Micro-blog where I jot down short random thoughts.
September 23, 2026
pi-voice has to be near-instant, so I added question variants — the same prompt can now flip between short phrasings without asking the model. Before that, the LLM was dressing up each announcement with extra context, which meant a whole model request every time. Way too slow.
September 22, 2026
pi-voice feedback is cheaper now. Random cue phrases replace the model summary, fewer model calls per feedback cycle — still audible, much less spendy.
September 21, 2026
Launched a new project today: pi-voice, a Pi extension that adds text-to-speech, speech-to-text, and an auto-send catchphrase to the coding agent. It now announces ask_user prompts out loud with a condensed version of the question topic.