Elon tweets "pure vision" will solve phantom braking

[mikes_fsd]

It is Karpathy's opinion that sensor fusion is the wrong tree because he is promoting the "vision-only" approach. That last slide is basically a hiring pitch: "Come work for Tesla because we can solve FSD!" Companies always try to lure new hires by claiming that they alone can solve everything.

First - vision - as in "company vision" - is how you win over great engineers. That is how SpaceX hires (come help us settle Mars by solving all these .... problems)
Second, a vision backed by real money and actual deliverables is how you keep those engineers engaged (a fleet of millions of cars, currently 5th best supercomputer in the world, shadow mode deployment and continual feedback, FSD a la "I Robot" - where driving manually seems not only odd but dangerous)

After watching ~6 hrs of presentations ( there were some that I just could not watch/listen to ) and hearing their pitches, the only vision that would entice this engineer is that of Tesla!

But AV companies using sensor fusion today all have real FSD.

You gotta get off that, there is no useful FSD deployed.
I've watched all the videos you've posted form the so-called competitors, and they are (mostly) all barking up the wrong trees!

 

[finman100]

ok. i did not know that. really seems wrong, because my auto high beams don't work well either! lights go high beam and to low beam whenever a reflective street sign comes up. plus they auto dim when an oncoming car is too close and too late (at least for me, I would auto dim when a car is pretty far away).

so there really is not a setting now in the car that let's you turn off the auto feature of high beams? or move the left stalk forward manually? ok then, i need to get with the new and "better". I'm so 3 years ago! not always a win for Tesla.

 

[Saabstory88]

ok. i did not know that. really seems wrong, because my auto high beams don't work well either! lights go high beam and to low beam whenever a reflective street sign comes up. plus they auto dim when an oncoming car is too close and too late (at least for me, I would auto dim when a car is pretty far away).

so there really is not a setting now in the car that let's you turn off the auto feature of high beams? or move the left stalk forward manually? ok then, i need to get with the new and "better". I'm so 3 years ago! not always a win for Tesla.

You still have full control over the state of your high beams, and whether or not to engage auto high beams but only when you are driving manually.

 

[gearchruncher]

so there really is not a setting now in the car that let's you turn off the auto feature of high beams?

This is only on cars sold in the last 4 weeks or so, which do not have radar. (These cars have other large changes as well)
As is normal with Tesla, your experience in your car is not always the experience of all vehicle configurations or software versions.
It may be that a future software update brings this behavior to all cars.
The fun thing about saying "read the manual" is that the last updated manual on Tesla's website is from 8+ months ago, so you can't really read the manual and believe it to be up to date or accurate for your vehicle. You'd need to sit in the vehicle after every software update and re-read it (assuming the in-car one actually keeps up to date with the rapid changes Tesla makes...)

 

[stopcrazypp]

I completely believe that sensor fusion was a potential roadblock to solving phantom breaking, but with reports on this forum and YouTube videos showing plenty of examples of non-radar cars doing the same thing, it’s clear that removing radar wasn’t the solution. Maybe a step towards it, maybe a reprioritized change based on chip availability, maybe a decision to save money; but not the fix.

There's a lot of causes for phantom braking though. A common cause even with no sensor perception is incorrect GPS based speed limit data. For this obviously neither radar nor the vision (other than speed limits sign detection) is really the cause.

 

[mikes_fsd]

First - vision - as in "company vision" - is how you win over great engineers. That is how SpaceX hires (come help us settle Mars by solving all these .... problems)
Second, a vision backed by real money and actual deliverables is how you keep those engineers engaged (a fleet of millions of cars, currently 5th best supercomputer in the world, shadow mode deployment and continual feedback, FSD a la "I Robot" - where driving manually seems not only odd but dangerous)

How convenient... this is what excites engineers... seeing your solution making a difference in real life!
Elon know that well:

https://twitter.com/x/status/1407040628530294791

Edit:
re: "this is what excites engineers... seeing your solution making a difference in real life!"
This is for ppl that like to see things get done, there is a type of "engineer" that is easily bogged down by analysis paralysis - so probably not for everyone. But EM's track record on recruiting great engineers seems to point that many/most engineers want to see the fruit of their labor in real life!

 

[gearchruncher]

A common cause even with no sensor perception is incorrect GPS based speed limit data.

I have never had the car decelerate in a way I would call it "braking" when it goes from a higher speed limit zone to a lower one. It's pretty smooth.

 

[gearchruncher]

How convenient... this is what excites engineers... seeing your solution making a difference in real life!
Elon know that well:

In general, anyone interested in working on physical-word AI problems, should consider joining Tesla. Fastest path to deploying your ideas irl.

Unless your ideas are based on sensor fusion, any kind of multi-modal sensing, or anything else that the Technoking has determined to be the wrong path.

Also, not such a good place if your own personal ethics say that you should test things before deployment or charging for them.

I'm sure Tesla is a great place form many, many autonomy engineers, but to claim it's the best and only is just fanboyisim that ignores the fact that many smart people believe different solutions are the right path, and those ideas would get crushed at Tesla, so a different place is quite rightly the pace for them. Given nobody has solved it yet, we don't yet know which is correct.

 

[stopcrazypp]

I have never had the car decelerate in a way I would call it "braking" when it goes from a higher speed limit zone to a lower one. It's pretty smooth.

Unexpected deceleration is still called "phantom braking" by how most people use it, no matter how "smooth" it may be. There's not really a hard definition and everyone has their own.

For example, I read this thread back when I looked for it in other cars, and the thing described as "phantom braking" was the car slowing down for a speed limit. Once they turned off the setting where the CC adjusts for speed limit, the problem was solved.
Phantom Braking

 

[mark95476]

I like how they causally show their supercomputer (#5 in number of flops) and their plans for Dojo.

These aren't vague science projects.
Tesla has a fleet. Collect real data and deploy new models (in shadow mode) to said fleet to test how their models would react.
Autonomy will require constant iteration and never be completely solved.
The sky's the limit--only limited by their imagination and ambitions.

How convenient... this is what excites engineers... seeing your solution making a difference in real life!
Elon know that well:
View attachment 675931

https://twitter.com/x/status/1407040628530294791

Edit:
re: "this is what excites engineers... seeing your solution making a difference in real life!"
This is for ppl that like to see things get done, there is a type of "engineer" that is easily bogged down by analysis paralysis - so probably not for everyone. But EM's track record on recruiting great engineers seems to point that many/most engineers want to see the fruit of their labor in real life!

 

[gearchruncher]

These aren't vague science projects.
Tesla has a fleet. Collect real data and deploy new models (in shadow mode) to said fleet to test how their models would react.

I agree- what you describe is not a "vague" science project. It's just a normal science project.

Nobody is saying Tesla hasn't invested heavily in this, and that they don't have a reasonable approach. But that doesn't mean it's the only approach, or that it is unquestionable Tesla will solve it, or that they will solve it faster than another approach.

The question is if they clearly sold participation in a science project to their customers as what they were buying ("deploying new models to said fleet to test how their models would react"), or did they sell a product?

The sky's the limit--only limited by their imagination and ambitions.

Physics. You forgot physics

 

[helvio]

I agree- what you describe is not a "vague" science project. It's just a normal science project.

Nobody is saying Tesla hasn't invested heavily in this, and that they don't have a reasonable approach. But that doesn't mean it's the only approach, or that it is unquestionable Tesla will solve it, or that they will solve it faster than another approach.

The question is if they clearly sold participation in a science project to their customers as what they were buying ("deploying new models to said fleet to test how their models would react"), or did they sell a product?


Physics. You forgot physics

Given that I can opt-in/opt-out, I would say they sold me a product which I can convert to a science project. I'm happy with that. Knew it from day 1.

 

[helvio]

The sky's the limit--only limited by their imagination and ambitions.

Disagree. Pretty sure there is (or was, not sure) a Roadster in orbit, a little beyond the sky.

 

[linux-works]

How convenient... this is what excites engineers... seeing your solution making a difference in real life!
Elon know that well:
View attachment 675931

https://twitter.com/x/status/1407040628530294791

Edit:
re: "this is what excites engineers... seeing your solution making a difference in real life!"
This is for ppl that like to see things get done, there is a type of "engineer" that is easily bogged down by analysis paralysis - so probably not for everyone. But EM's track record on recruiting great engineers seems to point that many/most engineers want to see the fruit of their labor in real life!

many of the engineers I work with have worked for tesla in the past and they tell me - basically - do NOT go work there. full stop. and I trust their judgement.

from all I hear about elon, he's not the guy you want to work for, IF you have strong feelings about how things should be done. he's not one to listen to reason and he'll fire you on the spot.

no one I know - who works in this field - is submitting resumes to tesla. take that as a data point - or ignore if you wish.

tesla is getting junior guys because the senior ones wont put up with elon's crap. (and the move to texas, but that's another thread)

 

[steve68]

What version is your car on?

I updated to 2021.4.18.2 sometime during the journey.

not enough info to jump to that conclusion.
It is only cars that were built after a certain date.
If his car was built before that date, but delivered now, he would be experiencing the same radar induced phantom braking.

Manufacture date is 06/21. I had to accept the Tesla Vision agreement before taking delivery.

It seemed to mostly occur when approaching an overpass, but not always. A few times I could not figure out any possible reason it braked.

 

[FloridaJohn]

Disagree. Pretty sure there is (or was, not sure) a Roadster in orbit, a little beyond the sky.

Still is. Been out there 3 years, 4 months, 14 days, 23 hours, and 46 minutes so far.

Where is Starman? Track Elon Musk's Tesla Roadster in Space! · Where is Starman?

Current location of Elon Musk's cherry red Tesla Roadster and Starman launched by SpaceX on the Falcon Heavy maiden flight

www.whereisroadster.com

 

[gearchruncher]

Given that I can opt-in/opt-out, I would say they sold me a product which I can convert to a science project. I'm happy with that. Knew it from day 1.

What do you mean you can opt in/out? You mean just not using AP when you don't like the current code after trying it? Tesla doesn't allow you to pick and choose what AP SW versions you use, or give you any real info about changes.

I'm glad you somehow knew this was a science project. Can you point me to where on the Tesla site it indicates this? Or did you learn this by talking to owners and their experiences, despite what Tesla told them?

 

[helvio]

What do you mean you can opt in/out? You mean just not using AP when you don't like the current code after trying it? Tesla doesn't allow you to pick and choose what AP SW versions you use, or give you any real info about changes.

I'm glad you somehow knew this was a science project. Can you point me to where on the Tesla site it indicates this? Or did you learn this by talking to owners and their experiences, despite what Tesla told them?

When you first play with the car, it asks you whether you want to contribute your telemetry to Tesla so they improve the fleet. You can still get to that section through your touchscreen, but since I enabled everything and I intend to keep things that way, I don't recall exactly how to get to that. I'd say it's in the Software section of the car but I might be wrong. Go ahead and uncheck those if you want a product that's not used as a science project. Nothing stops you from doing that.

As for the second paragraph, pretty much the "Coming later this year" section in the image below.
Note: I would not like to talk about how long it's taking. They say "this year" every year and that's accurate every year. /s

 

[mikes_fsd]

I updated to 2021.4.18.2 sometime during the journey.

Well, that would be pretty important, the version prior to 2021.4.18.x was reported to have a bunch of issues.
Did you see an improvement on the new version? Usually to update you have to be connected to wi-fi!

 

[mikes_fsd]

tesla is getting junior guys because the senior ones wont put up with elon's crap.

Damn, he can accomplish all this with just junior engineers?!??

Secondly, did you just call Karpathy a junior engineer? Oh, and Jim Keller too? (who designed the FSD chip that is in all the Tesla FSD equipped cars)
Real engineers want to solve problems (i.e. a challenge)
Then the other "engineers" "have strong feelings about how things should be done" instead of actually solving the problem.

Feel the difference?