· IDEAS · SIGNAL LOG
Concepts and thinking on voice, computing, and the tools we build. Some are RFCs, some are half-formed sketches, all are in flight.
How we redesigned the Talkie homepage by running six parallel subagents in git worktrees, comparing directions side by side, and picking a winner.
Apple's FoundationModels framework on macOS 26 lets Talkie run AI processing entirely on-device — no API key, no network call, no cost per request.
We tested Apple's Foundation Models against 24 real Talkie use cases. Here is what works, what needs a prompt, and what runs without an external model fee.
A draft RFC for making Talkie programmable from disk with rules, tools, workflows, and automations.
Turn a batch of voice memos into linked Obsidian topic files, daily notes, and an index with four reusable prompts.
A 0.6B model matches the 1.7B on accuracy, runs 2.4x faster, and fits in 350MB. At some point the task stops needing more parameters.
VLM vs text-only, where to draw the line between code and model, and why picking the right model size is harder than it sounds.
Fine-tuning a 1.5B model to reconstruct shell commands from voice. 97% accuracy, 3GB of RAM, under a second on a phone.
The Talkie CLI turns your voice memos and dictations into structured data that any script, pipeline, or AI agent can work with.
The same voice memo, transcribed three different ways. From Apple Speech on iOS to Parakeet v3 with post-processing — each layer improves readability while preserving the speaker's intent.
GUI apps are growing command lines. Here's why we built one for a voice app, and what it lets you do with your data after you've spoken it.
On where my best ideas actually happen, and why my phone keeps missing them.