4 ms·

"AI's", even in their specialized domain of pattern recognition, still can't distinguish a corgi from a loaf of bread or specially designed white noise. The job

by mjfl 8y ago

"AI's", even in their specialized domain of pattern recognition, still can't distinguish a corgi from a loaf of bread or specially designed white noise. The jobs he listed that are "in danger": truck driving, dish washing, fruit picking both require serious advances in robotics and do not require advances in general AI.

danans 8y ago
> require serious advances in robotics and do not require advances in general AI Which if anything makes the replacement of workers in those domains even more tractable in the near future, wouldn't you say? Dishwashing is an interesting case, because, assuming non-standardized dishes, glasses, and utensils, it requires solving a very general object recognition problem for a robot arm to handle different sizes and shapes of dishes. Andrew Ng's lab at Stanford was working on it some years ago, but I'm not sure if they solved it more generally: http://www.andrewng.org/portfolio/stair-stanford-artificial-intelligence-robot/ http://www.andrewng.org/portfolio/stair-stanford-artificial-... https://www.youtube.com/watch?v=SbipWOeqF1c https://www.youtube.com/watch?v=SbipWOeqF1c
- anonymous5133 8y ago
  And more importantly, if Andrew ng didn't figure it out then surely someone else will. The march of progress does not stop. It continues relentlessly. This is the mindset I see many fail at (less so on here since people here are smarter and more visionary). People see robotics and AI for what it is today and not what it will be in the future. The robotics and AI of today have their issues but in due time those issues will be solved to create something truly impressive. Truck driving, dish washing, order fullfilment and the myriad of other routine tasks will eventually be fully automated. Humans will be left to continue the march of progress as we should be doing.
  - oldgradstudent 8y ago
    Sounds more like a principle of faith than an empirical observation > I believe with perfect faith in the coming of the Messiah, and, though he tarry, I will wait daily for his coming.
- zhte415 8y ago
  Dishwashing can surely be done/automated now? A restaurant standardises a range of dish dimensions, these are mechanically sorted on a 'line' then cleaned according to their specification and surface scanned for error/reclean/escalation. No AI needed, decades old tech suffices.
  - oldgradstudent 8y ago
    > No AI needed, decades old tech suffices. This can be said of quite a lot of startup companies these days. Compare the Zume robot shown in this video https://www.bloomberg.com/news/articles/2018-08-07/softbank-in-talks-to-invest-up-to-750-million-in-zume-the-startup-that-sells-robot-made-pizza https://www.bloomberg.com/news/articles/2018-08-07/softbank-... to any number of videos showing factories making pizza such as https://www.youtube.com/watch?v=jdWftLcvmGM https://www.youtube.com/watch?v=jdWftLcvmGM
  - danans 8y ago
    In large-enough scale food serving operations (i.e. hotels) I imagine it is already automated to a large extent. It's the domestic dishwashing use case that Andrew Ng's robot was trying to solve. For smaller restaurants, the capital cost of standardized automation may be significantly higher than the marginal cost of low-wage dishwashing labor, who also can accomplish other tasks around the restaurant, like bussing, prepping, and cleaning, and therefore might not be worth it.
BucketSort 8y ago
"can't distinguish a corgi from a loaf of bread or specially designed white noise" You are talking about adversarial examples, which are in most cases extremely rare and not detrimental to the typical function of the image recognition system. Even humans have problems identifying some of these adversarial examples. The white noise is particularly troublesome from a technical standpoint. It goes to show that these learned manifolds are not perceptual manifolds.
- oldgradstudent 8y ago
  A road fatality is also an extremely rare event, once in a hundred lifetimes[1]. AI has to deal with extremely rare events to be better than human drivers. [1] average miles driven per year - 12,000, while a fatality happens every 86 million miles on average.

by mjfl 8y ago

danans 8y ago

> require serious advances in robotics and do not require advances in general AI Which if anything makes the replacement of workers in those domains even more tractable in the near future, wouldn't you say? Dishwashing is an interesting case, because, assuming non-standardized dishes, glasses, and utensils, it requires solving a very general object recognition problem for a robot arm to handle different sizes and shapes of dishes. Andrew Ng's lab at Stanford was working on it some years ago, but I'm not sure if they solved it more generally: http://www.andrewng.org/portfolio/stair-stanford-artificial-intelligence-robot/ http://www.andrewng.org/portfolio/stair-stanford-artificial-... https://www.youtube.com/watch?v=SbipWOeqF1c https://www.youtube.com/watch?v=SbipWOeqF1c

anonymous5133 8y ago
And more importantly, if Andrew ng didn't figure it out then surely someone else will. The march of progress does not stop. It continues relentlessly. This is the mindset I see many fail at (less so on here since people here are smarter and more visionary). People see robotics and AI for what it is today and not what it will be in the future. The robotics and AI of today have their issues but in due time those issues will be solved to create something truly impressive. Truck driving, dish washing, order fullfilment and the myriad of other routine tasks will eventually be fully automated. Humans will be left to continue the march of progress as we should be doing.
- oldgradstudent 8y ago
  Sounds more like a principle of faith than an empirical observation > I believe with perfect faith in the coming of the Messiah, and, though he tarry, I will wait daily for his coming.
zhte415 8y ago
Dishwashing can surely be done/automated now? A restaurant standardises a range of dish dimensions, these are mechanically sorted on a 'line' then cleaned according to their specification and surface scanned for error/reclean/escalation. No AI needed, decades old tech suffices.
- oldgradstudent 8y ago
  > No AI needed, decades old tech suffices. This can be said of quite a lot of startup companies these days. Compare the Zume robot shown in this video https://www.bloomberg.com/news/articles/2018-08-07/softbank-in-talks-to-invest-up-to-750-million-in-zume-the-startup-that-sells-robot-made-pizza https://www.bloomberg.com/news/articles/2018-08-07/softbank-... to any number of videos showing factories making pizza such as https://www.youtube.com/watch?v=jdWftLcvmGM https://www.youtube.com/watch?v=jdWftLcvmGM
- danans 8y ago
  In large-enough scale food serving operations (i.e. hotels) I imagine it is already automated to a large extent. It's the domestic dishwashing use case that Andrew Ng's robot was trying to solve. For smaller restaurants, the capital cost of standardized automation may be significantly higher than the marginal cost of low-wage dishwashing labor, who also can accomplish other tasks around the restaurant, like bussing, prepping, and cleaning, and therefore might not be worth it.

BucketSort 8y ago

"can't distinguish a corgi from a loaf of bread or specially designed white noise" You are talking about adversarial examples, which are in most cases extremely rare and not detrimental to the typical function of the image recognition system. Even humans have problems identifying some of these adversarial examples. The white noise is particularly troublesome from a technical standpoint. It goes to show that these learned manifolds are not perceptual manifolds.

oldgradstudent 8y ago
A road fatality is also an extremely rare event, once in a hundred lifetimes[1]. AI has to deal with extremely rare events to be better than human drivers. [1] average miles driven per year - 12,000, while a fatality happens every 86 million miles on average.