9 ms·

This is possibly a death spiral. GPT is only possible because it's been trained on the work humans have learned to do and then put out in the world. Now GPT is

This is possibly a death spiral. GPT is only possible because it's been trained on the work humans have learned to do and then put out in the world. Now GPT is as good as them and will put them all out of work. How can it improve if the people who fed it are now jobless?

jgust 4y ago
Presumably this problem is solved with technology improvements or the need is recognized to hire experts capable of generating high quality training material. In either situation, there's going to be extreme discomfort.
- mostlysimilar 4y ago
  GPT is good because of collective knowledge, lots of data. What do you have in mind by "hire experts"? Isn't that what we have now? Many experts in many fields, hired to do their work. Cut this number down and you reduce training data.
  - jgust 4y ago
    Let's assume that GPT eliminates an entire field of experts, runs out of training data, and whoever is at the helm of that GPT program decides that it's lucrative enough to obtain more/better data. One alternative is subsidizing these experts to do this type of work and plug it directly into the model. I don't expect the nature of the work to change, more likely it's the signature on the check and the availability of the datasets.
  - pixl97 4y ago
    You also may underestimate how quickly that AI could pass expert level. The experts out there still have many years of life left so they won't be disappearing soon. If we get self improving/self training AI sooner than later then, we'll humans won't be the experts.
  - HybridCurve 4y ago
    It's important to note however, that GPT does not itself have any knowledge, only information. Knowledge implies it has comprehension or understanding. It can just as easily produce bad information as good and it has little to no ability to self-assess the accuracy of information it provides.
- yoyohello13 4y ago
  There is a problem, how will people become experts in the field. If all entry level positions are taken by AI, nobody will be able to become an expert.
MonkeyMalarky 4y ago
Also what happens to the intuition and unwritten skills that humans learned and passed on over time? Sure, the model has probably internalized them implicitly from the training data. But what happens in a case where you need to have a human perform the task again (say after a devastating war)? The ones with the arcane knowledge are gone, and now humans are starting from scratch.
- mostlysimilar 4y ago
  Incredible that we've been writing speculative fiction about this for decades and still we sleepwalk right into it. I'd love to be wrong, but I think we're all still too divided and self-interested for this kind of technology to be successfully integrated. A lot of people are going to suffer.
  - salad-tycoon 4y ago
    It’s not just sci fi. It’s has already happened in past with construction. Things like pyramids and certain cathedrals and what not are no longer possible even with machines. At least this is what I’ve read and heard, I’m not actually an engineer or architect. Tangent, I’m looking for some sci fi about this topic. Any suggestions?
    - veqq 4y ago
      No. Things like Greek fire or Roman cement aren't possible - because we don't know the precise mixture or formulation involved. Many old descriptions mean we don't know how to do it, because they are very vague. But we can technically do much better waterproof concrete or whatever - however our incentives are also not aligned in the same ways. Here's a tangential link to monks building a Gothic cathedral with modern machines: https://carmelitegothic.com/ https://carmelitegothic.com/
      - pixl97 4y ago
        Actually we've recreated Roman concrete just recently. It appears a quicklime like substance in the concrete gives it a self healing property l
2OEH8eoCRo0 4y ago
Literally everything you do online is training data. This comment and discussion is future training data. Your browser history is logged somewhere and will be training data. Your OS probably spies on what you do...training data. It's training data all the way down. And they've hardly begun to take into account the physical world, video, music, etc. as training data.
xp84 4y ago
Presumably it will improve the same way humans did -- once it's roughly on par with us it'll be just as capable of innovating and trying new things. The only difference is that for humans, trying a truly new approach to something isn't really done that often by most. "GPT-9" might regularly and automatically try recomputing all the "tricky problems" it remembers from the past with updated models, or with a few tweaked parameters and then analyze whether any of these experiments provided "better" solutions. And it might do this operation during all idle cycles continuously. Honestly as a human who grasps how the economy works, this doesn't sound like a good thing, but I don't see any path to trying the fundamental changes that would be required for really good general AI to not be an absolute Depression generator. The only thing I'm wondering is, will the wealthiest ones, who actually have any power to influence these fundamental thing, figure this out before it's too late? I really doubt your Musks and Bezoses would enjoy living out their lives on ring-fenced compounds or remote islands while the rest of the world devolves into the Hunger Games.
- oska 4y ago
  > once it's roughly on par with us it'll be just as capable of innovating and trying new things This is complete nonsense. An LLM has no ability to innovate or 'try new things'.
  - pixl97 4y ago
    I keep seeing people post a line like this and it makes about zero sense to me... Just what the shit do you think Google/Microsoft/Amazon are pouring billions of dollars into machine learning/AI for? The first one that creates a self improving-self learning machine wins the game (or destroys the earth with a paperclip maximizer). You and your human intelligence are not magic. You're biological hardware and software that a lot of people are spending a lot of time and effort on reproducing in a digital format.
    - oska 4y ago
      > You're biological hardware and software This is (materialistic) nihilism. Materialism is a philosophy and not a new one, but an exceedingly sterile one, in my view. If you want to take that philosophical position you can, but others are free to reject it (and most people do) because it is only a philosophical position, not a proved description of reality (how can it be?).
WalterBright 4y ago
Imagine the devastation wrought by automatic looms, that put all the weavers out of a job! 97% of jobs used to be working on the farm. Now it's something like 2%.
- moffkalast 4y ago
  Can't wait for the economy that is 97% twitch streamers because that's all what humans are left qualified for. /s
  - msm_ 4y ago
    You joke, but an economy that is 97% artists (aka content creators) sounds... good? Isn't this the utopic end goal after we automate the scarcity out of our lifes?
    - jordanbeiber 4y ago
      Sure, with UBI. Otherwise not so much.
    - moffkalast 4y ago
      In theory it's great, in practice... who knows. The cynic in me would expect it to go worse than anyone could ever imagine. If everything is automated, why do you still need humans?
      - pixl97 4y ago
        We have more technology and more stuff than ever, yet it seems we've accomplished about zero in solving the human problems of greed and narcissism. Ted Faro is alive today. We just aren't sure which of the FRANK'S he's at.
        trafficante 4y ago
        Horizon: Zero Dawn was so compelling (to me) because of the outright horrifying plausibility of a gassed up tech CEO, convinced that software safeties were infallible, unleashes the consequences for their hubris upon the whole of humanity. Ted Faro was a horrible human being blinded by delusions of grandeur, but he wasn’t “evil” - he was even convinced he was saving humanity by ending the threat of both war and climate change. I don’t see GPT itself as representing a new Faro Plague, but I do see a lot of wannabe Ted Faros making the decisions at the top. If LLMs come even close to achieving their short-term potential, we’re unleashing a bigger destabilizing force on the world than the smartphone/social media combo - and the world of 202x seems blatantly incapable of absorbing that level of disruption.
acuozzo 4y ago
> GPT is only possible because it's been trained on the work humans have learned to do and then put out in the world. Which really makes you wonder if F/OSS was a good idea in the first place.

jgust 4y ago

Presumably this problem is solved with technology improvements or the need is recognized to hire experts capable of generating high quality training material. In either situation, there's going to be extreme discomfort.

mostlysimilar 4y ago
GPT is good because of collective knowledge, lots of data. What do you have in mind by "hire experts"? Isn't that what we have now? Many experts in many fields, hired to do their work. Cut this number down and you reduce training data.
- jgust 4y ago
  Let's assume that GPT eliminates an entire field of experts, runs out of training data, and whoever is at the helm of that GPT program decides that it's lucrative enough to obtain more/better data. One alternative is subsidizing these experts to do this type of work and plug it directly into the model. I don't expect the nature of the work to change, more likely it's the signature on the check and the availability of the datasets.
- pixl97 4y ago
  You also may underestimate how quickly that AI could pass expert level. The experts out there still have many years of life left so they won't be disappearing soon. If we get self improving/self training AI sooner than later then, we'll humans won't be the experts.
- HybridCurve 4y ago
  It's important to note however, that GPT does not itself have any knowledge, only information. Knowledge implies it has comprehension or understanding. It can just as easily produce bad information as good and it has little to no ability to self-assess the accuracy of information it provides.
yoyohello13 4y ago
There is a problem, how will people become experts in the field. If all entry level positions are taken by AI, nobody will be able to become an expert.

MonkeyMalarky 4y ago

Also what happens to the intuition and unwritten skills that humans learned and passed on over time? Sure, the model has probably internalized them implicitly from the training data. But what happens in a case where you need to have a human perform the task again (say after a devastating war)? The ones with the arcane knowledge are gone, and now humans are starting from scratch.

mostlysimilar 4y ago
Incredible that we've been writing speculative fiction about this for decades and still we sleepwalk right into it. I'd love to be wrong, but I think we're all still too divided and self-interested for this kind of technology to be successfully integrated. A lot of people are going to suffer.
- salad-tycoon 4y ago
  It’s not just sci fi. It’s has already happened in past with construction. Things like pyramids and certain cathedrals and what not are no longer possible even with machines. At least this is what I’ve read and heard, I’m not actually an engineer or architect. Tangent, I’m looking for some sci fi about this topic. Any suggestions?
  - veqq 4y ago
    No. Things like Greek fire or Roman cement aren't possible - because we don't know the precise mixture or formulation involved. Many old descriptions mean we don't know how to do it, because they are very vague. But we can technically do much better waterproof concrete or whatever - however our incentives are also not aligned in the same ways. Here's a tangential link to monks building a Gothic cathedral with modern machines: https://carmelitegothic.com/ https://carmelitegothic.com/
    - pixl97 4y ago
      Actually we've recreated Roman concrete just recently. It appears a quicklime like substance in the concrete gives it a self healing property l

2OEH8eoCRo0 4y ago

Literally everything you do online is training data. This comment and discussion is future training data. Your browser history is logged somewhere and will be training data. Your OS probably spies on what you do...training data. It's training data all the way down. And they've hardly begun to take into account the physical world, video, music, etc. as training data.

xp84 4y ago

Presumably it will improve the same way humans did -- once it's roughly on par with us it'll be just as capable of innovating and trying new things. The only difference is that for humans, trying a truly new approach to something isn't really done that often by most. "GPT-9" might regularly and automatically try recomputing all the "tricky problems" it remembers from the past with updated models, or with a few tweaked parameters and then analyze whether any of these experiments provided "better" solutions. And it might do this operation during all idle cycles continuously. Honestly as a human who grasps how the economy works, this doesn't sound like a good thing, but I don't see any path to trying the fundamental changes that would be required for really good general AI to not be an absolute Depression generator. The only thing I'm wondering is, will the wealthiest ones, who actually have any power to influence these fundamental thing, figure this out before it's too late? I really doubt your Musks and Bezoses would enjoy living out their lives on ring-fenced compounds or remote islands while the rest of the world devolves into the Hunger Games.

oska 4y ago
> once it's roughly on par with us it'll be just as capable of innovating and trying new things This is complete nonsense. An LLM has no ability to innovate or 'try new things'.
- pixl97 4y ago
  I keep seeing people post a line like this and it makes about zero sense to me... Just what the shit do you think Google/Microsoft/Amazon are pouring billions of dollars into machine learning/AI for? The first one that creates a self improving-self learning machine wins the game (or destroys the earth with a paperclip maximizer). You and your human intelligence are not magic. You're biological hardware and software that a lot of people are spending a lot of time and effort on reproducing in a digital format.
  - oska 4y ago
    > You're biological hardware and software This is (materialistic) nihilism. Materialism is a philosophy and not a new one, but an exceedingly sterile one, in my view. If you want to take that philosophical position you can, but others are free to reject it (and most people do) because it is only a philosophical position, not a proved description of reality (how can it be?).

WalterBright 4y ago

Imagine the devastation wrought by automatic looms, that put all the weavers out of a job! 97% of jobs used to be working on the farm. Now it's something like 2%.

moffkalast 4y ago
Can't wait for the economy that is 97% twitch streamers because that's all what humans are left qualified for. /s
- msm_ 4y ago
  You joke, but an economy that is 97% artists (aka content creators) sounds... good? Isn't this the utopic end goal after we automate the scarcity out of our lifes?
  - jordanbeiber 4y ago
    Sure, with UBI. Otherwise not so much.
  - moffkalast 4y ago
    In theory it's great, in practice... who knows. The cynic in me would expect it to go worse than anyone could ever imagine. If everything is automated, why do you still need humans?
    - pixl97 4y ago
      We have more technology and more stuff than ever, yet it seems we've accomplished about zero in solving the human problems of greed and narcissism. Ted Faro is alive today. We just aren't sure which of the FRANK'S he's at.
      - trafficante 4y ago
        Horizon: Zero Dawn was so compelling (to me) because of the outright horrifying plausibility of a gassed up tech CEO, convinced that software safeties were infallible, unleashes the consequences for their hubris upon the whole of humanity. Ted Faro was a horrible human being blinded by delusions of grandeur, but he wasn’t “evil” - he was even convinced he was saving humanity by ending the threat of both war and climate change. I don’t see GPT itself as representing a new Faro Plague, but I do see a lot of wannabe Ted Faros making the decisions at the top. If LLMs come even close to achieving their short-term potential, we’re unleashing a bigger destabilizing force on the world than the smartphone/social media combo - and the world of 202x seems blatantly incapable of absorbing that level of disruption.

acuozzo 4y ago

> GPT is only possible because it's been trained on the work humans have learned to do and then put out in the world. Which really makes you wonder if F/OSS was a good idea in the first place.