yg1112

6 ms·

1. 1.▲
   I read all 22 GB of my Codex conversations byte by byte. 71% was screenshots(mint.dzgapp.com)

   3 points by yg1112 9d ago|1 comments
2. 2.▲
   by yg1112 6mo ago

   The key difference is that MLX's array model assumes unified memory from the ground up. llama.cpp's Metal backend works fine but carries abstractions from the discrete GPU world — explicit buffer synchronization, command buffer bo
3. 3.▲
   by yg1112 6mo ago

   From the repo, it signs natively on macOS and falls back to a cloud signing server (shortcut-signing-server). That fallback matters -- without macOS you would have to reverse-engineer Apple signing format yourself, and it changes across iOS