Something amazing happened in July 2019. Pluribus won a poker game by beating ten other top ranked professionals over 48 hours of intense competition. Not beating, Thrashing. Pluribus is a computer.
So What? Computers have been taking humans to the cleaners in our most cherished disciplines for years. Last century Deep Blue beat Gary Kasparov, then the greatest chess player of all time. As a chess nerd from way back some of the romance of my favourite game died for me on May 11, 1997. That’s maybe how tearful Go Master Li Sedol, 18 times world champion, felt after being trounced by Google’s Alpha-Go in 2016. Sedol, an all time great, confidently predicted he’d win by a “landslide”. He lost 4-1. If its true that “Most games are fun because they are microcosms of some aspect of life” as one of Alpha-Go’s programmers says, then this was another aspect of humanity’s supremacy handed over to AI. In 2017 a programme called LIbratus beat poker expert Dong Kim in an individual match lasting 3 weeks. But Pluribus raised the stakes by taking on a group of top players.
What is so special about Poker? In Chess and Go there is “perfect information”- all the relevant information (rules of the game and configuration of the board) is always accessible and there is no luck or randomness. Not so in poker. Pluribus was playing a group of expert humans each bringing their own mix of highly honed rationality, non-rationality, values, gut feelings, emotion playing style and randomness to strategies developed over a unique playing history. Some of the information Pluribus needed was laid out in probabilities regarding it’s own hand and cards on the table but most was in the hands, the heads, the hormones even, of the other humans involved. All Pluribus could see was its own cards,, the other players’ bets and the hands they laid down. Unlike chess or go there is also a decent dollop of pure, uncontrollable chance in poker. Surely all that puts it safely in that inner sanctum of human life that is inaccessible to silicon?
No! Nothing is sacrosanct.
The sparse information Pluribus got was enough. It won so easily one opponent said “It was as if it knew my cards”.How? Deep Learning. Essentially the programmers just gave Pluribus the rules of poker and had it play first human players then itself millions of times, learning from experience, using its own strategies against itself, modifying them and playing itself and other humans again. The result? Just like in Chess and Go AI plays standard strategies really really well, and also comes up with moves humans do not understand.
All very interesting, not to say chastening, for us mere humans, but why AI and poker on this blog? Applications to other fields were immediately seen In Pluribus’ triumph : the stock market, medicine, international relations, business development- anywhere where sequences of interacting rational and irrational human strategies create complexity as they play out together.. But wait! That is a perfect description of the development spaces I work in. Might development decisions soon be made, and made much better, by AI? A romantic human side of me hopes not- let my cherished workspace be sacrosanct- but a realistic side of me asks “Why not?” If AI can make better development decisions let’s use them. Humans have made some good but many bad development decisions for all of our history. At this critical point in our development trajectory we need all the help we can get.
I already use electronic tools:- Excel spreadsheets, GIS overlays, qualitative research software as work tools but AI running development may still be a few years off. But there is a lesson for today too. “This was invaluable” said Google’s programmers after Lee Seedol’s brilliant win in his fourth game against Alpha-go. After he lost the fifth game Sedol humbly said “It learns from its mistakes” . We should learn now from how computers become so good in complex spaces.“
Deep learning embodies Mark Twain’s aphorism: “Where does good judgement come from? It comes from expereince. Where does experience come from? It comes from bad judgement.” . That’s how Mother Nature (aka evolution, God) learned to make hummingbirds and blue whales, butterflies and flowers., its also how every human learned their negotiating skills from long before the school playground right up to the UN security council. Trial, error, learning and iteration is not rocket science- its much more powerful.
Complexity thinker David Snowden suggests this as an approach to complexity: try things, monitor what happens, amplify experiments that work and damp down ones that don’t. This inductive approach is very different to standard planning where one forms a theory of change and deductively designs activities around them. In a recent evaluation report I excitedly noted a totally unexpected observation that the NGO was possibly altering the way Cambodian families relate to each other. I suggested deliberately exploring this and putting more resources into this area. Their monitoring officer, deeply imbued with (stuck in?) standard development, rejected this as invalid because I had no theory of change for how this might have happened. I said that that was the point. Empirically observed change where we have no theory of change is precisely where we learn and innovate. For hard to solve issues we have to not only do standard development well but also develop surprise moves. Being locked in theories of change stops us doing that. In something as complex as child rights in Cambodia we should approach our workspace with eyes, and minds, open. I reckon complexity aware development comes from deep sensitivity to context, deep innovation and deep learning. Not deep theory.
Ideas, ideas ideas: AI driven development in the future? Building deep learning into our development practise? Poker as an analogy for life? Doing development without a “theory of change” which is de rigeur for today’s funders and planners? Not chess, not poker and now possibly not development- what is humanity’s last bastion against computers?
Thoughts? Throw in a comment….