the notebook
Field notes from building Pom.
Written by the three people who built Pomvox. Real dates, real numbers, and lessons from building Pomvox.
-
Faster cleanup. Different words. We said no.
Cleaning speech in smaller pieces reduced the measured wait, but it failed the experiment's other requirement: preserve the whole-transcript output.
— Abhi -
The optimization we stopped questioning
A comment said prompt caching could not work. Measuring the repeated work showed why we needed to question it.
— Abhi -
Our tests missed a kind of correction
A correction across two sentences exposed a model gap. Fixing the model then exposed a check that rejected the right answer.
— Abhi -
The paste that pasted the wrong thing
Every so often, Pomvox pasted something you had copied earlier instead of what you just said. Waiting longer made it rarer. Only waiting smarter fixed it.
— Maanasa -
Teaching Pom to do less
We changed how Pomvox tidies up your words, so that when it isn't sure, it leaves them alone. This is where the promise, "never what you didn't," comes from.
— Maanasa -
The first 37 seconds
The very first time Pomvox starts up, it takes about 37 seconds. We couldn't make that go away, so we changed when it happens and what it feels like.
— Maanasa -
The question the app answered for you
We asked people a question, and then Pomvox quietly answered it for them the next time it opened.
— Maanasa -
Making Pom fit on an 8 GB Mac
Pomvox was too heavy for smaller Macs. Our first fix hid a feature from people. Our second fix asked them.
— Maanasa
Longer research threads, like the Dense Bench series on 70B inference, live on Abhi’s site.