The original note, as written at the time
A text input system that matches what you say with what you type and uses these patterns to aid your creative process.
Ongoing development
A text input system that compares what you say with what you type, learns the pattern of the difference, and uses that to help you write.
The comparison is the idea. Most dictation tries to be accurate, which is the wrong goal — the gap between what a person says and what they write is not an error, it is their editing. A writer speaks one way and types another, and the distance between the two is the most personal data they produce.
Learn that distance and you have something no model can guess from a corpus: this person’s own compression. They say “we need to have a conversation about the timelines” and type “let’s talk timelines”. They say “I think it would be better if” and type “better if”. The system that has watched this happen a thousand times can offer the second from the first, in their voice, because it is their voice.
That is a different product from an assistant. An assistant writes for you. This writes as you, and only where it has evidence.
The design constraint is that it must not learn the wrong thing. If a person types badly at speed and cleans up later, the model will learn the mess as style. So it has to compare speech to final text, not speech to first draft, which means it needs to know when a document is finished.
The privacy line is absolute. A record of the gap between what a person says and what they write is more revealing than either — it is a map of their hesitation. It belongs on the device and nowhere else.