My point wasn't that I had contradicting data. My point was that the data Tesla publishes isn't valuable. It's a different statement. The premise could be right or wrong.
> My point was that the data Tesla publishes isn't valuable.
That's not true at all, and you know it. Consider the extremes: if autopilot got in zero accidents over these however many million miles of travel, you'd have to admit it was safe to use, right? If it got in 10x as many accidents as the median vehicle, we'd all agree it was a deathtrap. There's value in those numbers. Well, the truth is somewhere in between, it gets in 4x fewer accidents (I think) than the median car. That still sounds good to me.
You're just nitpicking. You're taking a genuine truth (that the variables aren't independent because "miles travelled with AP" aren't randomly sampled) and trying to use it to throw out the whole data set. Probably because you don't like what it says.
But you can't do that. The antidote to incomplete or confounded analysis is better analysis. And no one has that. And I argue that if there was a signal there that says what you want ("Tesla AP is unsafe") that it would be visible given the number of cars on the road. And it's not.
Stop nitpicking and get data, basically. The time for nitpicking is when you don't have data.
Yes, a very strong signal (everytime AP is turned on the Tesla explodes or everytime AP is turned on the Tesla drives perfectly from parking spot at origin to parking spot at desitnation, and there has not been a crash in 10 billion miles) can overcome the noise in Tesla's published data. However, any signal there is not that strong and is so weak it is drowned in noise.
Meanwhile, I do have this one datum about how OTA updates caused a couple of accidents (see OP's story.) You cannot just dismiss it because it doesn't agree with a larger trend, if that larger trend is undetectable in the data.
You're just making stuff up. That's not a weak signal at all, and you know it. The argument you should be making is that it's a confounded signal. But that requires a level of sophistication that you aren't supplying (nor is anyone else). Which is exactly why I mentioned way upthread that these digressions were so tiresome.
There's literally no data in the world that will satisfy you, just admit it. Yours is an argument from priors only, to the extent that one anecdote ("A Tesla hit me!") is enough to outweigh a report like that.
