate for most values of X. If I understand what people mean by X, there's no need to quibble about whether it satisfies my personal definition; if I don't understand, I should try to clarify rather than just saying "it's not really X" and crossing my arms.
- bumby 3y ago
  I don't want to speak for the OP, but one of the issues they may be poking at is the transference of intelligence. E.g., you are "trained" to play a game, but at some point someone decides to change the rules of the game. Human children can transfer their previous knowledge to the new rule set fairly well. The most intelligent children can do so very quickly without supplemental "training" on the game played under the new rules. Humans certainly have flaws when it comes to this. I've heard some discussions about the success and failures of AI in this regard. Can someone in this domain elaborate on the current state-of-the-art performance in this regard?
  - famouswaffles 3y ago
    1. LLMs are general pattern machines. They can generalize and complete in context to a wide variety of complex non linguistic patterns.https://general-pattern-machines.github.io/ https://general-pattern-machines.github.io/ 2. LLMs see positive transfer in multi lingual capability. For example, an LLM trained on 500B tokens of English and 10B tokens of French will not speak in french like a model trained on only 10B tokens of French. What will happen is that the model will be nearly as competent in French as it is in English https://arxiv.org/abs/2108.13349 https://arxiv.org/abs/2108.13349 3. Language models of code reason better even if the benchmarks have nothing to do with code. https://arxiv.org/abs/2210.07128 https://arxiv.org/abs/2210.07128
    - bumby 3y ago
      This is interesting in the context of the other response that links to poor performance in terms of counterfactuals. I wonder if it is related to how well one domain maps to another? E.g., can they transfer to english to french well because both share a similar word classes (nouns, verbs, etc.). But I believe other languages change more based on social context (e.g., Japanese) than English does. Would a LLM transfer just as well to the latter? In that case, my guess is humans would find it more difficult to transfer as well, so it may not be a good measure. (Apologies to any linguists. Please correct anything above if I'm off).
      - famouswaffles 3y ago
- goatlover 3y ago
  "All words are made up." - Thor The word was made up to cover a range of cognitive abilities that humans and animals (to varying degrees) possess. And we're gradually figuring out how to design machines with similar abilities. The general idea of intelligence being you can figure out how to do things that aren't just instinctual. a generalized intelligence can do this across any number of domains without some fixed limit.
- AlexandrB 3y ago
  > Conversely, if you can do these things in the absence of any training or outside influence, you're intelligent enough to be called a god. So you're saying humans, collectively, are gods? After all, as a species we started with nothing - no "clean" training data, no paintings to replicate, not even language itself. And here we are - arguing about whether one creation of ours, trained on a bunch of other stuff we created, is as intelligent as us.
  - famouswaffles 3y ago
    We didn't continue with nothing. Evolution is training. LLMs "start with nothing" too
- prhn 3y ago
  > "Intelligence" is a word we made up So is all human knowledge. They're all things we "made up" to describe the world around us. That's a really cheap way to reduce something. There's a mountain of psychological literature that intends to describe and measure intelligence. Saying it doesn't mean much of anything is a bit of a postmodern hot take. > If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, or write a program that has never run before, it's intelligent enough to be called "intelligent." Painting pictures and composing poems aren't good ways to measure general intelligence. That's creativity. They're correlated but not the same. There are specific tests for measuring general intelligence, and there's absolutely no way current gen LLM would score highly on them because they don't even accept visual input. Yes I've seen that they can score highly on verbal-only tests. That's obvious anyway because they have seen, written down, and have access to the answers during the test. Even if they did accept visual input they would only be able to solve problems for which they've previously been given the answers (trained on the dataset). Even then I'd have my doubts given how terribly wrong ChatGPT4 is when I ask it to produce a very simple Kubernetes manifest which has a strictly defined spec.
  - CamperBob2 3y ago
    Even then I'd have my doubts given how terribly wrong ChatGPT4 is when I ask it to produce a very simple Kubernetes manifest which has a strictly defined spec. One problem is that ChatGPT4 has been degrading over time -- and no, I don't GAF about any opinions or assurances to the contrary. So we're seeing more and more people look at it for the first time and ask what the big deal is. Maybe the Code Interpreter feature will reverse that trend, we'll see. Those of us who were around (read: paying $20/month) when they first enabled GPT4 support are left waving our hands fecklessly, muttering "Yeah, but you should've seen it back in the old days, back in March of '23. Now you kids get off my lawn."
- Keegs 3y ago
  > If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, An LLM can write a poem that fools me into thinking it’s good, but not anyone who reads poetry as a hobby. Poetry is inspired in a way that’s hard to replicate with a prompt. Artists also take inspiration from other artists, but the “distance” between their influences and their work is so much greater than the distance between an LLM’s that I don’t fault anyone who think it’ll take time for LLMs to catch up.
  - CamperBob2 3y ago
    An LLM can write a poem that fools me into thinking it’s good, but not anyone who reads poetry as a hobby. And you don't grasp the fundamentally-incremental nature of this point?
    - jononomo 3y ago
      No. After you've asked ChatGPT to write 5-6 poems you'll get a sense for how profoundly uncreative it is as a poet. It's not going to "incrementally" be anything other than a dead-as-a-doornail algorithm.
      - Aeolos 3y ago
        That's ChatGPT version 3. Version 4 is already significantly better. What about version 5? 6? 10?
- mdp2021 3y ago
  > to be called "intelligent" quality is required.
disadvantage 3y ago
> we can't even build self-driving cars, let alone have a robot clear the table after dinner I think once you give AI working limbs and the ability to manipulate matter like humans, that's concerning, because if the AI becomes a runaway AI with no upper limits on the types of behavior allowed, we get into Paperclip Maximizer[0] territory very fast. [0] https://tanzanite.ai/paperclip-maximizer-experiment/ https://tanzanite.ai/paperclip-maximizer-experiment/
gnulinux 3y ago
I'm completely with you on this one. Sci-fi and AI hype are extremely strong on these threads. I bet we're not any closer to human level intelligence with ML than we were in 1500s with mechanical clocks. Both are impressive nonetheless.
thiago_fm 3y ago
Just as in finance everyday you see a hedge fund manager making some weird big bet of a stock or market for the next 20 years (like a prophet). Everyone wants to pose as the front(wo)man of that innovation, thought-leaders. Whatever. It's all marketing, for their own careers, their own personal agenda. The whole text looks like written by ChatGPT in a prompt like "Write about AI safety and relates it to the Age of Reason". This is a good reason why AI should replace us, so we no longer have human beings writing that stuff to self-promote themselves. The funniest part is in the end of the article: "Eric Ries has been my close collaborator throughout the development of this article" So the guy who wrote "The lean startup" is helping him with that article. drops mic
marvin 3y ago
A pretty influential and very well-financed subculture of AI philosophers are currently performing a comprehensive lobbying campaign for AI regulation. They generally go by Rationalists or Effective Altruism, and have a philosophy stemming from MIRI, Eliezer Yudkowsky, Nate Soares and others. They've had considerable success getting traction for these views with politicians in Europe, and that's the reason you see so much of this in the media currently. I'm hoping more sane voices will soon get organized, because significant parts of this are reminiscent of a well-funded doomsday cult. I think this article is an early attempt at creating a politically viable counterpoint.
PaulStatezny 3y ago
This reminds me of when I attended the first official conference for Facebook's popular React.js library. (2014 I believe?) Got to hang out with some "hotshot" Facebook developers after hours. I remember Lee Byron (of React, GraphQL, etc) stating very confidently that the roads would be teeming with self driving cars by 2020.
sp332 3y ago
When AIs kill people, it is not because they are intelligent, but because they are given lethal powers without appropriate judgement. The first autonomous weapons platform with a kill count was an accident. https://www.wired.com/2007/10/robot-cannon-ki/ https://www.wired.com/2007/10/robot-cannon-ki/ Self-driving cars kill people all the time. These things happen because the AIs are not super-intelligent, but people keep giving them too much power.
- cypress66 3y ago
  > Self-driving cars kill people all the time. Maybe L2 systems. But those are just a driver assistance, the driver must supervise and intervene. Actual L4/L5 systems haven't killed anyone as far as I know, except for the Uber accident, and even then the driver was supposed to be supervising it, I'm not sure if it counts as L4/L5
famouswaffles 3y ago
What testable definition of general intelligence does GPT-4 fail that a big chunk of humans don't also fail?
- SirMaster 3y ago
  Self reflection? Emotional intelligence? Empathy?
  - famouswaffles 3y ago
    I don't know how we'd go about comparing reflection to human baselines but GPT-4 can reflect just fine https://arxiv.org/abs/2303.11366 https://arxiv.org/abs/2303.11366 GPT-4's emotional intelligence seems to be very high by the looks of things https://arxiv.org/abs/2304.11490 https://arxiv.org/abs/2304.11490 Empathy =/ Intelligence
    - SirMaster 3y ago
      I really mean self-awareness and consciousness. Isn’t that a metric we use to determine the intelligence of animals and such? Does GPT love its creators the way we love our parents? Does it mourn a loss the way we do? Elephants are known to be of the more intelligent animals and apparently they mourn loss as well. So apparently there is something there linked with intelligence. Maybe it's not directly intelligence related, but what makes a human have a spontaneous emotional response like laughter for instance? The self-reflection link you sent isn't what I was thinking. I was thinking about things like self reflection on my existence. Like am I happy doing what I'm doing? Do I feel I am making a positive difference in the world? If not what should I start doing to work toward becoming what I want to be? The fact that I want to be something seems different. Nobody tells me what I want to be. But we still need to tell computers an awful lot about what they should be and what they should want to be. When will AI be able to decide for itself what it wants to be and what it wants to do? And to do all that sort of self reflection on it's own without us specifying a reward system that "makes" it want to do that? I don't strive to become a better in the world person for some reward.
ben_w 3y ago
> You know that these LLMs are not actually intelligent, right? Right??? Can a submarine swim? Does it matter if an LLM — which is far from the only kind of AI, and trivial to include in another more complex AI — does or doesn't match your definition of "intelligent" when it can write fluently in any language (human and code), and also getting near top-of-class results in the Bar exam and the biology olympiad? Corporations — sometimes used as another example of an unaligned optimising agent — can't drive cars, they have to hire humans to do that for them, but corporations can still be prosecuted for dangerous chemicals (3M); and AI, just like other computer programs, can have dangerous output that you should not rely on (e.g. Thule early warning radar incident where someone forgot to tell it that it's OK for the moon to not respond to IFF pings and it nearly started WW3).
- hoxtrader 3y ago
  [dead]

CamperBob2 3y ago

"Intelligence" is a word we made up. It doesn't mean much of anything at the end of the day. If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, or write a program that has never run before, it's intelligent enough to be called "intelligent." Conversely, if you can do these things in the absence of any training or outside influence, you're intelligent enough to be called a god. Assuming that this is not the case, don't hold machines to standards you're not willing to hold humans to.

throw47474777j 3y ago
“X is a word we made up. It doesn't mean much of anything at the end of the day.” This has to be a candidate for the most unhelpful statement possible. If you are sincere about it, then why not apply it to every word in your reply and see where that leaves you?
- SpicyLemonZest 3y ago
  I can't speak for the original commenter, but I would definitely say that "is this actually X" is a pointless debate for most values of X. If I understand what people mean by X, there's no need to quibble about whether it satisfies my personal definition; if I don't understand, I should try to clarify rather than just saying "it's not really X" and crossing my arms.
bumby 3y ago
I don't want to speak for the OP, but one of the issues they may be poking at is the transference of intelligence. E.g., you are "trained" to play a game, but at some point someone decides to change the rules of the game. Human children can transfer their previous knowledge to the new rule set fairly well. The most intelligent children can do so very quickly without supplemental "training" on the game played under the new rules. Humans certainly have flaws when it comes to this. I've heard some discussions about the success and failures of AI in this regard. Can someone in this domain elaborate on the current state-of-the-art performance in this regard?
- famouswaffles 3y ago
  1. LLMs are general pattern machines. They can generalize and complete in context to a wide variety of complex non linguistic patterns.https://general-pattern-machines.github.io/ https://general-pattern-machines.github.io/ 2. LLMs see positive transfer in multi lingual capability. For example, an LLM trained on 500B tokens of English and 10B tokens of French will not speak in french like a model trained on only 10B tokens of French. What will happen is that the model will be nearly as competent in French as it is in English https://arxiv.org/abs/2108.13349 https://arxiv.org/abs/2108.13349 3. Language models of code reason better even if the benchmarks have nothing to do with code. https://arxiv.org/abs/2210.07128 https://arxiv.org/abs/2210.07128
  - bumby 3y ago
    This is interesting in the context of the other response that links to poor performance in terms of counterfactuals. I wonder if it is related to how well one domain maps to another? E.g., can they transfer to english to french well because both share a similar word classes (nouns, verbs, etc.). But I believe other languages change more based on social context (e.g., Japanese) than English does. Would a LLM transfer just as well to the latter? In that case, my guess is humans would find it more difficult to transfer as well, so it may not be a good measure. (Apologies to any linguists. Please correct anything above if I'm off).
    - famouswaffles 3y ago
goatlover 3y ago
"All words are made up." - Thor The word was made up to cover a range of cognitive abilities that humans and animals (to varying degrees) possess. And we're gradually figuring out how to design machines with similar abilities. The general idea of intelligence being you can figure out how to do things that aren't just instinctual. a generalized intelligence can do this across any number of domains without some fixed limit.
AlexandrB 3y ago
> Conversely, if you can do these things in the absence of any training or outside influence, you're intelligent enough to be called a god. So you're saying humans, collectively, are gods? After all, as a species we started with nothing - no "clean" training data, no paintings to replicate, not even language itself. And here we are - arguing about whether one creation of ours, trained on a bunch of other stuff we created, is as intelligent as us.
- famouswaffles 3y ago
  We didn't continue with nothing. Evolution is training. LLMs "start with nothing" too
prhn 3y ago
> "Intelligence" is a word we made up So is all human knowledge. They're all things we "made up" to describe the world around us. That's a really cheap way to reduce something. There's a mountain of psychological literature that intends to describe and measure intelligence. Saying it doesn't mean much of anything is a bit of a postmodern hot take. > If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, or write a program that has never run before, it's intelligent enough to be called "intelligent." Painting pictures and composing poems aren't good ways to measure general intelligence. That's creativity. They're correlated but not the same. There are specific tests for measuring general intelligence, and there's absolutely no way current gen LLM would score highly on them because they don't even accept visual input. Yes I've seen that they can score highly on verbal-only tests. That's obvious anyway because they have seen, written down, and have access to the answers during the test. Even if they did accept visual input they would only be able to solve problems for which they've previously been given the answers (trained on the dataset). Even then I'd have my doubts given how terribly wrong ChatGPT4 is when I ask it to produce a very simple Kubernetes manifest which has a strictly defined spec.
- CamperBob2 3y ago
  Even then I'd have my doubts given how terribly wrong ChatGPT4 is when I ask it to produce a very simple Kubernetes manifest which has a strictly defined spec. One problem is that ChatGPT4 has been degrading over time -- and no, I don't GAF about any opinions or assurances to the contrary. So we're seeing more and more people look at it for the first time and ask what the big deal is. Maybe the Code Interpreter feature will reverse that trend, we'll see. Those of us who were around (read: paying $20/month) when they first enabled GPT4 support are left waving our hands fecklessly, muttering "Yeah, but you should've seen it back in the old days, back in March of '23. Now you kids get off my lawn."
Keegs 3y ago
> If an LLM can compose a poem found in no existing book, paint an attractive picture that no one has ever seen, An LLM can write a poem that fools me into thinking it’s good, but not anyone who reads poetry as a hobby. Poetry is inspired in a way that’s hard to replicate with a prompt. Artists also take inspiration from other artists, but the “distance” between their influences and their work is so much greater than the distance between an LLM’s that I don’t fault anyone who think it’ll take time for LLMs to catch up.
- CamperBob2 3y ago
  An LLM can write a poem that fools me into thinking it’s good, but not anyone who reads poetry as a hobby. And you don't grasp the fundamentally-incremental nature of this point?
  - jononomo 3y ago
    No. After you've asked ChatGPT to write 5-6 poems you'll get a sense for how profoundly uncreative it is as a poet. It's not going to "incrementally" be anything other than a dead-as-a-doornail algorithm.
    - Aeolos 3y ago
      That's ChatGPT version 3. Version 4 is already significantly better. What about version 5? 6? 10?
mdp2021 3y ago
> to be called "intelligent" quality is required.

disadvantage 3y ago

> we can't even build self-driving cars, let alone have a robot clear the table after dinner I think once you give AI working limbs and the ability to manipulate matter like humans, that's concerning, because if the AI becomes a runaway AI with no upper limits on the types of behavior allowed, we get into Paperclip Maximizer[0] territory very fast. [0] https://tanzanite.ai/paperclip-maximizer-experiment/ https://tanzanite.ai/paperclip-maximizer-experiment/

gnulinux 3y ago

I'm completely with you on this one. Sci-fi and AI hype are extremely strong on these threads. I bet we're not any closer to human level intelligence with ML than we were in 1500s with mechanical clocks. Both are impressive nonetheless.

thiago_fm 3y ago

Just as in finance everyday you see a hedge fund manager making some weird big bet of a stock or market for the next 20 years (like a prophet). Everyone wants to pose as the front(wo)man of that innovation, thought-leaders. Whatever. It's all marketing, for their own careers, their own personal agenda. The whole text looks like written by ChatGPT in a prompt like "Write about AI safety and relates it to the Age of Reason". This is a good reason why AI should replace us, so we no longer have human beings writing that stuff to self-promote themselves. The funniest part is in the end of the article: "Eric Ries has been my close collaborator throughout the development of this article" So the guy who wrote "The lean startup" is helping him with that article. drops mic

marvin 3y ago

A pretty influential and very well-financed subculture of AI philosophers are currently performing a comprehensive lobbying campaign for AI regulation. They generally go by Rationalists or Effective Altruism, and have a philosophy stemming from MIRI, Eliezer Yudkowsky, Nate Soares and others. They've had considerable success getting traction for these views with politicians in Europe, and that's the reason you see so much of this in the media currently. I'm hoping more sane voices will soon get organized, because significant parts of this are reminiscent of a well-funded doomsday cult. I think this article is an early attempt at creating a politically viable counterpoint.

PaulStatezny 3y ago

This reminds me of when I attended the first official conference for Facebook's popular React.js library. (2014 I believe?) Got to hang out with some "hotshot" Facebook developers after hours. I remember Lee Byron (of React, GraphQL, etc) stating very confidently that the roads would be teeming with self driving cars by 2020.

sp332 3y ago

When AIs kill people, it is not because they are intelligent, but because they are given lethal powers without appropriate judgement. The first autonomous weapons platform with a kill count was an accident. https://www.wired.com/2007/10/robot-cannon-ki/ https://www.wired.com/2007/10/robot-cannon-ki/ Self-driving cars kill people all the time. These things happen because the AIs are not super-intelligent, but people keep giving them too much power.

cypress66 3y ago
> Self-driving cars kill people all the time. Maybe L2 systems. But those are just a driver assistance, the driver must supervise and intervene. Actual L4/L5 systems haven't killed anyone as far as I know, except for the Uber accident, and even then the driver was supposed to be supervising it, I'm not sure if it counts as L4/L5

famouswaffles 3y ago

What testable definition of general intelligence does GPT-4 fail that a big chunk of humans don't also fail?

SirMaster 3y ago
Self reflection? Emotional intelligence? Empathy?
- famouswaffles 3y ago
  I don't know how we'd go about comparing reflection to human baselines but GPT-4 can reflect just fine https://arxiv.org/abs/2303.11366 https://arxiv.org/abs/2303.11366 GPT-4's emotional intelligence seems to be very high by the looks of things https://arxiv.org/abs/2304.11490 https://arxiv.org/abs/2304.11490 Empathy =/ Intelligence
  - SirMaster 3y ago
    I really mean self-awareness and consciousness. Isn’t that a metric we use to determine the intelligence of animals and such? Does GPT love its creators the way we love our parents? Does it mourn a loss the way we do? Elephants are known to be of the more intelligent animals and apparently they mourn loss as well. So apparently there is something there linked with intelligence. Maybe it's not directly intelligence related, but what makes a human have a spontaneous emotional response like laughter for instance? The self-reflection link you sent isn't what I was thinking. I was thinking about things like self reflection on my existence. Like am I happy doing what I'm doing? Do I feel I am making a positive difference in the world? If not what should I start doing to work toward becoming what I want to be? The fact that I want to be something seems different. Nobody tells me what I want to be. But we still need to tell computers an awful lot about what they should be and what they should want to be. When will AI be able to decide for itself what it wants to be and what it wants to do? And to do all that sort of self reflection on it's own without us specifying a reward system that "makes" it want to do that? I don't strive to become a better in the world person for some reward.

ben_w 3y ago

> You know that these LLMs are not actually intelligent, right? Right??? Can a submarine swim? Does it matter if an LLM — which is far from the only kind of AI, and trivial to include in another more complex AI — does or doesn't match your definition of "intelligent" when it can write fluently in any language (human and code), and also getting near top-of-class results in the Bar exam and the biology olympiad? Corporations — sometimes used as another example of an unaligned optimising agent — can't drive cars, they have to hire humans to do that for them, but corporations can still be prosecuted for dangerous chemicals (3M); and AI, just like other computer programs, can have dangerous output that you should not rely on (e.g. Thule early warning radar incident where someone forgot to tell it that it's OK for the moon to not respond to IFF pings and it nearly started WW3).

hoxtrader 3y ago
[dead]