We all know the Bitter Lesson. It basically says that as models get smarter, our instructions telling them how to accomplish things will look increasingly stupid compared to what they can do by themselves.
I’m a huge believer in this. It’s how I design my harness, and it’s how I train people to use AI.
But there’s a fundamental limitation to the Bitter Lesson. It doesn’t matter how smart a person or an AI is if it doesn’t know what you want.
The alien architect
Let’s say you want to build the perfect house for yourself, and you have an alien architect in front of you. He’s 1,000 times smarter than any human architect, and he can design and build pretty much anything.
So he asks you, “What would you like me to build?”
And you say, “A nice house.”
And he says, “No, I can build anything. Tell me exactly what you want.”
And because you’re not very good at communicating, you say,
I don’t know. I guess, like, a really nice house.
Hitting the wall

At some point, and I think we’re getting close, the problem is no longer execution. It’s knowing what should be executed.
I call this the Bitter Wall. It’s a wall for AI progress that has nothing to do with a scaling or intelligence limitation of the AI itself.
The problem is that AI still can’t read your mind.
If you can’t tell it exactly what you want in extreme detail, or it can’t go find out what you probably want by inferring it from a massive number of other sources, then it simply won’t be able to build what’s in your head.
Harnesses and context
Because we’re approaching this wall, we need to start thinking a lot more about our harnesses, and how they provide context to our models.
It’s one thing to say context is important. It gets a whole lot more real when you realize it doesn’t matter how smart you are if someone asks you to paint them the perfect picture but doesn’t tell you what they consider a good picture.
The Bitter Wall is that AI can’t read minds. So we’re nearing a point where all future gains in AI output quality will come from better articulation of what we want.

