3 ms·

A lot of commenters here are making a very common mistake with observing any burgeoning technology: assuming that its current weaknesses will be present indefin

A lot of commenters here are making a very common mistake with observing any burgeoning technology: assuming that its current weaknesses will be present indefinitely. Assuming that current models' problems with, say, details and specificity, will continue on is equivalent to looking at 19th century automobiles and scoffing that they'll ever largely replace horses. Or looking at "smartphones" from the 90's and determining that they'll never see much uptake from the average consumer. AI image generation models are going to continue getting better and better and better, and the flaws we see now will get less common and less severe.

Planktonne 4y ago
It's equally, if not more, utopian to assume that progress down any one route will automatically lead to the imagined pinnacle of the entire discipline. Progress isn't smooth or automatic, and we often hit dead ends, having to try new techniques to get over plateaus. It's equivalent to looking at 19th Century automobiles and asserting that they'll be able to travel to other stars by the end of the decade.
- TulliusCicero 4y ago
  > It's equivalent to looking at 19th Century automobiles and asserting that they'll be able to travel to other stars by the end of the decade. That's not actually equivalent at all. Looking at improvements to automobiles would mean them going faster, being more reliable, etc. Space is obviously a different domain. Your analogy is like if people were suggesting that AI image generation models were going to become advanced enough to start driving cars or folding your laundry.
Archelaos 4y ago
There is also this other common mistake with observing any burgeoning technology: the linear (or even exponential) extrapolation of early success into the future. Think of some technologies from fifty years ago that never realy took off, such as moon landings, super sonic passanger flights, maglev trains and (to a lesser extent) hydropneumatic suspension for cars. Instead of advancing further and further, these technologies reached plateaus and either stalled altogether or were adapted to a far lesser extent than originally anticipated. The fact that a lot of money has been poured into unsuccessful emerging technologies by very intelligent people in the past should be a warning that it may be impossible to predict the development of our current emerging complex technologies like AI beyond a few years. AI might as well soon reach a plateau -- or it might take off beyond most of us can imagine -- or something in between.
- TulliusCicero 4y ago
  Those aren't really equivalent situations though. Take moon landings, for example: the most obvious problem there is that it was ridiculously expensive for no real commercial gain. It was basically a dick waving contest for the government, so while there were steady improvements in space tech after, there was no reason to keep improving at moon landings specifically, or at least no commercially compelling reason. Now, if the original moon landings had already been reasonably priced, and there was a ton of super rare resources on the moon, different story. In contrast, training machine learning models like this, while not exactly cheap, is clearly cheap-er by a few orders of magnitude. And there's obvious commercial application: near-instant art in an absurd variety of styles, that can be created by anyone, no actual art skill required. That means there's the incentive and necessary resources to fund further improvements, and at least at the moment, there doesn't seem to be an obvious technological 'wall' that we won't be able to get past (though of course the better it gets and the fewer flaws there are, the less room there is for further improvement).

Planktonne 4y ago

It's equally, if not more, utopian to assume that progress down any one route will automatically lead to the imagined pinnacle of the entire discipline. Progress isn't smooth or automatic, and we often hit dead ends, having to try new techniques to get over plateaus. It's equivalent to looking at 19th Century automobiles and asserting that they'll be able to travel to other stars by the end of the decade.

TulliusCicero 4y ago
> It's equivalent to looking at 19th Century automobiles and asserting that they'll be able to travel to other stars by the end of the decade. That's not actually equivalent at all. Looking at improvements to automobiles would mean them going faster, being more reliable, etc. Space is obviously a different domain. Your analogy is like if people were suggesting that AI image generation models were going to become advanced enough to start driving cars or folding your laundry.

Archelaos 4y ago

There is also this other common mistake with observing any burgeoning technology: the linear (or even exponential) extrapolation of early success into the future. Think of some technologies from fifty years ago that never realy took off, such as moon landings, super sonic passanger flights, maglev trains and (to a lesser extent) hydropneumatic suspension for cars. Instead of advancing further and further, these technologies reached plateaus and either stalled altogether or were adapted to a far lesser extent than originally anticipated. The fact that a lot of money has been poured into unsuccessful emerging technologies by very intelligent people in the past should be a warning that it may be impossible to predict the development of our current emerging complex technologies like AI beyond a few years. AI might as well soon reach a plateau -- or it might take off beyond most of us can imagine -- or something in between.

TulliusCicero 4y ago
Those aren't really equivalent situations though. Take moon landings, for example: the most obvious problem there is that it was ridiculously expensive for no real commercial gain. It was basically a dick waving contest for the government, so while there were steady improvements in space tech after, there was no reason to keep improving at moon landings specifically, or at least no commercially compelling reason. Now, if the original moon landings had already been reasonably priced, and there was a ton of super rare resources on the moon, different story. In contrast, training machine learning models like this, while not exactly cheap, is clearly cheap-er by a few orders of magnitude. And there's obvious commercial application: near-instant art in an absurd variety of styles, that can be created by anyone, no actual art skill required. That means there's the incentive and necessary resources to fund further improvements, and at least at the moment, there doesn't seem to be an obvious technological 'wall' that we won't be able to get past (though of course the better it gets and the fewer flaws there are, the less room there is for further improvement).