the Foresight Institute

Let’s try to pull all the threads together, as futurists — which is the whole point here — and get some idea about when it might be reasonable to expect AI to show up.  When I say AI I want to look at the entire diahuman range, so the answer would still be a range even if we were historians looking back on the process from the vantage point of the far future.

I’ve claimed that “I think we have the techniques now to build an AI at the hypo/dia border, equivalent to a dull but functional human.”  That doesn’t mean we have one now, or even that one is possible next year.  What it means is that by the kind of techniques we can now use to program self-driving cars, we could, with a major development effort, program an AI that would be able to do as broad a range of that kind of task as a very dull human can, but which would need additional programming to do new tasks.

Commenter Alex Kilpatrick put forward a cogent objection to the “AI is near” thesis, writing:

All of the so-called gains in AI are still a million miles away from the “dull but functional human” There are some things like playing chess that computers do really well. And intelligent humans do those things too. But that in no way means the computer is remotely intelligent.
The whole AI field is nothing but clever programming. Some of those programs are quite clever indeed, but they represent the intelligence of their creators, not the programs. Some programs may appear intelligent in very narrow domains, but they are extremely brittle — they will not be useful at all even on the borders of the domains for which they were designed.

I agree with this strongly as a description of the state of AI today in general, but with one major reservation.  Not entirely all of the AI field is nothing but clever programming.  The AI programs that do the most impressive application tasks certainly are — because the efforts to build general learning machines are less than babies at the moment.

The key to moving up from the hypo/dia border into the diahuman range is imitation.  I’d guess that the state of the art would let us build a machine that would be able to watch someone sweeping a room and be able to sweep the same room with more or less the same series of strokes, being brittle to changes in the furniture positions and so forth. (Consider the kind of learning demonstrated in Ng’s helicopter.)  Building an AI that could watch lots of sweeping and then be able to figure out on its own how to sweep a new room — without having been programmed with any knowledge of sweeping ahead of time — is the kind of thing we need to advance the state of the art.

The difference is that in the second case the AI is inferring a model and a program from observations.  But this is what 21st century AI is (already) all about — typically, today, inferring statistical models from reams and reams of observations, but at least tackling the right problem.  The main thing that will determine the rate of advance is how much of the clever programming goes directly into end applications and how much goes into basic core learning.

Concept formation, model building, program inference, and so on are a quantum step harder than parameter tuning in a known ontology.  However, the math for that kind of thing is advancing, and the processing power to use techniques such as search and GAs is on its way in the next decade.  I don’t think we’ll have a superintelligent AI by 2020; indeed, I don’t think we’ll even have one that can educate itself by reading Wikipedia.  But I do think it’s at least a 50% chance we’ll have AIs that can learn something by a combination of imitation and careful verbal coaching.

Why we should suspect that the brain has a limited ability to recurse, but prefers to daisy-chain instead:

The house the malt the rat the cat the dog the cow with the crumpled horn the maiden all forlorn the man all tattered and torn the priest all shaven and shorn the cock that crowed in the morn the farmer sowing his corn kept waked married kissed milked tossed worried killed ate lay in was built by Jack.

… to boldly go where no man has gone before!

This final phrase of the classic Star Trek opening spiel had two problems with it, one as seen by people after the fact, and the other as seen by those who had gone before.

As seen by earlier generations, the phrase “to boldly go” is a split infinitive.  If E.E. Smith had written Star Trek in the ’20s, he would have written “boldly to go.”  Avoidance of split infinitives, like many elements of grammatical style, was a cognitively expensive signalling behavior that advertised, essentially, that the speaker or writer was in the educated classes in an era where being educated meant you knew Latin.  Infinitives in Latin are single words formed by inflection, rather than with a keyword such as “to” in English, so you can’t put an adverb in the middle of one.  Avoidance of split infinitives lasted in “proper” English at least until mid-20th century, but had begun to fade (to slowly fade ) away thereafter.

But there’s no real reason in English not to split infinitives.  They are completely understandable, and often less ambiguous than alternative constructions. Apart from being a cognitively expensive signalling behavior, they had no value, indeed a cost in cumbersome and ambiguous sentences.  Like many rules which were tacked onto the language by well-meaning grammarians, they were overly simplistic formalizations of a system, English grammar, which was and remains much deeper and more complex than anyone thought it was.

The lack of the ability of “hand-coded” grammar to handle real language is most clearly displayed in the early attempts at machine translation, which were an abject failure.  Only after 50 years of trying in natural language understanding, using statistically inferred models not formalized by humans, has serious progress been made (and there’s lots more progress needed before basic competence is achieved).

The other, retrospective, problem with the Star Trek blurb, that the phrasing was considered sexist, was corrected in later incarnations to the more politically correct “where no one has gone before.”  This, as it turns out, is another example where the unthinking application of a simple, formalized rule to “fix” something actually makes it worse.

In the ’60s, the use of “man” in such a context was standard and unambiguous.  It meant a human being (in fact, in my Websters from that era, “human being” is the first primary meaning, specialization to adult males being secondary).  Star Trek, if you remember or have studied these things at all, was in its time one of the most progressive, liberal science fiction shows ever.  It depicted a crew and implied a culture where barriers based on race and sex had been significantly lowered compared to the contemporary norm.  So in the original, “man” meant “human.”

Of course, the Enterprise went all over the galaxy seeking new life, new civilizations.  The citizens of these civilizations had been there before.  The distinction between “man” and these people is a fine one but one which can be reasonably be made consistent with the storyline.

But what happens if you switch from “man,” meaning human, to “one,” meaning, well, anyone. The term is intentionally more inclusive: it pushes the boundary from that between humans and non-humans, to that between someone and something.  But wait: doesn’t that mean that all the denizens of the strange new worlds, who have gone there before, are now not someones, but somethings?  Isn’t the new phrasing, taken out of the context of American academia and applied to Star Trek without thought or understanding, actually worse than before?  In classic Trek, the aliens are non-human people.  In PC Trek, they’re non-persons.

Just as was Victorian proper grammar, politically correct speech patterns are primarily a cognitively expensive signalling behavior.  They have exactly the same import: the speaker is educated, intelligent, and ambitious enough to pay the cognitive price to consciously modify the vernacular.  But as we have seen, PC speech is often yet another case of simplistic human-formalized rules, applied in a context-free way.  They fail on their own terms — implying just the wrong thing, as above — when context shifts.

In other words, simple human-formalized rules applied blindly to something as complex as grammar are brittle, a property they share with bureaucratic rules and AI programs.

Human ethics are similar to human language in their depth and complexity.  They are famously just as difficult to capture in simplistic formalism.  Indeed, given the examples of PC speech, it’s quite arguable that grammar and speech are a proper subset of ethics.  You can’t even reason about whether the Star Trek example is right or wrong without understanding language at a probably better than state of the art level.  And it’s certain that all the subtleties of ontology and epistemology are part of ethics, just as they are of language.

AI is just, in the past decade or so, beginning to get traction in the natural language field beyond the simple human-written formal rules stage.  As for ethics, we’ve just barely gotten into the simple human-written formal rules stage. But if you want a preview of what machine ethics will ultimately look like, study modern natural language processing.

Over the past ten to fifteen years, research in computational linguistics has undergone a dramatic “paradigm shift.” Statistical learning methods that automatically acquire knowledge for language processing from empirical data have largely supplanted systems based on human knowledge engineering. The original success of statistical methods in speech recognition has been particularly influential in motivating the application of similar methods to almost all areas of natural language processing. Statistical methods are now the dominant approaches in syntactic analysis, word sense disambiguation, information extraction, and machine translation.
Nevertheless, there is precious little research in computational linguistics on learning for “deeper” semantic analysis. …

–Raymond Mooney, in a paper at the AAAI 2004 Spring Symposium on Language Learning.

Although we have talked mostly about robotics in terms of how AI has been advancing, it’s instructive to look at developments in the other subfields as well. Natural Language Processing is among the oldest. Turing’s classic paper from 1950 laid out the ability to converse in ordinary, unstructured text as an unequivocal test of the point a machine could be said to think.

Although much can and has been written on the validity of the Turing Test as specified, it is clearly true that a computer with the ability to converse fluently in written and spoken English would be enormously more useful than the computers we have today.  I think it’s also reasonably clear that most people would assume that there was “somebody home,” i.e. begin to impute intelligence and a self to such a computer.

The paradigm shift in NLP has been a result of two things: the increasing willingness (and ability) of AI researchers to use statistics and other numerical methods from the scientist’s toolkit, and the increasing size and avalibility of databases and corpuses and the processing power to subject them to intense analysis.  The amateur AIer today can go on the web and obtain for free enough data and programs to cobble together an NLP system better than anything that existed in 1990.  The processing power in a high-end workstation is good enough for a reasonable amount of research, but a couple more orders of magnitude will help immensely.

What statistical methods have done, in essence, is replace the hand-written grammars that characterized classic-era AI NLP.  These are probabilistic, trained on huge corpuses, and are considerably more robust in use than the old ones.  On the other hand, they don’t reach up into the heights of semantics as well. I’d claim that there’s not a NLP system today that understands its entire vocabulary as well as SHRDLU did its.  The reason is that for its tiny vocabulary, SHRDLU could have a hand-written piece of code for each concept, and thus have a real understanding, in some sense, of the concept.  However, this will change over time.  To begin with, people will simply write code for the most important concepts.  People will come up with schemes to form new code for new concepts from fragments of old code and/or search methods like genetic programming.

Current leading-edge NLP systems (most of them proprietary, AFAIK) are surprisingly good at talking about whatever it is they actually know about, i.e. have a deep semantic model of, as long as you’re literal and prosaic (and expect them to be the same).  I think it’s a toss-up whether automatic programming of semantics makes it to hypohuman border this decade — but AI with hand-coded semantics such as Siri seems likely to be ubiquitous, and competent, by 2020.

In E.E. Smith’s famous Lensman series, the galaxy is the battleground between two races of superintelligent beings, the (good) Arisians and the (evil) Eddorians.  When I listen to people who worry that we are about to create a superintelligence which will take over the world, I get the impression they’ve come from reading “Galactic Patrol” and think that we are on the verge of disastrously creating an Eddorian unless we buckle down quick and figure out how to build a friendly Arisian instead.

In the books, the superintellects had lots of ESP powers but we can dismiss those.  The actual intellectual capability they were imputed to have was the ability to predict.  Prediction is of course the sine qua non of intelligence, but the Arisians were able to predict, e.g., five years ahead of time, that a certain man would be sitting in a barber’s chair and a kitten would jump onto his lap, jostling the barber’s arm and giving him a scratch.  All from the laws of physics and the knowledge of initial conditions.

There are many reasons why this is simply, completely, totally, always forever and truly impossible.

First of all the laws of physics are quantum and have a built-in probabalistic uncertainty. By the same token, it is impossible to know the initial conditions of any substantial portion of the universe to any very high precision: measuring a particle necessarily changes its state in a way do not completely know.

Second, huge parts of the phenomena of interest, a many levels of ontology, are in dynamic systems that are subject to chaotic behavior.  The Butterfly Effect reigns not only in weather, but in markets and politics and epidemiology and computers (one different bit out of a gigabyte can completely change the program’s behavior) and every human mind.

Computers are a particularly hard case of this.  Very basic theorems of computer science tell us that one cannot in general predict what a program will do without actually running it.  This is fine if your superintellect has plenty more processing power than the computer in question, and can emulate it.  But the closer the computer you’re trying to predict comes to having your own processing power, the more likely it will surprise you.

A weird special case of this is that you can’t even predict a universe if you yourself are part of it, because you are a computer with processing power equal to yourself.  (This, BTW, is where our notion of free will comes from: our world models must necessarily exempt our self-models from their general basis in determinism.) You could cheat and force yourself to act in the future according to a list of actions you prepared today, but you wouldn’t be acting all that intelligently; and you wouldn’t be acting with free will, either.

A more obvious case is simply a world with two (well-matched) superintellects, in which at least somewhere they are in competition, maybe even just a friendly game of chess.  In a game between two identical chess computers, each gets to see one ply deeper into the future than the other one did.  Neither can know enough to guess what the other one is going to do for sure.

In a world with lots of superintellects, no one will be able to predict any detail on which they compete.

“It was a true solar-plexus blow, and completely knocked out, Perkins staggered back against the instrument-board. His outflung arm pushed the power-lever out to its last notch, throwing full current through the bar, which was pointed straight up as it had been when they made their landing.”

My current research in AI, such as it is, is an attempt to build a system that’s capable of understanding the above quote. It’s from the middle of a book, and it is much much harder to understand, fully, than you might think. What I intend to do here is to unravel the process by which someone reading the book could be said to understand it. Largely the concern is about what kind of mental structures are being built and what structures must have been built by reading the previous half of the book for the passage to do what it does in the mind of the reader.

Without further ado, let us jump into the quote, which starts at the beginning of a paragraph:

“It was a true solar-plexus blow,”

There are two sources for the comprehension of this clause. First is the preceding paragraph, where a fight is described. A scene and script have been built up, like a movie in the mind. In particular, one man is holding a girl (who is struggling to escape) and another is trying to tie her feet. She kicks the second man, and that is the blow that’s being referred to.

Unlike a cinematic movie, however, much that would be evident on the screen has been left out. The specific positions of the bodies, the clothing in some cases, and many aspects of the background have been left to the imagination. In other words, the “movie” is a sequence of abstractions.

It is in no sense simply a pile of predicates, however. When I read this, I come away with a semi-visual motion script, such as could be used to orchestrate a re-enactment by action-figure dolls, even though the text doesn’t come close to specifying the actual positions or motions I imagine.

The second source is the reader’s memories of pertinent experiences, either of watching fights or having been in them. In the multi-level abstraction structure that’s being built, by and large, at least in the hands of a skillful writer, the things that get mentioned are the things that you’d pay attention to if watching the scene. It’s well established, for example in studies of eyewitness acounts in criminology, that people confabulate what happens between such points in their memory of actual events, much less from verbal stories. So to that extent, the structure of a story reflects that of memory.

If you’ve ever taken a hard blow to the solar plexus, you’ll have a much deeper understanding of this passage than someone who hasn’t. I have, and the sensation is unique; nothing else in my experience feels the same or has the same effects. If you have, note that among the few descriptions of clothing that were provided was that the girl was wearing riding boots.

At a higher level, the scene is part of an attempted abduction of the girl by the men. On this level, the reader is on tenterhooks to discover whether the abduction will succeed, given the girl’s spirited and at least partially efficacious resistance.

“and completely knocked out,”

Syntactically, this is a bit of a garden path; we expect it to be a conjunct of the previous predicate until we see the comma. It turns out to be a participle introducing the second clause. It appears to be for the benefit of those readers who have not experienced solar plexus blows. It describes the effect well enough to follow the action sensibly, but doesn’t really capture the experience.

This points out that there can be different amounts of actual understanding going on in different readers each of whom would claim to have understood the passage: there can be ties to emotions, sensations, memories, and/or mental models in various combinations and strengths.

“Perkins staggered back against the instrument-board.”

Perkins is the second man, and his staggering back is completely predictable from the description of the action so far. In fact, it’s predictable that staggering back is part of the process of his collapsing, which isn’t, and doesn’t need to be, stated explicitly. The ability to predict is one of the key elements of understanding, so we can propose that there is a model of collapsing after being knocked out (or struck in the solar plexus) that abstracts away from any particulars about the individual (or his specific position) that allows the extrapolation of the “movie,” if need be.

It’s the instrument-board that is the new item. In order to understand this, the reader has to pull into play a much broader background structure than heretofore. The action is taking place in the control room of a spaceship, and the “instrument-board” is its control panel. One is reminded of a programming language variable being looked up in a containing context after being found unbound in the local one.

“His outflung arm pushed the power-lever out to its last notch,”

Now we see that there are multiple disconnected levels of abstraction, as well as disconnected items of physical description, that need interpolation. We hear about the arm and the lever, which is local and concrete. We can imagine an arm striking a lever and pushing it. We still don’t know anything about how big the board is, where the lever is on the board, whether the board is horizontal, vertical, or tilted. We don’t know where Perkins is with respect to the pilot’s (and co-pilot’s?) seat(s). On the other hand, we do know much higher-level things about control panels, and power levers (I, for example, call to mind typical airplane cockpits as well as the spaceship control rooms in SF movies I’ve seen.) Although various things about the spaceship have been mentioned before in the book, no good description of the control room has been given; we have to assemble this as some useful level of abstraction as we read this passage.

“throwing full current through the bar,”

This is the key development, not only of the passage, but of the entire book. Note that without the model that the reader will have built up by that point, the phrase means virtually nothing. You might well think that there is a place where drinks are served on the ship.

The book is, as the perspicacious reader will have guessed, E. E. Smith’s 1928 space opera classic Skylark of Space, and its premise is that a bar of copper plated with a (fictional) transuranic metal can convert its mass to pure kinetic energy with the application of a current. The bar is the drive motor of the spaceship. As far as I know, there’s no other context, even in science fiction, where the motor of any vehicle is referred to as a bar. It certainly doesn’t happen in reality, and will not have been the case in any reader’s experience. Many of us have seen bars of copper, and to that extent the symbol is “grounded;” but in the salient sense of the story it is not. Its meaning comes completely from the model that has been built up out of pure words. (In simple terms, no AI with a static, pre-programmed ontology will be able to understand this century-old kids SF adventure story at even the most pedestrian level.)

The really remarkable thing about this phrase is that it throws implications across virtually every level at which the book is to be understood.

At the physical level, it describes the closing of a circuit and the application of voltage to a piece of metal.

At the technological level, using the (fictional) understanding of space drives built up before, that means that the ship will be placed under very heavy acceleration.

Back at the physical bodies level in which we were understanding the fight before, it means that the parties will be thrown to the floor and unable to move.

The fight will be at least suspended. At something closer to the plot level, the parties, pinned to the floor by acceleration, will be unable to stop the ship until the fuel runs out, stranding them far away in space.

This converts them from abductors and victim, where the conflict is all interpersonal, to fellow lifeboat passengers facing a common doom. There is room for various character development as they adjust to the shift in circumstances.

In a larger context, from the outside it will seem as if the abduction succeeded. This will force the girl’s fiance to give chase in his own spaceship (already a hackneyed plot by 1928, to be sure, but thereby all the more predictable for the reader).

But the fact that the ship will fly at top acceleration until its fuel is exhausted implies that the succeeding action will be far-removed from the familiar terrestrial scenes it has taken place in so far. In fact it converts the story from one of personalities and struggle in familiar circumstances (a la the Illiad) to a true voyage of the imagination (like the Odyssey).

If you are want to understand things deeply, you typically want to call in comparable things which can illuminate them by analogy. In this case the arm on the lever is like the tornado in Wizard of Oz (or of course the storm in the Odyssey); it not only throws the protagonists into a strange new world, but motivates their subsequent adventures with the quest to get home.

There is NO knowledge representation and inference scheme in the NLP field today that has even a snowball’s chance of capturing all this. But a human reader with a good grounding in the classics can see that this sentence is the turning point and spark of the whole book on something like five different levels simultaneously.

That’s quite a kick.

“… which was pointed straight up as it had been when they made their landing.”

In something of an anticlimax, Smith is keeping the reader up to date with the physical model of the ship, just in case someone wonders why, having gone down to land, it didn’t keep going down when the juice was turned back on. It had come down backwards, hanging on its thrust. Yet another model — and level of abstraction.

Phaedon Avouris, winner of the Feynman Prize in 1999, is head of the nanoscale science and technology group At IBM, which has recently reported significant advances in synthesizing transistors from graphene using conventional lithography methods.

IBM Demonstrates Graphene Transistor Twice as Fast as Silicon

Graphene transistors promise 100GHz speeds

Graphene Transistors that Can Work at Blistering Speeds

Big Blue demos 100GHz chip

Nanoclast interviews Avouris

and the Science paper,

100-GHz Transistors from Wafer-Scale Epitaxial Graphene

What does this all mean?  Basically, they have overcome a couple of substantial hurdles on the way to a carbon-based electronics, namely the bandgap issue and the ability to fab at wafer scale.  They still have a way to go: they need to bring gate length down by a factor of 10 or so to be in the range of silicon, and probably a few more hurdles and a lot of just plain legwork as well.  But if the research goes through to development, and the development goes through to manufacturing, we’ll have chips that are about two-and-a-half times as fast as the corresponding ones in silicon.

The bottom line, for my money, is that Moore’s Law is safe (in the sense that it will continue to hold true) for another decade at least.  I don’t see this as being a huge spike ahead of Moore’s Law, since graphene has a lot of catch-up to play, but in the long run it probably has more upside potential in speed and size, especially if/when they can get those nanoribbons atomically precise.

The first AI blog was written by a major, highly respected figure in the field. It consisted, as a blog should, of a series of short essays on various subjects relating to the central topic. It appeared in the mid-80s, just as the ARPAnet was transforming over into the internet.

The only little thing I forgot to mention was that it didn’t actually appear in blog form, which of course hadn’t been invented. The WWW didn’t appear until the next decade. It appeared in book form, albeit a somewhat unusual one since it was, as mentioned, a series of short essays, one to a page. It was, of course, Marvin Minsky’s Society of Mind.

Of course, you’re reading a blog about AI right now. The difference is that that was Minsky, and this is merely me. If you haven’t read SOM, put down your computer and go read it now.

Good. You’re back.  Here’s why SoM is relevant to our subject of whether and how soon AI is possible:

It remains a curious fact that the AI community has, for the most part, not pursued Society of Mind-like theories. It is likely that that Minsky’s framework was simply ahead of its time, in the sense that in the 1980s and 1990s, there were few AI researchers who could comfortably conceive of the full scope of issues Minsky discussed—including learning, reasoning, language, perception, action, representation, and so forth. Instead the field has shattered into dozens of subfields populated by researchers with very different goals and who speak very different technical languages. But as the field matures, the population of AI researchers with broad perspectives will surely increase, and we hope that they will choose to revisit the Society of Mind theory with a fresh eye.  (Push Singh — further quotes from the same source)

In other words, here’s a comprehensive theory of what an AI architecture ought to look like that is the summary of the lifework of one of the founders and leaders of the field, and yet no one has seriously tried to implement it.  (When I say serious, I mean put as much effort into it as has gone into, say, Grand Theft Auto.)  (There has been a serious effort to implement the theoretical approach of the CMU wing of classical AI, namely SOAR.)

Part of the reason for this is that SoM is in some sense only half a theory:

Minsky sees the mind as a vast diversity of cognitive processes each specialized to perform some type of function, such as expecting, predicting, repairing, remembering, revising, debugging, acting, comparing, generalizing, exemplifying, analogizing, simplifying, and many other such ‘ways of thinking’. There is nothing especially common or uniform about these functions; each agent can be based on a different type of process with its own distinct kinds of purposes, languages for describing things, ways of representing knowledge, methods for producing inferences, and so forth.
To get a handle on this diversity, Minsky adopts a language that is rather neutral about the internal composition of cognitive processes. He introduces the term ‘agent’ to describe any component of a cognitive process that is simple enough to understand, and the term ‘agency’ to describe societies of such agents that together performs functions more complex than any single agent could.

… but SoM doesn’t have a lot to say about what the individual functions are or how implemented, outside a few examples.  Since AI has for the past few decades concentrated on immediate results, most of the work has been on parts of the problem that could be described as stuff that would be inside a single agent, or at most an agency.

A good example of this happened a few years ago with the winning of the DARPA Grand Challenge and thus the development of the self-driving car. A few months after that happened, I was having a conversation with an AI researcher at a conference.  I maintained that the difference between the results of the first and second races — nobody got more than a mile or so, and then a couple years later several cars finished the whole 130-mile course –  represented real progress.  He pooh-poohed the idea.  All the techniques used in the cars had been previously known and published, he said.  All that had happened was that they had been integrated together into a working system.

I think this attitude goes a long way to explaining the lack of work on SoM and other overall cognitive architecture theories.  But as I reasoned previously:

The difference was, the Wright brothers knew an extra Good Trick, which was how to control the plane in the air once it was flying.
So to develop a working AI, we need the power, which we don’t think is going to be a problem. We need the lift, which is the kind of techniques found in narrow AIs and discussed above. And finally we need the control.

SoM represents a theory of how the control might work.  Where does that leave us?  Can we simply take Minsky’s books and papers and build an AI with all the existing narrow skill programs acting as agents? Hardly.  There’s a lot of work to be done, and probably several new Good Tricks left to be found.

The bottom line, though, is that we are not facing a blank wall.  We are facing a corridor with a sign reading “This way to the egress.”  Indeed we are partway down the corridor already; robotics and self-driving cars have required the development of integrated cognitive architectures along the lines that will probably lead to success.  Note that Brooks’ subsumption architecture had a lot in common with SoM.

So there is at least a case to be made that we are into the home stretch.  Of course that’s where the race really heats up and all the excitement happens…