Old wiring → new wiring
Old wiring: It was hard for me, so it should stay hard. New wiring: Keep the struggle that builds your judgment. Stop paying for the struggle that only builds your story.
The story
It’s 2018, and I’m in Barcelona at two in the morning.
Each episode was a piece of philosophical fiction, about ten thousand words, and I wrote, produced and voiced them myself. Six complete rewrites is roughly where “dedicated artist” and “undiagnosed obsessive compulsive” start to blur. My writing partner was asleep in another country and probably irritated, because I’d moved the ending again. I had a recording studio booked for eight in the morning before we left Spain, so it was now or never.
I made the studio. And when I finally closed the laptop and turned off the light, I had a thought that took me about eight years to answer. What if this is the wrong way to do it?
Under the hood: the Bitter Lesson
On March 13, 2019, a computer scientist named Rich Sutton posted a short essay called “The Bitter Lesson.” He’s one of the founders of the field that taught machines to learn by trial and error, and the essay runs about a page. People who build these systems still quote it at each other.
His claim covers seventy years of AI research. Over and over, researchers tried to make machines smart by building in what they themselves knew: how a grandmaster thinks about chess, how a linguist thinks about grammar. It always helped at first. Then a blunter approach, which was general methods plus a lot more computing power, went past it. In Sutton’s words, the general methods win “by a large margin.”
One sentence in the essay could have been written about me. Sutton says building your own knowledge into the machine “always helps in the short term, and is personally satisfying to the researcher.” The cost comes later, when the approach stalls. And when the other approach finally wins, he says the success is “tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.”
This matters to you even if you never read another word about how AI works, because it explains why these systems keep improving in ways that surprise the experts. Nobody is sitting there teaching the model the rules of your profession. It learns from an enormous amount of examples and an enormous amount of computing, and that recipe has a long record of beating the hand-built kind. So “it doesn’t do it the way we do it” is an argument that has been losing for seventy years.
The rewire
The old wiring is easy to spot once you know the sound it makes. It’s what you say to someone younger who gets there in an afternoon: I had to learn it the hard way.
That sentence is the tax bill. You paid it, and now you’d like to see everybody else’s receipt.
This week: difficulty or decision
Twenty minutes, and one task to try before Sunday.
- Pick the thing you’re proudest of making. A product, a deal, a book, a team.
- List the five parts you’re proudest of. For each one, ask two questions. Did it come from the difficulty or from a decision? And which draft did it show up in?
- Sort the difficulty pile once more. One stack is the struggle that made you better at judging. The other is the struggle that made you feel serious.
- Take one task from the second stack and do the fast version this week. When it’s done, paste the result and ask: Tell me what’s wrong with this. If you ask “is this good?” you get a cheerleader. This one gets you a colleague.
- Write down what you lost. If the answer is nothing, you just got some hours back. If the answer is something, now you know what the difficulty was for.
Sources
- Rich Sutton, “The Bitter Lesson,” March 13, 2019: http://www.incompleteideas.net/IncIdeas/BitterLesson.html