Vita
About
← All notes

AI insights, kept private

Vita can write you a short, readable summary of how a period went — where your wins were, where you slipped, what's trending. Useful at the end of a week. Quietly motivating most months.

The interesting part isn't the feature. It's where the writing happens.

What "AI feature" usually means

When most apps add an AI summary or AI coach, this is roughly the path your data takes:

A typical AI app your device AI company server your habits → leave the device Vita small model your habits stay on the device
Left: the usual path — your data takes a round trip. Right: Vita — the model lives inside your browser.

This is fine if you're OK with that. We weren't. So Vita's AI insights work differently.

What Vita does instead

The first time you turn on AI insights, Vita downloads a small language model to your browser. From then on, when you ask for a summary, the model runs right there in the browser, using your laptop's or phone's graphics card to do the math.

The summary you read was written on the same device you're reading it from. Nothing about your habits ever touched a network — not on the way out, not on the way back. There is no server-side log of what you asked for, because there is no server in this loop at all.

What an insight actually looks like

You hit "Generate insights" at the end of a week or a month. After a moment, you get a short, human-sounding read on the period:

This week, briefly May 18 – 24 Movement is the rock — walks landed every weekday, stretch held five of seven. Meditation slipped three days in a row — pair it with the morning walk next week? Done count by week w1 w2 w3 w4 this ↳ written on this device · never sent
A weekly insight. Short prose, a small chart underneath, a footer reminding you where it was written.

The shape is short on purpose. The point isn't to drown you in analysis; it's to tell you what jumps out and ask one useful question.

What it costs you, honestly

A model running on your own laptop has real tradeoffs. The big ones:

  • The first run downloads a model. A few hundred MB to a couple of GB depending on which one you pick. Cached after that, but the first time is slow if your network is slow.
  • It needs a relatively recent device. The model uses your graphics card via a browser feature called WebGPU. Most laptops and recent phones support it; older hardware and (today) Safari don't yet. If your browser can't run it, Vita just hides the feature — no half-broken fallback.
  • It's a smaller model than the ones you know by name. ChatGPT or Claude are trained on rooms full of servers; what runs in your browser is something you can fit on a phone. Summaries are shorter and less elaborate. Useful — but if you want a thousand-word essay on your year of pushups, you want a different tool.
  • You choose how big. Three options: a fast one (~370 MB) for slow devices, a balanced one (~900 MB) we recommend by default, and a best-quality one (~2.2 GB) if your laptop has the room and patience.

When you might not want this

Some honest reasons to leave AI insights off:

  • You don't trust LLMs to read your habits accurately. Reasonable. The summary is a starting point, not gospel. The numbers underneath (week view, summary) don't lie even when the prose oversimplifies.
  • Your device is old. A 2 GB model on a 2017 phone is going to be unhappy. Vita won't pretend otherwise.
  • You don't want to wait for the download. Skip it; the rest of Vita is fully usable without it.

You can turn AI insights on and off any time from Settings → AI insights, and clear the cached model with one click. There's no ambient AI in Vita that "runs in the background" — it only thinks when you ask it to.

Up next: Using Vita on more than one device — how to set up Vita on a phone and a laptop without giving up the "data never leaves my device" promise.