The parts of an AI feature that should not be AI

@OceanMind builds your five-minute practice at request time from two answers: what you want out of the next five minutes, and where your nervous system is right now. A model writes it. This month we moved a good part of that work into ordinary code, and the practices are steadier for it.

Why we changed it

A model is strong at language and at choosing between options. Holding a limit, doing arithmetic on seconds and remembering what you did last week are not that kind of problem. Those three moved to code, where they can simply be computed.

What we changed

The protocol limits are now worked out before the model is asked for anything, and what comes back is checked against them: how slow the breath goes, how long a hold may run at the activation you reported, how long the session is.

The bigger change is for people starting out. Hand a beginner a practice at five and a half breaths per minute and they tend to watch the pacer rather than follow it, so early practices now open nearer ordinary breathing and ramp into the slow pace across the session instead of repeating one cycle. That ramp is calculated from the target cycle rather than written freehand, so the tempo genuinely moves instead of only looking as though it does.

What you should notice

• Practices stop repeating. Six sessions in a row on one goal used to produce one or two different titles and similar patterns. Now it is five or six, and the pattern moves with them. Then more pattens user will try they more information for the best next suggestion the system will have.

• Your first sessions are followable. The pace you are given starts where you are and moves, rather than starting where we want you to end up.

• It learns what works for you. After enough sessions, the pattern that actually moved you toward your goal gets preferred, and it still never hands you the same thing twice.

What we took away

The parts of a practice that can be computed should be computed. What is left for the model is the part it is genuinely better at: writing something that reads like a person wrote it and work with history of the user to generate the proper pattern.

Practices are generated by AI and labelled as such in the app.

If you have been using it a while: has it stopped feeling repetitive to you? That is the change I am least sure we got right.