[discodave]

From the article:

> But zooming out from the specifics of Herzberg's crash, the more fundamental point is this: conventional car crashes killed 37,461 in the United States in 2016, which works out to 1.18 deaths per 100 million miles driven. Uber announced that it had driven 2 million miles by December 2017 and is probably up to around 3 million miles today. If you do the math, that means that Uber's cars have killed people at roughly 25 times the rate of a typical human-driven car in the United States.

I don't think there's enough evidence to say that self-driving cars are as safe as humans.

[jon_richards]

To be fair, you have to keep in mind that deploying 1000 cars would quickly make self-driving cars safer than humans. Yes, you lose some life in the interim, but you accelerate the creation and adoption of something that saves much more life in the long run.

But that is human experimentation, something we as a culture generally agree is abhorrent.

[AlexandrB]

> To be fair, you have to keep in mind that deploying 1000 cars would quickly make self-driving cars safer than humans.

This seems unreasonably optimistic.

First, this particular crash is an egregious counter-example. The car doesn't even seem to slow down when it first sees the pedestrian's foot. Nor does it try to swerve. This is basic stuff for a human driver, never mind more complex avoidance and risk mitigation a human driver can perform.

Second, we've had years of training various AI content curation algorithms on social networks, videos, blogs, etc - and the most advanced AI and search company in the world still can't keep adult-oriented conspiracy videos off of Youtube Kids. And while you might counter that content is a human problem, driving is too! Dangerous driving situations happen at the periphery of traffic rules, where someone is doing something the drivers around him don't expect. I've seen people run solid red lights, drive the wrong way down a one-way street, pedestrians start crossing when their light turns red, etc. Short of creating special roads for autonomous cars only - how does an autonomous car deal with all of this successfully?

[bonesss]

> the most advanced AI and search company in the world still can't keep adult-oriented conspiracy videos off of Youtube Kids

In fairness here: Google could do this so easily by throwing people at the problem. A whitelisted set of content producers could do this, with people dedicated to whitelist curation. Google can keep them off, and quite readily. Keeping that off automatically and cheaply is the issue...

On that front: automated content recognition that can parse the nuances between regular claymation Elsa and spooky claymation Elsa who talks about jamming things in her "happy spot" is AI-hard. These are not random videos, they are content tailor made and adapted to pass whatever filters are in place.

For data scientists there is a massive gap between working with concrete sensor data that can be managed reasonably, and undefined philosophical/moral/sexual boundaries in dirty noisy counter-programmed video content... One replaces our eyes and ears and reflexes, the other seeks to replace our fabulous brains.

[culturestate]

> First, this particular crash is an egregious counter-example. The car doesn't even seem to slow down when it first sees the pedestrian's foot. Nor does it try to swerve. This is basic stuff for a human driver, never mind more complex avoidance and risk mitigation a human driver can perform.

I'm not sure which side of this I fall on just yet, but something strikes me here: you're assuming a human driver with no impairment (e.g. eyesight or fatigue) paying complete attention to what they're doing. We know that this isn's always the case (and could even be a minority!), so this doesn't seem to be a great argument.

[sunir]

You shouldn’t be driving so fast that stopping distance is longer than visibility.

It is irrelevant what average humans do. That is the current rule set for all drivers.

[blt]

No. Machine learning is not magic. Machine learning has nothing to say when the objective function is not well understood.

[fourthark]

This is assuming that the learning actually happens.