Elon: FSD Beta tweets

[sleepydoc]

3. A supervised system is less safe than a human, which makes deploying it among VRU:s a bad idea.

<citation required>

This statement is dependent on the system and dependent on the human and is thus impossible to generalize.

 

[Knightshade]

No. English is not my first language, but "coming years" to me is at most 3-5 years. Again, let me know when you find a single scientist that says camera only autonomy will likely happen before 2030

Mobile eye says it, I even linked you to it.

You won't even find a Tesla employee

You don't think Elon Musk works for Tesla?

Your arguments continue to get sillier with each post.

Again, I linked to a random article This is a well-researched area

Then it's weird you linked to a random article that didn't actually support your claim.

Surely if it's well researched you can do better, right?

Let's see you do better.

. Do you dispute that automation complacency is a thing?

I dispute that it magically means "human supervised automation is worse than a human" because you've failed to support that claim in any way.

After being asked 3 times now.

See this text by Dr Phil Koopman: Tech Roadmap for Automakers Disillusioned With Robotaxis | The Ojo-Yoshida Report (see in particular the graph titled Lower Supervisablility in Unsafe Road testing).

What's hilarious is that graph shows higher safety than just a human at both ends

It's lower in the middle.

Which seems to disprove your claim it's "always" less safe.

Thanks for debunking your own claim!

I put 'never' in quotes, as was obviously talking about a system that has superhuman capacity or at least very high MTBF

Which is weird, because you were literally comparing it to normal human safety

So NOW claiming you "obviously" meant superhuman appears to be you once again realizing your argument is nonsense so you're moving the goalposts again

Carefully you're gonna throw your back out doing that so much!

 

[Daniel in SD]

It’s impossible not to use maps; the question is really how much and how they will be used. If you consider you a human drives, they have some sort of map in their head to know where they’re going. Similarly, the car needs to have an idea where it’s going, otherwise FSD would just be a random turn generator. The difference is how much it can interpret, adjust and accommodate ‘on the ground’ differences from the map data. One approach is to have hyper accurate data so the car never has to interpret anything. Since roads and conditions are constantly changing this is a fools errand, IMO. I think Tesla’s approach of being able to adjust based on visual data is likely better, just more difficult to implement.

Why not just use hyper accurate maps and then switch over to Tesla's approach if there's an inconsistency?
Though I do wonder if part of the reason Tesla hasn't deployed HD maps (or some euphemism for HD maps) is that it would be a huge improvement in performance but not solve the tough problems.

 

[sleepydoc]

You don't think Elon Musk works for Tesla?

He’s not exactly an employee. More like confabulator in chief

 

[sleepydoc]

Why not just use hyper accurate maps and then switch over to Tesla's approach if there's an inconsistency?
Though I do wonder if part of the reason Tesla hasn't deployed HD maps (or some euphemism for HD maps) is that it would be a huge improvement in performance but not solve the tough problems.

But then they still need to have a capable system for all the occasions when the ‘hyper accurate’ maps fail and they have the expense of maintaining hyper accurate maps AND developing a system that can manage when the maps fail.

 

[Bladerskb]

They appear to be using HD map data. The visualization is clearly rendering sections of the road that are not in view of the car. Tesla's feat is that they're using vision not only for object recognition, but also lanes and road edges.

And you think Mobileye isn't using vision for lanes and road edges? How do you think they are localizing to the HD Map?

 

[Bladerskb]

James Douma, Andrej Karpathy, Kilian Weinberger to name 3.

This is a perfect example of if you introduce someone as an "Expert" enough, you can fool your entire audience into actually believing they are.

As already pointed out.
James Douma has

0 Machine Learning Experience
0 Computer Vision Experience
0 Autonomous Driving Experience

Zero. Nil. Zip.

Let alone an "Expert". Do you know what being an expert consists of? Yann LeCun is a expert. Ian Goodfellow, Jim Keller, etc.
It takes decades with immense body of work to be an expert. Whether its ML/AI, or its for example C++. For someone to have absolutely ZERO experience and still have an entire community believe/call him an expert, speaks volume of that community.

It shows just how easy it is to fool people. Its incredible.

 

[EVNow]

Name one expert that has said camera-only autonomy will happen in the coming years.

I've not updated this in a while ... but I think all these are camera only.

 

[Daniel in SD]

But then they still need to have a capable system for all the occasions when the ‘hyper accurate’ maps fail and they have the expense of maintaining hyper accurate maps AND developing a system that can manage when the maps fail.

Nobody said AV development would be cheap!
Mobileye claims that they can build maps automatically using data from their fleet. When I use Apple Maps it tells me which lane I should be in, it's useful even for human drivers.

 

[Bladerskb]

Somebody is jealous ....

Literally no one is... You either have the experience or you don't. Its clear cut.
Why would I or anyone else be jealous of a guy who clearly took a ML 101 course and is pretending to be an ML/AI/AV expert in order to pump his favorite company which he has his life savings in?

I could make 100 videos of the things James Douma has said.

Who do I want to be? a guy with actual experience in ML, CV, AV or a guy with zero experience and won't be hired by any AV company even he offered to pay them to work?

It reminds me of James Wang who ArkInvest pumped up and used as credentials for their claim that Tesla was "5 years ahead of Nvidia". That ~"We have a guy that was at Nvidia for 8 year and he was blown away because of how Tesla's chip was 5 years ahead of anything Nvidia was doing or planning blah blah blah".

You would think that Wang was some senior chip engineer at Nvidia. A Jim Keller kind of guy in the chip world. Nope he was a freaking Website Editor at Nvidia. Yet he is paraded around like a chip expert and the TSLA community loved him.

Why? Because the only measurement of your expertise the TSLA community at large cares about is if you proclaim "Tesla/Elon is the greatest".

This is why there has been alot of in-fighting lately. The minute you say, well I disagree with this very small thing. You get blitzed by your very own.

If KS listed a guy like Alex Kendall, then you can say. Okay, atleast KS follows the industry.
But he didn't, he listed a guy whose only qualifications is being a Tesla investor and a youtuber.
Which means like a-lot of Tesla fans, he doesn't fellow the ML/AV industry. He just follows Tesla.
Now there isn't anything wrong with that, but Instead of admitting it, we end up with world salads.

 

[sleepydoc]

Nobody said AV development would be cheap!
Mobileye claims that they can build maps automatically using data from their fleet. When I use Apple Maps it tells me which lane I should be in, it's useful even for human drivers.

That ends up being circular reasoning. The system works as long as you have cars driving the streets to update the maps but you need updated maps for cars to drive the streets.

The practical fact is there will be streets that are driven infrequently and thus not updated. If you require hyperaccurate maps to drive then your range of drivable streets will continue to shrink unless you have software capable of navigating inadequately mapped streets. But then if you have that, you don’t need the hyperaccurate map data.

 

[spacecoin]

I've not updated this in a while ... but I think all these are camera only.


View attachment 878663

1) Marketing is not the same as being a credible researcher,
2) No one except Tesla and possible Waymo believe that camera only will lead to an MTBF that's safe enough to remove the driver from the DDT. So I think you need to research a bit better and update that diagram.

 

[Bladerskb]

That ends up being circular reasoning. The system works as long as you have cars driving the streets to update the maps but you need updated maps for cars to drive the streets.

The practical fact is there will be streets that are driven infrequently and thus not updated. If you require hyperaccurate maps to drive then your range of drivable streets will continue to shrink unless you have software capable of navigating inadequately mapped streets. But then if you have that, you don’t need the hyperaccurate map data.

This makes the ridiculous assumption that this whole thing is binary.
There's nothing binary about autonomous driving.

If there are 1 million miles of road and 1,000 miles changed due to construction. That's still remains 999,000 miles where you still benefit from the pros of HD mapping (forsight, low/partial visibility in adverse weather, redundancy).
While your system still uses its perception to drive the 1,000 miles.

 

[spacecoin]

(moderator edit)

People can make up their own minds from what I have shared so far.

Again, there is no one with a credible CV that believes that camera-only AUTONOMY is happening "in the coming years".

 

[clydeiii]

Lots of great tweets posted here recently…

 

[sleepydoc]

This makes the ridiculous assumption that this whole thing is binary.
There's nothing binary about autonomous driving.

If there are 1 million miles of road and 1,000 miles changed due to construction. That's still remains 999,000 miles where you still benefit from the pros of HD mapping (forsight, low/partial visibility in adverse weather, redundancy).
While your system still uses its perception to drive the 1,000 miles.

No, my point is that the remaining 1,000 miles still needs an accurate visual system that can function without the hyper accurate map data. There’s nothing binary about it.

 

[2101Guy]

Lots of great tweets posted here recently…

Elon usually posts his tweets on twitter

 

[Bladerskb]

No, my point is that the remaining 1,000 miles still needs an accurate visual system that can function without the hyper accurate map data. There’s nothing binary about it.

You can have an accurate visual system that is still limited by adverse weather conditions (heavy fog, moderate/heavy rain, snow) and visibility (blocked by other vehicles/objects, blind corners, sharp curves, hills and crests, steep, etc).

There are no 100%

What's so hard to understand? Having an HD map alleviates that problem and provides more safety and comfort then you do.

You're saying that because 0.1% of the road changes, then we should throw the baby out with the bathwater.

 

[spacecoin]

I mean, I was pointing out your claims were unsupported by facts, and that even the one source you originally tried to provide doesn't support them- then pointing out your second source actually proved your claim factually wrong

Then pointed out your response to having each claim you made debunked appears to factually be moving the goalposts further away to deny having to admit your being caught repeatedly in error.

Seems pretty factual to me.

So it's unsurprising you'd just dismiss the discussion-- you didn't have much other choice.

No. Wrt complacency, you clearly do not understand (or want to understand) the diagram and the research backing it. I showed it to my 14 year old son and he could figure it out the issue without guidance. Supervised systems are not as safe as the system + a human, nor safer than a human, as the human gets complacent. The system is as safe sa the system when the human is complacent. It's sad that this needs further discussion in 2022. It's obvious and almost self-evident.

No one is moving any goal posts here except you. You post random links which does not support your statements, such as to MobileEye. They do NOT claim autonomy on cameras. You refer to Weinberger without quotations. He does not believe in autonomy in the coming years for camera only either.

Not a single person outside of Tesla with a CV? Zero quotes shared. Zero links to dispute my statements. And your go to expert is James Douma. Really? Clearly you win!

 

[Daniel in SD]

That ends up being circular reasoning. The system works as long as you have cars driving the streets to update the maps but you need updated maps for cars to drive the streets.

The practical fact is there will be streets that are driven infrequently and thus not updated. If you require hyperaccurate maps to drive then your range of drivable streets will continue to shrink unless you have software capable of navigating inadequately mapped streets. But then if you have that, you don’t need the hyperaccurate map data.

There are going to be manually driven cars on the street for a long time. Mobileye claims to be getting mapping data from millions of vehicles. There are only 4 million miles of paved roads in the US. Once you've replaced the 100's of millions of manually driven vehicles with autonomous vehicles the cost of mapping will be such a small fraction of the total cost that it will be irrelevant.