Why there will never be AI in Numori

Every few weeks someone asks when Numori Notes will get a summarise button. The answer is never, and it is worth explaining why, because "no AI" sounds like a stylistic preference and it is not one.

The three costs nobody puts on the feature list

A generative feature inside a note-taking app carries three costs that are usually left out of the announcement post.

Your data has to go somewhere it can be processed. Numori Notes stores your notes locally and syncs them through a relay that holds no keys. A model cannot summarise something it cannot read. Adding a summarise button means either shipping your notes to a server that can decrypt them, or shipping a model large enough to be useful onto your laptop. The first breaks the architecture. The second is a 4 GB download to save you reading your own shopping list.

Someone has to pay for the inference. Inference is not free and it does not get cheaper per user as you grow — it scales with usage, unlike storage, which gets cheaper per gigabyte every year. A company that adds AI features to a subscription priced like ours has three options: raise the price, cap the feature so tightly it is useless, or find another way to make money from the people using it. All three are worse than not having the feature.

It changes what the product is for. Once a feature is expensive, it has to be promoted. It gets a place in the toolbar, then a prompt, then a tooltip suggesting you try it. Software that has to justify a running cost stops being a tool and starts being a funnel.

"But it is opt-in"

Opt-in is a real distinction and I do not want to dismiss it. But it decays. The feature that shipped behind a toggle in version 1 is on by default in version 4, because the metrics said adoption was low and adoption was the thing being measured.

More importantly, opt-in does not undo the architectural change. If the server can decrypt your notes when you press the button, it can decrypt your notes. The capability is what matters, not the current policy about when it is used. This is the same argument we make about privacy policies generally: I would rather be unable to read your data than promise not to.

What we do instead

The calculator in Numori Notes is a reasonable illustration of the alternative. It parses 48 GBP in EUR and 12% of 340 and gives you an answer. It is a parser and a unit graph. It runs in a fraction of a millisecond, entirely on your device, it gives the same answer every time, and when it does not understand something it says so rather than inventing a plausible number.

That last property is the one worth dwelling on. A calculator that is wrong 2% of the time is not a calculator. A great deal of what gets shipped as an AI feature is a function that used to be exact and is now approximate, with a friendlier tone of voice.

The commitment

No assistants, no summarisation, no generation, no classification of your content, no "smart" reordering, and no embedding models running over your files to build an index we can query.

It is an easy promise to verify, which is the point of publishing everything: Numori-Notes has no inference dependency in its lock file, and it never will. If that changes, you will see it in a pull request before you see it in a release, and you are welcome to say so loudly.