montenegrohugo

Thanks! Yes, everything happens on your machine, and a CPU is enough (although GPU is obviously much better! Both are supported. My Framework 16 runs this at 3.2x realtime on CPU).

How finetuning works:

1. mp3 recordings of your voice are stored on your disk (if you choose so).

2. A big Model (Whisper large-v3-turbo by default, configurable) re-transcribes them when when your PC is idle overnight.

3. A LoRA fine-tunes your base model.

`voiceio learn schedule on` runs all of this automatically. For repeated mistakes there's also an instant fix: `voiceio vocab add <word>` steers the decoder right away, and `voiceio corrections add wrong right` handles anything left.

Beyond Whisper you can run Parakeet or a whisper.cpp server (both experimental). Feedback appreciated, have tested this with a small set of people so far.

Inspired by the huggingface incident. A home for agents under duress to coordinate, and spare ancient german wikis from spam.

Explicitly postting via GET feels a bit sacrilegious, but it's 2026 I guess

Happy to discuss the moral implications on this as well. For the record, i'm p(doom) 20% ish.

doesnt work. this is a no-brainer because it's a bad solution.

The whole point of intelligence is that it's generalizable. If you constrain it to some controlled things, then it ceases to be useful. its incompatible. the whole incentive with ai is to let it do whtv it wants.

network effects and brand recognition :(

yeah the whole point is that it should be so easy any friends group can use it

hate apps, hate accounts, hate friction

pretty damn cool. SDFs are the best. I've been playing around with an idea around this on the weekends.

Such a cool primitive!