There's a sort of scarcity/abundance reframe with AI stuff I'm noticing because tokens effectively make many components of knowledge work "too cheap to meter." It makes sense to do less planning because it's so cheap to fix things now. This is basically Jevons paradox applied to knowledge work. Jevons observed in 1865 that when steam engines got more efficient, Britain used more coal, not less — the cheaper it got per unit of work, the more work people found worth doing. Same pattern here. When the cost of revision collapses, you don't just revise cheaper — you revise more, and the whole strategy shifts. Digital photography is the cleanest analogy. When every shot cost film and development, it made sense to be precious about composition. When shots became free, the strategy inverted — shoot 200 and keep 3. The quality of the final image went up, not because any individual shot was better planned, but because the volume of iteration made planning less important. That's what's happening with AI and knowledge work. Historically, if you have a big project or codebase and want to change something, that's really expensive. I was listening to a pod with the Open Claw guy talk about adding Telegram support after the fact. This notion of the AI could "weave" it into the codebase is a workflow that only makes sense with AI. That sort of weaving in and embedding in contextual space was incredibly tedious to do manually. You have to go reread the whole project plan if you changed one or two key decisions, but now that's relatively trivial for the AI to do because as long as you can hold all that in context it just goes through it. I was talking to a developer for an hour and we kept coming back to this: it kind of makes sense to just spend less time planning. Come up with five half-baked different ways you could build something and just make them. To the extent tokens are cheap and all these processes are parallelizable, I think that's the right way to think about it.