Michael Anissimov/Singularity Intro
Some random Singularity intro I was writing. Not sure where I was going with this. Meant to be casual, but not overly casual. Originally intended for a Live Journal, but I want to improve it beyond that standard, if possible. I've started using the phrase "Technological Singularity" all the time instead of just "Singularity" because if Singularitarians don't use the prior phrase, we'll end up being stomped on Google. Do any passages make you cringe? Awkward sentences? Straight-up poor phrasing? Comment immediately: Michael Anissimov/Singularity Intro Comments
You may have heard that I talk about the notion of a Technological Singularity a lot. What is this "Technological Singularity" I speak of?
Answer: the technological creation of significantly smarter-than-human intelligence.
Is such a thing really possible? To those who see the human brain as an actual physical object, not some sort of passive vessel containing mystical, nonphysical intelligence-dust all humans have by virtue of our specialness, the answer is obvious. Inference, memory, pattern-recognition, and so on, are information-processing routines, rooted in neurological machinery, just like our ability to run a mile is rooted in our muscles.
Muscles could be improved in theory, if we had synthetics that exceeded their specs in areas like strength and speed. Unless we blow ourselves up, it is entirely inevitable that we will eventually be able to make synthetic muscles, implant them into people, and produce runners that can finish a mile in two minutes or less.
There are even people who really desire augmentations like this - they're called "transhumanists". They're not insecure or weird or technology-obsessed or anything like that; they just want new experiences and better abilities than those they were born with. Perfectly straightforward. There are plenty of such people in the world today, probably several million. Hundreds of thousands of video games, movies, and books center around transhumanist themes. Humans have been dreaming of transhumanity ever since the earliest religions.
But the Technological Singularity isn't really about transhumanity in the traditional sense that transhumanists dream about it. Greater intelligence is totally different than bigger muscles. Greater ability to communicate information, to see patterns, abstract things, get the point, think outside the box, see the solution a human never could, shock and intrigue the greatest human geniuses with superhuman intelligence. We're talking a totally different domain here. Beings smarter than humans in the same way that humans are smarter than chimps, or dogs, or even rocks. Think that's impossible? If so, you're probably being anthropocentric.
Bigger muscles are relatively easy to imagine. Bigger intelligence is impossible to imagine. We're too dumb to imagine it. If we could, then we'd be that smart ourselves. Bigger intelligence can solve problems that we thought were impossible to solve. Aside from the possibility of bigger intelligence is also faster intelligence. Intelligence that experiences and ponders and laughs and cries and solves problems at millions or even billions of times the human speed.
Because our human brain is just a big network of neurons that spike roughly 100 times per second, it's kinda slow relative to the maximum possible speed an intelligence can be. Silicon transistors already process information millions of times faster. It's just that we haven't yet set them up in a way that matches or exceeds our human ability to invent, to get the picture, to compromise, to create, to tell a joke, whatever. Can we set silicon up to tell a creative joke, or invent a new superweapon, or improve its own design? My answer probably won't surprise you:
Yes, obviously. It's all just a matter of time. All intelligence is information processing. We've known that since the 50s and before.
So, we won't be able to do it until a million years from now, right? Wrong. We've already uncovered powerful, very broad algorithms which formally solve computational problems your average person presently thinks no machine will ever tackle. I'm not talking about computational problems like chess or how to configure a factory, but general problems like universal Inductive Inference, sequence prediction, and selecting decisions. Problems that are the foundations of intelligence. Problems that, if solved, produce something we could easily define as an intelligent - even superintelligent - type of entity.
How soon do I personally think that human programmers can code a superintelligent Artificial Intelligence? I would answer that question directly, but I hate to move arguments from technical grounds to cultural grounds. When Technological Singularity advocates start throwing around dates like 2010 or 2020 without going through the long chain of reasoning that lead to these dates, the audience rapidly switches from pondering a serious future possibility to paranoidly attempting to model the psychology of the advocate, wondering what bad sci-fi book devoured their mind.
When the audience models the underlying cause of the forecast as cultural rather than technical, they begin to formulate cultural counterarguments rather than technical ones. This problem is compounded by the fact that it is far easier to pull a cultural counterargument out of a hat than it is to carefully craft a valid technical counterargument against a given Technological Singularity forecast. Cultural counterarguments are funny and entertaining - they make you stop and laugh - and, incidentally, stop putting your mind to work on thinking seriously. They only make you look bad to people who are used to arguing almost exclusively in technical domains rather than cultural ones, and that's practically no one.
The feasibility or non-feasibility of human-surpassing Artificial Intelligence should be argued on technical grounds only. This is very difficult, in a cultural environment saturated with campy AI flicks and a murky multi-decade history of nonscientific speculation about AI. Arguing AI on technical grounds requires that the participants read actual books, and not just any book that claims to be about AI.
A large part of the problem is that software engineers of all stripes have co-opted the word "Artificial Intelligence" practically since day one. Any software program that does anything interesting is an Artificial Intelligence, according to them. "AI" is a sexy term; slap it onto everything and people will get excited.
But most programmers don't have a clue about what intelligence is. They don't know any mathematically rigorous definitions of intelligence, they haven't studied the human brain, they have seen plenty of bad sci-fi movies with AIs, they are soaked in cultures that use the term to describe anything from Google to the circuitry in their car.
Speaking about Artificial Intelligence in a mature way often requires trashing everything you thought you knew about AI and restarting from the beginning. It requires learning a good deal about the only working example of intelligence we currently see; the human brain. Not "the human mind" as in college psych class, but the human brain as in actual Cognitive Science and Evolutionary Psychology.
In the field of nanotechnology today, serious researchers are constantly being bombarded by idiotic criticisms and objections formulated on entirely political grounds. Self-replicating machines are impossible simply because they sound like they should be impossible. The actual physics is irrelevant. The fact that life itself is an example of a nanotechnological system is irrelevant. The fact that no error has been located in the primary text arguing the technical feasibility of nanotechology in over a decade since its original publication is irrelevant. The situation is simply sad, and its because something happened which has happened millions of times before this century - politics and culture is stepping on scientific fact.
Humans aren't built to deal with science. We're built to play politics, flow with culture, have healthy babies and get along with the tribe. Scientific evidence is selected to back up pre-existing political and cultural goals - a political or cultural goal could never derive from scientific facts alone. (Well, most of the time anyway, unless the scientific facts are already assimilated as common knowledge, but the delay between scientific discovery and public familiarity is sometimes measured in decades.)
The atom bomb didn't appear in any form of fiction prior to the explosion at Los Alamos. Everyday people never would have guessed. Particle physicists did. Einstein knew. Einstein begged Roosevelt not to build the bomb, but he did anyway. The bomb didn't become real when it detonated - the possibility of a bomb became solid and concrete as soon as some shocked physicist did some math and calculated the number of joules that would be released in a nuclear chain reaction involving uranium. "Nuclear chain reaction" is a phrase that would have made little sense to the so-called intellectuals filling up the best universities at the time - unless they were physicists.
Now I'm going to use a phrase you may not have heard of - "seed artificial intelligence". Seed AI is general problem-solving AI that improves upon its own source code. It's AI that shapes information - and matter - in accordance with its desires, or, if you will, utility function. It does this because the programmers specifically built it that way. It starts off really dumb, but gets smarter and smarter as the programmers improve the information structure that constitutes it.
As the AI gets smarter, it gets assigned self-improvement tasks that are increasingly sophisticated. At the beginning, the AI acts as no more than a compiler for itself - it automatically tidies up its own code using simple mathematical procedures. But over time, with programmer guidance, the AI begins to come up with new strategies and algorithms for self-improvement found in the code of no present-day compiler. The process snowballs, until one day the AI can make serious self-improvements without programmer assistance.
At some point, the AI crosses a critical line. It becomes smarter than its programmers. The AI is in luck. Since it lives on a silicon substrate, which is superfast, completely reprogrammable, integrates well with new hardware, and so on, it has a bunch of advantages that no intelligent being on Earth has ever had before. Such phenomena has been referred to as "the AI Advantage", advantages that come with the territory of AI, and they are very poorly acknowledged and studied among traditional AI communities. But the impact they have when truly general AI is created will be enormous.
One day you have an AI that is roughly human-equivalent, the next day you have an AI that is vastly superintelligent. That's because the AI operates on a different time scheme than you do - machine time, not biological time. Machine time is millions of times faster than biological time. It never gets bored. It doesn't sleep. It has huge portions of its cognitive complexity devoted specifically to improving itself, something humans can only dream about. Human self-improvement was largely an evolutionary accident and is extremely limited - this AI was specifically built for the purpose, and it can burn all available computing power specifically for the purpose of making itself more intelligent.
Some random person walks into the AI lab and sees a mainframe and screens, and a software program running on the mainframe. The AI programmers see an intelligent mind, at first an equal and quickly a superior, a child built a line of code at a time. The AI itself sees an environmental probability distribution, and a variety of possible futures based on the actions it quickly chooses.
With its vast intelligence and cognitive capabilities, the first superintelligent AI will not see the world as humans do. We look to one future and see a romantic relationship, another future with a completed home improvement project, and a third future teaching a language in a foreign country. AIs, intelligences built from scratch rather than by Homo sapiens genomes, see future possibilities in such a different light that putting ourself into their shoes will be cognitively impossible.
The future possibilities that the first superintelligent AI focuses upon will be based on one thing and one thing only - the information content of its utility function. Any intelligence will selectively process environmental information in accordance with its utility function. And there are far more possible utility functions in the universe than there are possible human goals. The entire space of human goals is an infintismal speck in the entire space of mathematically possible utility functions. Most utility functions have nothing to do with selfishness, jealousy, politics, culture, fear, hate, or any other goals that humans so intimately experience in our daily lives. Although human desires can be interpreted as utility functions, these functions are complicated mathematically, and very tangled, having been built in stages by the evolutionary process.
Most science fiction stories are overpopulated by anthropomorphic aliens with anthropomorphic psychologies. "Weird" aliens merely behave like weird humans, with few exceptions. Real aliens, beings that evolved on totally different planets and different evolutionary environments with different selection pressures would possess significantly different psychologies than humans do. Artificial Intelligences, not crafted by evolution at all but by other intelligences, could be even more alien than extraterrestrials. Basically, all our science fiction stories are wrong.
