j a c k *

Extropian Wake-up Call...

Charlie Stross has a post making the rounds among skeptics and curmudgeons about how the Singularity is not particularly near and may be impossible. I agree, but largely on the conclusion not necessarily on the particulars of Stross' argument.

It was nice of Stross to link to the 1993 article by SF author Vernor Vinge that first coined the term. Vinge looked ahead, as is proper for his profession, and he found his speculations on the human future were clouded by an event he imagined to be both world-altering and imminent. He rightly called this a "singularity" -- the point where the current rules no longer apply and further extrapolation is impossible. The future beyond the singularity is governed by principles so radically different that we cannot imagine it.

Nonetheless he imagined it anyway and described an apocalypse. This result is fairly common when people try to deal with potentially powerful historical processes. Malthus famously predicted much the same when he looked into the arithmetic of population growth. Likewise, those of us who grew up in the 70's and 80's can recall the widespread belief that once there were enough nuclear weapons to destroy the world multiple times over that it was just a matter of time before we did.

The event that Vinge cannot see beyond (or can only see as the end of the world) is the development of super-human artificial intelligence. AI is an old trope in SF. We've imagined a future full of wise-cracking robots or homicidal computers for decades and no one really gave them a second thought. We would always somehow defeat them with our human cleverness or some unique biological illogic that's always better than rationality.

The modern wrinkle is Moore's Law -- the observation that computer power is increasing exponentially. How can we continue to think we'll outsmart the robots if they are always getting smarter than we are?

Formally the argument goes like this: through heroic efforts or just by accident we develop super-human AI. That AI turns its efforts to improving both the algorithms that make it intelligent, and the hardware on which it runs. Any improvements in intelligence or performance result in a faster rate of improvement. Thus a runaway feedback loop produces super-super-human AI in short order and normal meat-based humans are toast.

One counter-argument is that Moore's Law will peter out. That's true eventually, of course, but I think we're a long way from the limits of how much computation can be wrung out of small amounts of matter. Others argue that hard AI is impossible. Either the brain must be made of meat for some reason, or that human attempts to understand the brain must necessarily fail.

I disagree. I believe that AI is possible; I don't think there's anything magical about the human brain that can't be figured out or converted to an algorithm. I do believe, however, that AI is hard. It isn't going to happen by accident, or just as a result of some threshold level of complexity. The Internet isn't just going to "wake up" one day and become conscious, like some fanatics fervently hope.

This will require a theory of mind. I believe we will eventually puzzle out the mystery of how neuronal interactions create mental phenomena, including awareness and consciousness. And it will be a singular event; people will understand the world and themselves very differently after that. On the other hand much will stay the same. People will have better insights into how their own minds work but that won't necessarily produce better behaviors. The problem is that because this development requires both research and insight, we cannot put a time estimate on when it might emerge. It could be a very long time to discover a theory of mind.

Once we understand how minds work it may still be a long time before we can put that knowledge to practical use. The path from science to engineering can be grueling and tortuous, especially if the subject is complex. We have a very good theory of life; we understand how DNA works, we have a large body of work on proteins and enzymatic action, we have great models of signaling inside and among cells, and we can manipulate all of these processes in the lab. And yet the grand results of biotechnology imagined decades ago are still decades away. Life processes aren't very amenable to engineering. Mental process may be just as complex if not more so.

But suppose they aren't. Suppose that engineering minds is relatively easy, or -- given enough time and research -- we've figured out how to do it. In that case do we get a runaway AI, a Frankenstein's Monster in silicon?

No. If we have a theory of mind, and if we have experience building artificial minds, then we'll know how to build them to omit or subsume anything that could be anti-social or dangerous. There's no reason to think that any of the features that come standard with natural human minds -- greed, pettiness or tribalism -- are required for the useful application of general intelligence. Even if an AI bent on domination did arise, we'd still have AI's on "our side" that would use their super-human intelligence to counter the threat.

Could we get super-human AI without a theory of mind? Possibly. The path in that case is to wait for technological development to reach the point where a human mind can be "digitized" by scanning the configuration of neurons and their interactions. We still wouldn't know how the brain works but we wouldn't have to -- we'd just be simulating it as a program, and if the computer runs faster than the equivalent neurons then you have a faster, and therefore super, human mind.

Except that because we don't know how it works, the only way the mind simulation can function is in a simulated world with a simulated body. That has to be a huge bottleneck. What you have is not really super-human AI but rather a sub-culture of normal humans living really fast. Sure, they can think faster than physical humans, but subjectively they are still living normal -- albeit virtual -- lives.

More importantly, there's very little that these uploaded fast-humans can do to improve themselves. They could work in the field of electronics to create faster computers, thus making themselves (and every other computer) faster, but that will still require them struggling through what for them will be years of virtual college. Without a theory of mind they cannot improve their own intellect. They are just a simulation of the off-the-shelf natural-selection model, and short of simulated evolution nothing can change that.

Put simply, I think SF writers can relax for a while.

- jack*

The Importance of Expertise ...

Editors of Scientific American, Dear Sirs:

Your recent article A Test for Consciousness was subtitled "How will we know when we've built a sentient computer? By making it solve a simple puzzle." I'm puzzled how an article lacking any substance could be published at all, let alone with such grandiose headlines.

To start with the authors seem entirely ignorant of the history of Artificial Intelligence (AI) research. They state, for example:

To your computer, all pixels are just a vast, disconnected tapestry of three numbers (corresponding to three colors) with no particular meaning. To you, an image is meaningful because it is chock-full of connections among its parts, at many levels.

In the 60's people used to take the more reductionist view that computers process images as a series of ones and zeros. I guess RGB is an improvement. But let me quote from my 1976 textbook on AI The Thinking Computer by Bertram Raphael:

Therefore the problem for the software is to translate confusing arrangements of numbers into simple, meaningful descriptions of visual scenes.

In other words, software engineers have known for at least 35 years that just dealing with raw data from the sensors is not enough. The goal is to interpret perceptual data in useful ways. Perhaps common knowledge and context can be important. Indeed the authors suggest:

Today's machines can pick out the face of a likely terrorist from a database of a million of faces, but they will not know his age, his gender or ethnicity, whether he is looking directly at the viewer or not, or whether he is frowning or smiling. And they will not know that if he is shaking hands with George Washington, the photograph is probably digitally doctored. Any conscious human can apprehend all these things and more in a single glance.

Wait -- what? Any conscious human can identify George Washington and knows the context in which he lived? What about William Kamkwamba, aka the Boy who Harnessed the Wind? He was about as ignorant as one can be of the modern world and still be a genius. I'm sure he had never seen George Washington during his life in tribal Malawi. Is he non-sentient?

What Koch and Tononi are talking about is the "general" knowledge or "common sense" knowledge problem. This is an old, well-known issue in AI research.

While researchers were aware that in an AI system, knowledge would have to be explicitly represented, they did not anticipate the vast amount of implicit knowledge we all share about the world and ourselves. [...] In retrospect, this is perhaps not surprising, because the implicit nature of this knowledge in humans means that we all take it for granted, and never have to state it or consider it explicitly.

I could excuse this lack of scholarship if the authors offered anything new or interesting. They do not. They propose a test for consciousness that involves ranking pictures based on cultural knowledge -- something only slightly better than IQ tests that require culturally contextual interpretations. Perhaps I know that a flowerpot can't be used as a computer keyboard. Or more likely I've never seen anything like either of those things.

The most disappointing part of this terrible article is that they never answer their own questions about "sentience" or "consciousness". These are very, very important questions, and yet they just wave their hands in response. Consciousness, they say, is related to complexity of wiring, while never considering software complexity. "Sentience" is never mentioned in the article at all.

There are a lot of smarter people who came before you. You are welcome back when you have something substantive to say.

- jack*

Drastic Aspect Ratios...

Dear Internet,

You know I love you. I obsess over you every day the way my parent's generation adored their newsprint data feeds -- perhaps more so. I need you to entertain me, to inform me, and to support my enterprise. I cannot do without you.

And yet you do stuff like this. What the hell, Internet?

Let me explain.

My computer, like most, has a wide-screen "high definition" monitor. This means it's short and stout. Look at your own computer display. Most likely it's also 80% wider than it is tall, and for good reason. All current video content is 16 by 9; thus all TV screens are 16:9; therefore all computer monitors -- made by the same companies and with the same equipment as flat screen TVs -- are also 16:9. This is a mathematical manufacturing reality mysterious to no one.

On the other hand most web content is oriented vertically. Look at this very website. You will see wide expanses of empty blue on the left and the right, while the actual content of the site inhabits a narrow strip down the middle of the page that demands scrolling. Even on more fully-populated pages, the useful material tends to run vertically down the center with the peripheral areas filled with secondary links and ads.

So we see that there is already tension between the pixels that are available to show web content, and the format of the content itself. At best we are left with a narrow band of useful pixels down the middle of the screen -- the intersection between the vertical content and the horizontal display. Perhaps we can use the remaining area fruitfully with other applications, perhaps not. What's clear in any case is that vertical real-estate is at a premium, while horizontal real-estate is going to waste.

Why then -- why oh why oh why -- do internet applications insist on eating up that precious vertical space? The "free web toolbar" previously linked -- that has started showing up on sites where I mostly want to read the content thank you very much -- sucks up at least 2 or 3 lines of what could be text. To what end? Those sites all have acres of empty space to the right or left that could be used for rarely-used toolbars. These are nearly as bad as the website headers and menus that persist in taking up space no matter how much you scroll. Are web programmers really incapable of making interfaces that use the sides of the screen instead of great swaths of vertical space?

Of course it's not just the web pages; it's the browser itself. My old browser reserved vertical space for: the window title, the application menu, the navigation toolbar, the bookmark toolbar, tabs, and a status info bar. The new one removes the status info and the menu -- incorporating the latter into the window title -- and yet it still makes considerable vertical space unavailable for content. Content which cannot, by the way, take advantage of the wide open frontiers of the sides of the browser.

So here's my question, Internet. Is it really impossible to make web content -- and web browsers for that matter -- that actually fit with and fully utilize widescreen computer displays? I think it is, and I don't feel like I should be having to ask this question.

- jack*