Sometimes what actually works is very different from what experts think will work. Richard Sutton, an AI researcher, wrote an article called The Bitter Lesson explaining how this shows up in AI research. Researchers try to make progress in two different ways. One group of researchers try to figure out how the human mind solves a problem, formulate it into rules, and get the computer to understand that. The other group gives the computer a lot of data and hopes that the computer makes sense of it by identifying which patterns show up more frequently. The first method is sophisticated, and the second one is crude.

But the Bitter Lesson that Sutton talks about is this: The crude methods with increased compute will always beat the sophisticated method with limited compute. That is, what actually matters is how much RAM, memory, processing power, etc. you can throw at the computer. The brilliance of your solution doesn’t matter. Brute force does.

The methods that I call “crude” here aren’t trivial. They are mathematical algorithms with complex logic that take some work to implement. But beyond a certain point, there is no “genius insight” that is going to illuminate a certain truth about how our brains work, and there is no completely new paradigm that is continually invented in this class of solutions. It’s a simple set of rules, and as long as you give it more compute, it gives you better results. This is commonly known as the scaling hypothesis.

What’s the consequence of this? As Sutton says:

over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation.

This means that the compute available to the project goes up even as the project is underway. If there are two researchers who started at the same time, the one who just implements the simple algorithm and concentrates on getting more data and more compute will beat the one who spends all his time trying to come up with a more beautiful algorithm. Abraham Lincoln’s motto of: If I have six hours to chop a tree, I’ll spend the first four sharpening my axe will actually fail in this case. The right thing to do is start chopping, hoping that a better axe will be delivered to you halfway through.

In other words, you’re hoping for the cavalry to arrive before you get to the end of the project. This is hope as a strategy, which is something generally frowned upon in the world. But it seems to pay off when it comes to pushing frontiers. This would work even in something like entrepreneurship. A founder who does the simple, stupid thing that will burn his runway because he knows with reasonable certainty (or hope) that funds will arrive at the end of the window will actually be in a position to make the best use of the funds when they do arrive. A founder who tries to optimize the funds he has, playing it safe, will get to the end of the runway with more money, but not have the momentum or resources to actually use the fresh funds he gets.

This is a gross and ugly oversimplification, and I’m sure this isn’t what Sutton meant at all, but I can’t help thinking that this boils down to: The writer who puts out more drafts, mechanically, bull-headedly, repeating the mundane process of sitting down and writing away will get further ahead compared to the writer who is procrastinating, studying, and stewing the best first draft in his head that he’ll get to on the last day.

You could call this idea “The unreasonable effectiveness of brute force.”

On a deeper level, there’s a trade-off you’re making. You’re sacrificing interpretability for competence. The guy who wants to understand the logic of every step of his process is trying to form a model in his head that he can explain to other people. The guy who is just “doing the thing” and course-correcting on the other hand is too involved in the process to document his steps and make them explainable. The explainer will always lose to the doer, when it comes to actually doing the thing.

Most competent people get results, but not all of them can explain to you how they get their results. The more they focus on dissecting and optimizing their process, the more time they’re taking away from repeating the simple, stupid steps, time they could be donating to the unreasonable effectiveness of brute force. Warren Buffett is a great investor, but he cannot distill his competence into a neat formula for you. Stephen King is a terrific writer, but he cannot create a model of what makes his writing tick. If you read every one of his books and come up with a model of why his books work, you will lose to the guy who has spent all that time relentlessly trying to imitate King’s style by writing the same story over and over.

I think it was Nassim Taleb who said that even if Warren Buffett lived the same life all over again, he would not be able to replicate the results he obtained. Maybe this isn’t beecause the “method” doesn’t work, but rather because he doesn’t “know” his own method in the way a philosopher would neatly tie things up in a thesis. Buffett’s genius is a set of simple principles applied with malleability to real situations. He cares only partially about interpretability, but he cares completely about competence and real results. There are a lot of things about his implicit knowledge that he cannot put into words, and that he won’t care to put into words. The same thing holds true for genius writers, or artists of any kind, probably. The more time you spend on one, the less you have for the other.

This has a real implication for you. It means that the games in which you can become really good at are the ones where you can do simple things over and over, get feedback, and course correct. Not the ones where you sit in a room figuring out your master thesis that will solve everything in one stroke.

Are you going to spend your life picking a game where you optimize for the genius move, or design a game where you iterate, see feedback, and improve?