dw2

Not everyone shares my view that AI is going to become a more and more important field during the coming decade.

I’ve received a wide mix of feedback in response to:

• my general description of the 2010’s as the decade of nanotechnology and AI,

• my mobile-specific prediction (part of the m2020 project) that the 2010’s will see Mobiles manifesting AI – fulfilling, at last, the vision of “personal digital assistants”,

• and my comments made in other discussion forums about the growth of AI.

Below, I list some of the questions people have raised – along with my answers.

Note: my answers below are informed by (among other sources) the 2007 book “Beyond AI: creating the conscience of the machine“, by J Storrs Hall, that I’ve just finished reading.

Q1: Doesn’t significant progress with AI presuppose the indefinite continuation of Moore’s Law, which is suspect?

There are three parts to my answer.

First, Moore’s Law for exponential improvements in individual hardware capability seems likely to hold for at least another five years, and there are many ideas for new semiconductor innovations that would extend the trend considerably further.  There’s a good graph of improvements in supercomputer power stretching back to 1960 on Shane Legg’s website, along with associated discussion.

Dylan McGrath, writing in EE Times in June 2009, reported views from iSuppli Corp that “Equipment cost [will] hinder Moore’s Law in 2014“:

Moore’s Law will cease to drive semiconductor manufacturing after 2014, when the high cost of chip manufacturing equipment will make it economically unfeasible to do volume production of devices with feature sizes smaller than 18nm, according to iSuppli Corp.

While further advances in shrinking process geometries can be achieved after the 20- to 18nm nodes, the rising cost of chip making equipment will relegate Moore’s Law to the laboratory and alter the fundamental economics of the semiconductor industry, iSuppli predicted.

“At those nodes, the industry will start getting to the point where semiconductor manufacturing tools are too expensive to depreciate with volume production, i.e., their costs will be so high, that the value of their lifetime productivity can never justify it,” said Len Jelinek, director and chief semiconductor manufacturing iSuppli, in a statement.

In other words, it remains technological possible that semiconductors can become exponentially denser even after 2014, but it is unclear that sufficient economic incentives will exist for these additional improvements.

As The Register reported the same story:

Basically, just because chip makers can keep adding cores, it doesn’t mean that the application software and the end user workloads that run on this iron will be able to take advantage of these cores (and their varied counts of processor threads) because of the difficulty of parallelising software.

iSuppli is not talking about these problems, at least not today. But what the analysts at the chip watcher are pondering is the cost of each successive chip-making technology and the desire of chip makers not to go broke just to prove Moore’s Law right.

“The usable limit for semiconductor process technology will be reached when chip process geometries shrink to be smaller than 20 nanometers (nm), to 18nm nodes,” explains Len Jelinek…

At that point, says Jelinek, Moore’s Law becomes academic, and chip makers are going to extend the time they keep their process technologies in the field so they can recoup their substantial investments in process research and semiconductor manufacturing equipment.

However, other analysts took a dim view of this pessimistic forecast, and maintain that Moore’s Law will be longer lived.  For example, In-Stat’s chief technology strategist, Jim McGregor, offered the following rebuttal:

…every new technology goes over some road-bumps, especially involving start-up costs, but these tend to drop rapidly once moved into regular production. “EUV [extreme ultraviolet] will likely be the next significant technology to go through this cycle,” McGregor told us.

McGregor did concede that the lifecycle of certain technologies is being extended by firms who are in some cases choosing not to migrate to every new process node, but he maintained new process tech is still the key driver of small design geometries, including memory density, logic density, power consumption, etc.

“Moore’s Law also improves the cost per device and per wafer,” added McGregor, who also noted that “the industry has and will continue to go through changes because of some of the cost issues.” These include the formation of process development alliances, like IBM’s alliances, the transition to foundry manufacturing, and design for manufacturing techniques like computational lithography.

“Many people have predicted the end of Moore’s Law and they have all been wrong,” sighed McGregor. The same apparently goes for those foolhardy enough to attempt to predict changes in the dynamics of the semiconductor industry.

“There have always been challenges to the semiconductor technology roadmap, but for every obstacle, the industry has developed a solution and that will continue as long as we are talking about the hundreds of billion of dollars in revenue that are generated every year,” he concluded.

In other words, it is likely that, given sufficient economic motivation, individual hardware performance will continue improving, at a significant rate (if, perhaps, not exponentially) throughout the coming decade.

Second, it remains an open question as to how much hardware would be needed, to host an Artificial (Machine) Intelligence (“AI”) that has either human-level or hyperhuman reasoning power.

Marvin Minsky, one of the doyens of AI research, has been quoted as believing that computers commonly available in universities and industry already have sufficient power to manifest human-level AI – if only we could work out how to program them in the right way.

J. Storr Hall provides an explanation:

Let me, somewhat presumptuously, attempt to explain Minsky’s intuition by an analogy: a bird is our natural example of the possibility of heavier-than-air flight. Birds are immensely complex: muscles, bones, feathers, nervous systems. But we can build working airplanes with tremendously fewer moving parts. Similarly, the brain can be greatly simplified, still leaving an engine capable of general conscious thought.

Personally, I’m a big fan of the view that the right algorithm can make a tremendous difference to a computational task.  As I noted in a 2008 blog post:

Arguably the biggest unknown in the technology involved in superhuman intelligence is software. Merely improving the hardware doesn’t necessarily mean the the software performance increases to match. As has been remarked, “software gets slower, more rapidly than hardware gets faster”. (This is sometimes called “Wirth’s Law”.) If your algorithms scale badly, fixing the hardware will just delay the point where your algorithms fail.

So it’s not just the hardware that matters – it’s how that hardware is organised. After all, the brains of Neanderthals were larger than those of humans, but are thought to have been wired up differently to ours. Brain size itself doesn’t necessarily imply intelligence.

But just because software is an unknown, it doesn’t mean that hardware-driven predictions of the onset of the singularity are bound to be over-optimistic. It’s also possible they could be over-pessimistic. It’s even possible that, with the right breakthroughs in software, superhuman intelligence could be supported by present-day hardware. AI researcher Eliezer Yudkowsky of the Singularity Institute reports the result of an interesting calculation made by Geordie Rose, the CTO of D-Wave Systems, concerning software versus hardware progress:

“Suppose you want to factor a 75-digit number. Would you rather have a 2007 supercomputer, IBM’s Blue Gene/L, running an algorithm from 1977, or an 1977 computer, the Apple II, running a 2007 algorithm? Geordie Rose calculated that Blue Gene/L with 1977’s algorithm would take ten years, and an Apple II with 2007’s algorithm would take three years…

“[For exploring new AI breakthroughs] I will say that on anything except a very easy AI problem, I would much rather have modern theory and an Apple II than a 1970’s theory and a Blue Gene.”

Here’s a related example.  When we think of powerful chess-playing computers, we sometimes think that massive hardware resources will be required, such as a supercomputer provides.  However, as long ago as 1985, Psion, the UK-based company I used to work for (though not at that time), produced a piece of software that played what many people thought, at the time (and subsequently) to be a very impressive quality of chess.  See here for some discussion and some reviews.  Taking things even further, this article from 1983 describes an implementation of chess, for the Sinclair ZX-81, in only 672 bytes – which is hard to believe!  (Thanks to Mark Jacobs for this link.)

Third, building on this point, progress in AI can be described as a combination of multiple factors:

1. Individual hardware power
2. Compound hardware power (when many different computers are linked together, as on a network)
3. Software algorithms
4. Number of developers and researchers who are applying themselves to the problem
5. The ability to take advantage of previous results (“to stand on the shoulders of giants”).

Even if the pace slows for improvements in the hardware of individual computers, it’s still very feasible for improvements in AI to take place, on account of the other factors.

Q2: Hasn’t rapid progress with AI often been foretold before, but with disappointing outcomes each time?

It’s true that some of the initial forecasts of the early AI research community, from the 1950’s, have turned out to be significantly over-optimistic.

For example, in his famous 1950 paper “Computing machinery and intelligence” – which set out the idea of the test later known as the “Turing test” – Alan Turing made the following prediction:

I believe that in about fifty years’ time it will be possible, to programme computers… to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification [between a computer answering, or a human answering] after five minutes of questioning.

Since the publication of that paper, some sixty years have now passed, and computers are still far from being able to consistently provide an interface comparable (in richness, subtlety, and common sense) to that of a human.

For a markedly more optimistic prediction, consider the proposal for the 1956 Dartmouth Summer Research Conference on Artificial Intelligence which is now seen, in retrospect, as the the seminal event for AI as a field.  Attendees at the conference included Marvin Minsky, John McCarthy, Ray Solomonoff, and Claude Shannon.  The group came together with the following vision:

We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.

The question for us today is: what reason is there to expect rapid progress with AI in (say) the next ten years, given that similar expectations in the past failed – and, indeed, the whole field eventually fell into what is known as an “AI winter“?

J Storrs Hall has some good answers to this question.  They include the following:

First, AI researchers in the 1950’s and 60’s laboured under a grossly over-simplified view of the complexity of the human mind.  This can be seen, for example, from another quote from Turing’s 1950 paper:

Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s? If this were then subjected to an appropriate course of education one would obtain the adult brain. Presumably the child brain is something like a notebook as one buys it from the stationer’s. Rather little mechanism, and lots of blank sheets. (Mechanism and writing are from our point of view almost synonymous.) Our hope is that there is so little mechanism in the child brain that something like it can be easily programmed.

Progress in brain sciences in the intervening years has highlighted very significant innate structure in the child brain.  A child brain is far from being a blank notebook.

Second, early researchers were swept along on a wave of optimism from some apparent early successes.  For example, consider the “ELIZA” application that mimicked the responses of a certain school of psychotherapist, by following a series of simple pattern-matching rules.  Lay people who interacted with this program frequently reported positive experiences, and assumed that the computer really was understanding their issues.  Although the AI researchers knew better, at least some of them may have believed that this effect showed that more significant results were just around the corner.

Third, the willingness of funding authorities to continue supporting general AI research became stretched, due to the delays in producing stronger results, and due to other options for how that research funds should be allocated.  For example, the Lighthill Report (produced in the UK in 1973 by Professor James Lighthill – whose lectures in Applied Mathematics at Cambridge I enjoyed many years later) gave a damning assessment:

The report criticized the utter failure of AI to achieve its “grandiose objectives.” It concluded that nothing being done in AI couldn’t be done in other sciences. It specifically mentioned the problem of “combinatorial explosion” or “intractability”, which implied that many of AI’s most successful algorithms would grind to a halt on real world problems and were only suitable for solving “toy” versions…

The report led to the dismantling of AI research in Britain. AI research continued in only a few top universities (Edinburgh, Essex and Sussex). This “created a bow-wave effect that led to funding cuts across Europe”

There were similar changes in funding climate in the US, with changes of opinion within DARPA.

Shortly afterwards, the growth of the PC and general IT market provided attractive alternative career targets for many of the bright researchers who might previously have considered devoting themselves to AI research.

To summarise, the field suffered an understandable backlash against its over-inflated early optimism and exaggerated hype.

Nevertheless, there are grounds for believing that considerable progress has taken place over the years.  The middle chapters of the book by J Storrs Hall provides the evidence.  The Wikipedia article on “AI winter” covers (much more briefly) some of the same material:

In the late ’90s and early 21st century, AI technology became widely used as elements of larger systems, but the field is rarely credited for these successes. Nick Bostrom explains “A lot of cutting edge AI has filtered into general applications, often without being called AI because once something becomes useful enough and common enough it’s not labeled AI anymore.” Rodney Brooks adds “there’s this stupid myth out there that AI has failed, but AI is around you every second of the day.”

Technologies developed by AI researchers have achieved commercial success in a number of domains, such as machine translation, data mining, industrial robotics, logistics, speech recognition, banking software, medical diagnosis and Google’s search engine…

Many of these domains represent aspects of “narrow” AI rather than “General” AI (sometime called “AGI”).  However, they can all contribute to overall progress, with results in one field being available for use and recombination in other fields.  That’s an example of point 5 in my previous list of the different factors affecting progress in AI:

1. Individual hardware power
2. Compound hardware power (when many different computers are linked together, as on a network)
3. Software algorithms
4. Number of developers and researchers who are applying themselves to the problem
5. The ability to take advantage of previous results (“to stand on the shoulders of giants”).

On that note, let’s turn to the fourth factor in that list.

Q3: Isn’t AI now seen as a relatively uninteresting field, with few incentives for people to enter it?

The question is: what’s going to cause bright researchers to devote sufficient time and energy to progressing AI – given that there are so many other interesting and rewarding fields of study?

Part of the answer is to point out that the potential number of people working in this field is, today, larger than ever before – simply due to the rapid increase in the number of IT-literate graduates around the world.  Globally, there are greater numbers of science and engineering graduates from universities (including China and India) than ever before.

Second, here are some particular pressing challenges and commercial opportunities, which make it likely that further research will actually take place on AI:

• The “arms race” between spam detection systems (the parts of forms that essentially say, “prove you are a human, not a bot”) and ever-cleverer spam detection evasive systems;
• The need for games to provide ever more realistic “AI” features for the virtual characters in these games (games players and games writers unabashedly talk about the “AI” elements in these games);
• The opportunity for social networking sites to provide increasingly realistic virtual companions for users to interact with (including immersive social networking sites like “Second Life”);
• The constant need to improve the user experience of interacting with complex software; arguably the complex UI is the single biggest problem area, today, facing many mobile applications;
• The constant need to improve the interface to large search databases, so that users can more quickly find material.

Since there is big money to be made from progressing solutions in each of these areas, we can assume that companies will be making some significant investments in the associated technology.

There’s also the prospect of a “tipping point” once some initial results demonstrate the breakthrough nature of some aspects of this field.  As J Storrs Hall puts it (in the “When” chapter of his book):

Once a baby [artificial] brain does advance far enough that it has clearly surpassed the bootstrap fallacy point… it might affect AI like the Wright brothers’ [1908] Paris demonstrations of their flying machines did a century ago.  After ignoring their successful first flight for years, the scientific community finally acknowleged it.  Aviation went from a screwball hobby to the rage of the age and kept that cachet for decades.  In particular, the amount of development took off enormously.  If we can expect a faint echo of that from AI, the early, primitive general learning systems will focus research considerably and will attract a lot of new resources.

Not only are there greater numbers of people potentially working on AI now, than ever before; they each have much more powerful hardware resources available to them.  Experiments with novel algorithms that previously would have tied up expensive and scarce supercomputers can nowadays be done on inexpensive hardware that is widely available.  (And once interesting results are demonstrated on low-powered hardware, there will be increased priority of access for variants of these same ideas to be run on today’s supercomputers.)

What’s more, the feedback mechanisms of general internet connectivity (sharing of results and ideas) and open source computing (sharing of algorithms and other source code) mean that each such researcher can draw upon greater resources than before, and participate in stronger collaborative projects.  For example, people can choose to participate in the “OpenCog” open source AI project.]

Appendix: Further comments on the book “Beyond AI”

As well as making a case that progress in AI has been significant, another of the main theme of J Storrs Hall’s book “Beyond AI: Creating the conscience of the machine” is the question of whether hyperhuman AIs would be more moral than humans as well as more intelligent.

The conclusion of his argument is, yes, these new brains will probably have a higher quality of ethical behaviour than humans have generally exhibited.  The final third of his book covers that topic, in a generally convincing way: he has a compelling analysis of topics such as free-will, self-awareness, conscious introspection, and the role of ethical frameworks to avoid destructive aspects of free-riders.  However, critically, it all depends on how these great brains are set up with regard to core purpose, and there are no easy answers.

Roko Mijic will be addressing this same topic in the UKH+ meeting “The Friendly AI Problem: how can we ensure that superintelligent AI doesn’t terminate us?” that it being held on Saturday 23rd January.  (If you use Facebook, you can RSVP here to indicate whether you’re coming.  NB it’s entirely optional to RSVP.)