Hacker News Reader

Comment by 0F | Hacker News Reader

It’s crazy how to truth can hide in plain site. This guys channel had raw footage of FSD driving through terrain deliberately chosen to be difficult and FSD absolutely shined. Just flicking through the channel it’s immediately apparent that FSD is the most advanced self driving system by a mile. But sometimes it made mistakes. And literally any time I have ever seen a video of FSD float to the top of a link aggregator site, it’s a video of one of the mistakes. If you didn’t deliberately look through the archives yourself, you might not ever know that FSD does anything besides make mistakes.

View on HN

That's pretty logical: 99.9% of self driving is easy. That last .1% is very, very hard.

You have a weird definition of easy :D But I agree with the 0.1%.

The reason the failure videos float to the top is because failure in self-driving can result in serious bodily harm for those that have little to no control of whether they’re participating in a review or test.

Make that simply driving. Human do not necessarily make the same mistakes as computer vision and deep learning systems, but they make mistakes too.

I genuinely don't think this is true. FSD is better than a lot of people make it out to be, but I'm pretty sure they're still leagues behind Cruise who is behind Waymo.

I don't think Tesla publishes their data, but from the 2021 report [1] Cruise is at 42,022 miles/disengagement and Waymo is at 7,800. In fairness, Waymo switched their car base car, which seems to be causing issues. Waymo was at 29,944 miles/disengagement in 2020.

The videos I've seen of Teslas appear to have a much lower miles/disengagement number. Even if we presume they're going 60mph the whole time (they're not), they would need to have 1 disengagement per 130 hours of video just to keep up with Waymo's new, much lower stats. They'd need 1 disengagement per 700 hours of video to keep up with Cruise. With a more realistic estimate of an average 15mph, you'd have to quadruple those numbers.

There's some wiggle room, because disengagements and whether they're reportable is somewhat subjective. The delta seems large enough to say that Tesla isn't particularly close, though.

1. https://thelastdriverlicenseholder.com/2022/02/09/2021-disen... (the dataset is available from the CA DMV, but it's in CSV so I used this because graphs)

Yes it can drive, but it does so on a very sketchy way. Yes it's advanced, but not nearly advanced enough. The worst part, sometimes you can't react fast enough if the car makes a mistake.

Getting this beta off the roads cannot happen quickly enough. I don't want to risk my life for the gain of some company. I don't care if FSD drives 99% correctly, the 1% is what can cost lives.

If you look careful, you also see FSD is not a smooth drive at all. I'm on the edge of the seat all the time and people who are comfortable in it, are because of ignorance.

But by doing the beta test, we can get a safe FSD out sooner, and it doesn’t take much to be safer than the average driver. So even though running a beta on public roads has its risks, it might save lives in the long run if it means we get an effective FSD out 5 years sooner for example.

I seen many fanboy made videos where shit just works, sometimes the fanboys live stream and the mistakes can't be hidden. When we get full data transparency( where full here mean 100% and not the fake "full" from the Tesla dictionary) just want to know if Tesla's PR stats include the mistakes too when they calculate safety , what I mean is if each time a human had to intervene is counted as a crash , and how corners or roads where the user has no choice then take over are also counted. Is like F in FSD seems to be mathematically { F, the subset of everything where our shit worked(we intentionally exclude stuff and have our fanboys defend them as micro mistakes or blame the roads, or blame the driver or blame anyone else/ in our dictionary FSD and autopilot means whatever we want to mean and yes the naming is very clever to be super missleading because $$$ ) }

I price motor insurance for a living and in my models the absolute worst 'insurable' risks (ones that pass underwriting scrutiny) will have some sort of impact every 3rd year. If we filter that to more severe cases where there's bodily injury it moves out to 8/9 years.

What this means is the absolute worst drivers will go ~100k miles between causing serious accidents so the threshold between good, bad and awful drivers is really far out into the tails of the distribution. Without actual driving data to work from we don't know how safe FSD is and youtube videos are purposefully selected for views. Even so the sheer number and 'randomness' of fails is concerning because nothing like alcohol, speeding or inexperience to point to.

Yes thank you. I was literally thinking "the worst driver I know has been in two accidents in the 20 years of our friendship, that must be pretty close to 99.9% safe."

I don't have very strong intuition about risks this small. I do know the thing that ultimately lead me to being able to beat nethack at will was internalizing the idea that "something that's 99% safe is almost certain to kill you over the course of a run." Feels similar.

This, as someone in tangentially related industry I see the very same issue - you need a LOT of tail events to be able to train the models on them and obviously tail events are rare by construction. So it will be all smooth sailing until you hit the 1/1e9 and kill someone with your model3.

So from viewing some videos it is apparent to you that Tesla's FSD is the most advanced? Sorry to say this, but this is as unscientific as it gets. There are data and metrics to determine this and there is so much to consider when evaluation and comparing the data from these kinds of systems. One person assessing this from on person's (and an employee even more so) youtube channel is just fanboy talk.

Almost as if the mistakes were the important part of self-driving. Crazy huh.

Reply on news.ycombinator.com