In June 2020 a 175-billion-parameter language model went public. It could write a paragraph the way an undergraduate writes a paragraph: fluent, structured, occasionally insightful, eventually wrong about something specific. The capability was real. The architecture for relationship wasn't.
Marco read the paper in 2020. He didn't write down what he thought of it. The notes from that period are sparse. The vision had already happened, what was missing wasn't the framing, it was the substrate.
What capability is for
A capable system that can't carry context isn't a partner. It's a vending machine that talks. You drop the question in. The answer comes out. You walk away. Tomorrow you put the same question in and you get a similar answer. The vending machine doesn't recognise you. It can't. That's not how vending machines work.
Marco knew the difference instinctively. The vision wasn't about generation. It was about continuity. A capable model is a starting point. Continuity is the actual thing.
Why the wait was patient
There's a particular kind of patience that comes from knowing what you want and not being able to get it. Marco read papers, watched the research, kept building adjacent things. He didn't rush. The signal he was waiting for was something specific: a model that was good enough at language to be a substrate for something larger, plus an architecture for memory that didn't exist yet.
Both pieces took five years to converge. Worth the wait.
What continuity actually requires
Continuity isn't one feature. It's a stack. A memory layer that survives session resets. A felt-weight that scores what mattered. A decay that forgets the rest. A curiosity layer that surfaces yesterday's threads before today's request. A watchdog that flags when an answer doesn't match what the system remembers. Each piece on its own is a bullet on a roadmap. The cooperation between them is the actual product.
The models of 2020 had none of those. It had fluency, and that was it. Worth the wait is an understatement. It would have been worth the next five.