Automation vs Augmentation

The best chess player in the world is Magnus Carlsen. He's been the world chamption since 2013. We have him play against a 2018 macbook pro running a good chess engine, like Stockfish. Who wins?

The computer, no contest. Off-the-shelf computers have outclassed humans for over a decade.

So let's give Carlsen a small handicap. We allow him to always undo his previous move. If he makes moves A, B, C, he can undo C and make move C', then undo C' and make move C'', but he can't undo move B. How does this change things?

At first, it wouldn't change much. Carlsen can fix obvious blunders. If he realizes after making a move that there was a better move, he can go back and do that instead. But that's not a huge advantage. He'd do a little better, but not a lot better.

But he'd quickly realize something. Normally, when he's thinking ten moves deep, he needs to predict how the computer will react to his first move. If the computer does something he didn't expect, that's a problem. But he no longer needs to do that: if he has three canditate moves, he can just make and undo all three moves and see how the computer reacts to each. Then he can settle on one based on that extra information.

If that's how he'll play, why not make that part of the UI? He selects his top N moves, and we show all N "next-boards", based on how the computer will react to each. Then he picks which of the N branches he likes and we jump directly to that state. To let him properly use his small edge, we've fundamentally changed what the game "is". He's not playing chess in the normal way, he's not presented with the normal UI, and he's not using his normal skills. He's playing a higher-level abstraction of chess.

But he's still worse than the computer. So let's give him unlimited undo. He can now see how the computer reacts six steps in, and then go back and try another six steps. But we quickly hit a new problem: his memory. A computer remembers everything, a human does not. But forgetting things doesn't make the human overall worse, as he can just undo back to a known point and replay the moves on the parallel timeline. At some point he pulls out a notebook and starts writing things down.

Let's improve the UI one more time. We now show the entire history of the game as a directed graph. If he makes moves A, B, C, then undoes C and tries C', then undoes C and B and does B', he'd see

A - B - C
      - C'
  - B'

But once he does B', what if he decides to go with C' after all? He could undo B' then do B, C'. But that's a deterministic series of steps, and he knows exactly where he's going. The UI should let him jump to any game state anywhere in the graph and continue playing from there.

Carlsen is now playing a fundamentally different game. The computer is playing a game of chess, but Carlsen is playing across games of chess, freely moving between all of the parallel timelines. But at the same time, it's still "just" a game of chess! The only difference is that Carlsen can undo moves and the computer can't. But we can present that functionality as a complete reframing of what means to play chess. It's now his game to lose.

Okay, so what's the point of all that? Plenty of chess games let you infinitely undo, and yet the computer still always wins. Because having that extra action doesn't make you dramatically better. But that action, combined with a reframing of how that action affects the game, and a better UI that lets you take advantage of that reframing, will radically change the game. You are now an augmented player. You need different skills to be a good augmented player than you do to be a good regular player, but your playing abilities are way, way higher.

In practice, you don't even need that infinite undo. Once we allow augmentation with the same rules of chess, we see the same seismic shift. In correspondence chess, players are allowed to use whatever they want to decide moves: other players, chess engines, whatever. You need different skills to be a good correspondence player. But once you have them, the game radically changes. In "over-the-board" chess, computers totally outclass regular humans. In correspondence chess, augmented humans totally outclass computers.

Whenever we talk about computers doing better than us, I wonder: how much of that is our skill with automation, and how much is our lack of skill with augmentation? We don't design games for augmented humans, we don't design our workflows around augmented humans, we don't really know how to economically produce augmentations. Maybe in that context, automations can beat unaugmented humans. What happens when starting playing games and performing tasks at a higher level?

I dunno, but it surprises and frustrates me that nobody else is thinking about it in this way. Then again, it's hard for me to think about it in this way, like I'm hitting walls on what I can conceive. That might be why it's neglected. Automation is easy, augmentation is hard, so we'll automate, even if the benefits of augmenting are so much greater.

If you're reading this on the web, you can subscribe here. Updates are once a week. My main website is here.

My new book, Logic for Programmers, is now in early access! Get it here.