Elon: "Feature complete for full self driving this year"

[diplomat33]

Feature complete does not mean you will want to actually use any of it; it will be a gimmick and will be maddening to attempt to actually use. It will remain so at least through the end of 2020, if not forever (on existing hardware).

I take a more optimistic view. I expect feature complete to be much better than the current NOA. And the current NOA is not a gimmick. Besides, "automatic city driving" has to be better than a gimmick or it would be dangerous. It has to be very reliable or Tesla will not release it.

So I expect "automatic city driving" to be much more than a gimmick.

And the advantage of "feature complete" where we get Highway NOA and City NOA will be that we will have a more seamless autopilot that works everywhere. It won't be like now were we can use AP on highwaya but have to disengage when we get to an intersection. So it will more than a gimmick.

 

[EVNow]

Andrej Karpathy | Multi-Task Learning in the Wilderness

This is a recent talk by Karpathy. This is more NN related - but gives a lot of hints on kind of things Tesla is working on.

For eg., they are working to figure out the road markings, I had not considered earlier.
- pedestrian crossing
- right / left turn signs

The other interesting thing is in terms of - is there a separate NN for FSD we haven't seen or whether the production branch has the dev NN from a few weeks back. Based on the configuration of networks that is described - do you guys think a particular network (say moving objects) can be put on production NN without putting rest of the network ? Would they be going through the complicated training process for a separate FSD NN and a separate production NN ?

ps : One more thing. The impression I get is when they take up a task they train it to get to six 9s (or whatever their target is). They don't just train to 3 nines mark the task done. This has implications on what they really mean by FC. My guess is they are going after few features but want to get to high reliability on those features. Ofcourse overall city NOA will remain unreliable because they haven't started on a number of "edge case" features. So, this is more of a MVP way of doing things rather than real FC for the entire set of features needed for robotaxi.

 

[diplomat33]

Andrej Karpathy | Multi-Task Learning in the Wilderness

This is a recent talk by Karpathy. This is more NN related - but gives a lot of hints on kind of things Tesla is working on.

For eg., they are working to figure out the road markings, I had not considered earlier.
- pedestrian crossing
- right / left turn signs

The other interesting thing is in terms of - is there a separate NN for FSD we haven't seen or whether the production branch has the dev NN from a few weeks back. Based on the configuration of networks that is described - do you guys think a particular network (say moving objects) can be put on production NN without putting rest of the network ? Would they be going through the complicated training process for a separate FSD NN and a separate production NN ?

Thanks for sharing. I will check it out.

I used to think that Tesla had two separate NN branches but now I think it is probably just one and Tesla will just release pieces as they go when those pieces work.

Of course, Tesla will need to feature limit AP for cars that dont have FSD.

 

[rnortman]

ps : For my current commute NOA is not very useful But I won't claim NOA is barely more than EAP. A lot of people use it and find it useful.

Um, NoA is included in EAP. Or was, when you could buy EAP, which you can't anymore.... because it's now way too close in scope to the new definition of FSD, which is sorta my point.

 

[EVNow]

Um, NoA is included in EAP. Or was, when you could buy EAP, which you can't anymore.... because it's now way too close in scope to the new definition of FSD, which is sorta my point.

Its very difficult to make out what your point is - except that you are not happy with Tesla, you are not interested in progress towards FSD - you just want FSD.

 

[EVNow]

Thanks for sharing. I will check it out.

I used to think that Tesla had two separate NN branches but now I think it is probably just one and Tesla will just release pieces as they go when those pieces work.

Of course, Tesla will need to feature limit AP for cars that dont have FSD.

Feature limit is done using procedural code.

Anyway, there are roughly two possibilities,
- Have separate NN for production and dev. The Dev NN is rapidly evolving and getting better for FC - but none of that is coming to production. The production NN is not being improved except perhaps to fix some issues. This is what people who think Tesla is further ahead on FSD than we realize say.
- Single NN the dev team is working on which gets merged to production once in a while, for larger releases. This is easier to manage and more efficient in terms dev hours usage.

 

[EVNow]

One more thing about separate NNs. This tweet seem to point to separate NNs.

This is what EM tweeted on Jul 7th.
​

Production fully switched over ~3 months ago. Functionality won’t diverge until Q4, as it’s limited by software validation. Will be later for Europe compared to rest of world due to regulatory constraints that were put in place years ago by big ICE companies.
​

So let's parse it.

"Functionality won’t diverge until Q4" = Same functionality for HW2 and HW3 until Q4. So, they will get the same NN.
"as it’s limited by software validation" = The new NN for HW3 will finish validation in Q4 and then it will be pushed out to fleet. At that time we will start upgrading the fleet to HW3. HW2 and HW3 cars will then have different NN.
​

 

[daniel]

Recognizing stop lights and stop signs is not super valuable nor super useful on its own

I actually think it would be really nice. Right now, the car will stop if there's a stopped car in my lane, but not otherwise. Having it stop for red lights and stop signs when there's no car ahead of me would be very nice. Probably not worth $5,000, but it would be nice.

Navigating city streets (intersections) is the big thing added to EAP. But it will be L2, and will be a useless gimmick, and therefore I still stand by it being "barely more than EAP" in terms of actual practical value.

I don't think it will be useless because it will have to be as reliable as autosteer is now. Obviously, if you had to intervene every 60 seconds it would be useless, but they'd never get away with that. If you have to intervene once every ten minutes it would be quite nice. Again, probably not worth $5,000 to me. (I think $5K is the upgrade cost? I could be full of beans on that.)

Holy cow, L3 on the highway even without NoA (so just stay in lane, don't take exits, alert driver with enough warning to take over if anything else is required) would be amazing if it actually worked. The original description of EAP implied L3 ("on ramp to off ramp with no intervention required by the driver" or something like that), which does admittedly give them plenty of wiggle room. But a lot of people who bought EAP read that as "L3 on the highway". In fact, on these forums, there were plenty of people arguing that EAP was L3 or even L4 on the highway when it was first announced.

I never imagined that EAP would be L3 any time soon. But maybe I'm just more skeptical or pessimistic than some folks are.

Wouldn't it be great for us all if I am wrong...

I would love to be wrong and see real FSD released this year. But for this to happen, Tesla would have to be years of development ahead of anything we've seen or have reason to think they have. There'd have to have been be a cosmic breakthrough in AI that they're keeping secret.

I'd be surprised if they could get NoA/city with stop sign and stoplight recognition working well enough to release to the public by the end of 2020. I really, really want them to succeed because in a decade I'll be at the age where most people have become road hazards and if I'm still ambulatory I want a car I don't have to "drive." I'm just thinking more and more that training a car for a long list of edge cases will never get us to L5, and that what's needed is an AI that can handle novel situations that it has not been specifically trained for. I think this is a decade away. I really want to be wrong!

 

[mongo]

I actually think it would be really nice. Right now, the car will stop if there's a stopped car in my lane, but not otherwise. Having it stop for red lights and stop signs when there's no car ahead of me would be very nice. Probably not worth $5,000, but it would be nice.

Plus I expect HW3 (and HW2?) cars will have red light/ stop sign detection as an additional safety feature.

 

[EVNow]

Again, probably not worth $5,000 to me. (I think $5K is the upgrade cost? I could be full of beans on that.)

$6,000 (AP is included in all but SR, which is a special order).

I bought EAP+FSD for when it was on sale for $5k.

 

[BigD0g]

Plus I expect HW3 (and HW2?) cars will have red light/ stop sign detection as an additional safety feature.

They already due, it's just disabled in software, and it's definitely not 100%.

 

[daniel]

$6,000 (AP is included in all but SR, which is a special order).

I bought EAP+FSD for when it was on sale for $5k.

I paid $5,000 for EAP at the time of purchase, and though I cannot find it now, I thought I remembered that the contract said I could buy FSD later for another $5,000. But maybe that was the then-current cost of FSD after the sale, and not a promise that it would continue to be the price. I feel kind of bad about not getting the new computer, but I'm not at all sure it would give me anything I really care about, since I expect to be ready for a new car before even Level 3 is available to the public. In 5 years they'll be on HW 5 and if it has features I really want I'll trade up.

 

[mongo]

They already due, it's just disabled in software, and it's definitely not 100%.

Right, I'm thinking that all cars/ drivers will get the additional safety features of the FSD code watching over their shoulder even if people don't pay for the FSD convenience functions.

I won't drive for you, but I will protect you. Will be interesting to see how that is implemented.

 

[EVNow]

Right, I'm thinking that all cars/ drivers will get the additional safety features of the FSD code watching over their shoulder even if people don't pay for the FSD convenience functions.

I won't drive for you, but I will protect you. Will be interesting to see how that is implemented.

Definitely they won't try to upgrade to HW3 people who haven't paid for FSD. Lot of cost for no revenue.

Its possible they will do minimum maintenance on HW2 - but unlikely they will spend a lot of resources trying to get them new features.

 

[daniel]

Right, I'm thinking that all cars/ drivers will get the additional safety features of the FSD code watching over their shoulder even if people don't pay for the FSD convenience functions.

I won't drive for you, but I will protect you. Will be interesting to see how that is implemented.

If those functions require HW 3, then HW 2.5 cars won't get them.

ETA: But new cars that come with HW 3 might.

 

[mongo]

Definitely they won't try to upgrade to HW3 people who haven't paid for FSD. Lot of cost for no revenue.

Its possible they will do minimum maintenance on HW2 - but unlikely they will spend a lot of resources trying to get them new features.

If those functions require HW 3, then HW 2.5 cars won't get them.

ETA: But new cars that come with HW 3 might.

Yah, but isn't red light/ stop light rolling out with 2.x now? I was thinking HW 2.x still had some growth left in it, so didn't want to limit safety improvements to only HW3.

 

[EVNow]

Yah, but isn't red light/ stop light rolling out with 2.x now? I was thinking HW 2.x still had some growth left in it, so didn't want to limit safety improvements to only HW3.

I never got it. Is it still a thing ?

Basically - what we don't know is how much work it is for Tesla to make something work on both HW3 & HW2. If making NN backward compatible makes the progress slower, I'm not sure they should do it. They should just concentrate on HW3 for FC - and then later try to bring it to HW2. Basically make the target HW3 rather than forcing backward compatibility.

ps : Other option is to offer to upgrade HW2 to HW3 for a price ($1000 or so).

 

[Sean Wagner]

I'm just thinking more and more that training a car for a long list of edge cases will never get us to L5, and that what's needed is an AI that can handle novel situations that it has not been specifically trained for.

Humans integrate data over time [ie. retain context] and manage attention by knowing what to expect in different circumstances.
Managing inter-vehicle communication then becomes easier too.

While forward thinking drivers take it a step further and are prepared for the unexpected and will, if required, focus attention so. And honk and brake when a soccer ball rolls out onto a residential street, but stay in their lane on a highway in rush hour. Kid not likely to follow down the overpass.

One of the reasons I was so astonished when the notion of drivable space was bandied about as the great new thing fairly recently. I think more and better real-time logic should lighten the load on image recognition.

 

[cutesheep]

Hi every automated vehicles fans,
I really need your help and input regarding my dissertation! I would really appreciate it if you could feel out my online survey regarding the features of automated vehicles in the online community. (Im postgraduate of University of Warwick)

Please click on the link below.. it shouldn't take more than 3 mins to complete it. Thank you!!!!!!
https://warwickwmg.eu.qualtrics.com/jfe/form/SV_7WdAPlhRJOSLDjT

 

[daniel]

Yah, but isn't red light/ stop light rolling out with 2.x now? I was thinking HW 2.x still had some growth left in it, so didn't want to limit safety improvements to only HW3.

I don't know anything about this. I have not noticed anything about it in the release notes of my latest firmware update.