looking back at o, ten years in
summary: ten years of theplatform. what i think now that i did not think then.
Core idea: O's durable lesson was that vectors, reactions, and grammars could share one dispatch model when the runtime was designed around judgments rather than subsystems.
Almost ten years. None of the ideas were new. K is forty years old. Join-calculus is from 1996. PEG is from 2004. What was new was insisting they share a runtime. That part I did not borrow from anywhere, and that part is where the leverage came from.
The second-order effects of the unification turned out to be bigger than the first-order ones. One debugger. One trace format. One extension protocol. One profiler. None of those were on the original argument for the design - they were just the consequences I did not have to fight for.
The other thing I would write down: performance is a design property, not a tuning property. The fast version of a primitive is the one that lives in the right crate, returns the right shape, and writes into the buffer the planner sized. I do not remember the last time I optimised a hot loop. I remember the times I changed a primitive's shape and the loop became fast.
If I had to hand the system to someone today and tell them one thing, it would be: keep the table small.