The coming AI phase change

How long will it be until we have real honest-to-goodness general human-level AI? This question is of at least some interest to a futurist (if for no other reason than that smart AIs might put futurists out of a job :-) ).

There are several polls of experts out there, summarized here, which can be summed up by saying the median guess is in the 2040s. (One quickie poll most current commentators missed was Wired's Reality Check in 1996, which predicted a C-3PO -like robot in 2047. The Macmillan Atlas of the Future (1998) calls for 2035.)

On the other hand, some people, e.g. Robin Hanson, think it will be much longer. Hanson informally surveyed various AI researchers and came to the conclusion that we still have a couple of centuries more work to do at the least.

On the other hand, Ray Kurzweil famously claims (with tongue at least a little in cheek, one presumes) it will be exactly 2029. He points out that the progress in a field like this is not linear, and gives the Human Genome Project as a recent, well-documented example of dramatic acceleration in terms of a naive figure of merit.

What is the chance for a dramatic acceleration of AI development? Do we even have a usable figure of merit for AI? How about IQ? Unfortunately, IQ was invented as a tool for tracking human intellectual development--the Q means the quotient of the mental to physical age, at least originally--so that it seems likely that by the time we have a machine for which it makes any sense at all to say it has any IQ at all, we will basically have already succeeded at the AI task.

I think to some extent AI is more like the major conceptual breakthroughs like building a flying machine. The hard part of developing working airplanes was not the solution of any of the various technical problems. It was understanding what the problems were. Everybody and his dog Muttley knew how to build a machine with lift which would get off the ground. It took the Wright brothers, with the point of view of bicycle makers, to realize that the problem was balance and attitude control. Once they succeeded, virtually nobody copied their technical solution (wing warping); ailerons quickly became the standard method. But beforehand, control was an "unknown unknown."

A decade ago in Beyond AI, I took up cudgels against those who were predicting a major takeoff in artificial intelligence by virtue of a self-improving super-AI. Long before that happens, I said, we will see something a bit more mundane but perfectly effective: AI will start to work, and people will realize it, and lots of money, talent, and resources will pour into the field. “... it might affect AI like the Wright brothers' Paris demonstrations of their flying machine did a century ago. After ignoring their successful first flight for years, the scientific community finally acknowledged it; and aviation went from a screwball hobby to the rage of the age, and kept that cachet for decades. In particular, the amount of development effort took off enormously.” That will produce an acceleration of results, which will attract more money, and there's your feedback loop. The amount of money going into aviation before 1910 was essentially nil (Langley's grants to the contrary notwithstanding). Once people caught on that airplanes really worked, though, there was a sensation and a boom. By the end of the 1920s, Pan American was flying scheduled international flights in the 8-passenger Ford Tri-motor. The ensuing exponential growth in capabilities continuing unabated right up to the Sixties.

Hanson points out that pouring more money into the field might not accomplish much:

Fourth, it is well known that most innovation doesn’t come from formal research, and that innovations in different areas help each other. Economists have strong general reasons to expect diminishing returns to useful innovation from adding more researchers. Yes, if you double the number of researchers in one area you’ll probably get twice as many research papers in that area, but that is very different from twice as getting much useful progress.

This is quite true, but there is another reason to think that we might be on the cusp of a phase change into a higher growth mode in AI research: it's moving out of academia and into industry. There's more emphasis on getting results and less on looking clever. I've been in both and I can personally attest. Academia is a lot more fun; in industry you get a lot more done that is useful.

In November 2005, attending a AAAI symposium, I found myself in casual conversation with one of the other artificial intelligence researchers at the reception. One of the hot topics of conversation in AI circles just then was the DARPA Grand Challenge. The previous month, five autonomous vehicles—self-driving cars and trucks—had successfully completed a grueling 131.2-mile course in competition for a $2 million prize offered by DARPA, the defense department's research agency. This was a major advance in the state of the art, since the previous Challenge, held just a year and a half earlier, had been a complete failure, with the best vehicle only managing to go 7.3 miles.

I had remarked as much to my AAAI friend, and he demurred. The apparent advance, he insisted, consisted of nothing but new ways of combining existing sensory, control, and navigation techniques. That seems to be a fairly common ivory-tower way of seeing things.

But that, of course, is exactly what the vast majority of actual technological progress consists of. And the Grand Challenge results show graphically what kind of a difference it can make in the real world. However much a specialist may recognize all the parts and elements of a new machine from earlier efforts, what the world at large notices is whether or not it works. And in the case of self-driving cars, a major watershed was crossed between March 2004 and October, 2005.

But there was no Clever New Trick that would have excited an academic, nothing even that an active AI researcher recognized as an advance.

The money makes a big difference too. I would guess that a machine that could run a human equivalent AI in real time would cost about $1 million today. One reason that AI progress has been slow and roundabout is that academic researchers were (a) underfunded and (b) sidetracked by finding ingenious solutions that ran on way-too-skimpy hardware, but which were brittle. In the brain, many things are done by brute force and are as a result robust.

All of which we may be in the process of shifting gears away from. Major feedbacks seem to be in place, and I surmise that we might be kicking into an exponential growth mode. The only problem is, I still don't know how to label the Y-axis.

How is a great nation like a flagellar bacterium?

Consider the humble E. Coli. It is worth studying for several reasons: it has atomically-precise electric motors to turn its flagella, for example. The flagella themselves are corkscrew-like hairs that act as propellers when turned one way, and induce random motions when turned the other. But the point of interest is what the flagella are used for.
The E. Coli can also sense the density of dissolved nutrients in the water it is swimming in. It swims along and notices that the level is climbing; it is swimming into a region of higher concentration; it's happy. So it keeps swimming.
On the other hand, if it swims through the region, it notices the concentration declining. So it reverses its flagella and tumbles. It winds up facing in a random direction, and sets out again. This may or may not get it on a course that takes it back into the food, but at least it has a chance. The result of the whole algorithm is a biased random walk that tends to keep it in the higher-concentration areas.
It turns out that an ordinary housefly does much the same thing. Its maddening buzzing around is a biased random walk through the concentration of food-like smells in the air. This causes a concentration of flies in the vicinity of things like garbage cans.

Like the bacterium or the fly, the population of a great nation cannot really see where it is going. (Needless to say, neither can the government.) All they can do, in general, is to tell if things are getting better or worse. And if things are getting worse, the people will call for Change. In a democracy, they have a mechanism to do this more or less peacefully, but they will do it the hard way if they must. Think of Romania; think of the Soviet Union.
And when the people call for Change, what they get is a tumble in the highly multidimensional space of policy options.
This makes it a bit hard for a futurist.

Gosper's hierarchy of needs

In yesterday's post I tried to point out that the intuitions about whether a machine implementation of our minds was really conscious (etc) seemed to depend on how much its internal mechanism resembled our own. In particular, a Chinese Room implemented as a lookup table seemed particularly resistant to the notion that there's "somebody home."

But that left unexamined the question of how the lookup table got filled in. In the case of HashLife, the answer is straightforward: take the patch of cellular automata space you are trying to skip forward, run the Life algorithm on every possible configuration, and fill in your lookup table. But equally obviously, you don't actually have to do this ahead of time: you run your Life simulation as usual, looking for speedups in your table, and every time you see a situation you don't have listed, run it as normal a step at a time and then insert the results in the table. That's why it's a hash table, sparsely populated in the address space of starting patches.

In practice, the big systems in Life that the experimenters were trying to run were highly stylized, with glider guns and sinks and mirrors and similar gadgetry, to construct circuitry and Turing machines or even to emulate (!) more complex cellular automata. In such a case HashLife essentially creates a direct table-driven implementation of the higher-level machine.

How would we apply this scheme to running a human mind? We don't have hash tables in our heads, and whats more, the address space a human experiences is so vast and finely divided that we never experience exactly the same input or situation.

We don't have hash tables in our heads, but we do have circuitry that looks suspiciously like associative memory, a point I first ran across in Pentti Kanerva's thesis. What's more, as we know from our experience with neural networks, it is reasonably straightforward to arrange such circuitry so that it will find the nearest stored memory to the requested address. With a bit more work, you can make a memory that will interpolate, or extrapolate, two or more stored memories near a requested address.

Do you remember learning to walk, or tie your shoes, or tell time from an analog clock dial, or read and write? These were all significant cognitive challenges, and at one time you were heavily concerned with the low-level details of them.  But now you do them unconsciously, having essentially hashed them out to operate at a higher level.

Thus it seems not unreasonable to claim that the authentic human experience includes the HashLife-like phenomenon of losing direct consciousness of lower-level details in the process of becoming concerned with higher ones. Indeed I would claim that you cannot have the authentic human experience without it.

The remaining question is, how high up can we go?

HashLife and Permutation City

I've been thinking a bit about the Hanson/Caplan disconnect on ems. To recap and oversimplify, Hanson makes the assumption that an emulation or upload should count as a person; in a world with many times as ems as biological humans, the happiness of the ems matters as much or more than that of the meat. Caplan thinks the opposite: only we wild-type folks are real people, and the ems are just computer programs, with the straightforward result that they can't really be happy, or presumably even conscious of any real emotion at all.

Where you fall on the spectrum between these two beliefs would seem to depend very strongly on your intuition of what you really are. As a lifelong AI researcher, for example, I have always seen myself (and of course everybody else) as a computational process that just happens to be running on a n evolved biological substrate, albeit one of phenomenal sophistication and computational power. On this view, running the same computation on another substrate would not make any difference. It would not only still be conscious, and still be human; it would still be me.  Hanson explicitly endorses this view by using the term "emulation" for what would otherwise be called an "upload."

What if, however, instead of simulating the brain on a neuron-by-neuron level, you started working out the functionality of various pieces of it, as we have begun to do for the pre-processing in the retina, various pathways in visual and auditory cortex, and so forth. Many of these are perfectly understandable data-processing tasks reducible to algorithms, and others might be modeled to an acceptable precision by, e.g., neural nets trained on traces from real brains.

One can take this process further, reducing larger and larger parts of the implementation of "me" to algorithmic black boxes, and losing more and more of the information processing structure that is explicitly parallel to my brain. Let us suppose that we can do this in such a way that the result continues to act just like me from the outside, doing and knowing the same things, having all my memories, quirks, personality, and so forth.

The obvious endpoint of this is that we get to the point that the whole mind is one algorithmic black box, only related to the original person by input/output correspondence. It the resulting program conscious, human, me?

The problem is that this is not the real endpoint. The conceptually simplest way to implement that i/o black box is not a mysterious machine that might intuitively be comparable to the mysterious machinery of a human brain, but as a lookup table. We input one large number, say every tenth of a second, that encodes every sensory input, combine it with another large number that represents your memory, look up the corresponding line in the table, where we find two more large numbers, one of which encodes all the nerve signals to your muscles, and the other one is the new memory. That's all the mechanism; the rest is just the table of numbers.

Somewhere in the depths of cyberspace, John Searle is smiling.

But it gets worse. I first heard this scheme from Hans Moravec at a conference, but it also got folded into one of the weirder and more thought-provoking SF books, Greg Egan's Permutation City. It begins by considering one of the more sophisticated implementations of Conway's Game of Life cellular automaton universe. In an ordinary implementation, you compute the contents of each cell at each generation by a lookup table that encodes its state and interactions with its neighbors. But in HashLife, you don't have to do that on such a limited, atomic scale in either time or space. You have a giant hash table that stores the mapping from chunks of space to their configuration several steps of time later. The bigger and more complete your hash table, the wider the areas and more steps at a time you can skip.

Let us now return to Em City, and imagine what HashEms would be like, given precisely the strong economic pressures to efficiency that Robin depends on for the central prediction of the book. Not only do HashEms begin to skip internal steps: predictable interactions between multiple ems get folded into the table and optimized away.  Ultimately, no human-level day-to-day interaction (or recreation) is explicitly computed; only the I/O behavior of Em City as a whole emerges from the table-driven black box.

Are the ems still conscious? Human? Me?