If that’s what it sees when it’s still, idk how to trust it when it’s moving.
I mean, try this yourself! Pull up to a stop light, close your eyes, and recite the distances to and orientations of the nearest five or six vehicles. Now read it off the screen. Think you'll beat the car? Really? In fact, no, it's quite clear that the car, with 360 degrees of camera coverage at 30 Hz and no fovea or neck scanning required, has a vastly better picture of its surroundings than the driver. It's not even close.
But yeah, the esimates aren't precise and yeah, they do jump around[1]. You're absolutely correct about the way the modelling works. You're just disastrously wrong about the way that the competing systems perform.
[1] I will grant that there's probably a human factors issue here. In fact the human driver really can't do much to process that display and it's probably best that details like just be left out or filtered down to a minimal subset.
Other commenter "My coworker gave me a ride in his Tesla ... FSD tried to cross over into the opposing lanes, ran a red, narrowly avoided hitting a cyclist"
Your reply "Huh. FSD (Mine) took my kid to diving practice this afternoon. Weird world. Just please stop with the hate."
I'm not sure what you hope to prove by replying to anecdotes about unacceptable autopilot with anecdotes about acceptable autopilot? It doesn't mean everyone else is a hater any more than you're a shill. It just means autopilot performance across hardware and environmental factors is not within the envelope it needs to be.
Maybe try replacing "FSD" in these comments with the names of two fictional friends.
"Steve tried to cross over into the opposing lanes, ran a red, narrowly avoided hitting a cyclist"
"Huh. Rick took my kid to diving practice this afternoon. Weird world. Just please stop with the hate."
Should we trust Steve because Rick is a good driver?
Speaking of risk acceptance, I wouldn't trust my life or anyone else's to FSD. I don't have FSD in my Model 3, but I have used all of the other driver assist features. There have been too many times when those features have made serious mistakes, so I don't use them. Even at it's best, FSD appears to be about as good as the worst drivers on the road. One anecdote of poor performance, for me, is one too many.
We should accept that all drivers make mistakes, even you, and demand measurable safety metrics from all our control systems. Is that really so weird?
What we should not do is start thousand-comment flame wars on HN about how Steve is unsafe and that we should take away his license. In fact, those three mistakes are of course routine for human drivers. I bet you personally, statop, have done them in the past. And yet, no lynch mob.
That's the "hate" part I'm talking about. Please stop. Please. It's really hurting the discourse about a product area that's actually very subject to measurement, and about a specific product that is objectively very safe.
