Posted: 31/10/2017
What is it like to make something more intelligent than yourself? My early experiments showed how AI could work – and answered:
Who would play better – God or the Devil?
The God strategy
In 1979, I was lucky enough to have access to a TI-58 programmable calculator. What should I program it to do? AI obviously.
I cast around for a worthy AI problem. Playing noughts and crosses would be AI, my childhood peers were happy to proclaim.
A few weeks later, my new program played well. With no persistent storage, I had to re-enter the entire program each time I switched off. But it always won if it could, and was unbeatable.
While people were quite surprised the program could actually play, the effect was slightly diluted by the fact that five year olds could also play this game.
And then the verdict: it was not intelligent after all because it was “merely playing moves I had taught it”. Intelligence was apparently not what the decision was, but how you came to that decision.
Learning to play like God
Octopawn is played on a 4×4 chess board with eight pawns – an upgrade to Hexapawn. Too complex to go through all the board permutations, my friends declared that, unlike my previous game, winning at this would indeed be true AI.
So with O levels behind me, my Casio FX-501p started out with just the rules. It played kids at school, never making the same mistake twice. By afternoon it had worked out how to force a win as white. Friends who beat it early on were gleeful when it later slaughtered their rivals.
This program had learned a strategy that even I did not know. And it played like God – always making the best move. AI at last.
Moving goalposts sounded in the background. Despite its ability to learn its own strategy, as a deterministic game which could be crunched through on a fast computer without AI, octopawn was not even an AI class problem.
The Devil strategy
By 1983, with A levels behind me, I was determined to create an AI before I left school . With my new FX-602P, I had to move so far that even moving goalposts couldn’t reach me. My program would have to out-think my friends on human territory.
In my undercut variant, two players simultaneously choose a number from {1,2,3,4,5,6}. The person with the highest number wins that number of points – except in the case where the other player undercuts them by one. In that case the player with the smaller number wins the sum of the two numbers. The first to 200 wins.
As both players can choose any number, you can’t just go through all the combinations. To win, you actually have to out-think your opponent.
Solving this game mathematically gives you the God strategy – an optimal strategy which cannot be beaten. But accidently I created a Devil strategy, which beats people much more convincingly than this “optimal” solution. My program would out-think opponents by learning their strategies, and then subverting them.
After about 100 points, the calculator would suddenly understand, and charge decisively through to victory. Despite the agonising thought processes and discussions they went to, it beats both individuals, and groups of friends working together.
I knew how it thought – but not what it was thinking. The Devil strategy exploited weaknesses no one even knew they had – and still don’t understand.
Finally, my friends agreed, this was AI.
I wish I had known how far the world would come now when:
- Even a 35 year old calculator can out-think people
- Computer hardware grows in power continuously
- Computers can learn and can handle uncertainty
- Understanding beats brute force to intelligence
- Play a Devil strategy AI only if you like being slaughtered.
- Making something which out-thinks people is scarily exhilarating.
Are you ready to play the Devil?
Stephen B Streater
Founder and Director of R&D
Blackbird is best-of-breed
Jon Hanford - Group CTO, Deltatre