4 ms·

Seems like a better testing suite is in order. How many times since then has that happened, and have we gotten better at figuring out issues before they become

Seems like a better testing suite is in order. How many times since then has that happened, and have we gotten better at figuring out issues before they become this big of an issue. I ask this because we've had everything from chernobyl to the Columbia rocket explosion. We didn't hamstring those things. We got better at anticipating them.

johnnujler 6y ago
Remember, it doesn't matter if you've survived 100 Chernobyls, 50 Tsunamis or 20 Covids, all it takes is one bad one to wipe you out of the gene pool(kill). The worse is called worse precisely because it did not happen until then. Yes, testing helps but only when you know what to test, a hastily designed system where the workers are put under pressure and made to ssh into a fleet(See: https://twitter.com/atomicthumbs/status/1032939617404645376 https://twitter.com/atomicthumbs/status/1032939617404645376) to fix bugs, is not a mere matter of testing. It is a matter of culture, autonomy, prudence, care, and add other things that would help build a well-designed system from ground up.
DrScump 6y ago
The issue isn't size of the testing suite, it's covering all the edge cases. Some rather common edge cases go ignored. Unless they've updated Waymo in the last couple of months, they still fail to distinguish when they have a reserved destination lane when making right turns, as exhibited when going from eastbound Central to southbound Castro every damn day (before Castro was closed). Self driving system vendors can claim that fewer deaths will occur with their system vs the average driver, but that mostly reflects how dangerous the "average"/distracted driver is. Those numbers are little comfort to the loved ones of someone killed in an "edge case" that a competent driver handles without injury.
Balgair 6y ago
> Seems like a better testing suite is in order The cry of every neuroscientist! The trick is how to make an ex vivo brain simalog at all. Currently we use the brains of other animals. It gets you somewhere. I mean, you can deal with how the heart moves the brain about, how grooming and eating move the brain, you can even get seizures and ischemic events working with the tech. But, those last few percent is where all of it matters. Kinda like viewing an eclipse: the difference between 99.99% and 100% totality is everything. And that's kinda the problem. Figuring out all the edge cases is really hard and time consuming. And making a faux-brain simalog is probably harder than just going slowly.

johnnujler 6y ago

Remember, it doesn't matter if you've survived 100 Chernobyls, 50 Tsunamis or 20 Covids, all it takes is one bad one to wipe you out of the gene pool(kill). The worse is called worse precisely because it did not happen until then. Yes, testing helps but only when you know what to test, a hastily designed system where the workers are put under pressure and made to ssh into a fleet(See: https://twitter.com/atomicthumbs/status/1032939617404645376 https://twitter.com/atomicthumbs/status/1032939617404645376) to fix bugs, is not a mere matter of testing. It is a matter of culture, autonomy, prudence, care, and add other things that would help build a well-designed system from ground up.

DrScump 6y ago

The issue isn't size of the testing suite, it's covering all the edge cases. Some rather common edge cases go ignored. Unless they've updated Waymo in the last couple of months, they still fail to distinguish when they have a reserved destination lane when making right turns, as exhibited when going from eastbound Central to southbound Castro every damn day (before Castro was closed). Self driving system vendors can claim that fewer deaths will occur with their system vs the average driver, but that mostly reflects how dangerous the "average"/distracted driver is. Those numbers are little comfort to the loved ones of someone killed in an "edge case" that a competent driver handles without injury.

Balgair 6y ago

> Seems like a better testing suite is in order The cry of every neuroscientist! The trick is how to make an ex vivo brain simalog at all. Currently we use the brains of other animals. It gets you somewhere. I mean, you can deal with how the heart moves the brain about, how grooming and eating move the brain, you can even get seizures and ischemic events working with the tech. But, those last few percent is where all of it matters. Kinda like viewing an eclipse: the difference between 99.99% and 100% totality is everything. And that's kinda the problem. Figuring out all the edge cases is really hard and time consuming. And making a faux-brain simalog is probably harder than just going slowly.