When will AI produce an intelligent machine?

Scott Fahlman,   March 26, 2008
Categories:  AI    

“When will AI research produce an intelligent machine?” This is a favorite question of journalists and people I meet at parties. And it’s a question that almost all AI experts try to avoid. (I usually say something like, “Let’s see… today is Wednesday?” and then start counting slowly on my fingers: “Thursday… Friday…Saturday…” Usually, by time I get to Sunday or so, the questioner understands that I don’t want to give him a serious answer.)[1]

There are several reasons why this is a difficult question. The first is that predicting the future is hard, and predicting when a scientific breakthrough will occur is very hard indeed.

In 1978, John McCarthy said, “human level AI might require 1.7 Einsteins, 2 Maxwells, 5 Faradays and .3 Manhattan Projects, the project coming after the conceptual breakthroughs”.[2] One can certainly quibble about the exact numbers, but the point is that, in McCarthy’s opinion, we don’t yet have all the key ideas required to produce human-level AI, and that coming up with these new ideas will require a breakthrough – and then some very hard work.

Once the key ideas are in hand, it’s relatively easy to predict how long an engineering effort will take (assuming an adequate, steady level of funding). Going to the moon in the 1960’s was a project of this nature – it required a lot of very creative engineering, but basically the scientific/engineering community knew what had to be done and how to do it. But the arrival of new fundamental ideas is not so predictable: the breakthrough ideas in AI might arrive tomorrow, or 20 years from now, or 100 years from now. Perhaps the key ideas are already present somewhere in the AI community, but we, as a field, haven’t yet recognized them and put them together. Or perhaps the current ideas about AI are all we’re going to get. (I very much doubt this, but it is possible.) If that is the case, it looks like the engineering effort will be a very long and tedious one, even taking into account the recent progress on statistical machine learning.

But there’s a more fundamental reason why the “when” question is difficult: as stated, the question is meaningless. Even if we set aside the rather pointless argument about whether we are willing to call a machine “intelligent” just because it behaves in an intelligent manner[3], there is no agreed-upon threshold above which we would accept a machine (or person or animal) as “intelligent”.

In fact, intelligence is not a scalar.  It is not a single attribute to which we can assign a numerical score; it is a complex bundle of more-or-less independent attributes. A machine or person can score well on some of these attributes and poorly on others. We see this in humans: some have a great talent for mathematics, but are hopeless when it comes to inter-personal skills. Some people have an encyclopedic knowledge of facts, but have barely enough common sense to function in the real world. Some people have musical talent, some have talent for drawing and painting, and some can write well – and many people lack all of these talents, but get along just fine.

So who is more intelligent: Einstein, Shakespeare, Mozart, Confucius, or the guy who can figure out how to repair my furnace? It depends on how you measure it and what task you have in mind. On a cold winter night, with a smell of gas in the house and no ignition, this last guy seems a lot more usefully intelligent than these four world-renowned geniuses.

With this multi-faceted nature of intelligence in mind, let’s return to the original question: “When will AI produce an intelligent machine?” Well, the very first electronic computers were already super-human in one small aspect of intelligence: the ability to do arithmetic. And, as time went on, the machines surpassed us in other aspects of intelligence: the ability to store and regurgitate large amounts of information (as in a database); the ability to solve problems requiring a lot of brute-force search (it turns out that chess is, more or less, one of those problems); the ability to find optimal solutions in simple spaces; the ability to find likely matches in a huge universe of text files; the ability to mine subtle patterns from a huge, noisy dataset; and on and on.

But there are still some very important aspects of intelligence where our machines lag far, far behind typical human performance. I wrote, sometime in the mid-1970’s, that our machines can do many things well, but they cannot begin to match the common sense of a five-year-old child or the sensory-motor performance of a rat. That is still true. With recent progress in robotics – as seen, for example in the recent DARPA Urban Challenge – the rats may be feeling a bit nervous about their position relative to robo-rats, but the five-year-old kids are still safe: as of now, no machine can match them in the ability to acquire, store, and make effective use of vast amounts of “common sense” knowledge; no machine can match them in the flexibility and versatility of their non-optimal but very effective problem solving; and the machines are left in the dust by a child’s ability to handle natural language. On top of that, kids can learn complex new concepts by being told or by observing just one or two real-world examples.

So, as of today, AI is both super-human and sub-human, depending on what aspects of intelligence you are looking at. A lot of AI research is focused on making the super-human parts more super-human. That’s useful, but in my view the challenge of improving the sub-human abilities – and ultimately bringing them up to a human-like level – is far more interesting. To achieve that, I believe that we will need some new breakthrough ideas – and then a lot of work. How long will that take? Let’s see… today is Wednesday?…

[1] The closely related question, “Will AI research ever produce an intelligent machine?”, is, for me, much easier to answer: “Yes!” I’ll say more about that in some future posts.

[2] I have heard many slightly different versions of this McCarthy quote. The version quoted here is from a talk given by Raj Reddy in 2000. The full text of Reddy’s very interesting talk is here.

[3] See Alan Turing’s classic 1950 paper on this: “Computing Machinery and the Mind“