Tesla runs their new releases for millions of miles on the wide fleet in shadow mode before they release to limited beta. This is not an untested release in the way the article makes it seem.

That's not as useful as it might sound. If a particular action the driver didn't take would lead to results that didn't happen, then shadow mode will have a lot of trouble knowing that. In general it's a problem called "off policy evaluation" and it's really hard. Tesla has to estimate what would have happened counterfactually, had the bot been driving, instead of seeing what happens with the bot driving. A million miles of this shadow mode evaluation isn't worth nearly as many miles of actual self driving.

> If a particular action the driver didn't take would lead to results that didn't happen

I'm having a hard time parsing your double-negative, but when shadow mode is running, it's choices are back-tested in real-time with the driver's choices, so I'm not sure how this fails to replicate what a real-driver would have done in such scenarios.

Here an example. I’m on the freeway in my Tesla and instead of staying behind a car, at a safe following distance as predicted, I change lanes and speed past. Was it because, 1) the car was doing something I perceived as dangerous and I moved to a safer position, 2) I’m in a freaking Tesla and I can

Tesla's shadow mode is collecting real data up until the point where the car's decisionmaking process and the human driver disagree. Everything after that is simulated fantasy. And since human drivers aren't necessarily going to do "the right thing" in all circumstances (there isn't necessarily a single 'right thing') you can't treat 'car-disagrees-with-human' as an error case.

Other cars respond to your actions, so the moment your actions deviate from the bot the reactions of the other cars are no longer valid to test against.

Backtesting a driving model, it’s very hard to tell what would have happened, had the car deviated at all from what the driver did.

e.g. You see the driver swerve for no apparent reason, did the driver simply drive more recklessly than the car would, or did they avoid an obstacle the model didn’t catch?

And yet there are several recent videos of Tesla cars doing crazy stuff on the roads with this system.

Wait till you see the videos of humans doing crazy stuff on roads in cars!

If humans did the crazy stuff on the road at the same rate as we've seen from the few select hours of FSD driving uploaded by Tesla owners (who are fans and more likely trying to show it at it's best) this week, we'd all be dead.

Do people randomly brake at high speeds for no reason ? Tesla cars do:

https://m.youtube.com/watch?v=xj73Q3lWvFM

All the time. Witness the effects of cellphones/children/food/spilled food/makeup/jarring news/exciting music/significant others/attractive people/gawking at random bits of interest. Also, brake-checking just cause' and for insurance fraud.

If what you're proposing is that "success" in this experiment is that every Tesla now drives at least as well as the absolute worst-case human driver, then you've pretty much articulated why this experiment is so dangerous.

The thing is these are not the "absolute worst-case human driver", these are the absolute average human drivers.

People get easily distracted, get excited, get tired, all of which are propitious to bad decision-making.

Many human drivers are downright dangerous, but that's not a reason to add dangerous machine drivers on the road, either.

The adaptive cruise control in my Honda is also bad about braking without good cause. Of course, they don’t market it as “self-driving”.

Automatic emergency braking also has false positives.

Who is legally responsible for the car doing something crazy? Will engineers or executives go to jail?

How does that explain the persistent issue with phantom breaking ?

https://m.youtube.com/watch?v=xj73Q3lWvFM

It's been a problem for years now.

And it's scary as hell. Had it happen to me twice in a 600-mile highway roundtrip this summer. Was very happy nobody was following me closely, because the combination of regen breaks and disc breaks really stops the car quickly.

Still magnitudes safer than human drivers.

Drunk drivers and human mistakes are two things I accept in traffic (or rather - laws and human behavior both show we don’t accept them). It’s factored into my risk.

And FSD isn’t orders of magnitude safer. They are barely as good as human drivers and only in some conditions.

For the worst situations they aren’t comparable at all (since they won’t drive at all)

This must be repeated over and over for some reason: people will never accept self driving cars that are just as good or bad as human drivers. Nor should we. It’s a much too low bar.

Self driving cars will need to be orders of magnitude safer than human drivers to be even remotely acceptable.

If it's consistently as good as an average human that's still better than an average human. That's because robots don't get tired, hungover, angry, bored, drunk, or get an eyelash in their eye at a really bad time.

Yes. But again people won’t accept that. People will rather be killed more often by drivers that feel guilt, go to prison, have strokes or poor eyesight than less often by a machine that does not have all of those flaws but also none of the feelings and responsibilities.

I’d be thrilled if self-driving cars could meet that bar, because then I can feel at least as comfortable napping in the car as I do driving it. And unlike humans, I have the expectation that my car learns when other cars crash, so I have a lot of reason to expect that once the system rolls out it will continue to improve.

I hate almost everything about Tesla and their business model but the idea of full automatic driving, once vetted to even a bare minimum level of approximate human parity, is enough to make me consider one anyway.

Lets do a simple test:

You are in between two walls and there are two straight lanes. Lane 1 has Tesla FSD coming in at full speed. Lane 2 has a human driver coming in at full speed. You can choose to be in one lane. The FSD and the human can both see you only 50 mts out and if they brake as soon as they see you, you will survive.

Which lane will you choose?

Well, a Model S can just barely stop in 50 meters at 60mph, while some other luxury and sporty cars in the segment can stop in 35 meters or less. And that’s only 60mph, hardly “top speed” or even a normal highway speed. So regardless of driver assistance features, I’d choose a car that at least has a chance to even physically stop assuming instant reflexes. 50 meters is probably a much too small distance for this thought experiment.

I don't know my cars that well, but an alert driver will take at least 300ms to begin reacting. That's over 8m at 60mph (not sure why we're combining unit systems here). So I'll stand in front of the Model S in this particular scenario and hope to get away with 'mere' broken legs/pelvis. Give the human more time and is probably take the human, not because that's necessarily the more reliable option, but because its distribution of outcomes is more well known to me.

Most, if not essentially all, cars in that category have emergency automatic braking, so the Model S still isn't the best choice.

It absolutely isn't.