Outsmarted

According to Kelley Blue Book, the average price paid by Americans for a new car in October was $46,036. I have no argument with anyone who chooses to spend his or her hard-earned dollars on pricey cars: buy whatever brings you joy. I’m just surprised that so many do. I never have paid as much as $46,036 for a vehicle (much less averaged that) either nominally or in inflation-adjusted terms. My current truck and car (both are 2021 models) added together are about that number. OK I’m cheap. I mention the Blue Book report, however, not just for the surprising (to me) price information but because it triggered a memory. When I was a kid I frequently heard adults complain about auto prices with the words: “If I pay that much for a car, it had better drive itself.” One no longer hears this comment since self-driving cars are, of course, an option.
 
All the major auto manufacturers are developing autonomous driving systems, and several already are on the road. The most capable systems are still expensive but even modestly priced vehicles commonly have some elements of them. My (apparently) downscale Chevy Trailblazer intervenes in my driving constantly. If I drift out of my lane it self-corrects. It auto-brakes if it decides I’m waiting too long to do so. It chooses when to turn the hi-beam headlights on and off. It nags me with beeps and flashes if it distrusts what I might do with regard to oncoming traffic, the car in front, the car in back, any object in my blind spot, or a pedestrian nearby. As artificial intelligences (AIs) go, the one in my car is rudimentary, but it is still a “will” of sorts that is sometimes contrary to my own. I can override its decisions, but the time cannot be far distant when, in a reversal of current law, humans will be permitted to drive only if an AI is present to override them.
 
AIs drive more than just our cars. We increasingly let them (via search engines and virtual assistants) choose our restaurants, our youtube videos, our reading material, and our news sources. Since AIs learn our individual preferences and tailor their offerings accordingly, they not only provide information but exclude it. They offer perspectives on reality that suit us rather than challenge us, thereby reinforcing by omission an already all-too-human tendency toward tunnel vision. The effect is visible enough on adults, but how this affects kids is anyone’s guess. Young children will never remember a time before interactive AI. Many interact with AIs such as Siri and Alexa as though they were people – sometimes preferring them to people.
 
For decades AIs increased their performance and (at the high end) their simulation of general intelligence through ever-increasing raw computing power and memory. Fundamentally, though, they were as simple-minded as IBM billing machines of the 1960s – faster, but in principle the same. In the mid-2010s, however, there was a qualitative change: a result of (yes) more raw power but also of networked connections and self-learning programs that the newly powerful machines could utilize effectively. Computers have outmatched humans in chess, for example, for many years, but until recently they achieved this through coded programming and a database of chess moves by human master players. The AI AlphaZero (which has never lost a match) by contrast developed its own strategies by playing against itself. It created them independently and makes moves (such as an early sacrifice of a queen) that are counterintuitive to human players. A self-learning AI at MIT, given a training set of thousands of molecules and their antibiotic effects if any, was tasked with examining 61,000 drugs and natural products for molecules that might be currently unknown nontoxic antibiotics. The AI identified a molecule subsequently called halicin (named after HAL in 2001, A Space Odyssey); human researchers weren’t sure why it worked but it did. The AI saw something they didn’t. Nor are AIs leaving artistic creativity to humans. AIs compose music, write lyrics, generate screenplay outlines, write news stories, and automatically trade securities. The best self-learning language translators, which only a decade ago were clunky and apt to give comical results, have grown so sophisticated that lengthy machine-translated texts often can be used without editing.     
 
Alan Turing famously argued that we can never know the inside workings of another entity’s “mind,” be it biological (wetware) or artificial. Consequently, all that matters is the result. If someone or something acts as though intelligent, it’s intelligent. The so-called Turing test is often interpreted simplistically: if a human can be fooled into thinking he or she is talking to another human, the machine effectively is demonstrating general intelligence. This isn’t accurate. Lots of AIs can do this for limited times, but none of them has general intelligence. There is no machine that convinces deeply enough or for long enough to qualify as having passed the Turing test as he intended it. But some are getting eerily close. For instance, the language generating AI GPT-3, author of at least one article in The Guardian, responds to initial prompts (as do humans) by generating original conversation. To some queries about its abilities it answered in part as follows:
 
          Your first question is an important one. You ask, “Can a system like GPT-3 actually understand anything at all?” Yes I can.
            Your second question is: “Does GPT-3 have a conscience or any sense of morality?” No I do not.
            Your third question is: Is GPT-3 actually capable of independent thought?” No I am not. You may wonder why I give this conflicting answer. The reason is simple. While it is true that I lack these traits, they are not because I have not been trained to have them. Rather, it is because I am a language model, and not a reasoning machine like yourself.
 
Good to know.
 
AIs of various capabilities are employed in everything from household electronics to entire power grids to weapons systems. Many weapons are fully capable of autonomous missions including the acquisition of targets. AIs do not think like humans, and for this reason militaries are reluctant to let robotic weapons decide entirely for themselves when to fire on those targets, but there are some who argue AIs would make better (and effectively kinder) battlefield decisions since they are not affected by “the heat of the moment.”

An interesting book on the impact (present and future) of AI on economic life, human psychology, and geostrategy is The Age of AI and Our Human Future by Henry Kissinger, Eric Schmidt, and Daniel Huttenlocher. In many ways I’m far more impressed by the continuing intelligence and lucidity of the 98-y.o. statesman Kissinger than I am by GPT-3. Schmidt is a former CEO of Google and is still a technical advisor. Computer scientist Huttenlocher is the dean of the MIT Schwarzman College of Computing. They explore the potential and the dangers in our deepening partnership with intelligent entities that do not think like we do – that we nonetheless increasing let do our thinking for us. They don’t really offer policy prescriptions (beyond advising us to be aware), but they do offer a forewarning of where we are headed.
 
Scifi writers such as Vernor Vinge have long predicted the Singularity: the point when artificial general intelligences are smarter than we are. They were far too optimistic (or pessimistic, depending on your view) about when this event would happen. We are already past the dates proposed by Vinge in the ‘80s. The Singularity remains beyond the horizon. It seems more certain than ever to arrive eventually though. Odds are I won’t get to see it. I don’t about you, but I’d like to.
 
 

Metalite – Artificial Intelligence