The Path to Autonomy
Driverless cars will not arrive all at once. They will arrive one boring intervention at a time, until nobody remembers the last one.
Every few months someone asks me when the cars will finally drive themselves. The honest answer is that they already do, most of the time, and that “most of the time” is the whole problem. A system that handles ninety-nine percent of drives is a product. A system that handles all of them is a utility. The distance between the two is not a feature. It is a long, unglamorous grind of edge cases.
What the last mile actually looks like
The public imagines the end of the road as a single announcement. In practice it looks like a graph of interventions per mile that keeps bending toward zero, and a set of arguments about what counts as an intervention.
- A tap on the accelerator at a stale red light.
- A nudge around a delivery truck that stopped in the lane.
- A takeover in a construction zone drawn by someone with a grudge against geometry.
None of these are crashes. All of them are the reason a safety driver is still in the seat. Removing the driver means the system has to be right about the weird stuff, not just the common stuff, and the weird stuff has a very long tail.
The hard part of self-driving is not driving. It is proving, to yourself and then to everyone else, that you no longer need to watch.
Why data changes the shape of the curve
The reason I stay optimistic is that the tail is not infinite. Every car on the road is a sensor for the ones that come after it. A rare situation seen by a million vehicles is not rare anymore; it is a training example with a thousand variations. That is a different regime from the one where a handful of test cars drove the same loops in one city.
Scale does not make the problem easy. It makes it finite.
What to watch instead of the headlines
If you want to know where this is going, ignore the demos and watch three numbers:
- Miles between critical interventions, on ordinary routes chosen by ordinary people.
- The size of the geofence, and how often it grows.
- How long a new city takes to open compared with the last one.
When the third number stops shrinking, the software is done and the rest is politics. My guess is that we are closer to that moment than the discourse suggests, and that it will feel anticlimactic when it comes. The cars will simply keep driving, and one day you will notice that nobody has needed to take over in a very long time.