The Long Pause

Carmakers are leaping over each other to roll out cars that meet SAE Level 3 requirements, whereby under some conditions drivers can let go of the steering wheel. Getting to Level 5 will take a lot longer, and there is some debate about where and even whether Level 4 will ever happen (see Fig. 1).

Fig. 1: Levels of autonomy. Source: Auto Alliance

There are two big gaps that need to be addressed here. First, AI systems are nowhere close to being able to handle a nearly infinite number of possible interactions and road conditions that might crop up. AI is very good at recognizing a pothole in the road, and it can identify the position of traffic lights. It cannot handle unexpected driving situations as well or adapt as quickly as an alert human. (Editor’s Note: Tesla just modified its claims about Level 5 availability.)

AI is not an intuitive system like the human brain. It doesn’t actually think. It responds in mostly predictable ways, and within certain parameters, to conditions that it has been programmed to recognize. For example, AI systems still can’t differentiate a male chick from a female chick, no matter how many photographs are fed into its memory from every different angle. That’s a hard problem for computers. It takes a worker in a poultry plant a couple weeks to figure that out. It may be years before machines figure that out—if they ever do.

No one has figured out yet how to program intuition into machines. Lots of very smart people are working on this problem all over the world. But it’s unlikely to happen fast enough to roll out a fully autonomous vehicle capable of dealing with unexpected hazards involving people, animals, floods, volcanic eruptions, or unexpected events that programmers fail to plan for.

The second big gap involves bringing this future technology down to an affordable cost. A tree branch breaking off in a storm or a car swerving on black ice are unpredictable. And while those kinds of behaviors can be programmed into a car’s logic (assuming, of course, that anyone has placed a camera on the roof of a car facing up), they require instantaneous response. But taking all of these possible scenarios into account requires massive amounts of processing power, memory, bandwidth, sophisticated pre-processing, and real-time prioritization of signals to figure out how to save the occupants of the car without injuring other people around the car.

While it’s not clear how much this ultimately will cost, it’s almost certainly not going to fit into the financial budget of the average car buyer. It also won’t fit into the power budget of an autonomous vehicle. It may not even fit into the power budget of certain regions of the world, where electricity grids already are running at maximum capacity and proposals for new data centers are routinely rejected. This is certainly true in developing countries, but it’s also true for the Northeast corridor of the United States, which has been power-constrained ever since decommissioning multiple nuclear plants.

While very smart people, backed by huge investments in R&D, are working on all of these issues, none of this is going to be fixed overnight. Drivers will continue to be distracted by smart phones for the foreseeable future, perhaps even more than in the past if they are driving a Level 2 or Level 3 vehicle and forced to respond to unexpected interruptions. And people will still make mistakes or take risks that cause accidents.

That may leave people clamoring for the next couple levels of driving technology. But it’s not obvious that any of this will happen quickly. In fact, all indications are that we are headed for a long pause with lots of engineering work before the next phase of technology is usable, affordable, and reliable enough.