Rich and Strange Aeons

When self-driving cars first started getting talked about, like 10 years ago, many people were enthusiastic and anticipatory. I was skeptical, because as someone who walks around dense cities, driving safely and effectively in such felt like a human-complete AI problem, needing theory of mind, social interaction, and a large amount of adaptation to unforeseen circumstances.

Also because while in some things like chess or Go, rather dumb computers beat humans through powerful search, a more common AI pattern is that a fairly simple system can get 60-90% of human performance, but then stalls despite a lot of effort. Which is fine when you're making models for targeting direct mailing, and poorer performance can be balanced by much faster turnaround time and it's just moderate amounts of money at stake anyway. Less fine when even a missing 1% of performance may mean people die, or alternatively that traffic is frozen as cars can't figure out how to safely push through busy streets.

(The direct mailing example is from my first full-time job; we could build a decision tree, to predict response rates to a direct mailing, that was said to be 60% of a hand-crafted model but took a few hours instead of a few months to create. A machine translation course in grad school included various systems that could do 60-95% as well as humans, on fairly narrow word tests, but improving that was Hard. Statistical translation, rule-based, hybrid, all stalled.)

Basically an application of the Pareto principle: 20% of the work can get you 80% of the performance. Except it might be more like 1% of the work gets you 80% of the performance; since we don't *have* human-equivalent AI in most of these domains, we can't even say how much work it actually takes.

Early articles were along the lines of "we're making lots of progress! (but can't drive in the rain or snow and are tested mostly in low-density sunlight)", which for some people sounded like "we're almost there but for a bit more work" but to me sounded like "we're already spending years on the *easy* stuff, imagine what the hard stuff will be like."

More recent articles have been more like "wow, this is harder than we thought", with even the executives in charge of developing and selling this stuff saying like "thirty years away" or "never" or "far in the future", or "decades away".

Singapore reportedly has deployed them, as someone on Facebook likes to keep saying, but a friend there observed various caveats: 10 MPH, a bounded area, not mixed with other cars, safety driver, and attendants trying to shoo pedestrians out of the way. Also see. And this is the state of the art!

So, "ha ha!"

I'll also include a FB thread I made two years ago about predictions, and include just one example of receding predictions:

2014: Volvo promises fully self-driving cars by 2017, 3 years later.
2017: Volvo promises partial self-driving cars by 2021, 4 years later.