4 ms·

He may be right about the quality of the current state of the art. But like most of us I think he struggles to comprehend the implications of exponential impro

He may be right about the quality of the current state of the art. But like most of us I think he struggles to comprehend the implications of exponential improvement. What if 2 or 4 years now the LLMs have improved by a factor of 10 or 100? Will they still not be as good as the best humans at writing a movie script? What about 2 years after that with similar rate of improvement?

techpineapple 2y ago
Are there enough examples of "best humans at writing a movie script?" Is there an objective measure of "best movie scripts"? I think LLM's are probably already at the level where they can generate enough quality to grab people's attention, but I think there is a human component, not to sort of the quality of a film, but the ability to sell people on something that is probably just not common, consistent, or replicable enough to emulate. Like, I don't think objective quality can tell you why Taylor swift sold 3 times as many records as the rolling stones. Nor do I think that kind of fame would be replicable by AI. And actually here's what makes me somewhat sure about this. If AI is good enough to replicate the popularity of Taylor Swift, it's probably good enough to do it 1000 times, and microtarget, diluting it's own popularity. Is an artist that's 1/1000 times less popular than Taylor Swift as good as Taylor Swift if there's no objective measure of "good"? Now, could AI capture 98% of the attention on Tik Tok through experiments and algorithms? Absolutely. If not today, very soon.
AaronRossPowell 2y ago
Trend lines don't always keep going up, in absolute terms or in their rate. And I'm skeptical that the nature of LLM technology specifically can achieve AGI, because it's at base a prediction model (though an incredibly sophisticated and powerful one) and not actual intelligence. To think that LLMs can grow into AGI is, thus, a category mistake, even if they get better at what they do, and even if they prove super useful in all kinds of ways.
- jimbokun 2y ago
  Current LLMs are already far beyond what a "prediction model" might be capable of.
- dinfinity 2y ago
  Until you define "actual intelligence" and how the (human) brain produces it, it is nonsense to state that a "prediction model" is not "actual intelligence". It's just a form of the no true Scotsman fallacy. This line of thinking is also very closely related to the similarly misguided Chinese Room thought experiment by Searle. "It just inputs and outputs symbols" (or "it just predicts the next word") misses the enormous wealth of complexity that goes into how to get from input to output and just handwaving it away is intellectual poverty. With regard to the rate of progress and potential: Don't forget that the biological difference between humans and animals is tiny, yet the difference in ("actual") intelligence is enormous. This is not to say that LLMs (and its offspring) are guaranteed to only need small tweaks, but it very well might suffice to reach AGI as you would define it. Sidenote: Stop using the word "techbro". It is sexist, polarizing, and childish. Just because they're the bad guys doesn't make it okay.
mcphage 2y ago
> But like most of us I think he struggles to comprehend the implications of exponential improvement. Or he successfully comprehends the implications of sigmoid improvement. Consider self-driving cars: there was a few years of huge progress, and it looked like the end of human driving was right around the corner. And then… the exponential part of the curve stopped, and it reverted to a much slower rate of improvement. We’ll get there eventually, but… it’s not around the corner.
- jimbokun 2y ago
  Where are we now on the LLM sigmoid? How can you tell?
  - mcphage 2y ago
    That’s the neat thing: you can’t. You get explosive improvement, and then suddenly you hit an invisible inflection point and they slow to a crawl.
    - ben_w 2y ago
      Indeed, and worse: actual improvements along the curve are somewhat random, meaning you can't even tell that you passed the equivalent inflection point (returns became diminishing) until significantly after.

techpineapple 2y ago

Are there enough examples of "best humans at writing a movie script?" Is there an objective measure of "best movie scripts"? I think LLM's are probably already at the level where they can generate enough quality to grab people's attention, but I think there is a human component, not to sort of the quality of a film, but the ability to sell people on something that is probably just not common, consistent, or replicable enough to emulate. Like, I don't think objective quality can tell you why Taylor swift sold 3 times as many records as the rolling stones. Nor do I think that kind of fame would be replicable by AI. And actually here's what makes me somewhat sure about this. If AI is good enough to replicate the popularity of Taylor Swift, it's probably good enough to do it 1000 times, and microtarget, diluting it's own popularity. Is an artist that's 1/1000 times less popular than Taylor Swift as good as Taylor Swift if there's no objective measure of "good"? Now, could AI capture 98% of the attention on Tik Tok through experiments and algorithms? Absolutely. If not today, very soon.

AaronRossPowell 2y ago

Trend lines don't always keep going up, in absolute terms or in their rate. And I'm skeptical that the nature of LLM technology specifically can achieve AGI, because it's at base a prediction model (though an incredibly sophisticated and powerful one) and not actual intelligence. To think that LLMs can grow into AGI is, thus, a category mistake, even if they get better at what they do, and even if they prove super useful in all kinds of ways.

jimbokun 2y ago
Current LLMs are already far beyond what a "prediction model" might be capable of.
dinfinity 2y ago
Until you define "actual intelligence" and how the (human) brain produces it, it is nonsense to state that a "prediction model" is not "actual intelligence". It's just a form of the no true Scotsman fallacy. This line of thinking is also very closely related to the similarly misguided Chinese Room thought experiment by Searle. "It just inputs and outputs symbols" (or "it just predicts the next word") misses the enormous wealth of complexity that goes into how to get from input to output and just handwaving it away is intellectual poverty. With regard to the rate of progress and potential: Don't forget that the biological difference between humans and animals is tiny, yet the difference in ("actual") intelligence is enormous. This is not to say that LLMs (and its offspring) are guaranteed to only need small tweaks, but it very well might suffice to reach AGI as you would define it. Sidenote: Stop using the word "techbro". It is sexist, polarizing, and childish. Just because they're the bad guys doesn't make it okay.

mcphage 2y ago

> But like most of us I think he struggles to comprehend the implications of exponential improvement. Or he successfully comprehends the implications of sigmoid improvement. Consider self-driving cars: there was a few years of huge progress, and it looked like the end of human driving was right around the corner. And then… the exponential part of the curve stopped, and it reverted to a much slower rate of improvement. We’ll get there eventually, but… it’s not around the corner.

jimbokun 2y ago
Where are we now on the LLM sigmoid? How can you tell?
- mcphage 2y ago
  That’s the neat thing: you can’t. You get explosive improvement, and then suddenly you hit an invisible inflection point and they slow to a crawl.
  - ben_w 2y ago
    Indeed, and worse: actual improvements along the curve are somewhat random, meaning you can't even tell that you passed the equivalent inflection point (returns became diminishing) until significantly after.