Comment by Almondsetat
15 hours ago
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.