Does AP2's NN trickle down to AP1 development?

[fasteddie7]

This is a question for those who know way more than me. Does AP2's NN and development of Tesla vision translate to improvements in AP1 where the team can easily translate the learning and improvements to AP1 for whatever hardware will allow or does Tesla have dedicated resources for AP1 development which will obviously dwindle over time. Do they evolve in tandem or is the AP1 and AP2 development different with not much to share between the two?

 

[MP3Mike]

This is a question for those who know way more than me. Does AP2's NN and development of Tesla vision translate to improvements in AP1 where the team can easily translate the learning and improvements to AP1 for whatever hardware will allow or does Tesla have dedicated resources for AP1 development which will obviously dwindle over time. Do they evolve in tandem or is the AP1 and AP2 development different with not much to share between the two?

AP1 is based on MobileEye hardware and isn't likely to benefit from any of the Tesla developed vision/NN.

AP1 could benefit some from driving logic updates, which we don't think are NN based, that are done for AP2, but I wouldn't expect AP1 to change much going forward.

 

[fasteddie7]

Thank you, I only ask because I honestly don't know. What about things that Mobileye hardware is capable of, such as stop sign detection? Who are we at the mercy of with the chance of features being added, Tesla, mobileye, or both?

 

[MP3Mike]

Thank you, I only ask because I honestly don't know. What about things that Mobileye hardware is capable of, such as stop sign detection? Who are we at the mercy of with the chance of features being added, Tesla, mobileye, or both?

Tesla for sure, but possibly MobileEye as well. My guess is that the hardware in AP1 isn't capable of much more and Tesla isn't going to invest significant resources in trying to add features to it.

 

[MarcusMaximus]

Tesla for sure, but possibly MobileEye as well. My guess is that the hardware in AP1 isn't capable of much more and Tesla isn't going to invest significant resources in trying to add features to it.

Unfortunately for AP1 owners, neither company has much reason to continue work. MobilEye certainly doesn’t, given the public, messy breakup between the two, and Tesla has every reason to favor AP2.

 

[whitex]

Unfortunately for AP1 owners, neither company has much reason to continue work. MobilEye certainly doesn’t, given the public, messy breakup between the two, and Tesla has every reason to favor AP2.

Even if Mobile Eye has more capabilites it would make no sense for Tesla to enable them unless AP2 develops those capabilities as well, or AP1 will just make look AP2 even more inferior. If ME has for example stop sign detection, then only if AP2 also has that and it's implemented in the driving logic, then maybe, if it doesn't take much effort, Tesla could enable it on ME.

 

[robby]

It's very unlikely that Tesla's vision neural network will apply to MobilEye chips, but an AP1 EyeQ chip can recognize more things than current AP2 processing and current AP1 programming. So it's likely that there is some overlap between what an AP1 chip can (but does not) do today and what Tesla intends for an AP2 chip to do tomorrow. I would expect them to deliver functionality to AP1 cars where such overlap exists.

Improvements may also be possible beyond what is vision-based. One example of this came last year when Tesla discovered the Bosch radar could be used to see two cars ahead by bouncing radio waves under the car immediately in front. That had little if anything to do with the EyeQ chip. Elon has said that AP1 cars will eventually be able to take offramps between highways without disengaging Autopilot; that appears to be purely programmatic at this point.

We are certainly past the prime of AP1 development and there are no doubt diminishing returns on investment there, but I expect a handful of small, helpful improvements in the years to come.

 

[dss33]

As much sense as it does indeed make to focus on AP 2+ development now, Tesla has more incentive than you'd otherwise assume to update AP1 with little improvements for the next couple of years. 1) It maintains AP1 car values longer and 2) they're going to have a good number of AP1 car leases coming up over the next two years so it's in their best interest to maintain high residual values. It isn't a major factor of course - development is expensive and time is precious - but it ain't nothing.

 

[MarcusMaximus]

but an AP1 EyeQ chip can recognize more things than current AP2 processing and current AP1 programming

I’d urge caution on that. You’re comparing unreleased potential functionality against actual end user functionality. It’s unknown what Tesla’s system is capable of but not yet enabled to do. We know there are NN output units that aren’t currently used to control end user car behavior.

 

[robby]

I’d urge caution on that. You’re comparing unreleased potential functionality against actual end user functionality. It’s unknown what Tesla’s system is capable of but not yet enabled to do. We know there are NN output units that aren’t currently used to control end user car behavior.

Oh, for sure. I just mean what customer cars can do today.

 

[whitex]

they're going to have a good number of AP1 car leases coming up over the next two years so it's in their best interest to maintain high residual values.

The cars coming off the leases come back to the leasing company (or a bank) which owns them, not Tesla. Tesla may buy them but if their value has dropped they will just offer less for them. While Tesla may have other reasons to want residual values to stay high, lease returns is not one of them.

 

[dss33]

The cars coming off the leases come back to the leasing company (or a bank) which owns them, not Tesla. Tesla may buy them but if their value has dropped they will just offer less for them. While Tesla may have other reasons to want residual values to stay high, lease returns is not one of them.

I'm not sure what the breakdown of leases is, though Tesla does do its own lease financing. We were declined by US Bank (still not sure why) so Tesla Financing stepped in and approved. So at least for those cars, it's particularly worth it.

Still probably not a primary factor - I think Elon and the company also see it as another way to thank early owners, but in the end, it's only wise to invest far more time and energy into AP2+ as I fully support that as an AP1 owner.

 

[ev-now]

I see very little major functionality being add to AP1 due to two factors; cost of development and cost of testing. Seriously, it makes no sense to maintain value in the older fleet when AP2 is barely if at all ahead of AP1. Would love to be proved wrong, but I doubt it will happen - it's not like the new system is remotely close to the old sensor system so any new functionality is all added cost.

I just hope we do, finally, get a usable Nav - or give up and allow my phone a two-way connection to the screen!

 

[wolfson292]

There have been no changes to the AP1 firmware since the end of 2016.

 

[BigD0g]

AP1 is EOL in terms of AP, maybe some better radar algorithms or such, but doubtful. I'd only expect feature updates in software is all, maybe that elusive better web browser!

Since, AP1, AP2 and AP2.5 (x/s) all use the same old 10 yr tegra processor and CID. So, any UI / Features should be cross all platforms I would expect. Exept m3, since it's intel based and peppier I suspect, but don't know for sure.

 

[AnxietyRanger]

I believe on there was some quote on TMC recently from someone at Tesla (or was it rumor?) that Tesla is looking to combining AP1/AP2 (or maybe just EAP?) codebase in a manner that can be maintained uniformly. This may already be happening to some extent, we know EAP emulates MobilEye in code anyway...

So it certainly seems possible Tesla is trying to combine various hardware iterations (AP2/2.5 too) within code to packages that can serve a multitude of different target systems - and thus allow bringing new features to even old systems to the extent of their capabilities. Then again, there is also some precedent that Tesla tends to abandon systems no longer in latest production cars (e.g. pre-AP parking sensors never got any improvements after AP1 came, even though technically they probably could have...).

Just now pre-AP Teslas didn't get the Chill mode... speculatedly because Tesla didn't bother to add Chill text to the aging Classic dial on pre-AP 7.x/8.x firmware... P85 can't CHILL?

 

[BigD0g]

I believe on there was some quote on TMC recently from someone at Tesla (or was it rumor?) that Tesla is looking to combining AP1/AP2 (or maybe just EAP?) codebase in a manner that can be maintained uniformly. This may already be happening to some extent, we know EAP emulates MobilEye in code anyway...

So it certainly seems possible Tesla is trying to combine various hardware iterations (AP2/2.5 too) within code to packages that can serve a multitude of different target systems - and thus allow bringing new features to even old systems to the extent of their capabilities. Then again, there is also some precedent that Tesla tends to abandon systems no longer in latest production cars (e.g. pre-AP parking sensors never got any improvements after AP1 came, even though technically they probably could have...).

Just now pre-AP Teslas didn't get the Chill mode... speculatedly because Tesla didn't bother to add Chill text to the aging Classic dial on pre-AP 7.x/8.x firmware... P85 can't CHILL?

No, it’s fundamentally not possible, AP2 emulates AP1 currently as a stop gab, until the “new” stuff is up and running. It’s all custom and it’s all running on a dedicated ape box.
Yes, AP2/2.5 are the same codebase but they are all the same architecture with simply more processing and redundancy on 2.5, but AP1. Is not gettting merged in. Heck, even Elon tweeted they currently don’t have any more processing power available on AP1.

 

[AnxietyRanger]

No, it’s fundamentally not possible, AP2 emulates AP1 currently as a stop gab, until the “new” stuff is up and running. It’s all custom and it’s all running on a dedicated ape box.
Yes, AP2/2.5 are the same codebase but they are all the same architecture with simply more processing and redundancy on 2.5, but AP1. Is not gettting merged in. Heck, even Elon tweeted they currently don’t have any more processing power available on AP1.

I get what you're saying, but IMO you may be a tad bit excessive with your comment. There are many levels at which the merging of codebases could happen.

Who remembers where this comment was recently from a Tesla employee commenting on the challenge/need of merging AP1/AP2 codebases?

 

[AnxietyRanger]

@verygreen @lunitiks do you recall ^ ?

 

[Dax279]

I don’t recall, but are some the the AP1 features not still Beta? Perhaps they could finalize the software and remove the Beta warnings.