Humans Fought Back

If you’re even vaguely interested in AI

You probably know how big a deal it was when Deep Blue beat world champion chess player Garry Kasparov in 1997.

Even before that, Deep Thought beat grand master Bent Larsen in 1988.

Then AlphaZero in 2017 was the first to learn chess from self-play alone

But there’s an interesting story of when a human beat AI at a game despite AI being better than humans at that game.

Go is a 2,500 year old board game that is simpler than chess to learn and much harder to play well. KataGo is an open-source Go AI that is superhuman and beats every human on even terms.

So when Shin Jin-seo beat it KataGo? It caused quite the uproar.

But what was interesting was how Shin changed strategy across three games to ensure victory.

He was given a handicap - a headstart to help his chances.

Despite that, in game one he fought KataGo head on and lost by resignation.

In games 2 and 3, he won by simplifying completely, picking one or two fights (or no fight at all!) Using not just what he knew about Go, but what he knew about his competitor.

Because to Shin’s mind, the perfection of the AI competitor was also a weakness - it shied away from lines that were difficult and maximised on probability of win, not the size of a win.

He knew he didn’t have to outplay it, but just stop the game becoming the kind of game where being outplayed is possible. KataGo had chances to beat him, but its risk-aversion meant it ignored the kind of chaos needed to overturn the deficit.

He won by modelling the engine’s habits, whereas the engine knew nothing about Shin specifically as a human.

When we think about coaching - that’s the advantage.

Yes apps can do a lot of what coaches can - they can give a board, data, information on what someone is doing

It can’t look at the person and read the person behind it. Their history, their behaviours, what changes when they’re put under stress etc.

I think this can help us understand our work better, and our place in work.

AI has incredible specialisation in closed-domains - And we have jobs that need deep, verifiable, closed-domain knowledge. Some of that we can do, a lot of it we will use AI.

But we also want the Shin Jin-seo mode - the ability to openly read people in an unformalisable way.

Most AI misuses come from using it for things we should do, and using us for things AI should do.