I do think that self-driving cars will be safer, but it's upon it's proponents to prove that.

Each of the crashes that self-driving cars can be fixed and prevented from happening again. The list I gave are human flaws that will almost certainly never be fixed.

I further agree with you it's up to the proponents to prove that. It's a good thing to force a really high bar for self-driving cars. Then assuming the technology is maintained once AI passes the bar it should only ever get better.

Not if you put neural networks / deep learning in the equation. This stuff is black boxes connected to black boxes, that work fine until they don't, and then nobody knows why they failed - because all you have is bunch of numbers with zero semantic information attached to them.

The only questionable parts will be where the vision system fails, and those are similar actually to human problems. Because human vision also often fails (sunlight on windshield, lack of attention, darkness, etc.)

Are you in very vague words implying that AGI has been invented? AI might have matched humans in image recognition, but it is far away in general decision making.

And finally, I am tired of listening to "safer than a human". That should never be the comparison, but a human at the helm and an AI running in the background which will take over when the human does an obvious mistake -- you know, like a emergency braking system,

If those situations recur exactly as they happened the first time, sure they can be prevented from happening again.

That is, if a car approaches the exact same intersection as the exact same time of day, and a pedestrian that looks exactly like the pedestrian in this accident crosses the street in exactly the same way, with exactly the same other variables (like all the other pedestrians and cars around there that the sensors can see), the data could be enough the same that the algorithm will detect it at close enough to the original situation to avoid the accident this time.

But it's not at all clear how well their improvements will generalize to other situations which humans would consider to be "the same" (ie. when any pedestrian in any intersection crosses any street).

If self-driving cars are, at their best, roughly as capable as a human driver.

This is a big 'if'.

The solution to not killing people is a kind of rocket science. In fact, it's probably harder than rocket science[0]. It's predicated on a lot of things that are very, very, very hard. The fact is that humans, who are already pretty capable of most of these very very hard things, often choose to reduce their own capabilities.

If the best self-driving tech is no better than a drunk human, however, then we haven't gained much.

---

[0] though perhaps not harder than brain surgery.

However, the process your describing, of collecting human-level performance data, requires the ability to gather all of the data relevant to the act of driving in a manner consumable by the algorithm in question. This is the simulation problem, and it's very, very, very hard (it's why genetic algorithms have traditionally not gotten much further than toy examples, in spite of being a cool idea). Perhaps it is the case that it is very important to have an accurate model of the intentions of other agents (e.g., pedestrians) in order to take preventative action rather than pure reaction. Perhaps it is very important to have a model of what time of day it is, or the neighborhood you're driving in. The likelihood that it is going to rain some time in the next hour. Whether the stock market closed up or down that day.

It also assumes that neural networks (or the more traditional systems used elsewhere) are sufficiently complex to model these behaviors accurately. Which we do not yet have an answer to yet.

So, when I say, 'a big if', I mean for the foreseeable future, barring some massive technological/biological breakthrough. That could be a very long time.