Firmware 9 in August will start rolling out full self-driving features!!!

[MelaniainLA]

Everyone loves man made neural nets until some dude with ipsilateral right nasal hemianopia keeps drifting left on the highway, then you are going to pray that that man has some macular sparing before he drifts all up in your grill and causes a 4 car pile up.... just saying....

You’re an ophthalmologist too?!

 

[conman]

You guys charge too much, that's why you can both afford these cars...
/me says nothing about his own fees.

 

[MarcusMaximus]

You guys charge too much, that's why you can both afford these cars...
/me says nothing about his own fees.

On a salary, working for The Man, here.

 

[buttershrimp]

You’re an ophthalmologist too?!

No, but I sure did love myself some neuro visual fields during med school. Fascinating stuff.

 

[buttershrimp]

You guys charge too much, that's why you can both afford these cars...
/me says nothing about his own fees.

Well, when you factor in the 12-13 years of 100 hour week learning and debt, and the work hours, and the life saving/life changing.... I vote to raise the fees in the new year.
Feel free to cut ceo and pharma Pay. Also please cut insurance exec pay too.

Yours truly,
Buttershrimp

 

[conman]

Well, when you factor in the 12-13 years of 100 hour week learning and debt, and the work hours, and the life saving/life changing.... I vote to raise the fees in the new year.
Feel free to cut ceo and pharma Pay. Also please cut insurance exec pay too.

I'm a doctor of sorts too... but on the other side of the planet.

 

[MelaniainLA]

You guys charge too much, that's why you can both afford these cars...
/me says nothing about his own fees.

After Obamacare I was making $60-90k a year. They killed my profession. I quit for profit medicine. Went into biotech. And now care for homeless/prisoners/illegals at a public hospital for free. Not happy parking a Tesla there however!!

 

[conman]

After Obamacare I was making $60-90k a year. They killed my profession. I quit for profit medicine. Went into biotech. And now care for homeless/prisoners/illegals at a public hospital for free. Not happy parking a Tesla there however!!

I've only ever worked in public hospitals down here (.au) Parking was always an issue - the model S barely made it through the parking gates they were so tight D: I refused to go through one of the gates as I was sure it wouldn't fit. But now I've taken us way offtopic... on the other hand, I've looked at those gates MANY times and wondered just how TF full self drive would ever bring the car down the multilevel carpark and through those gates to meet you at the front door of the hospital and concluded it NEVER would.

 

[rnortman]

Also we have have necks so we can turn our heads, that's where the multiple cameras come in.

But yeah you are absolutely right. There should be enough sensors. Elon's thought is like this: If a human can see and do it, the deep learning neural net should be able to do the same.

HW3 supposedly processes 2000 FPS which means you have a tick of 500µs, which is pretty amazing for this kind of system. Not sure if that's for one camera or all 8. But worst case you would have 2000 / 8 = 4ms tick rate, which is more than good enough.

My guess it that progress will seem slow, but once they nail it, it will be amazing, and the competitors will be left in the dust. I am convinced that the deep learning + "normal" cameras is the way to go.

Here is the problem -- you assume that deep learning technology, given sufficient data to train on and fast enough hardware to run inference on, can solve the autonomy problem. That remains an unproven and highly dubious assertion -- we know the human brain operates on more than just gradient descent training but we don't yet understand how it does everything it does. There is no good reason to be confident that we can just throw hardware at this problem and make it go away. (History shows that Elon does not need good reasons to be confident.)

It does seem that deep learning, given sufficient data and hardware, has almost completely solved the detection and recognition problems, which is a big part of the puzzle. But it's not the whole thing.

And you'll still always be limited by what your sensors can see -- As a concrete example, my Tesla seems to have a lot of problems seeing cars that are on the other side of a valley as I'm headed down into the valley and the other cars are sitting on the uphill side. (I'm talking small "valleys" that you encounter in normal driving -- on-ramps, off-ramps, overpasses, etc. -- not giant mountain valleys.) They are completely invisible to it until I start uphill -- I suspect they're outside the FOV of the forward cameras. How is a new AI chip going to solve that problem? Maybe if they start using the wide-angle camera I guess they could fix this. There are a lot of "ifs" here.

Yes it will be level 3-4 not 5. It says "not all circumstances" on the FSD option. So that is probably to not be legally liable. I don't believe it has limitations as I stated above.

We can debate what L3 and L4 mean all over again (let's not) but let me put it this way: FSD, like EAP, will forever be a driver assistance system. Maybe a very, very good driver assistance system. It may operate the car 99% of the time without the human driver doing anything. But you will not be able to operate the car without a human driver in it on the current hardware suite, and the human will be legally responsible for what the car does. But Tesla has clearly implied that the car will not require a driver -- the whole "Tesla Network" thing implies the car can drive itself around and pick up passengers, drop them off, and then go pick up another passenger. And you can summon the car across the country. This will not happen on current model vehicles without substantial hardware upgrades. (Which might, maybe, in 4-5 years time, mean just compute upgrades using the existing sensors, once they make major advances in the software, but I suspect new sensors will also be required for any practical autonomous use.)

This is why I say that FSD is a lie. I do believe it will become a very good driver assistance system, at least by today's standards. Unfortunately, it will not be today by the time they finally get it figured out, and other makers will have their own kickass driver assistance systems available by then. (Probably with rear corner radar for blind spot monitoring, and maybe with lidar.) And the real AV companies will have truly autonomous commercial fleets operating as well by then. Tesla's toy FSD will seem very ho-hum at that point.

 

[rnortman]

There’s nothing in parallax vision that requires matched FOV. So long as the FOV on each individual camera is fixed, stereo vision is fully possible. The only potential difficulty with that with Tesla’s sensor suite is the wide angle-lens producing spatial distortions on the outer edge, which doesn’t tend to work well with standard convolutions.

Yes and 3D information can be reconstructed by the lateral movement of the single camera. It essence, the forward movement of several feet between frames creates a stereo pair. This is how single-camera matchmoving software works in film production.

Sure, in pinciple they can do stereo. And yes, in a moving vehicle you can do temporal parallax even with a single camera using consecutive frames. But it's much harder and more limited (esp if the things you're looking at are also moving). Right now, today, if you have a matched and aligned camera pair, you can get a cheap off-the-shelf ASIC that gives you a depth field in real time, no need to develop your own sophisticated software or run it on an expensive and power-hungry CPU/GPU. But Tesla has cut off this option and made it hard on themselves by not including a stereo pair.

 

[BigD0g]

In fairness, the AP program has changed hands a couple times since then. We had an AP director who guided the whole program down a useless rewrite that had to be completely re-written again to fix it. That whole thing took, what? A year? Under Karpathy things have completely stabilized, especially starting in 10.4. Unless he leaves anytime soon, I wouldn’t expect the pace to slow down again like it had last year.

I'm sorry, as a customer, I fail to see how that is my fault, and how I contributed to that failure in any way shape or form. I bought based on a video sold as working software under going validation in Dec 2016, to realize it was all smoke an mirrors and frankly lies. Also, if KP leaves, what happens then?

 

[dbldwn02]

If I’m the company taking eventual responsibility for traffic violations, I’d want the system to be based on sign reading AND maps database.

Ideally with some checks and balances to identify changes and ensure correct results.

A car driving through a small speed trap town could easily miss a sign if driving next to a truck on a Texas 2 lane each direction “highway”.

So I'm driving a loaner AP1 and just realized it has sign detection. Mind blown! Haha. AP2 software isn't even at parity with AP1 and we're talking about FSD features? It's gonna be a while.

 

[brkaus]

So I'm driving a loaner AP1 and just realized it has sign detection. Mind blown! Haha. AP2 software isn't even at parity with AP1 and we're talking about FSD features? It's gonna be a while.

And the best news is AP1 will also detect speed limit stickers on the back of some rental trailers, speed limit signs intended for exit ramps or frontage roads while on the main highway, etc. It also ignores complicated school speed limit signs. It's going to have to be a combined solution. Yup, its gonna be a while.

 

[rnortman]

I'm sorry, as a customer, I fail to see how that is my fault, and how I contributed to that failure in any way shape or form. I bought based on a video sold as working software under going validation in Dec 2016, to realize it was all smoke an mirrors and frankly lies. Also, if KP leaves, what happens then?

And this is why companies should not sell and accept payment for products that they have not yet developed, at least not without being very upfront about the fact that it's still in the R&D stage. Instead Tesla's marketing did everything they could to imply that it was pretty much ready to go.

 

[buttershrimp]

I've only ever worked in public hospitals down here (.au) Parking was always an issue - the model S barely made it through the parking gates they were so tight D: I refused to go through one of the gates as I was sure it wouldn't fit. But now I've taken us way offtopic... on the other hand, I've looked at those gates MANY times and wondered just how TF full self drive would ever bring the car down the multilevel carpark and through those gates to meet you at the front door of the hospital and concluded it NEVER would.

Can you help buttershrimp move there? I’m ready.

 

[MarcusMaximus]

I'm sorry, as a customer, I fail to see how that is my fault, and how I contributed to that failure in any way shape or form. I bought based on a video sold as working software under going validation in Dec 2016, to realize it was all smoke an mirrors and frankly lies. Also, if KP leaves, what happens then?

Didn’t intend to say it was. I’m not excusing what was done in the past, just saying that, assuming nothing crazy happens, we’ve seen lately and can expect to continue to see a more rapid pace of improvement. Up to now, that’s been just really solidifying and perfecting the existing features, with small new ones(like automatically changing speed on freeway exits).

 

[emmz0r]

Here is the problem -- you assume that deep learning technology, given sufficient data to train on and fast enough hardware to run inference on, can solve the autonomy problem. That remains an unproven and highly dubious assertion -- we know the human brain operates on more than just gradient descent training but we don't yet understand how it does everything it does. There is no good reason to be confident that we can just throw hardware at this problem and make it go away. (History shows that Elon does not need good reasons to be confident.)

It does seem that deep learning, given sufficient data and hardware, has almost completely solved the detection and recognition problems, which is a big part of the puzzle. But it's not the whole thing.

Not everything has to be done with deep learning. You do normal software programming (if else then). Current hardware should be sufficient. You need logics on top to deal with the context and transitions between scenarios. That is fully solvable.

And you'll still always be limited by what your sensors can see -- As a concrete example, my Tesla seems to have a lot of problems seeing cars that are on the other side of a valley as I'm headed down into the valley and the other cars are sitting on the uphill side. (I'm talking small "valleys" that you encounter in normal driving -- on-ramps, off-ramps, overpasses, etc. -- not giant mountain valleys.) They are completely invisible to it until I start uphill -- I suspect they're outside the FOV of the forward cameras. How is a new AI chip going to solve that problem? Maybe if they start using the wide-angle camera I guess they could fix this. There are a lot of "ifs" here.

One of the situations that can be coded / handcrafted. If it is classified as a hill based on tilt, horizon, be ware that there might be cars. Common sense has to be coded in. Maybe even that can be learned in the long run through positive reinforcement.

We can debate what L3 and L4 mean all over again (let's not) but let me put it this way: FSD, like EAP, will forever be a driver assistance system. Maybe a very, very good driver assistance system. It may operate the car 99% of the time without the human driver doing anything. But you will not be able to operate the car without a human driver in it on the current hardware suite, and the human will be legally responsible for what the car does. But Tesla has clearly implied that the car will not require a driver -- the whole "Tesla Network" thing implies the car can drive itself around and pick up passengers, drop them off, and then go pick up another passenger. And you can summon the car across the country. This will not happen on current model vehicles without substantial hardware upgrades. (Which might, maybe, in 4-5 years time, mean just compute upgrades using the existing sensors, once they make major advances in the software, but I suspect new sensors will also be required for any practical autonomous use.)

This is why I say that FSD is a lie. I do believe it will become a very good driver assistance system, at least by today's standards. Unfortunately, it will not be today by the time they finally get it figured out, and other makers will have their own kickass driver assistance systems available by then. (Probably with rear corner radar for blind spot monitoring, and maybe with lidar.) And the real AV companies will have truly autonomous commercial fleets operating as well by then. Tesla's toy FSD will seem very ho-hum at that point.

I agree that Tesla Network and level 5 is unachievable in 2018, but we'll see. Other automakers will run into the same problems as Tesla have, but with the deep learning approach they are ahead of the curve since it will yield a more "organic" response to situations. Chevy's SuperCruise is geofenced to death, has a high precision GPS, and records your face while driving. If Tesla is a long way off, then other makers are wayyyyyyyyy off. Also LIDAR is not the way either since it doesn't perform in rough conditions, It emits light and that gets disturbed by water droplets etc.

 

[MarcusMaximus]

Sure, in pinciple they can do stereo. And yes, in a moving vehicle you can do temporal parallax even with a single camera using consecutive frames. But it's much harder and more limited (esp if the things you're looking at are also moving). Right now, today, if you have a matched and aligned camera pair, you can get a cheap off-the-shelf ASIC that gives you a depth field in real time, no need to develop your own sophisticated software or run it on an expensive and power-hungry CPU/GPU. But Tesla has cut off this option and made it hard on themselves by not including a stereo pair.

Ah, that. Sure, they could have, but that’d also be much more limited in its capabilities. There’s a couple problems with doing that:

1. If this can be done, then it’ll take some heavy duty machine learning to do it. Just putting some hard coded logic on top of a 3D mask isn’t going to do it(otherwise, everyone would already have it).

2. In general with these networks, you don’t want crazy preprocessing like you describe on the input. Take the raw input, do some basic proprocessing to limit the range(typically, mean variance normalization) and let the network learn what it needs to pick out from it. You may *think* you know what the network needs, but as Karpathy says: SGD can write better code than you.

 

[dbldwn02]

And the best news is AP1 will also detect speed limit stickers on the back of some rental trailers, speed limit signs intended for exit ramps or frontage roads while on the main highway, etc. It also ignores complicated school speed limit signs. It's going to have to be a combined solution. Yup, its gonna be a while.

Oh snap! I didn't know that. With that in mind, I revise my previous estimate...3 decades maybe...6 definitely.

 

[daktari]

I'm a doctor of sorts too... but on the other side of the planet.

Hey, guess what my nick translates too!