Why FSD Is Harder Than We Think

It's now been over a year since Elon Musk started saying Full Self Driving (FSD) is coming "next month." If you are wondering what's taking so long maybe check out this recent paper "Why AI is Harder Than We Think" by Melanie Mitchell. She is the Davis Professor of Complexity at the Santa Fe Institute, and Professor of Computer Science (currently on leave) at Portland State University. Her research focuses on conceptual abstraction, analogy-making, and visual recognition in artificial intelligence systems. Given that Tesla FSD relies on visual AI it would seem that she may be qualified to explain the delay in FSD rollout. Her paper does not mention Tesla or FSD at all, but I think her observations apply to the current status of FSD.

Mitchell points to four classic fallacies in the predictions made by AI developers:

1. Narrow intelligence is on a continuum with general intelligence

It's easy to assume that if you make some incremental progress in an AI problem that it's just a matter of time before you solve the whole thing. I.e. just a few more months to FSD. But Mitchell says that's like claiming that the first monkey that climbed a tree was making progress towards landing on the moon. Ain't gonna happen. Plus there's this unexpected obstacle in the assumed continuum of AI progress. "The problem of common sense," she says, which humans have subconsciously but AI systems lack completely. Nobody knows how to code for common sense, which comes in handy when you're driving a car.

2. Easy things are easy and hard things are hard

In fact easy things for us are hard for computers. She quotes Hans Moravec the computer scientist who came up with one of the first algorithms for computer vision. He once wrote, "It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, yet difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility." Unfortunately FSD is all about perception and mobility. We simply don't appreciate the complexity of our own thought processes and we overestimate how easy it is to give these abilities to a computer. Mitchell says Moravec put it this way: "Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it." FSD doesn't have any of this. Mitchell then quotes the grandfather of AI Marvin Minsky who said, "In general, we're least aware of what our minds do best."

3. The lure of wishful mnemonics or metaphors

The computational technique underlying FSD is a neural network, a metaphor loosely inspired by the brain, but with major differences. Mitchell says, "Machine learning or deep learning methods do not really resemble learning in humans (or in non-human animals). Indeed, if a machine has learned something in the human sense of learn, we would expect that it would be able use what it has learned in different contexts. However, it turns out that this is often not the case." Computer scientist Drew McDermott first used the term "wishful mnemonics" in 1976, pointing out that by labeling some computer code "Full Self Driving," for instance, we are imbuing it with the "wish" that it will actually do what it says. He said a better idea would be to label it "G0034" and then see if the programmers can convince themselves or anyone else that G0034 implements some part of self driving. We all seem to be caught in a sort of wishful FSD state at the moment. But wishing it won't make it happen.

4. Intelligence is all in the brain

This fallacy assumes intelligence is disembodied and lives only in the brain, the so called "information processing model of mind." It's the old idea that if you had enough computing power you could "upload" a mind into a machine. But a growing number of cognitive scientists now believe in a sort of "embodied cognition." Mitchell says, "Nothing in our knowledge of psychology or neuroscience supports the possibility that "pure rationality" is separable from the emotions and cultural biases that shape our cognition and our objectives. Instead, what we've learned from research in embodied cognition is that human intelligence seems to be a strongly integrated system with closely interconnected attributes, including emotions, desires, a strong sense of selfhood and autonomy, and a common sense understanding of the world. It's not at all clear that these attributes can be separated." While one could argue that Tesla's FSD also embodies the car via sensors and cameras it may take more than a few weeks before Tesla's programmers pull off this trick that took Nature a billion years to integrate.

Ultimately Mitchell uses these four fallacies to explain the cyclic nature of AI research since its inception in the 1970s. It tends to blossom in Spring time with magnificent overconfident predictions at first, but then when the scale of the challenge is realized, a sort of AI Winter descends and progress can stall for up to a decade.

Obviously we all want Tesla to be successful and pull off Full Self Driving next week, next month, or even next year. And Elon Musk has pulled a rabbit out of a hat more than once. Who can forget when those two returning Falcon Heavy booster rockets landed perfectly and simultaneously in 2018? So fingers crossed. But if you're ever craving an explanation for why FSD is taking so long devote 20 minutes to Melanie Mitchell's paper.

Barry Shell is a freelance writer in Vancouver, Canada. He created www.science.ca, the top Google hit for any search on Canadian science. He has written four books, and has published in magazines and newspapers including the Globe and Mail and the New York Times. Originally from Winnipeg, Barry has a BSc in Organic Chemistry from Reed College in Portland, OR and an MSc in Resource Management Science from UBC. His book, "Sensational Scientists" profiling 24 of Canada's greatest scientists and published by Raincoast Books, won a national book award in 2005. If you enjoyed this article please consider using Barry’s Tesla referral code to get 1k miles of Supercharging for free: https://ts.la/barry73962