The assistant knows you. It doesn't know the water.

An AI assistant is remarkably good at some things and structurally bad at others. Worth knowing which is which, especially at sea, where a confident wrong answer is the expensive kind.

What it can’t do. Compute the sea distance between two points. Plot a route that avoids land and respects your draft. Give you the precise coordinates of a marina it’s never been to. Know whether tomorrow’s wind will round the cape at 11:00. These aren’t rare failures: they sit outside the model’s reach. A language model predicts plausible text; none of those answers live in the training set.

What it does well. It knows you. You’ve been talking to it for months. It remembers that you prefer anchorages to marinas, that you hate motor sailing, that your draft is 1.8 m, that your partner gets seasick in short chop. It knows how you ask questions, what level of detail you want, which compromises you’ll accept.

That part isn’t marketing. It’s the piece the model actually does, and does better than a purpose-built marine app could. A cockpit instrument doesn’t know you. Your assistant does.

The division of labor. You don’t have to pick. Keep the assistant that knows you. Give it something real to reach for when the question touches the water. That’s what a connector does: the model stays the model; the data comes from where data lives, forecast models, tidal stations, OSM, a routing graph that respects coastlines.

Nausika is that connector, for the sea.

A hallucinated coordinate lands you on a rock. A well-connected assistant asks the forecast, plots around land, remembers how you like to sail, and tells you what it doesn’t know.

That’s the pairing worth having.