FSD Beta Videos (and questions for FSD Beta drivers)

[mhan00]

This one will be fun:

Check out this path. It'd give me PTSD:

View attachment 683369

At about 2:00 it shows a graphic of a person on a scooter for a girl on a scooter. Tesla has clearly added some new graphics.

Man, this is looking like it’s a rough vid. Not five minutes in and one major disengagement and another where the car started scraping its Side against some overgrown bushes.

5:10 - damn, on a left turn the car randomly jerks to the right with traffic to its right, scaring the driver and no doubt scaring the driver of the adjacent vehicle. Driver was paying attention but didnt have his hands on the wheel, so if the car hadn’t fixed itself he wouldn’t have been able to stop the car from hitting the adjacent car. That’s why I hate watching vids where the driver doesn’t keep their hands on the wheel (cough, Mars, cough). I’ve had instances like that on AP crossing some intersections, but since I keep my hands on the wheel I can feel the torque and tighten my grip to stop the wheel immediately instead of fumbling for it like the driver in this vid.

And then another major screw up not 30 seconds later. Lmao. This is terrible. I would think it would do better in SF where there are a ton of Tesla’s and where the HQ and many employees are located, not horrifically worse.

7:25 - most of the incidents are time tagged by the vid poster, but this one was not. Car randomly decides to maneuver to the left into an oncoming car, driver saves it.

There are some nice maneuvers in the vid, but it’s difficult to overstress how HORRIBLY the beta performed in this video. My excitement from watching other videos has dulled considerably.

 

[powertoold]

At about 2:00 it shows a graphic of a person on a scooter for a girl on a scooter. Tesla has clearly added some new graphics.

https://twitter.com/x/status/1413894747995922436

 

[powertoold]

Man, this is looking like it’s a rough vid. Not five minutes in and one major disengagement and another where the car started scraping its Side against some overgrown bushes.

Near the end, he says V9 is much improved over 8.2. It's definitely not ready for SF, but this guy says V9 seems to be many steps forward.

 

[mhan00]

Near the end, he says V9 is much improved over 8.2. It's definitely not ready for SF, but this guy says V9 seems to be many steps forward.

First vid I’ve seen of his. I shudder to think how badly his car performed on 8.2 then if V9 is a big upgrade for him. Gonna add him along with Brandon and Chuck as one of my bench markers, I guess, even though it’s stressful to watch him keep his hands off the wheel even after so many near misses.

 

[Matias]

Elon isn't wrong about the "step change" or "mind-blowing"-ness that V9 represents. Like I said, I still think V9 is more convincing evidence for human level+ FSD, soon.

What exactly is mind-blowing or step change in 9 vs 8.2?

 

[Matias]

The vids are impressive, but clearly still a lot of work to be done, especially with unprotected lefts across highway traffic like Chuck has.

Could someone explain me, why the unprotected left that Chuck tests, is so difficult for FSD? Car can see oncoming traffic and apparently their speeds (aforementioned is clearly hard to solve with vision only, but Tesla has apparently solved it). When car knows projectiles vectors, isn’t it trivial to solve, when there is enough time to turn?

 

[GrimRe]

Could someone explain me, why the unprotected left that Chuck tests, is so difficult for FSD? Car can see oncoming traffic and apparently their speeds (aforementioned is clearly hard to solve with vision only, but Tesla has apparently solved it). When car knows projectiles vectors, isn’t it trivial to solve, when there is enough time to turn?

The theory is that the B pillar cameras can only resolve cars about 80 metres away. Because of this, if the cross traffic is moving at 60 miles per hour, you only have 3 seconds between detection and impact.

Humans will see a car about 200 metres away but have a very poor understanding of its velocity. So we naturally wait for larger gaps.

In theory, FSD could thread the needle and very confidently take a small gap, but its perception would need perfect precision.

Most Beta testers (and rightfully so) are not going to take the chance of the car taking a gap that's smaller than what they are comfortable with.

 

[Matias]

This one will be fun:

Check out this path. It'd give me PTSD:

View attachment 683369

That ride was epic fail. I lost count how many disengagements and near collisions it had.

With this video and the Seattle video trying to hit monorail pillars and planters I would say that this build was seriously overhyped.

Yes, there are some incremental improvements and eye candy, but the most basic functionality, i.e. “do not hit objects“ still lacks.

 

[diplomat33]

Some of these drives are approaching Waymo performance in areas more complicated than Chandler (IMO):

What about this video where V9 fails miserably and is way worse than Waymo?

This why I warn against making judgments based on videos. It is too easy to cherry pick. Yes, you can find videos where V9 appears to perform better or appears to perform as good as Waymo in certain cases but there are also videos of V9 performing way worse.

 

[emmz0r]

What's a bit unsettling is that cars pop in and out of existence, even if they are parked. Shouldn't they be tracked and persisted?
That's what the company I was hired to was doing at least, and that's soon 10 years ago...

 

[Matias]

What's a bit unsettling is that cars pop in and out of existence, even if they are parked. Shouldn't they be tracked and persisted?
That's what the company I was hired to was doing at least, and that's soon 10 years ago...

At least in visualizations fsd doesn’t have any memory. Immediately when something is not visible to cameras, it is forgotten.

 

[emmz0r]

At least in visualizations fsd doesn’t have any memory. Immediately when something is not visible to cameras, it is forgotten.

The visualizations probably reflect the model 100% honest.
And add the fact that it doesn't see pillars. They can see edges and curbs, but not pillars. Wonder where the problem is. I would seriously had a hybrid model with classical CV that sees edges and identifies walls and stuff, and ML in parallel until a better ML alternative replaces classical CV just to get it out the door.

 

[qdeathstar]

@powertoold - Yep, that's pretty cool. A nice incremental improvement, like the youtuber says. 100X mind blowing though because it manages to not slow down as much for a car in front? It seems so hyperbolic to say that a normal person watching that would say they never ever thought a computer could drive like that until they saw it, or that it means Tesla is only a few months away from 5 orders of magnitude improvement on disconnect rate.

Take me up on my bet.

you are gonna have to offer up better odds. 5000k today for 10000k tomorrow isn’t as attractive as you say it is because I can probably put that same amount of money in Tesla stock and get the same if not greater return...

 

[ZeApelido]

@powertoold @mikes_fsd and @ZeApelido has some explaining to do.

Too funny. What explaining do I have to do? Maybe you should review my 4+ year post history in this forum as you seem to be confused...again.

 

[jeewee3000]

Fine. I'll do it with Tesla stock price. $5K worth of shares today, I pay out 4X the Tesla stock price on the day Tesla has public L5 in a 2016-2021 car. You literally cannot lose if Tesla ships L5 in a current car.


Of course it doesn't. But current accident rate is 1:2M miles per Tesla, and 1:100M fatalities per NHTSA, with all that drinking and distracted driving included. These are actually very, very good numbers. So you're saying 1:10M accident, 1:500M fatality, with only a child in the back seat in 2023, and it can drive on a snowy road at night like a human can?

I'm guesstimating just as much as you.

Your "bet" puts odds in your favor greatly since the taker of the bet can only win with L5 being achieved, which is indeed a monumental task that might not happen in our lifetime (you never know, it hasn't happened yet. Completely uncharted territory).

For you to win on the other hand, nothing has to happen. I.e. L5 is not reached or it's not proven to exist. So of course you are ahead.

Add to that the fact that it's unclear what the benchmarks will be. The accident data only comes up after x amount of months or even a year so by the time the bet expires we won't have the data to know who won.

Betting with strangers on the internet about sugar like this is is pure gambling (duh), which I don't do. I'm guessing TSLA has a shot at L5 by 2023 but of course I know that's a long shot and more likely it's much later than that. Not that I'm backing down, just saying it's easy to be Negative Nancy but not adding any value. If L5 doesn't happen by then (most likely, I admit), TSLA probably still will be up since the rest of the business is also solid.

If you want to bet someone, sell TSLA 1400 calls expiring JUN2023 against all your shares. Free money. You can even buy puts, shares or whatever with it to take the bet further. But let's just stick to healthy discussion here regarding what current FDS beta can and cannot manage.

 

[FloridaJohn]

I haven’t watched every video, but it seems like there is still a need for an intervention every couple miles. That seems at about the same rate as the previous version.

 

[Bladerskb]

Too funny. What explaining do I have to do? Maybe you should review my 4+ year post history in this forum as you seem to be confused...again.

Your idea that real world data is the solution and simulation is worthless and overhyped. Clearly the opposite is happening as companies (waymo) who invested more in SOTA sim is thriving while Tesla is struggling to break the 10 mile safety barrier.

 

[powertoold]

What about this video where V9 fails miserably and is way worse than Waymo?

This why I warn against making judgments based on videos. It is too easy to cherry pick. Yes, you can find videos where V9 appears to perform better or appears to perform as good as Waymo in certain cases but there are also videos of V9 performing way worse.

That's why I said some of these "drives" are definitely approaching Waymo performance. If we were to put V9 in Chandler doing the same routes (mostly protected turns and detours through easy streets), I'm sure it'd be similar performance. Waymos do well when the environment is simple and predictable (don't go showing me some video provided by Waymo to show it does well, lol. I just watched a JJRicks video where Waymo got stuck again in a simple situation).

It's unsurprising that V9 fails in SF, mostly because of its poor lane choices / confusion and poor lane changing speed.

 

[diplomat33]

That's why I said some of these "drives" are definitely approaching Waymo performance. If we were to put V9 in Chandler doing the same routes (mostly protected turns and detours through easy streets), I'm sure it'd be similar performance. Waymos do well when the environment is simple and predictable.

It's unsurprising that V9 fails in SF, mostly because of its poor lane choices and poor lane changing speed.

FSD Beta V9 only approaches Waymo in very simple driving tasks. It does not approach Waymo in more complicated tasks. V9 is L2 "door to door", it is not autonomous driving. Waymo is actual L4 autonomous driving. And yes, V9 could probably handle easy routes in Chandler. Waymo's 5th Gen can handle more complicated environments than Chandler. Waymo's 5th Gen FSD is light-years ahead of V9.

 

[powertoold]

It does not approach Waymo in more complicated tasks

Sure, post some non-affiliated videos of anything complicated Waymo does in the Waymo thread.