5 ms·

“Autonomous miles without accident” is not the only metric. What matters is also how the algorithm fails. To give an example. If you have autonomous miles with

by j7ake 4y ago

“Autonomous miles without accident” is not the only metric. What matters is also how the algorithm fails. To give an example. If you have autonomous miles without accident that is two times better than humans, but every accident tends to involve dark-skinned children and women, then the algorithm can have better statistical numbers in safety but still be rejected by humans. Safety is a high dimensional problem that cannot be collapsed into one metric.

spywaregorilla 4y ago
So if there's fewer deaths overall, including among black kids, but all the deaths are black kids that's not ok? I get your larger point but this one seems like a really poorly reasoned example.
- j7ake 4y ago
  It would depend on the details. If black kids are dying because the car accelerates in school zones and swerves directly into black kids at high speeds, you can imagine that would obviously be not okay no matter what the statistics are. My point is that a single metric like “deaths overall” is insufficient. The failure modes need to be reasonable and sympathetic to a human.
  - sokoloff 4y ago
    The premise of GP may have been slightly different. If a safety improvement reduces child pedestrian deaths by 100% in groups A, B, C, and D and by 50% in group E, that safety improvement should be welcomed, even though all child pedestrian deaths are now in group E (even though that group is also better off).
    - efitz 4y ago
      That is completely reasonable and would be completely unacceptable politically.
      - sokoloff 4y ago
        Perhaps it could be improved by being explicitly programmed to kill 25% as many children in groups A-D vs the original baseline. Then let activists argue why that figure should be doubled.
- paulryanrogers 4y ago
  Computer vision systems often fail to identify people with darker skin.
- throw_nbvc1234 4y ago
  That's what "equality of outcome" is; it's not about better it's about equality. When equality of outcome is implemented by raising everyone to a better outcome then it's a fine thing but that's not inherent to the underlying concept.
pwinnski 4y ago
Testing shows that Teslas specifically have trouble identifying young (short) children as people to be avoided[0]. To your point, if Tesla were able to provide in the future that they actually do result in fewer deaths, but that the deaths which remain were heavily concentrated in young children, I think people would have a serious problem with that, regardless of whether the overall count of children dying was higher or lower. It's irrational, but that's how humans are. 0. https://dawnproject.com/wp-content/uploads/2022/08/The_Dawn_Project___Tesla_FSD_Test__8_.pdf https://dawnproject.com/wp-content/uploads/2022/08/The_Dawn_...

by j7ake 4y ago

spywaregorilla 4y ago

So if there's fewer deaths overall, including among black kids, but all the deaths are black kids that's not ok? I get your larger point but this one seems like a really poorly reasoned example.

j7ake 4y ago
It would depend on the details. If black kids are dying because the car accelerates in school zones and swerves directly into black kids at high speeds, you can imagine that would obviously be not okay no matter what the statistics are. My point is that a single metric like “deaths overall” is insufficient. The failure modes need to be reasonable and sympathetic to a human.
- sokoloff 4y ago
  The premise of GP may have been slightly different. If a safety improvement reduces child pedestrian deaths by 100% in groups A, B, C, and D and by 50% in group E, that safety improvement should be welcomed, even though all child pedestrian deaths are now in group E (even though that group is also better off).
  - efitz 4y ago
    That is completely reasonable and would be completely unacceptable politically.
    - sokoloff 4y ago
      Perhaps it could be improved by being explicitly programmed to kill 25% as many children in groups A-D vs the original baseline. Then let activists argue why that figure should be doubled.
paulryanrogers 4y ago
Computer vision systems often fail to identify people with darker skin.
throw_nbvc1234 4y ago
That's what "equality of outcome" is; it's not about better it's about equality. When equality of outcome is implemented by raising everyone to a better outcome then it's a fine thing but that's not inherent to the underlying concept.

pwinnski 4y ago

Testing shows that Teslas specifically have trouble identifying young (short) children as people to be avoided[0]. To your point, if Tesla were able to provide in the future that they actually do result in fewer deaths, but that the deaths which remain were heavily concentrated in young children, I think people would have a serious problem with that, regardless of whether the overall count of children dying was higher or lower. It's irrational, but that's how humans are. 0. https://dawnproject.com/wp-content/uploads/2022/08/The_Dawn_Project___Tesla_FSD_Test__8_.pdf https://dawnproject.com/wp-content/uploads/2022/08/The_Dawn_...