5 ms·

Yeah, I recall Waymo bragging they could do this years ago... a few months before a Google engineer reported to a Slate reporter that a Google Self-Driving Car

Yeah, I recall Waymo bragging they could do this years ago... a few months before a Google engineer reported to a Slate reporter that a Google Self-Driving Car would happily run any red light that wasn't in its map in advance... The long tail of edge cases in self-driving is vast, and each one needs to be individually addressed.

giancarlostoro 6y ago
So long as it requires an internet or similar connection I think its mainly useless. I rather a self driving car that works whether Google is damned and none existent or not. That would be impressive, you download OSM or better maps and terrain details of the world and the car based on where it travels knows where you are despite not getting a GPS signal cause reasons. We should be able to do this sort of fauly tolerance in software.
vmception 6y ago
> The long tail of edge cases in self-driving is vast, and each one needs to be individually addressed. But when a single car encounters an issue, all cars get updated at once. Just keep solving those captchas people, they really need to know what a crosswalk looks like.
- perl4ever 6y ago
  My searches produce an inordinate number of captchas, thus proving that Google can't identify humans, even in the very limited environment of Google searches. Which doesn't bode well for the cars. It appears that reasonably rapid use of inurl: makes it think you're a bot.
  - vmception 6y ago
    Captcha's aren't to identify humans, they are to train AI. It is easy for me to understand the tolerance level, I know which images it understands and which images it likely doesn't. I can easily finish captchas without selecting all images of crosswalks. Anyway, your point seems to have the purpose mixed up. Pointing out images of crosswalks trains the same system that some cars use or will use.
    - perl4ever 6y ago
      >Captcha's aren't to identify humans I mean, that is their primary purpose, and there's a message that says so. Unless you are saying "they aren't to identify humans, they're to exclude bots" in which case I'm not sure what that means.
      - vmception 6y ago
        their implementation is such that it is more to train AI, the reason I think this way is that you can predict how much of a Captcha that you can get wrong, because you know which parts it doesn't know the answer to. It is presented as "select the right answers, we know the right answers just like your teacher did in school, if you don't select the right answers you are a robot" but in reality is is "select the right answers, so that we learn the right answers, the right answers are an average of what other people selected as well as completely new information that we aren't sure about"
        drewbug01 6y ago
        The purpose of a captcha, and the purpose of the implementation of a captcha don’t need to be the same thing. Said differently, a given thing can fulfill multiple roles and that isn’t an inherent contradiction nor is it wrong. In this case: yes, the stated purpose of a captcha is to identify humans - ie, legitimate users. The unstated purpose of ReCaptcha (at least) is to train AI. But it does both simultaneously. The funniest thing to me is that we’re all in on it. :)
- dogma1138 6y ago
  That’s assuming cars could update their model or that the new model won’t still have the same edge case but now it will be even more hidden. This is the issue with blackboxes while you might be able to figure out which model exactly in your system caused a weird behavior figuring out what exactly in it that caused it and then what training data was responsible is currently near impossible afaik.
dogma1138 6y ago
The more you add to this the more edge cases you create, if you for example add a model that when a pedestrian is recognized near the edge of the side walk the car then tries to figure out their intent rather than always slowing down or signaling the driver to take over you aren’t resolving an edge case you are creating a bunch of new ones.
- drewbug01 6y ago
  Honestly, this is a good example for why our industry shouldn’t be pursuing this particular tech. Tech (writ large) thrives on shipping incomplete solutions and iterating on them - for better or worse. And that’s not something that’s acceptable for self-driving cars. Better to put our efforts towards other things, in my opinion. Personal autos aren’t sustainable even if they drive themselves. Feels like this is the sunk-cost fallacy (or as I prefer, “throwing good money after bad”) but played out across the intersection of two industries.

giancarlostoro 6y ago

So long as it requires an internet or similar connection I think its mainly useless. I rather a self driving car that works whether Google is damned and none existent or not. That would be impressive, you download OSM or better maps and terrain details of the world and the car based on where it travels knows where you are despite not getting a GPS signal cause reasons. We should be able to do this sort of fauly tolerance in software.

vmception 6y ago

> The long tail of edge cases in self-driving is vast, and each one needs to be individually addressed. But when a single car encounters an issue, all cars get updated at once. Just keep solving those captchas people, they really need to know what a crosswalk looks like.

perl4ever 6y ago
My searches produce an inordinate number of captchas, thus proving that Google can't identify humans, even in the very limited environment of Google searches. Which doesn't bode well for the cars. It appears that reasonably rapid use of inurl: makes it think you're a bot.
- vmception 6y ago
  Captcha's aren't to identify humans, they are to train AI. It is easy for me to understand the tolerance level, I know which images it understands and which images it likely doesn't. I can easily finish captchas without selecting all images of crosswalks. Anyway, your point seems to have the purpose mixed up. Pointing out images of crosswalks trains the same system that some cars use or will use.
  - perl4ever 6y ago
    >Captcha's aren't to identify humans I mean, that is their primary purpose, and there's a message that says so. Unless you are saying "they aren't to identify humans, they're to exclude bots" in which case I'm not sure what that means.
    - vmception 6y ago
      their implementation is such that it is more to train AI, the reason I think this way is that you can predict how much of a Captcha that you can get wrong, because you know which parts it doesn't know the answer to. It is presented as "select the right answers, we know the right answers just like your teacher did in school, if you don't select the right answers you are a robot" but in reality is is "select the right answers, so that we learn the right answers, the right answers are an average of what other people selected as well as completely new information that we aren't sure about"
      - drewbug01 6y ago
        The purpose of a captcha, and the purpose of the implementation of a captcha don’t need to be the same thing. Said differently, a given thing can fulfill multiple roles and that isn’t an inherent contradiction nor is it wrong. In this case: yes, the stated purpose of a captcha is to identify humans - ie, legitimate users. The unstated purpose of ReCaptcha (at least) is to train AI. But it does both simultaneously. The funniest thing to me is that we’re all in on it. :)
dogma1138 6y ago
That’s assuming cars could update their model or that the new model won’t still have the same edge case but now it will be even more hidden. This is the issue with blackboxes while you might be able to figure out which model exactly in your system caused a weird behavior figuring out what exactly in it that caused it and then what training data was responsible is currently near impossible afaik.

dogma1138 6y ago

The more you add to this the more edge cases you create, if you for example add a model that when a pedestrian is recognized near the edge of the side walk the car then tries to figure out their intent rather than always slowing down or signaling the driver to take over you aren’t resolving an edge case you are creating a bunch of new ones.

drewbug01 6y ago
Honestly, this is a good example for why our industry shouldn’t be pursuing this particular tech. Tech (writ large) thrives on shipping incomplete solutions and iterating on them - for better or worse. And that’s not something that’s acceptable for self-driving cars. Better to put our efforts towards other things, in my opinion. Personal autos aren’t sustainable even if they drive themselves. Feels like this is the sunk-cost fallacy (or as I prefer, “throwing good money after bad”) but played out across the intersection of two industries.