George Hotz Interview: Comma.ai, OpenPilot, Autonomous Vehicles

[diplomat33]

Lex Friedman interviews George Hotz:

At the 34 mn mark, Lex says that he thinks Tesla is wrong to pursue L4 because he thinks lane keeping is the only thing that really adds value to customers. I don't agree. Sure, lane keeping does add value. I am sure there are lot of drivers who would be perfectly happy with a car that just centers the lane perfectly and let's them to the rest. But there are plenty of other features that will add value too. To limit yourself to just offering lane keeping seems very short sighted. Even Hotz corrects Lex and says that you need ACC (Adaptive Cruise Control) too, not just lane keeping. And eventually, L4 will become mainstream and you will be left behind if all you offer is lane keeping.

I do like Hotz' answer that you absolutely need driver monitoring, especially when your self-driving starts to get good. In other words, it's that twilight zone in between L2 and L3, where your self-driving car is good enough that the driver thinks they don't need to pay attention anymore but not quite good enough for them to actually stop paying attention, that's when you need driver monitoring the most.

 

[diplomat33]

Discussion on autonomous vehicles starts around the 1 hour mark. Pretty interesting stuff.

Hotz feels that NOA is too robotic in how it does auto lane changes. He believes end to end machine learning is the only way to do L5 autonomy. His best advice to Tesla is to add a driver monitoring system in their cars ASAP.

Hotz feels that Waymo can probably "hack" their way to a L4 geofenced system by next year but he doubts their approach will get to L5.

 

[EVNow]

He also thinks FSD is 2 years away ?

 

[diplomat33]

He also thinks FSD is 2 years away ?

No Hotz does not think FSD is 2 years away. He thinks any company promising FSD in 2 years is way too optimistic.

 

[EVNow]

No Hotz does not think FSD is 2 years away. He thinks any company promising FSD in 2 years is way too optimistic.

I haven't listened to it yet - someone posted wrong info in the investment thread, then.

ps : What he meant was, they will be two years behind Tesla in L5. @39 minutes. He thinks Tesla gets there first … doesn't know when.

 

[diplomat33]

I haven't listened to it yet - someone posted wrong info in the investment thread, then.

ps : What he meant was, they will be two years behind Tesla in L5. @39 minutes. He thinks Tesla gets there first … doesn't know when.

At the 1:08 mark, he also laughs when Lex says that other companies are promising FSD in 2020-2022 and he says he would bet against them. He adds that he thinks Waymo can "hack" their way to L4 geofence that only works in Chandler, AZ by 2020-21 but Waymo won't get to "L5 New York City".

Hotz also says that it was a mistake for Elon to promise L5 robotaxis by end of next year.

 

[EVNow]

That interview was interesting (and then some Wikipedia linked articles) - didn't know his history with Musk.

 

[Soda Popinski]

Finally got around to watching this. I don't know enough to have an informed opinion, but my B.S. Peter Tingle was going off at times.

Do we know if Tesla is planning on using NN for driving policy, or hand coding like the current NoA? I'd guess using NN would require the HW3/FSD computer. If this is the case, we can't really base what FSD will be like based on either NoA or Smart Summon.

 

[BigD0g]

Do we know if Tesla is planning on using NN for driving policy, or hand coding like the current NoA? I'd guess using NN would require the HW3/FSD computer. If this is the case, we can't really base what FSD will be like based on either NoA or Smart Summon.

Kaparthy, appears to be advocating for this approach yes. Google his software 2.0 approach talk

 

[diplomat33]

I don't know enough to have an informed opinion, but my B.S. Peter Tingle was going off at times.

Probably because Hotz, like Musk, has a way of speaking very confidently about what is or might be possible in the future where you are not sure if they are crazy or crazy genius.

Do we know if Tesla is planning on using NN for driving policy, or hand coding like the current NoA? I'd guess using NN would require the HW3/FSD computer. If this is the case, we can't really base what FSD will be like based on either NoA or Smart Summon.

We don't know at this point. I think Tesla wants to use NN for driving policy but they are not doing it yet. I think that is a point that Hotz brings up actually in his criticism of NoA. He says that the vision is NN based but that the auto lane changes are hard coded probably by a different team. The hard coded nature of the auto lane changes is why they feel so identical and robotic. NN based driving policy would make NoA much smarter because it could do auto lane changes differently and react differently to different situations rather than just doing the same auto lane change every time.

I do agree that if Tesla does implement NN based driving policy, taking advantage of HW3, that FSD has the potential to be radically better than what we have now. And no, we can't really use the current AP or NOA on AP2 to judge FSD.

 

[mspohr]

Here's his system running on a 2014 Model S

Early Tesla Model S gets 'Autopilot' retrofit and it works surprisingly well

George Hotz of comma.ai sees his open-sourced driver-assist system openpilot as the Android of the full self-driving market. While Tesla’s Autopilot system is much like Apple’s iOS in the way that it is confined to the electric car maker’s parameters, openpilot allows owners to dig deeper into their vehicles, and this results in some cars, including a pre-Autopilot 2014 Tesla Model S, gaining capabilities that are nearly comparable to AP1 cars.