I doubt we’ll see Level 5 self-driving cars anytime soon. Level 5 is considered a level of sophistication where the car is as good as a human driver under all situations. Such a car needs no driver’s seat, no steering wheel. Fully autonomous.
Why won’t this happen soon?
Because this is an AI-complete problem, i.e. if we solve this, we solve all other outstanding problems in artificial intelligence. In some ways, it is almost axiomatic because we want the car to take care of situations, known and unknown, like a human would and that requires us to understand how human intelligence works.
But even so, why can’t this happen soon?
This is because our efforts in AI are a bit misplaced. Almost all the research is focused on data and statistics. We’re fitting curves to data, though with more and more sophistication every year, and extrapolating the models to get answers. Everything is quite empirical.
And why’s that not good enough?
Because empirical is how we start a scientific journey into an unknown territory. And we need to see the end of this journey because self-driving seems to be an AI-complete problem.
To solve this problem, we need to step outside of the empirical realm and focus on the principles. Just to be clear, empirical and data based engineering shouldn’t be taken lightly — for example, thermodynamics wasn’t well understood even after some fairly sophisticated steam engines had been built.
But on the contrary, the GPS satellite system couldn’t be built without first understanding the underlying physics of gravity-induced time dilation or the positions would have been too inaccurate to be of much value. As an analogy to GPS, it’s as if we don’t understand relativity, but we still want to build the GPS network. Nothing wrong with this except if there’s little effort in understanding the underlying cause of the discrepancies.
So how long before we can buy a true self-driving car?
From a principles and theory point of view, my belief is that this will be solved in the next 25 years and not sooner.
However, bringing these principles to reality may never happen — just like weather prediction may never become 100% accurate.
Our current computational systems are too good at things where humans don’t even play — like high precision arithmetic or repeated calculations. Clearly something’s not in sync.
We might need a new type of computer which is perhaps less precise, where parallelism is an inherent part of the architecture and requires less energy. It needs to be less like a calculator and more human-like.
Give it 50+ years.
