What is the singularity and just how close is it?
The short answers are “it depends who you ask” and “nobody knows.” The longer answers are, well…you’ll see.
Wikipedia provides a good basic definition: “The technological singularity—or simply the singularity—is a hypothetical point in time at which technological growth will become radically faster and uncontrollable, resulting in unforeseeable changes to human civilization.”
The technological growth in question usually refers to artificial intelligence (AI). The idea is that an AI capable of improving itself quickly goes through a series of cycles in which it gets smarter and smarter at exponential rates. This leads to a super intelligence that throws the world into an impossible-to-predict future.
Whether this sounds awesome or awful largely depends on your view of what a superintelligence would bring about, something that no one really knows.
The impossible-to-predict nature is an aspect of why, in fact, it’s called a singularity, a term that originates with mathematics and physics. In math, singularities pop up when the numbers stop making sense, as when the answer to an equation turns out to be infinity. It’s also associated with phenomena such as black holes where our understandings of traditional physics break down. So the term, as applied to technology, suggests a time beyond which the world stops making sense (to us) and so becomes impossible to forecast.
How Many Flavors Does the Singularity Come In?
Is a runaway recursively intelligent AI the only path to a singularity? Not if you count runaway recursively intelligent people who hook their little monkey brains up to some huge honking artificial neocortices in the cloud.
Indeed, it’s the human/AI interface and integration scenario that folks like inventor-author-futurist Ray Kurzweil seem to be banking on. To him, from what I understand (I haven’t read his newest book), that’s when the true tech singularity kicks in. At that point, humans essentially become supersmart, immortal(ish) cyborg gods.
But there are other possible versions as well. There’s the one where we hook up our little monkey brains into one huge, networked brain to become the King Kong of super intelligences. Or the one where we grow a supersized neocortex in an underground vat the size of the Chesapeake Bay. (A Robot Chicken nightmare made more imaginable by the recent news we just got a cluster of braincells to play pong in a lab–no, really).
Singularity: Inane or Inevitable?
The first thing to say is that maybe the notion is kooky and misguided, the pipedream of geeks yearning to become cosmic comic book characters. (In fact, the singularity is sometimes called, with varying degrees sarcasm, the Rapture for nerds.)
I’m tempted to join in the ridicule of the preposterous idea. Except for one thing: AI and other tech keeps proving the naysayers wrong. AI will never beat the best chess players. Wrong. Okay, but it can’t dominate something as fuzzy as Jeopardy. Wrong. Surely it can’t master the most complex and challenging of all human games, Go. Yawn, wrong again.
After a while, anyone who bets against AI starts looking like a chump.
Well, games are for kids anyway. AI can’t do something as slippery as translate languages or as profound as unravel the many mysteries of protein folding. Well, actually…
But it can’t be artistic…can it? (“I don’t do drugs. I am drugs” quips DALL-E).
Getting Turing Testy
There’s at least one pursuit that AI has yet to master: the gentle art of conversation. That may be the truest assessment of human level intelligence. At least, that’s the premise underlying the Turing test.
The test assumes you have a questioner reading a computer screen (or the equivalent). The questioner has two conversations via screen and keyboard. One of those conversations is with a computer, the other with another person. If questioner having the two conversations can’t figure out which one is the computer, then the computer passes the test because it can’t be distinguished from the human being.
Of course, this leaves us with four (at least!) big questions.
First, when will a machine finally pass that final exam?
Second, what does it mean if and when a machine does? Is it truly intelligent? How about conscious?
Third, if the answer to those questions seems to be yes, what’s next? Does it get driver’s license? A FOX News slot? An OKCupid account?
Fourth, will such a computer spark the (dun dun dun) singularity?
The Iffy Question of When the Singularity Arrives
In a recent podcast interview, Kurzweil predicted that some soon-to-be-famous digital mind will pass the Turing Test in 2029.
“2029?” I thought. “As in just 7-and-soon-to-be-6-years-away 2029?”
Kurzweil claims he’s been predicting that same year for a long time, so perhaps I read about it back in 2005 when his book The Singularity Is Near (lost somewhere in the hustle bustle of my bookshelves). But back then, of course, it was a quarter of a decade away. Now, well, it seems damn near imminent.
Of course, Kurzweil may well turn out to be wrong. As much as he loves to base his predictions on the mathematics of exponentials, he can get specific dates wrong. For example, as I wrote in a previous post, he’ll wind up being wrong about the year solar power becomes pervasive (though he may well turn out to be right about the overall trend).
So maybe a computer won’t pass a full blown Turing test in 2029. Perhaps it’ll be in the 2030s or 2040s. That would be close enough, in my book. Indeed, most experts believe it’s just a matter of time. One survey issued at the Joint Multi-Conference on Human-Level Artificial Intelligence found that just 2% of participants predicted that an artificial general intelligence (or AGI, meaning that the machine thinks at least as well as a human being) would never occur. Of course, that’s not exactly an unbiased survey cohort, is it?
Anyhow, let’s say the predicted timeframe when the Turing test is passed is generally correct. Why doesn’t Kurzweil set the date of the singularity on the date that the Turing test is passed (or the date that a human-level AI first emerges)? After all, at that point, the AI celeb could potentially code itself so it can quickly become smarter and smarter, as per the traditional singularity scenario.
But nope. Kurzweil is setting his sights on 2045, when we fully become the supercyborgs previously described.
What Could Go Wrong?
So, Armageddon or Rapture? Take your pick.
What’s interesting to my own little super-duper-unsuper brain is that folks seem more concerned about computers leaving us in the intellectual dust than us becoming ultra-brains ourselves. I mean, sure, our digital super-brain friends may decide to cancel humanity for reals. But they probably won’t carry around the baggage of our primeval, reptilian and selfish fear-fuck-kill-hate brains–or, what Jeff Hawkins calls our “old brain.”
In his book A Thousand Brains, Hawkins writes about about the ongoing frenemy-ish relationship between our more rational “new brain” (the neocortex) and the far more selfishly emotional though conveniently compacted “old brain” (just 30% of our overall brain).
Basically, he chalks up the risk of human extinction (via nuclear war, for example) to old-brain-driven crappola empowered by tech built via the smart-pantsy new brain. For example, envision a pridefully pissed off Putin nuking the world with amazing missiles built by egghead engineers. And all because he’s as compelled by his “old brain” as a tantrum-throwing three-year-old after a puppy eats his cookie.
Now envision a world packed with superintelligent primate gods still (partly) ruled by their toddler old-brain instincts. Yeah, sounds a tad dangerous to me, too.
The Chances of No Chance
Speaking of Hawkins, he doesn’t buy the whole singularity scene. First, he argues that we’re not as close to creating truly intelligent machines as some believe. Today’s most impressive AIs tend to rely on deep learning, and Hawkins believes this is not the right path to true AGI. He writes,
Deep learning networks work well, but not because they solved the knowledge representation problem. They work well because they avoided it completely, relying on statistics and lots of data instead….they don’t possess knowledge and, therefore, are not on the path to having the ability of a five-year-old child.
Second, even when we finally build AGIs (and he thinks we certainly will if he has anything to say about it), they won’t be driven by the same old-brain compulsions as we are. They’ll be more rational because their architecture will be based on the human neocortex. Therefore, they won’t have the same drive to dominate and control because they will not have our nutball-but-gene-spreading monkey-brain impulses.
Third, Hawkins doesn’t believe that an exponential increase in intelligence will suddenly allow such AGIs to dominate. He believes a true AGI will be characterized by a mind made up of “thousands of small models of the world, where each model uses reference frames to store knowledge and create behaviors.” (That makes more sense if you read his book, A Thousand Brains: A New Theory of Intelligence). He goes on:
Adding this ingredient [meaning the thousands of reference frames] to machines does not impart any immediate capabilities. It only provides a substrate for learning, endowing machines with the ability to learn a model of the world and thus acquire knowledge and skills. On a kitchen stovetop you can turn a knob to up the heat. There isn’t an equivalent knob to “up the knowledge” of a machine.
An AGI won’t become a superintelligence just by virtue of writing better and better code for itself in the span of a few hours. It can’t automatically think itself into a superpower. It still needs to learn via experiments and experience, which takes time and the cooperation of human scientists.
Fourth, Hawkins thinks it will be difficult if not impossible to connect the human neocortex to mighty computing machines in the way that Kurzweil and others envision. Even if we can do it someday, that day is probably a long way off.
So, no, the singularity is not near, he seems to be arguing. But a true AGI may, in fact, become a reality sometime in the next decade or so–if engineers will only build an AGI based on his theory of intelligence.
So, What’s Really Gonna Happen?
Nobody know who’s right or wrong at this stage. Maybe Kurweil, maybe Hawkins, maybe neither or some combination of both. Here’s my own best guess for now.
Via deep learning approaches, computer engineers are going to get closer and closer to a computer capable of passing the Turning test, but by 2029 it won’t be able to fool an educated interrogator who is well versed in AI.
Or, if a deep-learning-based machine does pass the Turing test before the end of this decade, many people will argue that it only displays a façade of intelligence, perhaps citing the famous Chinese-room argument (which is a philosophical can of worms that I won’t get into here).
That said, eventually we will get to a Turing-test-passing machine that convinces even most of the doubters that it’s truly intelligent (and perhaps even conscious, an even higher hurdle to clear). That machine’s design will probably hew more closely to the dynamics of the human brain than do the (still quite impressive) neural networks of today.
Will this lead to a singularity? Well, maybe, though I’m convinced enough by the arguments of Hawkins to believe that it won’t literally happen overnight.
How about the super-cyborg-head-in-the-cloud-computer kind of singularity? Well, maybe that’ll happen someday, though it’s currently hard to see how we’re going to work out a seamless, high-bandwidth brain/supercomputer interface anytime soon. It’s going to take time to get it right, if we ever do. I guess figuring all those details out will be the first homework we assign to our AGI friends. That is, hopefully friends.
But here’s the thing. If we ever do figure out the interface, it seems possible that we’ll be “storing” a whole lot of our artificial neocortex reference frames (let’s call them ANREFs) in the cloud. If that’s true, then we may be able to swap ANREFs with our friends and neighbors, which might mean we can quickly share skills I-know-Kung-Fu style. (Cool, right?)
It’s also possible that the reticulum of all those acquired ANREFs will outlive our mortal bodies (assuming they stay mortal), providing a kind of immortality to a significant hunk of our expanded brains. Spooky, yeah? Who owns our ANREFs once the original brain is gone? Now that would be the IP battle of all IP battles!
See how weird things can quickly get once you start to think through singularity stuff? It’s kind of addictive, like eating future-flavored pistachios.
Anyway, here’s one prediction I’m pretty certain of: it’s gonna be a frigging mess!
Humanity will not be done with its species-defining conflicts, intrigues, and massively stupid escapades as it moves toward superintelligence. Maybe getting smarter–or just having smarter machines–will ultimately make us wiser, but there’s going to be plenty of heartache, cruelty, bigotry, and turmoil as we work out those singularity kinks.
I probably won’t live to the weirdest stuff, but that’s okay. It’s fun just to think about, and, for better and for worse, we already live in interesting times.
Addendum: Since I wrote this original piece, things have been moving so quickly In the world of AI that I revisited the topic in The Singularity Just Got Nearer…Again…And How!
Featured image by Adindva1: Demonstration of the technology "Brain-Computer Interface." Management of the plastic arm with the help of thought. The frame is made on the set of the film "Brain: The Second Universe."