3 ms·

For now. I don't think it's a law of nature that GPT-generated text must be this discordant and obvious. This is simply the current state of the art, soon to b

For now. I don't think it's a law of nature that GPT-generated text must be this discordant and obvious. This is simply the current state of the art, soon to be eclipsed by improved models that are more difficult or impossible to discern, because we have the combination of mathematics, underlying technology, and economic incentive to improve the models.

palata 2y ago
> because we have the combination of mathematics, underlying technology, and economic incentive to improve the models Still we'll have to see how much the new models improve. Tesla also has all those incentives to get to actual full self-driving... yet they seem to have reached a plateau years ago.
- LeifCarrotson 2y ago
  I think that in both cases it's a question of when the economics run out: There's a Pareto principle plateau where further investment stops making sense when the project is 80% of the way to functional with 80% of the work remaining. (That's the general Pareto 80/20 rule, I suspect the actual numbers for FSD are far worse.) If Tesla can sell a "Full Self Driving" package with 20% of the work of the real thing, some cameras, and some empty promises for $8k at 0.99%, I think their actual incentives to go from Level 3 to Level 5 are pretty weak. In contrast, GPTs currently have a small window of utility in marketing, search, and other text generation processes where accuracy isn't that important and weird tics are tolerable. The general public isn't yet accustomed to their hallucinations and foibles, so people are willing to pay to use them in contexts where the current state of the art is inadequate, but there will be recoil from those uses. And there's a huge gold rush underway as Google/Microsoft/Meta/others all try to dominate the future regardless of present economics.

3 ms·

For now. I don't think it's a law of nature that GPT-generated text must be this discordant and obvious. This is simply the current state of the art, soon to b

by LeifCarrotson 2y ago

palata 2y ago
> because we have the combination of mathematics, underlying technology, and economic incentive to improve the models Still we'll have to see how much the new models improve. Tesla also has all those incentives to get to actual full self-driving... yet they seem to have reached a plateau years ago.
- LeifCarrotson 2y ago
  I think that in both cases it's a question of when the economics run out: There's a Pareto principle plateau where further investment stops making sense when the project is 80% of the way to functional with 80% of the work remaining. (That's the general Pareto 80/20 rule, I suspect the actual numbers for FSD are far worse.) If Tesla can sell a "Full Self Driving" package with 20% of the work of the real thing, some cameras, and some empty promises for $8k at 0.99%, I think their actual incentives to go from Level 3 to Level 5 are pretty weak. In contrast, GPTs currently have a small window of utility in marketing, search, and other text generation processes where accuracy isn't that important and weird tics are tolerable. The general public isn't yet accustomed to their hallucinations and foibles, so people are willing to pay to use them in contexts where the current state of the art is inadequate, but there will be recoil from those uses. And there's a huge gold rush underway as Google/Microsoft/Meta/others all try to dominate the future regardless of present economics.