Content by Cory Doctorow (first e reader txt) 📕

as transcendently complex, and we think of it as being the bit that makes us us.

But brains aren’t that complex, Kurzweil says. Already, we’re starting to unravel their mysteries.

“We seem to have found one area of the brain closely associated with higher-level emotions, the spindle cells, deeply embedded in the brain. There are tens of thousands of them, spanning the whole brain (maybe eighty thousand in total), which is an incredibly small number. Babies don’t have any, most animals don’t have any, and they likely only evolved over the last million years or so. Some of the high-level emotions that are deeply human come from these.

“Turing had the right insight: base the test for intelligence on written language. Turing Tests really work. A novel is based on language: with language you can conjure up any reality, much more so than with images. Turing almost lived to see computers doing a good job of performing in fields like math, medical diagnosis and so on, but those tasks were easier for a machine than demonstrating even a child’s mastery of language. Language is the true embodiment of human intelligence.”

If we’re not so complex, then it’s only a matter of time until computers are more complex than us. When that comes, our brains will be model-able in a computer and that’s when the fun begins. That’s the thesis of Spiritual Machines, which even includes a (Heinlein-style) timeline leading up to this day.

Now, it may be that a human brain contains n logic-gates and runs at x cycles per second and stores z petabytes, and that n and x and z are all within reach. It may be that we can take a brain apart and record the position and relationships of all the neurons and sub-neuronal elements that constitute a brain.

But there are also a nearly infinite number of ways of modeling a brain in a computer, and only a finite (or possibly nonexistent) fraction of that space will yield a conscious copy of the original meat-brain. Science fiction writers usually hand-wave this step: in Heinlein’s “Man Who Sold the Moon,” the gimmick is that once the computer becomes complex enough, with enough “random numbers,” it just wakes up.

Computer programmers are a little more skeptical. Computers have never been known for their skill at programming themselves — they tend to be no smarter than the people who write their software.

But there are techniques for getting computers to program themselves, based on evolution and natural selection. A programmer creates a system that spits out lots — thousands or even millions — of randomly generated programs. Each one is given the opportunity to perform a computational task (say, sorting a list of numbers from greatest to least) and the ones that solve the problem best are kept aside while the others are erased. Now the survivors are used as the basis for a new generation of randomly mutated descendants, each based on elements of the code that preceded them. By running many instances of a randomly varied program at once, and by culling the least successful and regenerating the population from the winners very quickly, it is possible to evolve effective software that performs as well or better than the code written by human authors.

Indeed, evolutionary computing is a promising and exciting field that’s realizing real returns through cool offshoots like “ant colony optimization” and similar approaches that are showing good results in fields as diverse as piloting military UAVs and efficiently provisioning car-painting robots at automotive plants.

So if you buy Kurzweil’s premise that computation is getting cheaper and more plentiful than ever, then why not just use evolutionary algorithms to evolve the best way to model a scanned-in human brain such that it “wakes up” like Heinlein’s Mike computer?

Indeed, this is the crux of Kurzweil’s argument in Spiritual Machines: if we have computation to spare and a detailed model of a human brain, we need only combine them and out will pop the mechanism whereby we may upload our consciousness to digital storage media and transcend our weak and bothersome meat forever.Indeed, this is the crux of Kurzweil’s argument in Spiritual Machines: if we have computation to spare and a detailed model of a human brain, we need only combine them and out will pop the mechanism whereby we may upload our consciousness to digital storage media and transcend our weak and bothersome meat forever.

But it’s a cheat. Evolutionary algorithms depend on the same mechanisms as real-world evolution: heritable variation of candidates and a system that culls the least-suitable candidates. This latter — the fitness-factor that determines which individuals in a cohort breed and which vanish — is the key to a successful evolutionary system. Without it, there’s no pressure for the system to achieve the desired goal: merely mutation and more mutation.

But how can a machine evaluate which of a trillion models of a human brain is “most like” a conscious mind? Or better still: which one is most like the individual whose brain is being modeled?

“It is a sleight of hand in Spiritual Machines,” Kurzweil admits. “But in The Singularity Is Near, I have an in-depth discussion about what we know about the brain and how to model it. Our tools for understanding the brain are subject to the Law of Accelerating Returns, and we’ve made more progress in reverse-engineering the human brain than most people realize.” This is a tasty Kurzweilism that observes that improvements in technology yield tools for improving technology, round and round, so that the thing that progress begets more than anything is more and yet faster progress.

“Scanning resolution of human tissue — both spatial and temporal — is doubling every year, and so is our knowledge of the workings of the brain. The brain is not one big neural net, the brain is several hundred different regions, and we can understand each region, we can model the regions with mathematics, most of which have some nexus with chaos and self-organizing systems. This has already been done for a couple dozen regions out of the several hundred.

“We have a good model of a dozen or so regions of the auditory and visual cortex, how we strip images down to very low-resolution movies based on pattern recognition. Interestingly, we don’t actually see things, we essentially hallucinate them in detail from what we see from these low resolution cues. Past the early phases of the visual cortex, detail doesn’t reach the brain.

“We are getting exponentially more knowledge. We can get detailed scans of neurons working in vivo, and are beginning to understand the chaotic algorithms underlying human intelligence. In some cases, we are getting comparable performance of brain regions in simulation. These tools will continue to grow in detail and sophistication.

“We can have confidence of reverse-engineering the brain in twenty years or so. The reason that brain reverse engineering has not contributed much to artificial intelligence is that up until recently we didn’t have the right tools. If I gave you a computer and a few magnetic sensors and asked you to reverse-engineer it, you might figure out that there’s a magnetic device spinning when a file is saved, but you’d never get at the instruction set. Once you reverse-engineer the computer fully, however, you can express its principles of operation in just a few dozen pages.

“Now there are new tools that let us see the interneuronal connections and their signaling, in vivo, and in real-time. We’re just now getting these tools and there’s very rapid application of the tools to obtain the data.

“Twenty years from now we will have realistic simulations and models of all the regions of the brain and [we will] understand how they work. We won’t blindly or mindlessly copy those methods, we will understand them and use them to improve our AI toolkit. So we’ll learn how the brain works and then apply the sophisticated tools that we will obtain, as we discover how the brain works.

“Once we understand a subtle science principle, we can isolate, amplify, and expand it. Air goes faster over a curved surface: from that insight we isolated, amplified, and expanded the idea and invented air travel. We’ll do the same with intelligence.

“Progress is exponential — not just a measure of power of computation, number of Internet nodes, and magnetic spots on a hard disk — the rate of paradigm shift is itself accelerating, doubling every decade. Scientists look at a problem and they intuitively conclude that since we’ve solved 1 percent over the last year, it’ll therefore be one hundred years until the problem is exhausted: but the rate of progress doubles every decade, and the power of the information tools (in price-performance, resolution, bandwidth, and so on) doubles every year. People, even scientists, don’t grasp exponential growth. During the first decade of the human genome project, we only solved 2 percent of the problem, but we solved the remaining 98 percent in five years.”

But Kurzweil doesn’t think that the future will arrive in a rush. As William Gibson observed, “The future is here, it’s just not evenly distributed.”

“Sure, it’d be interesting to take a human brain, scan it, reinstantiate the brain, and run it on another substrate. That will ultimately happen.”

“But the most salient scenario is that we’ll gradually merge with our technology. We’ll use nanobots to kill pathogens, then to kill cancer cells, and then they’ll go into our brain and do benign things there like augment our memory, and very gradually they’ll get more and more sophisticated. There’s no single great leap, but there is ultimately a great leap comprised of many small steps.

“In The Singularity Is Near, I describe the radically different world of 2040, and how we’ll get there one benign change at a time. The Singularity will be gradual, smooth.

“Really, this is about augmenting our biological thinking with nonbiological thinking. We have a capacity of 1026 to 1029 calculations per second (cps) in the approximately 1010 biological human brains on Earth and that number won’t change much in fifty years, but nonbiological thinking will just crash through that. By 2049, nonbiological thinking capacity will be on the order of a billion times that. We’ll get to the point where bio thinking is relatively insignificant.

“People didn’t throw their typewriters away when word-processing started. There’s always an overlap — it’ll take time before we realize how much more powerful nonbiological thinking will ultimately be.”

It’s well and good to talk about all the stuff we can do with technology, but it’s a lot more important to talk about the stuff we’ll be allowed to do with technology. Think of the global freak-out caused by the relatively trivial advent of peer-to-peer file-sharing tools: Universities are wiretapping their campuses and disciplining computer science students for writing legitimate, general purpose software; grandmothers and twelve-year-olds are losing their life savings; privacy and due process have sailed out the window without so much as a by-your-leave.

Even P2P’s worst enemies admit that this is a general-purpose technology with good and bad uses, but when new tech comes along it often engenders a response that countenances punishing an infinite number of innocent people to get at the guilty.

What’s going to happen when the new technology paradigm isn’t song-swapping, but transcendent super-intelligence? Will the reactionary forces be justified in razing the whole ecosystem to eliminate a few parasites who are doing negative things with the new tools?

“Complex ecosystems will always have parasites. Malware [malicious software] is the most important battlefield today.

“Everything will become software — objects will be malleable, we’ll spend lots of time in VR, and computhought will be orders of magnitude more important than biothought.

“Software is already complex enough that we have an ecological terrain that has emerged just as it did in the bioworld.

“That’s partly because technology is unregulated and people have access to the tools to