Elon: FSD Beta tweets

[qdeathstar]

What if this is a regulatorily issue? Elon always believes in FSD. Most FDS beta videos looks great and is impressive with a few exceptions. I believe we'll always see the word Beta but I wonder if their concerned about the legal aspects.

killing someone is a pretty big legal issue, so I agree....

 

[qdeathstar]

A bit more of improvements is required! On July 14th, I have experienced a potentially catastrophic failure of FSD on a good, straight 30 m/h road outside of Tenby (Wales). I tried FSD features during my trip from Birmingham (UK) to Wales and in general liked it, although a few times I nervously took the steering myself as it seemed that my car was overtakeing obstacles too closely. Eventually, I decided that the autopilot knows better - a grave mistake. It passed too closely to a parked BMW resulting in two crashed mirrors. Now I think of this as a lucky escape - such a failure could have resulted in much worse. I have reported the accident to Tesla and wait for their engineers reaction (not expecting, naturally, that Tesla admits any liability.

You are responsible for your car. I really am surprised people are shocked by this. I feel like anyone paying even a modicum of attention would have swerved to avoid the car....

 

[qdeathstar]

Of course you do. But Waymo, that's Nirvana to you. In spite of total fails in construction zones.

Waymo does a lot better on the small number of roads it operates on that Tesla does on the very large number of roads it operates on. Waymo is not scalable though, or at they very least in the 15 years or so it has been in developement it hasn’t shown any sign of being scalable.

Waymo limits its edge cases by limiting the environment in which if operates in. From FSD beta videos I’ve seen, most of the errors seem like they would be easy to correct. They are just a sceneiro the car isn’t programmed for, not a limitation of the cars hardware, so I think that is a good sign.

On the other hand, there are almost an infinite number of edge cases.... how many edge cases do you need to solve in order to achieve 1-2m miles without disengagement. My guess is quite a few. So you need artificial intelligence that can problem solve on its own, but it doesn’t seem like fsd beta 9 has any of that to me.... just decision trees....

 

[rxlawdude]

Why? I did too until I watched the video of them operating. It sounds worse than the escalator to nowhere, at least at the end of the pointless endeavor you are dead....

You must have tunnel vision.

 

[rxlawdude]

Waymo does a lot better on the small number of roads it operates on that Tesla does on the very large number of roads it operates on. Waymo is not scalable though, or at they very least in the 15 years or so it has been in developement it hasn’t shown any sign of being scalable.

Waymo limits its edge cases by limiting the environment in which if operates in. From FSD beta videos I’ve seen, most of the errors seem like they would be easy to correct. They are just a sceneiro the car isn’t programmed for, not a limitation of the cars hardware, so I think that is a good sign.

On the other hand, there are almost an infinite number of edge cases.... how many edge cases do you need to solve in order to achieve 1-2m miles without disengagement. My guess is quite a few. So you need artificial intelligence that can problem solve on its own, but it doesn’t seem like fsd beta 9 has any of that to me.... just decision trees....

And Waymo does not handle those geofenced areas when there is construction or any significant changes to previously mapped streets. Regardless of what our resident Waymo promoter constantly misstates. Geofencing works by understanding the streets as mapped. Tesla's approach is more "human" like in that it does not rely on a geofence, but as you point out it opens up a Pandora's Box of edge cases.

 

[Mardak]

Hopefully that means the "Enhanced AP without FSD" folks with highway NOA will benefit from the new stack

I would guess the new stack requires HW3+ for birds-eye-view style neural networks. It looks like production Tesla Vision might have visualizations based on the old stack (i.e., using individual camera detections to render adjacent-lane vehicles) while FSD Beta 9 has some highway driving visualizations partially using the new stack (i.e., rendering fused 360° view of all cameras in any lane even for highway NoA).

Here's a @DirtyT3sla screenshot with production Tesla Vision before getting FSD Beta 9 where the adjacent truck likely wasn't visible enough to be predicted from the main camera nor repeater camera:

And a screenshot from FSD Beta 9 showing an incorrectly placed truck (it's not actually halfway out of its lane) but at least it didn't disappear:

If the new stack makes it to production for vehicles even with radar, it seems like one functional benefit is better lane changes, which have relied (exclusively?) on the repeater camera. Often times lane changes have been conservative especially around large trucks because it's not exactly sure if it's in the adjacent lane or 2 lanes away, so it assumes it's closer and sometimes results in an aborted lane change, but the new stack able to also use the pillar camera to help locate the large truck in a 360° view should be more confident and consistent.

Similarly, production Autopilot sometimes fails to make a lane change because it can't determine that there's an adjacent lane to switch to perhaps relying only on a single camera with insufficient view of the (wider?) adjacent lane, so hopefully the new stack reaching production fixes this for another user noticeable improvement.

@DirtyT3sla, have you noticed these issues from your extensive use of Autopilot before FSD Beta? And if so, perhaps you might be able to notice these differences as the new stack is applied to highway driving?

 

[diplomat33]

And Waymo does not handle those geofenced areas when there is construction or any significant changes to previously mapped streets. Regardless of what our resident Waymo promoter constantly misstates.

I am not misstating anything. You are spreading FUD. You are taking one example where it failed and wrongly generalizing. Waymo would not have 1 disengagement per 30,000 miles if it could not handle construction or significant changes to mapped streets.

Waymo Driver is designed to handle changes to the map:

We designed our self-driving system so it can navigate safely if something new or unexpected is encountered, such as construction.

Our system can detect when a road has changed by cross-referencing the real-time sensor data with its on-board map. If a change in the roadway is detected, our vehicle can identify it, reroute itself, and automatically share this information with our operations center and the rest of the fleet in real time.

We can also identify more permanent changes to the driving environment, such as a new crosswalk, an extra vehicle lane squeezed into a wide road, or a new travel restriction, and quickly and efficiently update our maps so that our fleet has the most accurate information about the world around it at all times.

And here is an example where the Waymo did handle a significant change to the map with no issue:

My Arizona trip was right before July 4. In some parking lots, fireworks vendors had erected makeshift tents with a perimeter of traffic cones surrounding them. In one trip, Driver was circling a lot’s outer edge and making a final end run to the drop-off. Just one problem: It had to pass through an ultra-narrow berth between a fireworks stand and the curb. The car laid on the brakes after inching toward the slot. The fireworks vendor manning the stand, named Austin, stood up, approached the car, and watched on curiously. After a few seconds, the car glided through the ~7.5-foot berth without incident. Afterwards, Austin told me he was surprised by how the van negotiated the scenario.

 

[rxlawdude]

I am not misstating anything. You are spreading FUD. You are taking one example where it failed and wrongly generalizing. Waymo would not have 1 disengagement per 30,000 miles if it could not handle construction or significant changes to mapped streets.

Waymo Driver is designed to handle changes to the map:





And here is an example where the Waymo did handle a significant change to the map with no

Let's watch it when it's in play other than the simplest geofenced limited beta area of suburban Phoenix.

And the cost of running those support vehicles seems to me is a fiscal disaster.

So we'll see who prevails in their respective markets, which do not realistically intersect in the foreseeable future.

 

[linux-works]

killing someone is a pretty big legal issue, so I agree....

one person, yes.

600,000: you get a free pass.

 

[qdeathstar]

one person, yes.

600,000: you get a free pass.

Do we have to bring the D-man into every thing? One death is a tragedy, a million deaths a statistic.

 

[JJRicks]

This is completely false. There are lots of videos of Waymo rides from the public, not marketing videos from Waymo.

And I have spent hundreds of hours of my life and terabytes of disk space to make this a reality. Nobody had be better saying there's a lack of Waymo videos from independent third parties

 

[rxlawdude]

And I have spent hundreds of hours of my life and terabytes of disk space to make this a reality. Nobody had be better saying there's a lack of Waymo videos from independent third parties

Let's try this challenge: TODAY, let's take a Waymo vehicle with all its sensors on non-mapped streets, vs an FSD Tesla that hasn't been on that same roaute. Let's compare how the two drive over a couple of miles.

Oh, wait. This can't happen, because Waymo can't do it. Am I exaggerating? Waymo guys?

 

[diplomat33]

Let's try this challenge: TODAY, let's take a Waymo vehicle with all its sensors on non-mapped streets, vs an FSD Tesla that hasn't been on that same roaute. Let's compare how the two drive over a couple of miles.

Oh, wait. This can't happen, because Waymo can't do it. Am I exaggerating? Waymo guys?

That's a false metric. The metric should not be can the AV drive without HD maps. The metric should be can the AV drive reliably and safely without driver supervision. If it takes HD maps to make the FSD more reliable and safer, that's a good thing, not a bad thing. Waymo drives with HD maps and it makes it way more reliable than Tesla's FSD that drives without HD maps.

 

[qdeathstar]

That's a false metric. The metric is not can the AV drive without HD maps. The metric is can the AV drive reliably and safely without driver supervision. If it takes HD maps to make the FSD reliable and safe, that's a good thing, not a bad thing. Waymo drives with HD maps and it makes it way more reliable than Tesla's FSD that drives without HD maps.

that is the metric. L5 driving in a city I don’t live in is useless to me. We aren’t going to be able to scale hd maps. Waymo can only maintain one city

 

[diplomat33]

that is the metric. L5 driving in a city I don’t live in is useless to me. We aren’t going to be able to scale hd maps. Waymo can only maintain one city

HD Maps are scalable. See Mobileye. They are already doing it in multiple cities. Waymo has also HD mapped several cities in the US and the process is almost entirely automated.

 

[rxlawdude]

That's a false metric. The metric should not be can the AV drive without HD maps. The metric should be can the AV drive reliably and safely without driver supervision. If it takes HD maps to make the FSD more reliable and safer, that's a good thing, not a bad thing. Waymo drives with HD maps and it makes it way more reliable than Tesla's FSD that drives without HD maps.

So when can we bring a Waymo down to Irvine and have it go head to head with Beta 9 FSD?

Serious question: when?

 

[diplomat33]

So when can we bring a Waymo down to Irvine and have it go head to head with Beta 9 FSD?

Serious question: when?

Waymo is testing in CA and they are making sure it is safe enough before deployment. When Waymo does come to Irvine, you will be able to ride in the back seat as the car drives fully autonomously.

It's cool that Tesla's FSD Beta works everywhere but it requires driver interventions and it is not safe and reliable everywhere yet. Waymo is focused on safe autonomous driving first before expanding.

 

[DanCar]

... We aren’t going to be able to scale hd maps. Waymo can only maintain one city

Maps isn't the issue. Waymo can generate the 3d centimeter accurate maps fairly quickly. The issue is risk taking. Waymo won't take risk of harming people. Tesla has and will continue to do so. Although Tesla tech is inferior it will win in the end against everyone else because of ability to take risk.

 

[idriveacar]

Maps isn't the issue. Waymo can generate the 3d centimeter accurate maps fairly quickly. The issue is risk taking. Waymo won't take risk of harming people. Tesla has and will continue to do so. Although Tesla tech is inferior it will win in the end against everyone else because of ability to take risk.

There has to be a sweet spot somewhere. Everyone knows autonomous cars have and will continue to kill people. The morality of forging ahead with building technology that will end people's lives who otherwise would have lived is tough - regardless of how many other lives will be saved (which, eventually, will be a very large number).

 

[rxlawdude]

That's a false metric. The metric should not be can the AV drive without HD maps. The metric should be can the AV drive reliably and safely without driver supervision. If it takes HD maps to make the FSD more reliable and safer, that's a good thing, not a bad thing. Waymo drives with HD maps and it makes it way more reliable than Tesla's FSD that drives without HD maps.

The other metric is the cost of what Waymo's doing in creating their geofence. Add the cost of roving support vehicles (staffed with humans) and that's a helluva sunk and ongoing cost proposition to sell a service at or below taxi/Lyft/Uber per passenger mile.