I would challenge two parts of Musk's argument: that a computer camera system can cost effectively emulate human driving and vision performance and the idea that humans are safe drivers with only eyes.
I’m interested in comparing the ratio of (accidents / miles driven) between the two.
Tesla seems like it has an engineering culture built around shipping whatever version of product they have at the date Elon announces, regardless of the quality (based on the engineers I have met who have worked there).
To me, self driving seems like the opposite of Moore's law.
In the last decades, one of the principle of SWE was to take into account how much computing power will have improved when planning multi-year project. Meaning, you could write something that was too heavy for today's machines, but would be bleeding edge in 5 years.
IMHO, self-driving is actually the inverse situation.
Piloting a car in an human-centered environment is difficult and requires the machine to behave humanly. This of course requires an absurd amount od data and training to pull off. But what happens when self driving adoption increases? At a certain point, so many driverless cars will be roaming the streets that most daily interactions will be automated by letting the cars negotiate in a nice, deterministic, algorithmic way. Thus, reliance on predictive and opaque systems like neural networks will be needed less and less, actually reducing the complexity of self driving.
My main point is: what should win is the tech that makes cars drive as well as trained professional drivers. Once that's done, adoption will drive down human driving and thus unpredictable behavior on the road, reducing the computational load needed to correctly perform tasks. Next, cars will start to behave more programmatically and deterministic and will need less sensors and tech. Car companies will have accurate maps of everything, and cars will mostly become shuttles which can rely more on predetermined routines and less on world models, especially as smart cities gain a foodhold too.
There might still be significant pushback at least for one more generation. Although we can see things move quickly now that the war on general purpose computation is moving quickly(the recent 3D printer ban proposals, introducing age verification at the OS level etc.) so many things might move slowly for a long time and then move fast all at once.
>Car companies will have accurate maps of everything, and cars will mostly become shuttles which can rely more on predetermined routines and less on world models, especially as smart cities gain a foothold too.
Car companies won't have squat. The whole point of GM buying Cruise and others trying to get into self driving was that they will be relegated to white box manufacturers if they dont try and bring this tech in house. Its funny how the MBAs at these companies tried to outsource all manufacturing to 'suppliers' such that all they really wanted to do was stick the badge on the car at the end. Now they realize this thinking is going to take themselves out of the equation as well....whoops. If your vision comes to pass why would anyone care what badge is on the front of the car?
This is probably why Waymo had to use Jaguar i-Paces: only companies desperate to offload their unsold inventory would cooperate with them.
Depends on what non-car things you are talking about. If we're talking about an obstacle on the road or a pedestrian, a model to take care of that is almost trivial, since the car would simply signal everything around it that there is a problem and it's performing a manouver, and it will not have to worry about anything other than the obstacle and the road boundaries.
The idea that hardware costs matter most is idiotic^H^H^H^H^H^H^H silly. If you are buying a car, it matters, somewhat. If you are riding in a taxi or leasing a car, you are paying for a service. When you go to a restaurant, is their profitability predicated on how much they payed for their ovens?
LIDAR costs $1000, whoops $100, whoops $10, whoops it's on a chip module that costs a tenth of a cent. Building your business model on hardware NOT getting cheaper seems silly. Sure, software is cheap and getting cheaper, but
Sorry, I have to say it: "It's a service, stupid!"
A well-written article that is a pleasure to read, in our slop driven day and age.
I have been skeptic of the no extra sensors approach Tesla took since they announced it . It is obvious that extra sensors is what can make self-driven cars outperform human drivers, not merely match them.
I would challenge two parts of Musk's argument: that a computer camera system can cost effectively emulate human driving and vision performance and the idea that humans are safe drivers with only eyes.
Certainly, ears and touch (both steering wheel and whatever is on the seat) play an important role gathering information.
I'm not sure drivers in Bangaloru even use their eyes--i believe they mostly drive by echolocation
I’m interested in comparing the ratio of (accidents / miles driven) between the two.
Tesla seems like it has an engineering culture built around shipping whatever version of product they have at the date Elon announces, regardless of the quality (based on the engineers I have met who have worked there).
To me, self driving seems like the opposite of Moore's law.
In the last decades, one of the principle of SWE was to take into account how much computing power will have improved when planning multi-year project. Meaning, you could write something that was too heavy for today's machines, but would be bleeding edge in 5 years.
IMHO, self-driving is actually the inverse situation.
Piloting a car in an human-centered environment is difficult and requires the machine to behave humanly. This of course requires an absurd amount od data and training to pull off. But what happens when self driving adoption increases? At a certain point, so many driverless cars will be roaming the streets that most daily interactions will be automated by letting the cars negotiate in a nice, deterministic, algorithmic way. Thus, reliance on predictive and opaque systems like neural networks will be needed less and less, actually reducing the complexity of self driving.
My main point is: what should win is the tech that makes cars drive as well as trained professional drivers. Once that's done, adoption will drive down human driving and thus unpredictable behavior on the road, reducing the computational load needed to correctly perform tasks. Next, cars will start to behave more programmatically and deterministic and will need less sensors and tech. Car companies will have accurate maps of everything, and cars will mostly become shuttles which can rely more on predetermined routines and less on world models, especially as smart cities gain a foodhold too.
There might still be significant pushback at least for one more generation. Although we can see things move quickly now that the war on general purpose computation is moving quickly(the recent 3D printer ban proposals, introducing age verification at the OS level etc.) so many things might move slowly for a long time and then move fast all at once.
>Car companies will have accurate maps of everything, and cars will mostly become shuttles which can rely more on predetermined routines and less on world models, especially as smart cities gain a foothold too.
Car companies won't have squat. The whole point of GM buying Cruise and others trying to get into self driving was that they will be relegated to white box manufacturers if they dont try and bring this tech in house. Its funny how the MBAs at these companies tried to outsource all manufacturing to 'suppliers' such that all they really wanted to do was stick the badge on the car at the end. Now they realize this thinking is going to take themselves out of the equation as well....whoops. If your vision comes to pass why would anyone care what badge is on the front of the car?
This is probably why Waymo had to use Jaguar i-Paces: only companies desperate to offload their unsold inventory would cooperate with them.
Isn’t there always the fact that non-car things happen near and on a road, thus forever requiring high amounts of compute?
Depends on what non-car things you are talking about. If we're talking about an obstacle on the road or a pedestrian, a model to take care of that is almost trivial, since the car would simply signal everything around it that there is a problem and it's performing a manouver, and it will not have to worry about anything other than the obstacle and the road boundaries.
The idea that hardware costs matter most is idiotic^H^H^H^H^H^H^H silly. If you are buying a car, it matters, somewhat. If you are riding in a taxi or leasing a car, you are paying for a service. When you go to a restaurant, is their profitability predicated on how much they payed for their ovens?
LIDAR costs $1000, whoops $100, whoops $10, whoops it's on a chip module that costs a tenth of a cent. Building your business model on hardware NOT getting cheaper seems silly. Sure, software is cheap and getting cheaper, but
Sorry, I have to say it: "It's a service, stupid!"
A well-written article that is a pleasure to read, in our slop driven day and age.
I have been skeptic of the no extra sensors approach Tesla took since they announced it . It is obvious that extra sensors is what can make self-driven cars outperform human drivers, not merely match them.