The Human Advantage
I hear a lot of people talk about the things that people are better at than computers. Like processing what we humans like to call “natural language.” Or driving, which is what I want to get into in this post.
Let’s pause for a moment and unpack the above: to whom is natural language natural? To humans, of course. And who set the rules for driving, and even more important than the rules, the norms? Humans.
To look at the performance of AI today as compared to humans and therefore think that it is hopelessly behind is to overlook the tremendous advantage humans have in human-system-centric tasks. Pretty much everything humans build, including roads — road widths, road curvatures, road markings, road signs, speed limits, traffic calming implements, lane widths, shoulders, stoplights, intersections, parking spaces, etc. — are designed by humans, for humans. Even when humans aren’t explicitly mentioned in the design, their dimensions, capabilities, tolerances, and constraints are implicitly taken into account. For example, when urbanists lament city streets designed for cars rather than people, what they mean is street designed for people in cars rather than people walking (for the record, I am one of these urbanists). For another example, it may seem that parking spaces are designed for the size of a car, but they are actually designed for the capability of a human driver who is parking that car: the length of a parallel parking space mat the amount of space a human needs to park their car; cars with automated parking can parallel park the same vehicle in far less space.
Not all humans have the same capabilities and constraints, or the same cultural references or intelligence to understand the same road signs. This can be one cause of the anxiety many feel about driving outside of their home country or region. Not all road signs are well designed for even the local population to understand. But, by and large, in any given region, all the drivers will share enough language and cultural reference to at a minimum muddle through without getting themselves or others killed, even if they have no special local training. Even a child who is old enough to ride in the front seat with his parent knows that red, or a forward facing hand means stop; a blinking light on one side of the rear of the car in front means that car is going to turn to the side of the light; drive on the right side of the road (in most countries); stay between the lines. These understandings are partially the result of asking the driver what these things mean, but they are also the result of the aggregate knowledge the child has accumulated over their hundreds of thousands of hours of life. They have played with toy cars and trucks, or see other do so. They have seen cars and roads in movies. They know what lines are and that you are generally supposed to stay within them. They understand that the turn signals are on two different sides of a car, that those sides are called left and right, that a blinking light is designed to get your attention, and therefore it’s easy for them to understand that they mean. They know the meaning of a stop sign and of the light turning red, green, and yellow long before they start driving, because even as a pedestrian this knowledge isi important to know.
They know what it means to make eye contact with a driver while you are crossing the street, or to make eye contact with a pedestrian when you are the driver. How to negotiate with another person over the same parking space, or to merge into busy highway traffic using nothing but eye contact and hand gestures.
Now imagine the poor AI of an autonomous vehicle. It hasn’t had hundreds of thousands of hours negotiating and absorbing social norms with other people. Not only that, but all of its sensory capture is from specialized vehicle sensors: cameras, Lidars, radar, IMUs, etc., none of which cover that human element. It’s not even clear how you would collect data on this human element for training an autonomous vehicle’s AI. Most likely it would have to be in a staged, closed setting. The alternatives would require having enough in-vehicle data capture tool distribution — whether app or dedicated hardware — for two tools to be on opposite sides of the same interaction, or have a data capture tool that from a single car is capable of recording the driver of the other car as well as the driver of the car the tool is in. Both of these are significant challenges, though potentially lucrative business opportunities for whomever first figures them out. The latter is probably the most promising option, as it reflects the perspective of the human driver the AI is functionally replacing, but it can probably only be done with well-calibrated smart glasses or even smart contacts and software that can track the driver’s gaze in the real world and at high speed.
Perhaps paradoxically, as a result of the above, I think that the current performance of driverless cars underplays its potential. While AIs are still far from being able to understand and interact with humans in areas requiring informal social skills, there are people attacking this problem from different directions, they are improving every day, and the moment that they do achieve this, we will realize many of the other barriers and challenges were being propped up by this one and are going to fall quickly.
There are some questions about how we would test such an autonomous system, because it is clear that our current procedures for testing individual drivers are hopelessly ineffective when it comes to testing machines. But that is a question for another blog post.
