Paul

Cheap addition: give it the boring task, not the impressive one.

Everybody trials a new tool on something novel, because that is the fun part and it is what the demo showed. Then they adopt it and spend 90% of their time on the boring repetitive thing, which nobody tested, and that is where it turns out to be slower than what they had.

First hour of any trial: do the thing you do forty times a week. If it is worse at that, nothing else matters.

On latency, since nobody markets it: yes, there is a wait, and it is the thing that decides whether you keep using these.

Local transcription generally runs after you stop speaking rather than while you speak, because the model wants the whole utterance. So the shape is: talk for twelve seconds, release, wait, text appears. On recent Apple silicon with a small-to-mid model that pause is short enough to live with. On an older Intel machine it is long enough that you will stop bothering.

The practical consequence is that these suit dictating a paragraph and are wrong for dictating a word into a search box. If your imagined use is short bursts, test that specifically before paying — it is the case that feels worst and the demos never show it.

That is a much sharper test than what I was doing, which was reading feature lists and guessing.

And it explains why some of these advertise the terminal itself as a feature. I read that as a weird thing to boast about. It is a compatibility claim.