Hacker News Reader

Comment by sashank_1509 | Hacker News Reader

Lots of “Tesla FSD Beta almost killed me” in the comments here but then I see hundreds of videos like this: https://youtu.be/Mj6vXnEYAUg

I choose to believe my eyes, it seems like FSD Beta can see almost everything a human sees. In fact in the latest version I don’t think I’ve found a single example of it missing something clearly visible to a human.

It’s planning needs some work, it seems to lack longer term planning, frequently putting itself in positions where it needs to do 3-4 quick consecutive lane changes. You may hate Elons politics, but the FSD Beta has been released to 100k+ Tesla owners and it’s not inconceivable that no one will be in the drivers seat in the next 5 yrs.

Edit: as mentioned in the comments below FSD can’t handle parking lots which is annoying because that was the one place I hoped it would work the most.

View on HN

I recently did a 70 mile drive on FSD with no interventions from local suburb roads, to highway, to urban roads, to some odd roads near a zoo.

That being said, often times I've had to intervene going <10 miles on local roads because it's so cautious that I'm afraid other drivers may react poorly and cause an accident. Rarely, it has done some dangerous actions that I'd characterize as "bad driving." So no, it's not there yet and it is dangerous because of the rare situations, but it's also occasionally perfect and usually "okay."

I agree, many of my interventions are because:

1. It hangs at an intersection too long that the person behind honks at me

2. It refuses to lane change out of fear leading me to miss highway exits or turns (sometimes but rarer, it tries a turn from the wrong lane)

3. It just gets bumpy some rides with a lot of phantom breaking that can be dangerous at times (I’ve seen the phantom breaking reduce in V11 though)

I don’t think the standard fear of if I didn’t intervene I would hit so and so obstacle is something I ever experienced in my FSD, though I did start using FSD in the much later versions of FSD 10+ where a lot of kinks were ironed out

> In fact in the latest version I don’t think I’ve found a single example of it missing something clearly visible to a human.

Is this from personal experience of using FSD Beta or are you relying purely on YouTube videos to assess its performance? Because not everyone drives around recording videos for YouTube.

I use it every time I take my Tesla out unless I feel like driving. I rarely intervene on my drive to work. But honestly I’m not impressed by this because this was true a year ago, FSD has no problem with the light traffic roads of the Bay Area. FSD also I’ve found rarely has issues in SF, it can infamously do the Lombard Street in SF with no problems but I’m guessing it’s overfit to SF at this point since the FSD team is in the bay area. I’m more intrigued by the YouTube videos of people trying it out in Florida, Texas etc.

I’ve been using FSD in my 3 for months now and I’d miss a car without it. It’s still definitely beta quality, but it’s a clear step up from enhanced autopilot and can take me from the on-ramp to my office without intervention (about thirty miles in California’s central coast). It would be door-to-door, but it doesn’t work well in parking lots and there’s a couple complicated intersections in my small town that it can’t handle yet.

Lots of anecdotes, not a lot of bad accidents. But lots of miles driven. And lots of incidents involving people getting killed by other cars. Many of which incidents involve non artificial intelligent drivers being impaired by stupidity, alcohol, distractions provided by their phone or onboard infotainment system, or the fact that they just aren't very competent drivers. Or all of those combined.

I get it, it's easy to make fun of Tesla and the plenty mess-ups and incidents. But the argument that those incidents add up to super dangerous situations seems to consistently not translate into a whole lot of real world injuries and deaths. You'd struggle to make that point. Teslas continue to have excellent safety metrics. And the more people they roll this out to and the more these metrics fail to degrade, the weaker that argument gets.

Nice little thought experiment: Ask yourself would you rather have an AI drive your car than any random owner of a US drivers license (you don't get to choose)? IMHO The barrier for AI is pretty low if you think about it like that.

Reply on news.ycombinator.com