HN Theater @HNTheaterMonth

From Tesla's release of their full self-driving beta and its performance in a recent CGP Grey video where it successfully navigates the most dangerous road in America[1] I think a lot of people are going to have to eat their hats on this stance very soon.

I consider myself bullish on this tech, and I have worked closely with autonomous robotics, but I didn't expect it this soon.

[1]: https://www.youtube.com/watch?v=d6tgmGqXysM

⬐ lm28469

Lmao, their tech is faaaar away from anything remotely usable in real life. Sure it can follow a double yellow in the middle of the woods, I could do that when I was 10, as soon as you put it in city traffic it's game over in 15 seconds.

We will not have fully autonomous vehicles in our street until we ban human driven vehicles and completely rework our infrastructure, aka probably not in our lifetime.

Look at that video, it's comically bad: https://www.reddit.com/r/IdiotsInCars/comments/rgiu8m/the_fu...

⬐ hellomyguys

Following a well-marked winding road is not a difficult challenge for self-driving cars. Driving in a city, with bus lanes, light rail, bikers, and pedestrians is much harder.

⬐ aasasd

BTW, if anyone wants to experience the boredom of 18 km of turns among the trees, without actually going to NC/TN—there are custom circuits of Deals Gap/Tail of the Dragon for rFactor, Assetto Corsa and perhaps other sims. (Possibly ‘Grand Prix Legends’ was the first one to get the mod.)

⬐ somerandomqaguy

I'm really not sure how that is at all challenging for computer vision to handle. Conditions like these are probably more along the lines to challenge self driving tech:

- British Columbia Highway 5 (Coquihalla Highway) Daytime: https://www.youtube.com/watch?v=fIF79ZlWcOI

- The Coquihalla Highway at night: https://www.youtube.com/watch?v=hM3BPOLVoUg

- Alberta Highway 2: https://www.youtube.com/watch?v=Re-Ap5fboi8

⬐ leereeves

That actually didn't look like much of a challenge. All the Tesla had to do was drive slowly and stay in its lane on curves.

There were no obstacles that Teslas have had trouble with in the past, like stopped firetrucks, ending lanes, or people changing tires.

⬐ bombcar

Whilst impressive, that's also a "pretty good setup" for the Tesla - the yellow line is well marked as is the white line; and the Tesla isn't speeding.

⬐ gonzoflip

That is the most boring Tail of the Dragon driving vid I have ever seen.

⬐ root_axis

I own a Tesla and it's the best car I've ever owned. However, the FSD software is garbage, I'd sooner trust a drunk human over the Tesla FSD any day of the week.

⬐ wolrah

> where it successfully navigates the most dangerous road in America

Twistiest, maybe. Most dangerous, not even close unless you or someone driving towards you do something dumb.

Historically Autopilot's weak spots have been crappy/ambiguous road markings, low speed corners on high speed roads, and large vehicles stopped on/across the road. This video hit none of those points. The lane markings are clear, the posted limit is low, and large vehicles aren't even supposed to be there.

⬐ seoulmetro

Not to mention it's dangerous because of the amount of people acting irrationally and speeding there. There were barely any cars on the road with him (and he did that on purpose, possibly to save people the hassle or actually not kill others).

⬐ maxdo

your historical observations are irrelevant for FSD beta. It's a new system, it works amazing without markings on the road at all.

⬐ wolrah

> your historical observations are irrelevant for FSD beta. It's a new system

1. I find it very unlikely that Tesla completely scrapped the previous system to start from scratch for the FSD beta. 2. If they had, that would make me even more suspicious of the new system because that would mean no one would understand its weaknesses yet.

Either way, being most interested in where the previous generation system and other semiautonomous systems have historically failed makes sense.

> it works amazing without markings on the road at all.

I have yet to see any videos of the new system handling poorly marked high speed roads one way or another, but the crash where it tried to veer in to oncoming traffic doesn't provide confidence. All the unmarked/poorly marked roads I've seen so far have been low speed residential/urban, and even there performance has been unpredictable at best. Let's see it on a highway where night time rain has made the markings hard to see. Let's see it when snow covers the road and your only hing or lane

The state of the art is such that it can simultaneously be amazing and terrifying to think that normal people are allowed to "test" it wherever they feel like trying on public roads. What it does with the limited inputs it has is incredible, but at the same time it has dangerous failure modes that do not seem to be predictable even to those who use the system regularly.

---

Don't get me wrong, there is nothing I'd love more than never having to manually handle a traffic jam, or "go home car, I'm drunk", but the way this is being released will kill people.

⬐ reaperducer

Twistiest, maybe. Most dangerous, not even close

It's just CGP Grey, not NHTSA. Hyperbole is to be expected.

I know he's a darling among the chattering tech class, but it's been several years since he jumped the shark from "insightful, carefully researched, and well thought out" to "just another vlogger doing and saying what is needed to get views."

⬐ deep-root

That road is so well suited to self-driving it almost feels like a paid advertisement. Here is a typical urban stress test of FSD 10.6 beta in San Jose: https://youtu.be/2ub2F-UnXIU