What is Watson? Strong AI and Higher Education

Watson, for those of you who have spend the last week under a rock, is an IBM computer which soundly trounced two long standing human champions in the US quiz show Jeopardy last week. The Watson story is a good hook for me to jump out of mundane Higher Education policy and get back to some bright splangly futurism!

Watson is one of the newest incarnations of a weak AI - an artificial intelligence with limited scope and capacity, below human levels. These are increasingly abundant things. They beat us at chess, decide on our creditworthiness, keep our cars going, or even drive them for us, trade on stock markets and so on. A great many human jobs only needed 'Weak AI' levels of function anyway, and they have simply vanished, or were never created. Our world economy runs on a vast network of invisible switchboard operators, filing clerks and so on, invisible in the machines. There would be billions of them, but for the machines.

Strong AI - Artificial intelligence on a human equivalent level is a different matter. Like Moon holidays and Aircars, Science Fiction promised it to us a half a century ago, and it never came. Moon holidays and Aircars were disbarred by economics and physics - they could be made to work, but never at a useful price. But the same forces, economics and physics, that stole these dreams from us brought Moore's Law. This rule of thumb predicts the doubling of the available processing power, at a given price, every 18 months. That makes strong AI inevitable. You can argue when, but not if.

Strong AI will mean the end of Universities as we know them, but perhaps also their rebirth as we dreamed them. To understand why, we need to unpack the economics of first decade or two of a world with strong AI.

One fine day, in our lifetimes, IBM, or HP, or some tech giant unborn, will unveil a strong AI. It will be able to pass a Turing test, and will do so for our entertainment on Oprah, The Late Late show, or wherever. It will hold it's own at Go, write a technically competent Sonnet and then quickly fade from the news cycle. Kurzweil predicts a date of around 2029, others later (there is a famous bet on it). It's development will have cost it's company around US$100 million in today's money, that being about as big a budget as a high risk project can justify and sustain. Most of that cost will have been payroll, the hardware will only be a fraction of that, perhaps US$10m (the Watson hardware will cost you about US$3m).

Let's assume that strong AI is about equivalent to a new graduate. It will have relative strengths and weaknesses compared to us 'meatbags' of course. It can read the manual quickly, but might not be so good at charming potential clients. But it probably won't sleep, take holidays, lunchbreaks, or gossip by the water cooler either, so in terms of raw hours it should be about 10 times as effective as a human. If we take a graduate salary of say, $30,000, and an initial cost for a strong AI hardware at US$10m, it's not economic. But Moores law will halve the cost of that power every 18 months. So in a decade or so, a strong AI is going to be cost competitive with a graduate hire, with a hardware cost of around US$300,000, equivalent to the first year wages of ten graduate hires that do the same work.

There many, many assumptions here. I haven't factored in software licensing (Open Source Strong AI anyone?), recruitment and training costs. I've not considered overheads for the humans or AI's, or AI downtime (will AI's need to spend 8 hours powered off a day sorting out our memories as we do?).  Nor have I considered out year salaries beyond year 1. It's all order of magnitude guesses, but with exponential growth in available power, an order of magnitude error makes only 5 years difference. I'm dancing past an enormous debate on whether Moore's Law will hold or not, and taking the probable outcome it that it will.  Early in the second decade after you see a strong AI interviewed on the telly, it will be a cheaper alternative to hiring human graduates.

Now, a human graduate takes 4 years to train (on average, assuming a short MSc after a 3 year degree), and another year before that to get college entry exams sorted out. That only leaves five years after graduation to earn back the cost of your University education, if a strong AI exists before you start. Even allowing a few years slack to uptake of AI's, unless you are already in college when you see that strong AI launched on the news, don't bother going. If you planned to do so to help you get a job, it's too late. Even if you get a job,  you won't make your degree investment back in time before you are replaced. At best you'll spend a couple of years as a human buddy to an AI, until the HR AI figures out that your presence is no longer reducing the error rate, and you are gone. They'll hire a human to fire you. There's a sensitivity subroutine. They're nice like that.

You can still go to college, but go to have a good time. Study Fine Art, or Ancient Persian. Whatever interests and stimulates you. Do Social Work, or Teaching - people centred jobs will be the last to go. Chase your dreams. Learn to Paint, or dance. Meet people. Make friends. Study comparative literature, and sociology. Forget about Business, or IT, or Law, or any of the bankable professions of the olden days. You can't compete.

Our Universities long and often stormy relationship with practicality will be at an end. No longer will they need to bow before Mammon, and produce MBA's and degrees in Marketing or computational Finance. They will return to our dream of them, playgrounds of the mind, where we pursue knowledge for the joy of it, for it's own sake, and not for profit.

(The end of our day to day involvement in economic life may, of course, present other difficulties, which remain out of scope for this blog).

Why Predications Fail.

You can't talk about the future of Tertiary education without making predictions. If you are not prepared to put your neck on the block and say what you think is going to happen, it's a pointless exercise. Predicting puts you in harm way - almost all forecasts about the future are wrong.

Before I make too many predictions, It's worth reviewing why predictions fail.

1. Failure to account for economics as a key driver, rather than technology. 

Flying cars are a great example here. People think because a thing can be done, it will. We don't go to work in flying cars because they are impossible - they're not. They're just to expensive, and it's cheaper to go by road.

2. Failure to consider human factors and rates of change. 

A thing can be done a better way, but you have to wait for the old ones who do it the old way to die off first. This is especially important in Universities, where the great old ones live long.The paperless office comes to mind as an example. I've worked in one, it was great. There weren't any old people there who liked to print things off and scribble on them.

3. Predicting out of area of expertise

People who are experts in one field assume expertise in others. Artists imagining space travel is a nice example. There are nice paintings, but I'm not flying in that, thanks. A lot of people fail on technology driven predictions here, often the devil is in the details. Another reason we don't have many paperless offices is that until recently, the screens just weren't good enough, and the software tools for easy annotation weren't either. Generalists sweeping past miss those kinds of details.

4. Failure to account for changes out of area of expertise

This is really the converse of the above reason. There is a tendency to assume that changes you know about will dominate, and changes you do not know about are unimportant. For comparison, consider Ray Kurzwiel Singularity work - focused on technology with George Friedman's 'The Next 100 years' book - focused on geopolitics. Both are fine pieces of work by experts in their field, well argued and, like all predictions, probably wrong. Both focus heavily on developments in the authors own area of expertise, and miss, or err, on key topics outside the field.

5. Wishful thinking. 

Confusing 'We Can' 'We Should' and 'We aught" with what will probably occur. A prediction is not a wish, or a hope, it is a cold, rational analysis of what is likely to occur, whether we like it or not.

6. Predicting the Weather, not the Climate

When people think of predicting, they think of predicting earthquakes, or the stock market, or the weather. You can't predict these in any detail. They are essentially random noise in a pattern. You can predict where earthquakes are likely, that stock markets will exist and be useful, and that there will be weather. Very specific predications often fail not (just) because they are more specific bets on a random future, but because they are attempting to predict things on too fine a scale. Big trends have a mass, an inertia to them that is often the elephant in the room, too big to see. How the big trends collide and play out is important, but as humans we get lost in the human scale details. It's said no one predicted the First World War (and yet, every general staff in Europe had a plan for it for decades). It's true we couldn't predict the details of it, but history tells us that Great Power wars happen a lot, and technology and economics could tell us they would get bigger, and meaner. In the big scale of things, who fought, who won and who lost, the "Battles and Kings" school of history, don't matter so much. What was important to predict was that there would be battles and kings.

Number 5, Wishful thinking, is about the only one I'm confident of avoiding. Point 2 (rates of change) and point 6 (scale of prediction) are always going to be tricky, and the others all rely on having the right spread and depth of expertise - knowing enough about enough things to get the big picture right but not miss sneaky details.

The Singular Future

There's a useful summary of Ray Kurzweil's predictions on Wikipedia. If you haven't heard of him, there's a TED talk where he presents his ideas. Kurzweil is a little over hyped (There's a movie - The Transcendant Man, and a University, in collaboration with Google and NASA), and widely criticised, but that doesn't make him wrong.

Kurzweil's basic idea is that technological change, in some key areas, is exponential, not linear. Moores law, that processing power per dollar doubles every 18 months is an example of this kind of technological rule of thumb that has held good for many years. Exponential processes, in their late stages, tend to get a little strange, and Kurzweils predictions, inferred from that, rapidly get wierd. That's trouble with exponential change. Humans can't intuitively grasp it. Our minds, evolved for counting bananas and holding grudges, tend to be unable to get a grip on it. The pond might be a quarter full of weed that doubles every day, but we still expect to be able to leave clearing it to next week.

Where Kurweil breaks from many other futurists is the prediction that computers will reach a point where they are smart enough to improve their own design. At this point, their development and intelligence will rapidly accelerate and exceed ours, and the chart of scientific development goes off the scale. Anything is possible at that point, and the machines will send us an eMail to tell us about it, if they have remembered to feed us. According to this picture, few, if any, of the institutions we know of would remain relevant, let along Universities.

Right or wrong in the long term, Kurzweil's predications in the nearer term are a useful cribsheet for the kinds of technological changes Universities must weather in the next century. True immersive virtual worlds and Artificial Intelligances smarter than us are not outrageous predictions for the 21st century, and will have serious implications for Universities as we know them. If you are a young academic, by the time you have fought your way up to a professorship, you'll be at the sharp end in dealing with these things in teaching. Just when you thought you were clever for mastering powerpoint animations and signing up to Twitter, it's going to get a whole lot harder and meaner.

The impact of technology on the structure of the University is a huge topic, and I'll return to it in coming posts where I'll be looking at the implications of specific potential technologies for the University in detail.