Karpathy talk today at CVPR 2021

[Knightshade]

Really? So if the assist motor fails how does AP steer the car. (Note: It can't.)

It still can operate the brakes though to avoid hitting something.

So again your example sucks.

BTW speaking of the steering motor- did you know there are two independently powered controllers that feed the motor?

FOR REDUNDANCY.

Revealed: Tesla Model 3's secret technology | Autocar Professional

When it comes to Tesla, attention is usually focused on its production struggles, the latest actions of company founder Elon Musk or, from a technology perspective, its pure-electric powertrains. But with the Model 3, the firm is also aiming to pioneer a new approach to chassis and suspension...

www.autocarpro.in

Three- phase power steering motor has built-in redundancy.
...The electric power steering system has a rapid 10:1 ratio, giving two turns lock-to-lock. The system has full redundancy with separate power feeds taken directly from the high-voltage battery, two electronic control modules and two inverters providing ‘hot backup’ if one fails.

Man- Tesla spent a LOT of money and engineering work on stuff you claim they don't have and never needed!

And if the FSD computer fails the car doesn't become an unguided missile, it would just stop. Just like if the motor fails.

Again- no.

If the drive motor fails the brakes AND steering still work so the car can fail quite safely.

If the driving computer fails it just keeps going in a straight line at full speed until it hits something or eventually stops due to gravity or regen.


Like I said- increasingly desperate arguments from you.

 

[alsetym]

It still can operate the brakes though to avoid hitting something.

So again your example sucks.

BTW speaking of the steering motor- did you know there are two independently powered controllers that feed the motor?

FOR REDUNDANCY.

Revealed: Tesla Model 3's secret technology | Autocar Professional

When it comes to Tesla, attention is usually focused on its production struggles, the latest actions of company founder Elon Musk or, from a technology perspective, its pure-electric powertrains. But with the Model 3, the firm is also aiming to pioneer a new approach to chassis and suspension...

www.autocarpro.in

Again- no.

If the drive motor fails the brakes AND steering still work so the car can fail quite safely.

If the driving computer fails it just keeps going in a straight line at full speed until it hits something or eventually[ stops due to gravity or regen.


Like I said- increasingly desperate arguments from you.

You seem quite sure this will never work...there are hundreds of possible ways that tesla can implement safety considerations. One possible is that they have partial fail over, if one chip dies it goes into a simplified mode on the second chip that just looks for where it is safet to pull over. They could decide to add a second board in the future, upgrade to HW4, recompile to a smaller program that fits on a single chip (after they have cracked it). Suffice it to say, they have some of the top people in the word working on this, just because you don't know what the plan is doesn't mean they don't have one!

 

[MP3Mike]

Again, you are stating things as facts, that you are only assuming are true.

If the driving computer fails it just keeps going in a straight line at full speed until it hits something or eventually[ stops due to gravity or regen.

Wow, I had no idea Tesla could design something so horribly that if the command signal for toque was lost that the motor would just supply 100% torque. (Note: It won't.) And there are other system in the car that could apply the brakes if the FSD computer fails.

And you can't even agree with yourself. You say it will go at "full speed" but in the same sentence you also say it will stop because of regen. Which is it?

 

[Knightshade]

Again, you are stating things as facts, that you are only assuming are true.



Wow, I had no idea Tesla could design something so horribly that if the command signal for toque was lost that the motor would just supply 100% torque.

You appear unfamiliar with Newtons first law of motion.

(among a ton of other things you appear unfamiliar with)

No torque is required to be supplied after failure for what I described to happen.

(Note: It won't.) And there are other system in the car that could apply the brakes if the FSD computer fails.

With no human in the car, since we're discussing self driving that doesn't require one? Which ones?

And you can't even agree with yourself. You say it will go at "full speed" but in the same sentence you also say it will stop because of regen. Which is it?

Do you also not understand the word "or"?


Since what I actually said was

"If the driving computer fails it just keeps going in a straight line at full speed until it hits something or eventually stops due to gravity or regen."



I'm embarrassed for you at this point dude.

 

[alsetym]

You seem quite sure this will never work...there are hundreds of possible ways that tesla can implement safety considerations. One possible is that they have partial fail over, if one chip dies it goes into a simplified mode on the second chip that just looks for where it is safet to pull over. They could decide to add a second board in the future, upgrade to HW4, recompile to a smaller program that fits on a single chip (after they have cracked it). Suffice it to say, they have some of the top people in the word working on this, just because you don't know what the plan is doesn't mean they don't have one!

Or, this is much smarter, each node does every second frame. So if you lose one you have half the frame rate but all the decision making power. That's probably what I would do.

 

[Knightshade]

You seem quite sure this will never work...there are hundreds of possible ways that tesla can implement safety considerations. One possible is that they have partial fail over, if one chip dies it goes into a simplified mode on the second chip that just looks for where it is safet to pull over.

Since you'd potentially have a surviving node that doesn't have any data about what the car is currently seeing, there'd be a significant safety risk in trying to spin up something like that on the fly with no environmental info..... this would maybe be ok on the highway but pretty horrific for urban settings with pedestrians, bikes, etc everywhere.


Plus- even if they were able to make it entirely safe- unless they drastically reduce how often a node crashes, this would be nightmarish for the robotaxi program.... "This self driving cab is great except for how it keeps having to pull over and stop and wait for a reboot every couple of rides...."

They could decide to add a second board in the future, upgrade to HW4

Yes- that's specifically how I expect it'll be resolved.

Assuming HW4 has enough compute--- which we (and Tesla) still don't actually know- and won't, until they actually solve Vision entirely.

The IDEAL case for Tesla at this point is they CAN do it with all the compute available on both HW3 nodes-- if that's the case then the 3x performance of HW4 means they can run everything in one node, and have a second for failover redundancy, and they're gold other than needing to upgrade potentially a million+ cars to HW4 (if everyone gets FSD, which once solved is likely to be close to the truth).

. Suffice it to say, they have some of the top people in the word working on this, just because you don't know what the plan is doesn't mean they don't have one!

Those same top people also don't know how much compute is needed to solve FSD.

nobody does.

Those same top people have entirely changed direction on how to get to FSD multiple times.

In fact one of Teslas towering strengths is their willingness to realize their "top people" have gone down a dead end, and decide they need to do something different.



This is also why the ideal case is them solving with the HW3 both-nodes compute.... because if they can't, then they still have no idea how much compute they need... and no idea if HW4 has enough to do it... and no idea how they should be speccing compute for HW5.

They'd just be guessing. Again.


Remember- the "top people" at Tesla thought HW2 was sufficient for FSD. It wasn't.

Then they thought HW2.5 was-- it wasn't.

Then they thought HW3 was- it's increasingly appearing it also wasn't.

 

[Daniel in SD]

I actually did, it was in the thread about discussing LIDAR.

Search the forums.

Nope, you disagreed with his take on "loss of face" and argued that lidar and high resolution radar are completely different. You certainly did not call his prediction that Tesla was switching radar vendors "delusional".

 

[mikes_fsd]

Nope, you disagreed with his take on "loss of face" and argued that lidar and high resolution radar are completely different. You certainly did not call his prediction that Tesla was switching radar vendors "delusional".

Yes, I was pointing out his delusion about what radar vs lidar is.
We are still talking about whether green is delusional about the conclusions he jumps to right? Specifically this post from you "If it's so delusional why didn't you, or anyone else, bring it up in your initial rant about his tweets?"

Or do you want to get hung up on a specific delusion, i.e. the fact and timeline of a new 4d radar introduced for end-user cars?

 

[Daniel in SD]

Or do you want to get hung up on a specific delusion, i.e. the fact and timeline of a new 4d radar introduced for end-user cars?

Yeah, I just think it's odd that you got into an argument about the differences between radar and lidar instead of pointing out that Green's assumption that Tesla was going to switch radar vendors (or use radar at all!) was delusional.

I know some folks expect technology to stand still.... but that is a silly expectation and with Tesla that is just downright dumb.

The idea was that the core approach (camera's, radar and ultrasonics) are the Tesla way.

Not -

• only 720p cameras produced in 2016
• radar from Bosch or some other vendor
• ultrasonics that can only see 2 lanes over.

Each component of the approach will get refined as time goes.

It seems that you were also under the "delusion" that Tesla would upgrade the radar unit.

 

[mikes_fsd]

Yeah, I just think it's odd that you got into an argument about the differences between radar and lidar instead of pointing out that Green's assumption that Tesla was going to switch radar vendors (or use radar at all!) was delusional.

Yes, how the hell would I -- or green -- know if Tesla was going to go with a new radar? (in production, not just testing)
Just because they are testing one does not guarantee they would deploy one to the fleet.

But showing the track record of delusional posturing from green should make you pause in the future.

"Hey I found this in the firmware" -- observable fact
everything else from him is delusional speculation and he even cements it with "Otherwise why show it to me, right?"
"... and this thing means there will be new hardware on the cars down the line" -- pure speculation
"... and look this thing is like the other thing that they claim is not needed for FSD" -- hallucinations? (if you missed this, this is where he said 4D radar "resembling output of a sensor that Tesla cannot include due to perceived loss of face"... ie lidar

src:

https://twitter.com/x/status/1319363517338759168

It seems that you were also under the "delusion" that Tesla would upgrade the radar unit.

I am not the one making these delusional statements! Green is!
But good attempt at trying to twist this into "you agreed with it because you didn't question it back then!"

 

[ohmman]

Moderator Note: Ideally, arguments should be about the subject matter and include less references to "you." The latter is how we end up with personal attacks, which aren't permissible here. The former hopefully allows us to argue in order to learn. Thanks for being nice to each other moving forward.

 

[dgatwood]

Apart from the fact LIDAR is better for that exact task, there would, again be no reason whatsoever to put that in the production firmware if that was what they were doing. Dev code running on the test car would be all you need.

LIDAR is not necessarily better; "better" depends on the criteria.

• If one of the criteria is "can add it into 1% of random production vision-only cars without anyone noticing or caring, to allow in-the-field verification of our assumptions", then a visually obvious LIDAR attachment is a non-starter.
• If one of the criteria is "can install it on employee-owned test cars without it looking ugly and making the employee mad," then LIDAR is out.
• If one of the criteria is "can reasonably ship this someday without looking like fools for dissing LIDAR," then LIDAR is out.

And running a development branch on a car is fine if it's a company-owned car, but for employee-owned cars, it is problematic. Nobody wants to run an internal version of the software (which may have implications for what passengers can ride in the car) just to be able to collect telemetry for your employer.

They could literally upload a dev build that was exactly the production build plus extra radar support, specifically to those VINs.

Pushing that code to production, going wide to the fleet, is nonsensical if it's for a small group of internal-only testers.

Just because you can do something doesn't mean it's the right thing. Where I work, we can deploy a separate version to 1% of users, but we don't. We deploy the same build, then turn on individual behaviors for 1% of users via configuration flags. Why? Because every extra build that you put out is one more configuration that your release engineering team has to manage.

And when you want to keep releasing that functionality for release after release over multiple months, that sort of change represents a lot of overhead, because suddenly those users won't be able to upgrade to any of the newer public builds unless they remember to turn a new variant of that public build with the extra features. Suddenly you're badly clogging up the works by adding a full round of extra testing to cover a second build of every single OS release you ship.

So it's really not as cut and dried as you think. It's WAY easier to just ship the binary to everybody and turn it on only if the configuration flag is set to indicate that the hardware exists.

This new radar code showed up in fall of 2020.

6+ months before there were 10,000 cars waiting for radar on quality hold, and 6+ months before Tesla announced they were ditching radar.

6+ months before they had built those 10,000 cars, not 6+ months before they knew that they were going to fall short. Tesla almost certainly places their hardware orders a year in advance. Odds are pretty good that they realized that they were going to be in trouble many months before they actually ran out of hardware.

So that can't be it.

(also "having to update all those thousands of cars" is literally a button press, thanks to OTA--- this isn't Volkswagen)

For a fully built car, that's true. However, Tesla's software updates fail if they are unable to update the firmware for any of the expected components. It isn't clear whether it is possible to properly update the firmware on a car that is missing key safety components without manual intervention.

 

[dgatwood]

It doesn't say what an adequate level of redundancy is or what requires any at all. What if an adequate level is none?

Otherwise no current Tesla could ever qualify:

• A single cooling system: no redundancy.

I've had a coolant failure in my car. Tesla determined it was safe to drive five or six miles to the service center. This is not a catastrophic failure at all.

• A single power steering motor: no redundancy.

Assuming you mean the electric steering rack, that is a concern, and various companies are working on that:

Steering Control Unit

With the new modular built Steering Control Unit (control unit with an electro motor) the EPS supports all driver assistance functions and autonomous driving.

www.bosch-mobility-solutions.com

But the failure rate is likely measured in millions of years per failure. The average failure rate from cosmic rays is 4 bits per gigabyte of RAM per month. Whether those bit flips cause a critical failure or not is pure luck. These are not comparable.

• A single brake iBooster: no redundancy.

In addition to using regenerative braking to slow the vehicle down, there's also a separate set of parking brake calipers on the rear wheels. There's plenty of redundancy there.

• A single HV contactor: no redundancy.
• A single set of HV wiring: no redundancy.
• In RWD vehicles: a single drive unit: no redundancy.
• In AWD vehicles: a single primary inverter which if it fails takes the other inverter offline: no redundancy.
• The list goes on and on.

A loss of power does not typically result in a crash. So from a safety perspective, all of those things are a non-issue unless the 12V system somehow goes down suddenly.

Where did Elon say it was required? Just because they were going to do it, or wanted to do it doesn't mean it is required.

If Tesla shipped a system without redundancy, it would literally be the only company to do so. The entire industry decided long ago that having a single computer with no redundancy wasn't an acceptable risk. There's no way regulators would approve a system that lacks redundancy. No chance whatsoever.

 

[stopcrazypp]

There's no way regulators would approve a system that lacks redundancy. No chance whatsoever.

The thing is the argument is over if there is any proof of regulations that require such redundancy. So far there is not a single example given. You are putting a lot of faith in regulators understanding a lot of the nuances of autonomous vehicles, when even tech or vehicle focused journalists can get something as basic as the SAE levels wrong.

We know there are many states that are taking a laissez faire approach. For example, Arizona's recently passed law doesn't require any reporting of how the cars work to the state nor even how often they fail, so obviously redundancy isn't even a concern at all.

Self-driving car bill approved by Gov. Doug Ducey

The first bill to impose any regulation on the self-driving car industry was signed by the Gov. Doug Ducey on March 24.

www.12news.com

On the federal front, although NTSB has pushed for more regulation (even of L2 vehicles) they have no enforcement or rulemaking powers. NHTSA has ignored their suggestions and has not pushed rules yet related to it (only a few voluntary guidances for it, which a quick search through didn't mention redundancy), leaving it to the states for now.

 

[RobDickinson]

Legislation often moves slowly. Theres no guarantee anyone's systems will match future requirements.

Tesla does have two redundant computers both able to cope on their own to bring the car to safety.

 

[dgatwood]

The thing is the argument is over if there is any proof of regulations that require such redundancy. So far there is not a single example given. You are putting a lot of faith in regulators understanding a lot of the nuances of autonomous vehicles, when even tech or vehicle focused journalists can get something as basic as the SAE levels wrong.

I'm putting faith in them having common sense. Every car likely exhibits slightly over one spontaneous bit flip per day. That's at least five or six orders of magnitude too high to be in control of a car.

Right now, there are no examples because the safety regulations don't exist yet. States with autonomous vehicle laws have generally just covered the basic functionality requirements, and left the safety aspects up to NHTSA.

NHTSA's rulemaking plans are described here:

https://www.nhtsa.gov/document/framework-automated-driving-system-safety-advance-notice-proposed-rulemaking

Note that redundancy is mentioned on p. 24 as one of the criteria expected to be addressed in the actual rulemaking. When the rules get nailed down, redundancy will be a requirement.

 

[stopcrazypp]

Legislation often moves slowly. Theres no guarantee anyone's systems will match future requirements.

Tesla does have two redundant computers both able to cope on their own to bring the car to safety.

I've touched on this before many years ago. Basically even when Tesla releases the software that makes the car "FSD" there's no guarantee a regulation is not passed that makes it so that car can't be driven in FSD mode, due to regulations that might for example require lidar.

That said, this discussion is talking about current regulations.

 

[RobDickinson]

Teslas FSD3.0 has ECC memory and afik never has 1 NN making decisions on its own. Its always a consensus between multiple NN.

 

[DanCar]

Teslas FSD3.0 has ECC memory and afik never has 1 NN making decisions on its own. Its always a consensus between multiple NN.

What makes you say that?

 

[RobDickinson]

What makes you say that?

Which part?

The first is from the FSD3.0 specs, the second is from information Karpathy etc have said.